1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86ShuffleDecodeConstantPool.h"
22 #include "X86TargetMachine.h"
23 #include "X86TargetObjectFile.h"
24 #include "llvm/ADT/SmallBitVector.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/ADT/StringSwitch.h"
29 #include "llvm/Analysis/EHPersonalities.h"
30 #include "llvm/CodeGen/IntrinsicLowering.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/WinEHFuncInfo.h"
38 #include "llvm/IR/CallSite.h"
39 #include "llvm/IR/CallingConv.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/GlobalAlias.h"
44 #include "llvm/IR/GlobalVariable.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/MC/MCAsmInfo.h"
48 #include "llvm/MC/MCContext.h"
49 #include "llvm/MC/MCExpr.h"
50 #include "llvm/MC/MCSymbol.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/ErrorHandling.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Target/TargetOptions.h"
56 #include "X86IntrinsicsInfo.h"
62 #define DEBUG_TYPE "x86-isel"
64 STATISTIC(NumTailCalls, "Number of tail calls");
66 static cl::opt<bool> ExperimentalVectorWideningLegalization(
67 "x86-experimental-vector-widening-legalization", cl::init(false),
68 cl::desc("Enable an experimental vector type legalization through widening "
69 "rather than promotion."),
72 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
73 const X86Subtarget &STI)
74 : TargetLowering(TM), Subtarget(&STI) {
75 X86ScalarSSEf64 = Subtarget->hasSSE2();
76 X86ScalarSSEf32 = Subtarget->hasSSE1();
77 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
79 // Set up the TargetLowering object.
81 // X86 is weird. It always uses i8 for shift amounts and setcc results.
82 setBooleanContents(ZeroOrOneBooleanContent);
83 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
84 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
86 // For 64-bit, since we have so many registers, use the ILP scheduler.
87 // For 32-bit, use the register pressure specific scheduling.
88 // For Atom, always use ILP scheduling.
89 if (Subtarget->isAtom())
90 setSchedulingPreference(Sched::ILP);
91 else if (Subtarget->is64Bit())
92 setSchedulingPreference(Sched::ILP);
94 setSchedulingPreference(Sched::RegPressure);
95 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
96 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
98 // Bypass expensive divides on Atom when compiling with O2.
99 if (TM.getOptLevel() >= CodeGenOpt::Default) {
100 if (Subtarget->hasSlowDivide32())
101 addBypassSlowDiv(32, 8);
102 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
103 addBypassSlowDiv(64, 16);
106 if (Subtarget->isTargetKnownWindowsMSVC()) {
107 // Setup Windows compiler runtime calls.
108 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
109 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
110 setLibcallName(RTLIB::SREM_I64, "_allrem");
111 setLibcallName(RTLIB::UREM_I64, "_aullrem");
112 setLibcallName(RTLIB::MUL_I64, "_allmul");
113 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
117 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
120 if (Subtarget->isTargetDarwin()) {
121 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
122 setUseUnderscoreSetJmp(false);
123 setUseUnderscoreLongJmp(false);
124 } else if (Subtarget->isTargetWindowsGNU()) {
125 // MS runtime is weird: it exports _setjmp, but longjmp!
126 setUseUnderscoreSetJmp(true);
127 setUseUnderscoreLongJmp(false);
129 setUseUnderscoreSetJmp(true);
130 setUseUnderscoreLongJmp(true);
133 // Set up the register classes.
134 addRegisterClass(MVT::i8, &X86::GR8RegClass);
135 addRegisterClass(MVT::i16, &X86::GR16RegClass);
136 addRegisterClass(MVT::i32, &X86::GR32RegClass);
137 if (Subtarget->is64Bit())
138 addRegisterClass(MVT::i64, &X86::GR64RegClass);
140 for (MVT VT : MVT::integer_valuetypes())
141 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
143 // We don't accept any truncstore of integer registers.
144 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
146 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
147 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
148 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
149 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
151 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
153 // SETOEQ and SETUNE require checking two conditions.
154 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
156 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
159 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
161 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
163 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
165 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
167 if (Subtarget->is64Bit()) {
168 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
169 // f32/f64 are legal, f80 is custom.
170 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
173 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
174 } else if (!Subtarget->useSoftFloat()) {
175 // We have an algorithm for SSE2->double, and we turn this into a
176 // 64-bit FILD followed by conditional FADD for other targets.
177 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
178 // We have an algorithm for SSE2, and we turn this into a 64-bit
179 // FILD or VCVTUSI2SS/SD for other targets.
180 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
183 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
185 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
186 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
188 if (!Subtarget->useSoftFloat()) {
189 // SSE has no i16 to fp conversion, only i32
190 if (X86ScalarSSEf32) {
191 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
192 // f32 and f64 cases are Legal, f80 case is not
193 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
196 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
200 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
203 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
205 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
208 if (!Subtarget->useSoftFloat()) {
209 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
210 // are Legal, f80 is custom lowered.
211 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
212 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
214 if (X86ScalarSSEf32) {
215 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
216 // f32 and f64 cases are Legal, f80 case is not
217 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
219 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
220 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
224 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand);
225 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand);
228 // Handle FP_TO_UINT by promoting the destination to a larger signed
230 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
234 if (Subtarget->is64Bit()) {
235 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
236 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
237 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
238 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
240 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
241 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
243 } else if (!Subtarget->useSoftFloat()) {
244 // Since AVX is a superset of SSE3, only check for SSE here.
245 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
246 // Expand FP_TO_UINT into a select.
247 // FIXME: We would like to use a Custom expander here eventually to do
248 // the optimal thing for SSE vs. the default expansion in the legalizer.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
251 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
252 // With SSE3 we can use fisttpll to convert to a signed i64; without
253 // SSE, we're stuck with a fistpll.
254 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
256 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
259 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
260 if (!X86ScalarSSEf64) {
261 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
262 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
263 if (Subtarget->is64Bit()) {
264 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
265 // Without SSE, i64->f64 goes through memory.
266 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
270 // Scalar integer divide and remainder are lowered to use operations that
271 // produce two results, to match the available instructions. This exposes
272 // the two-result form to trivial CSE, which is able to combine x/y and x%y
273 // into a single instruction.
275 // Scalar integer multiply-high is also lowered to use two-result
276 // operations, to match the available instructions. However, plain multiply
277 // (low) operations are left as Legal, as there are single-result
278 // instructions for this in x86. Using the two-result multiply instructions
279 // when both high and low results are needed must be arranged by dagcombine.
280 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
281 setOperationAction(ISD::MULHS, VT, Expand);
282 setOperationAction(ISD::MULHU, VT, Expand);
283 setOperationAction(ISD::SDIV, VT, Expand);
284 setOperationAction(ISD::UDIV, VT, Expand);
285 setOperationAction(ISD::SREM, VT, Expand);
286 setOperationAction(ISD::UREM, VT, Expand);
288 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
289 setOperationAction(ISD::ADDC, VT, Custom);
290 setOperationAction(ISD::ADDE, VT, Custom);
291 setOperationAction(ISD::SUBC, VT, Custom);
292 setOperationAction(ISD::SUBE, VT, Custom);
295 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
296 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
297 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
298 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
299 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
300 setOperationAction(ISD::BR_CC , MVT::f128, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
303 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
304 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::f128, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
310 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
311 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
312 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
313 if (Subtarget->is64Bit())
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
315 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
316 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
317 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
318 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
320 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
321 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
322 // is. We should promote the value to 64-bits to solve this.
323 // This is what the CRT headers do - `fmodf` is an inline header
324 // function casting to f64 and calling `fmod`.
325 setOperationAction(ISD::FREM , MVT::f32 , Promote);
327 setOperationAction(ISD::FREM , MVT::f32 , Expand);
330 setOperationAction(ISD::FREM , MVT::f64 , Expand);
331 setOperationAction(ISD::FREM , MVT::f80 , Expand);
332 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
334 // Promote the i8 variants and force them on up to i32 which has a shorter
336 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
337 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
338 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
339 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
340 if (Subtarget->hasBMI()) {
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
342 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
343 if (Subtarget->is64Bit())
344 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
347 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
348 if (Subtarget->is64Bit())
349 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
352 if (Subtarget->hasLZCNT()) {
353 // When promoting the i8 variants, force them to i32 for a shorter
355 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
356 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
358 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
361 if (Subtarget->is64Bit())
362 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
364 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
365 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
366 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
367 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
368 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
369 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
370 if (Subtarget->is64Bit()) {
371 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
372 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
376 // Special handling for half-precision floating point conversions.
377 // If we don't have F16C support, then lower half float conversions
378 // into library calls.
379 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
380 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
381 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
384 // There's never any support for operations beyond MVT::f32.
385 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
386 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
387 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
388 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
390 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
391 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
392 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
393 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
394 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
395 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
397 if (Subtarget->hasPOPCNT()) {
398 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
400 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
401 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
402 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
403 if (Subtarget->is64Bit())
404 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
407 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
409 if (!Subtarget->hasMOVBE())
410 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
412 // These should be promoted to a larger select which is supported.
413 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
414 // X86 wants to expand cmov itself.
415 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
416 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
417 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
418 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
419 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
420 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
421 setOperationAction(ISD::SELECT , MVT::f128 , Custom);
422 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
423 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
424 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
425 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
426 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
427 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
428 setOperationAction(ISD::SETCC , MVT::f128 , Custom);
429 setOperationAction(ISD::SETCCE , MVT::i8 , Custom);
430 setOperationAction(ISD::SETCCE , MVT::i16 , Custom);
431 setOperationAction(ISD::SETCCE , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
434 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
435 setOperationAction(ISD::SETCCE , MVT::i64 , Custom);
437 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
438 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
439 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
440 // support continuation, user-level threading, and etc.. As a result, no
441 // other SjLj exception interfaces are implemented and please don't build
442 // your own exception handling based on them.
443 // LLVM/Clang supports zero-cost DWARF exception handling.
444 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
445 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
448 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
449 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
450 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
451 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
452 if (Subtarget->is64Bit())
453 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
454 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
455 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
456 if (Subtarget->is64Bit()) {
457 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
458 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
459 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
460 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
461 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
463 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
464 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
465 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
466 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
467 if (Subtarget->is64Bit()) {
468 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
469 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
470 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
473 if (Subtarget->hasSSE1())
474 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
476 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
478 // Expand certain atomics
479 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
480 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
481 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
482 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
485 if (Subtarget->hasCmpxchg16b()) {
486 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
489 // FIXME - use subtarget debug flags
490 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
491 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
492 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
495 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
496 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
498 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
499 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
501 setOperationAction(ISD::TRAP, MVT::Other, Legal);
502 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
504 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
505 setOperationAction(ISD::VASTART , MVT::Other, Custom);
506 setOperationAction(ISD::VAEND , MVT::Other, Expand);
507 if (Subtarget->is64Bit()) {
508 setOperationAction(ISD::VAARG , MVT::Other, Custom);
509 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
511 // TargetInfo::CharPtrBuiltinVaList
512 setOperationAction(ISD::VAARG , MVT::Other, Expand);
513 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
516 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
517 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
519 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
521 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
522 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
523 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
525 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
526 // f32 and f64 use SSE.
527 // Set up the FP register classes.
528 addRegisterClass(MVT::f32, &X86::FR32RegClass);
529 addRegisterClass(MVT::f64, &X86::FR64RegClass);
531 // Use ANDPD to simulate FABS.
532 setOperationAction(ISD::FABS , MVT::f64, Custom);
533 setOperationAction(ISD::FABS , MVT::f32, Custom);
535 // Use XORP to simulate FNEG.
536 setOperationAction(ISD::FNEG , MVT::f64, Custom);
537 setOperationAction(ISD::FNEG , MVT::f32, Custom);
539 // Use ANDPD and ORPD to simulate FCOPYSIGN.
540 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
541 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
543 // Lower this to FGETSIGNx86 plus an AND.
544 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
545 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
547 // We don't support sin/cos/fmod
548 setOperationAction(ISD::FSIN , MVT::f64, Expand);
549 setOperationAction(ISD::FCOS , MVT::f64, Expand);
550 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
551 setOperationAction(ISD::FSIN , MVT::f32, Expand);
552 setOperationAction(ISD::FCOS , MVT::f32, Expand);
553 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
555 // Expand FP immediates into loads from the stack, except for the special
557 addLegalFPImmediate(APFloat(+0.0)); // xorpd
558 addLegalFPImmediate(APFloat(+0.0f)); // xorps
559 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
560 // Use SSE for f32, x87 for f64.
561 // Set up the FP register classes.
562 addRegisterClass(MVT::f32, &X86::FR32RegClass);
563 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
565 // Use ANDPS to simulate FABS.
566 setOperationAction(ISD::FABS , MVT::f32, Custom);
568 // Use XORP to simulate FNEG.
569 setOperationAction(ISD::FNEG , MVT::f32, Custom);
571 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
573 // Use ANDPS and ORPS to simulate FCOPYSIGN.
574 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
575 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
577 // We don't support sin/cos/fmod
578 setOperationAction(ISD::FSIN , MVT::f32, Expand);
579 setOperationAction(ISD::FCOS , MVT::f32, Expand);
580 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
582 // Special cases we handle for FP constants.
583 addLegalFPImmediate(APFloat(+0.0f)); // xorps
584 addLegalFPImmediate(APFloat(+0.0)); // FLD0
585 addLegalFPImmediate(APFloat(+1.0)); // FLD1
586 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
587 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
589 if (!TM.Options.UnsafeFPMath) {
590 setOperationAction(ISD::FSIN , MVT::f64, Expand);
591 setOperationAction(ISD::FCOS , MVT::f64, Expand);
592 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
594 } else if (!Subtarget->useSoftFloat()) {
595 // f32 and f64 in x87.
596 // Set up the FP register classes.
597 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
598 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
600 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
601 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
602 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
603 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
605 if (!TM.Options.UnsafeFPMath) {
606 setOperationAction(ISD::FSIN , MVT::f64, Expand);
607 setOperationAction(ISD::FSIN , MVT::f32, Expand);
608 setOperationAction(ISD::FCOS , MVT::f64, Expand);
609 setOperationAction(ISD::FCOS , MVT::f32, Expand);
610 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
611 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
613 addLegalFPImmediate(APFloat(+0.0)); // FLD0
614 addLegalFPImmediate(APFloat(+1.0)); // FLD1
615 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
616 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
617 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
618 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
619 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
620 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
623 // We don't support FMA.
624 setOperationAction(ISD::FMA, MVT::f64, Expand);
625 setOperationAction(ISD::FMA, MVT::f32, Expand);
627 // Long double always uses X87, except f128 in MMX.
628 if (!Subtarget->useSoftFloat()) {
629 if (Subtarget->is64Bit() && Subtarget->hasMMX()) {
630 addRegisterClass(MVT::f128, &X86::FR128RegClass);
631 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat);
632 setOperationAction(ISD::FABS , MVT::f128, Custom);
633 setOperationAction(ISD::FNEG , MVT::f128, Custom);
634 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
637 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
638 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
639 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
641 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
642 addLegalFPImmediate(TmpFlt); // FLD0
644 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
647 APFloat TmpFlt2(+1.0);
648 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
650 addLegalFPImmediate(TmpFlt2); // FLD1
651 TmpFlt2.changeSign();
652 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
655 if (!TM.Options.UnsafeFPMath) {
656 setOperationAction(ISD::FSIN , MVT::f80, Expand);
657 setOperationAction(ISD::FCOS , MVT::f80, Expand);
658 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
661 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
662 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
663 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
664 setOperationAction(ISD::FRINT, MVT::f80, Expand);
665 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
666 setOperationAction(ISD::FMA, MVT::f80, Expand);
669 // Always use a library call for pow.
670 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
671 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
672 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
674 setOperationAction(ISD::FLOG, MVT::f80, Expand);
675 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
676 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
677 setOperationAction(ISD::FEXP, MVT::f80, Expand);
678 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
679 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
680 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
682 // First set operation action for all vector types to either promote
683 // (for widening) or expand (for scalarization). Then we will selectively
684 // turn on ones that can be effectively codegen'd.
685 for (MVT VT : MVT::vector_valuetypes()) {
686 setOperationAction(ISD::ADD , VT, Expand);
687 setOperationAction(ISD::SUB , VT, Expand);
688 setOperationAction(ISD::FADD, VT, Expand);
689 setOperationAction(ISD::FNEG, VT, Expand);
690 setOperationAction(ISD::FSUB, VT, Expand);
691 setOperationAction(ISD::MUL , VT, Expand);
692 setOperationAction(ISD::FMUL, VT, Expand);
693 setOperationAction(ISD::SDIV, VT, Expand);
694 setOperationAction(ISD::UDIV, VT, Expand);
695 setOperationAction(ISD::FDIV, VT, Expand);
696 setOperationAction(ISD::SREM, VT, Expand);
697 setOperationAction(ISD::UREM, VT, Expand);
698 setOperationAction(ISD::LOAD, VT, Expand);
699 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
701 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
702 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
703 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
704 setOperationAction(ISD::FABS, VT, Expand);
705 setOperationAction(ISD::FSIN, VT, Expand);
706 setOperationAction(ISD::FSINCOS, VT, Expand);
707 setOperationAction(ISD::FCOS, VT, Expand);
708 setOperationAction(ISD::FSINCOS, VT, Expand);
709 setOperationAction(ISD::FREM, VT, Expand);
710 setOperationAction(ISD::FMA, VT, Expand);
711 setOperationAction(ISD::FPOWI, VT, Expand);
712 setOperationAction(ISD::FSQRT, VT, Expand);
713 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
714 setOperationAction(ISD::FFLOOR, VT, Expand);
715 setOperationAction(ISD::FCEIL, VT, Expand);
716 setOperationAction(ISD::FTRUNC, VT, Expand);
717 setOperationAction(ISD::FRINT, VT, Expand);
718 setOperationAction(ISD::FNEARBYINT, VT, Expand);
719 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
720 setOperationAction(ISD::MULHS, VT, Expand);
721 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
722 setOperationAction(ISD::MULHU, VT, Expand);
723 setOperationAction(ISD::SDIVREM, VT, Expand);
724 setOperationAction(ISD::UDIVREM, VT, Expand);
725 setOperationAction(ISD::FPOW, VT, Expand);
726 setOperationAction(ISD::CTPOP, VT, Expand);
727 setOperationAction(ISD::CTTZ, VT, Expand);
728 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
729 setOperationAction(ISD::CTLZ, VT, Expand);
730 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
731 setOperationAction(ISD::SHL, VT, Expand);
732 setOperationAction(ISD::SRA, VT, Expand);
733 setOperationAction(ISD::SRL, VT, Expand);
734 setOperationAction(ISD::ROTL, VT, Expand);
735 setOperationAction(ISD::ROTR, VT, Expand);
736 setOperationAction(ISD::BSWAP, VT, Expand);
737 setOperationAction(ISD::SETCC, VT, Expand);
738 setOperationAction(ISD::FLOG, VT, Expand);
739 setOperationAction(ISD::FLOG2, VT, Expand);
740 setOperationAction(ISD::FLOG10, VT, Expand);
741 setOperationAction(ISD::FEXP, VT, Expand);
742 setOperationAction(ISD::FEXP2, VT, Expand);
743 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
744 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
745 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
746 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
747 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
748 setOperationAction(ISD::TRUNCATE, VT, Expand);
749 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
750 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
751 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
752 setOperationAction(ISD::VSELECT, VT, Expand);
753 setOperationAction(ISD::SELECT_CC, VT, Expand);
754 for (MVT InnerVT : MVT::vector_valuetypes()) {
755 setTruncStoreAction(InnerVT, VT, Expand);
757 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
758 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
760 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
761 // types, we have to deal with them whether we ask for Expansion or not.
762 // Setting Expand causes its own optimisation problems though, so leave
764 if (VT.getVectorElementType() == MVT::i1)
765 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
767 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
768 // split/scalarized right now.
769 if (VT.getVectorElementType() == MVT::f16)
770 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
774 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
775 // with -msoft-float, disable use of MMX as well.
776 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
777 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
778 // No operations on x86mmx supported, everything uses intrinsics.
781 // MMX-sized vectors (other than x86mmx) are expected to be expanded
782 // into smaller operations.
783 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
784 setOperationAction(ISD::MULHS, MMXTy, Expand);
785 setOperationAction(ISD::AND, MMXTy, Expand);
786 setOperationAction(ISD::OR, MMXTy, Expand);
787 setOperationAction(ISD::XOR, MMXTy, Expand);
788 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
789 setOperationAction(ISD::SELECT, MMXTy, Expand);
790 setOperationAction(ISD::BITCAST, MMXTy, Expand);
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
794 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
795 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
797 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
798 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
799 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
800 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
801 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
802 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
803 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
804 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
805 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
806 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
807 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
808 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
809 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
810 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
813 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
814 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
816 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
817 // registers cannot be used even for integer operations.
818 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
819 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
820 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
821 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
823 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
824 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
825 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
826 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
827 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
828 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
829 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
830 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
831 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
832 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
833 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
834 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
835 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
836 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
837 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
838 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
839 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
840 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
841 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
842 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
843 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
844 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
845 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
847 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
848 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
849 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
850 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
852 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
853 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
854 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
855 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
857 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
858 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
859 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
860 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
861 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
863 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
864 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
865 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
866 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
868 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
869 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
870 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
871 // ISD::CTTZ v2i64 - scalarization is faster.
872 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
873 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
874 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
875 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
877 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
878 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
879 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
880 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
881 setOperationAction(ISD::VSELECT, VT, Custom);
882 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
885 // We support custom legalizing of sext and anyext loads for specific
886 // memory vector types which we can load as a scalar (or sequence of
887 // scalars) and extend in-register to a legal 128-bit vector type. For sext
888 // loads these must work with a single scalar load.
889 for (MVT VT : MVT::integer_vector_valuetypes()) {
890 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
891 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
892 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
893 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
894 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
895 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
896 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
897 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
898 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
901 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
902 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
903 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
904 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
905 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
906 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
907 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
908 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
910 if (Subtarget->is64Bit()) {
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
915 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
916 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
917 setOperationAction(ISD::AND, VT, Promote);
918 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
919 setOperationAction(ISD::OR, VT, Promote);
920 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
921 setOperationAction(ISD::XOR, VT, Promote);
922 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
923 setOperationAction(ISD::LOAD, VT, Promote);
924 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
925 setOperationAction(ISD::SELECT, VT, Promote);
926 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
929 // Custom lower v2i64 and v2f64 selects.
930 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
931 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
932 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
933 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
935 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
936 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
938 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
940 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
941 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
942 // As there is no 64-bit GPR available, we need build a special custom
943 // sequence to convert from v2i32 to v2f32.
944 if (!Subtarget->is64Bit())
945 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
947 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
948 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
950 for (MVT VT : MVT::fp_vector_valuetypes())
951 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
953 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
954 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
955 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
958 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
959 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
960 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
961 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
962 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
963 setOperationAction(ISD::FRINT, RoundedTy, Legal);
964 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
967 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
968 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
969 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
970 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
971 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
972 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
973 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
974 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
976 // FIXME: Do we need to handle scalar-to-vector here?
977 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
979 // We directly match byte blends in the backend as they match the VSELECT
981 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
983 // SSE41 brings specific instructions for doing vector sign extend even in
984 // cases where we don't have SRA.
985 for (MVT VT : MVT::integer_vector_valuetypes()) {
986 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
988 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
991 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
997 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1004 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1006 // i8 and i16 vectors are custom because the source register and source
1007 // source memory operand types are not the same width. f32 vectors are
1008 // custom since the immediate controlling the insert encodes additional
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1013 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1015 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1018 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1020 // FIXME: these should be Legal, but that's only for the case where
1021 // the index is constant. For now custom expand to deal with that.
1022 if (Subtarget->is64Bit()) {
1023 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1024 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1028 if (Subtarget->hasSSE2()) {
1029 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1031 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1034 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1036 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1037 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1039 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1040 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1042 // In the customized shift lowering, the legal cases in AVX2 will be
1044 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1045 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1047 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1048 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1050 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1051 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1054 if (Subtarget->hasXOP()) {
1055 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1062 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1065 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1066 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1071 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1073 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1074 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1075 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1077 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1087 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1088 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1090 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1100 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1101 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1103 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1104 // even though v8i16 is a legal type.
1105 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1106 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1107 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1109 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1110 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1111 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1113 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1114 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1116 for (MVT VT : MVT::fp_vector_valuetypes())
1117 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1119 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1120 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1122 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1123 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1125 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1126 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1128 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1131 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1133 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1135 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1137 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1139 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1140 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1142 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1143 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1145 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1146 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1148 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1150 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1153 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1155 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1158 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1162 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1164 if (Subtarget->hasAnyFMA()) {
1165 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1166 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1167 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1168 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1169 setOperationAction(ISD::FMA, MVT::f32, Legal);
1170 setOperationAction(ISD::FMA, MVT::f64, Legal);
1173 if (Subtarget->hasInt256()) {
1174 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1175 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1176 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1177 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1179 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1180 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1181 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1182 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1184 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1185 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1186 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1187 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1189 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1190 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1191 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1192 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1194 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1196 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1197 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1199 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1200 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1202 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1203 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1205 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1207 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1208 // when we have a 256bit-wide blend with immediate.
1209 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1211 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1217 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1224 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1226 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1227 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1228 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1229 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1231 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1232 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1233 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1234 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1236 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1237 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1238 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1239 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1241 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1243 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1244 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1246 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1247 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1249 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1250 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1252 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1255 // In the customized shift lowering, the legal cases in AVX2 will be
1257 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1258 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1260 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1261 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1263 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1264 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1266 // Custom lower several nodes for 256-bit types.
1267 for (MVT VT : MVT::vector_valuetypes()) {
1268 if (VT.getScalarSizeInBits() >= 32) {
1269 setOperationAction(ISD::MLOAD, VT, Legal);
1270 setOperationAction(ISD::MSTORE, VT, Legal);
1272 // Extract subvector is special because the value type
1273 // (result) is 128-bit but the source is 256-bit wide.
1274 if (VT.is128BitVector()) {
1275 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1277 // Do not attempt to custom lower other non-256-bit vectors
1278 if (!VT.is256BitVector())
1281 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1282 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1283 setOperationAction(ISD::VSELECT, VT, Custom);
1284 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1285 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1286 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1287 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1288 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1291 if (Subtarget->hasInt256())
1292 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1294 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1295 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1296 setOperationAction(ISD::AND, VT, Promote);
1297 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1298 setOperationAction(ISD::OR, VT, Promote);
1299 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1300 setOperationAction(ISD::XOR, VT, Promote);
1301 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1302 setOperationAction(ISD::LOAD, VT, Promote);
1303 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1304 setOperationAction(ISD::SELECT, VT, Promote);
1305 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1309 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1310 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1311 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1312 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1313 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1315 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1316 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1317 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1319 for (MVT VT : MVT::fp_vector_valuetypes())
1320 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1322 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1323 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1324 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1325 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1326 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1327 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1328 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1329 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1330 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1331 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1332 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1333 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1335 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1336 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1337 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1338 setOperationAction(ISD::XOR, MVT::i1, Legal);
1339 setOperationAction(ISD::OR, MVT::i1, Legal);
1340 setOperationAction(ISD::AND, MVT::i1, Legal);
1341 setOperationAction(ISD::SUB, MVT::i1, Custom);
1342 setOperationAction(ISD::ADD, MVT::i1, Custom);
1343 setOperationAction(ISD::MUL, MVT::i1, Custom);
1344 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1345 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1346 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1348 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1350 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1351 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1352 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1355 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1356 setOperationAction(ISD::FABS, MVT::v16f32, Custom);
1358 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1359 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1363 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1364 setOperationAction(ISD::FABS, MVT::v8f64, Custom);
1365 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1366 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1368 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1369 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1371 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1376 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1377 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1381 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1382 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1383 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1385 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1387 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1388 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1389 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1390 if (Subtarget->hasVLX()){
1391 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1393 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1394 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1395 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1397 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1399 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1400 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1401 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1403 setOperationAction(ISD::MLOAD, MVT::v8i32, Custom);
1404 setOperationAction(ISD::MLOAD, MVT::v8f32, Custom);
1405 setOperationAction(ISD::MSTORE, MVT::v8i32, Custom);
1406 setOperationAction(ISD::MSTORE, MVT::v8f32, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1409 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1410 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1411 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1412 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1413 if (Subtarget->hasDQI()) {
1414 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1415 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1417 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1418 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1419 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1420 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1421 if (Subtarget->hasVLX()) {
1422 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1423 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1424 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1425 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1426 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1427 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1428 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1429 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1432 if (Subtarget->hasVLX()) {
1433 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1434 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1435 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1438 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1439 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1440 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1442 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1443 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1444 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1445 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1446 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1447 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1448 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1449 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1452 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1453 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1454 if (Subtarget->hasDQI()) {
1455 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1456 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1458 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1459 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1460 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1461 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1462 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1463 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1464 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1465 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1466 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1467 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Custom);
1475 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1476 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1478 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1480 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1482 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1484 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1485 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1486 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1490 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1491 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1493 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1494 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1495 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1496 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1497 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1498 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1499 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1500 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1502 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1503 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1505 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1506 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1508 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1510 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1511 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1513 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1514 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1516 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1517 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1519 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1520 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1521 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1522 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1523 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1524 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1526 if (Subtarget->hasCDI()) {
1527 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1528 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Expand);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Expand);
1532 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1533 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1535 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1536 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Expand);
1537 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Expand);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Expand);
1539 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Expand);
1541 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1542 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1544 if (Subtarget->hasVLX()) {
1545 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1546 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1547 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1548 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1549 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1550 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1551 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1552 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1554 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1555 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1556 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1557 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1559 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1560 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1561 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1562 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1563 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Expand);
1564 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Expand);
1565 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Expand);
1566 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Expand);
1568 } // Subtarget->hasCDI()
1570 if (Subtarget->hasDQI()) {
1571 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1572 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1573 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1575 // Custom lower several nodes.
1576 for (MVT VT : MVT::vector_valuetypes()) {
1577 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1579 setOperationAction(ISD::AND, VT, Legal);
1580 setOperationAction(ISD::OR, VT, Legal);
1581 setOperationAction(ISD::XOR, VT, Legal);
1583 if ((VT.is128BitVector() || VT.is256BitVector()) && EltSize >= 32) {
1584 setOperationAction(ISD::MGATHER, VT, Custom);
1585 setOperationAction(ISD::MSCATTER, VT, Custom);
1587 // Extract subvector is special because the value type
1588 // (result) is 256/128-bit but the source is 512-bit wide.
1589 if (VT.is128BitVector() || VT.is256BitVector()) {
1590 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1592 if (VT.getVectorElementType() == MVT::i1)
1593 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1595 // Do not attempt to custom lower other non-512-bit vectors
1596 if (!VT.is512BitVector())
1599 if (EltSize >= 32) {
1600 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1601 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1602 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1603 setOperationAction(ISD::VSELECT, VT, Legal);
1604 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1605 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1606 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1607 setOperationAction(ISD::MLOAD, VT, Legal);
1608 setOperationAction(ISD::MSTORE, VT, Legal);
1609 setOperationAction(ISD::MGATHER, VT, Legal);
1610 setOperationAction(ISD::MSCATTER, VT, Custom);
1613 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1614 setOperationAction(ISD::SELECT, VT, Promote);
1615 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1619 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1620 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1621 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1623 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1624 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1626 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1627 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1628 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1629 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1630 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1631 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1632 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1633 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1634 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1635 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1636 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1637 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1640 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1641 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1644 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1645 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1646 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1647 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1648 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1649 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1650 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1651 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1652 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1655 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1656 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1660 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1661 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1662 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1663 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1664 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1665 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1669 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1670 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1671 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1672 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1673 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1674 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1675 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1676 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1678 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1679 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1680 if (Subtarget->hasVLX())
1681 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1683 if (Subtarget->hasCDI()) {
1684 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1685 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1686 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Expand);
1687 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Expand);
1690 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1691 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1692 setOperationAction(ISD::VSELECT, VT, Legal);
1693 setOperationAction(ISD::SRL, VT, Custom);
1694 setOperationAction(ISD::SHL, VT, Custom);
1695 setOperationAction(ISD::SRA, VT, Custom);
1697 setOperationAction(ISD::AND, VT, Promote);
1698 AddPromotedToType (ISD::AND, VT, MVT::v8i64);
1699 setOperationAction(ISD::OR, VT, Promote);
1700 AddPromotedToType (ISD::OR, VT, MVT::v8i64);
1701 setOperationAction(ISD::XOR, VT, Promote);
1702 AddPromotedToType (ISD::XOR, VT, MVT::v8i64);
1706 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1707 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1708 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1710 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1711 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1712 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1713 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1714 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1715 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1716 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1717 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1718 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1719 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1720 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1721 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1723 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1724 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1725 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1726 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1727 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1728 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1729 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1730 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1732 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1733 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1734 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1735 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1736 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1737 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1738 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1739 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1742 // We want to custom lower some of our intrinsics.
1743 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1744 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1745 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1746 if (!Subtarget->is64Bit()) {
1747 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1748 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i64, Custom);
1751 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1752 // handle type legalization for these operations here.
1754 // FIXME: We really should do custom legalization for addition and
1755 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1756 // than generic legalization for 64-bit multiplication-with-overflow, though.
1757 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1758 if (VT == MVT::i64 && !Subtarget->is64Bit())
1760 // Add/Sub/Mul with overflow operations are custom lowered.
1761 setOperationAction(ISD::SADDO, VT, Custom);
1762 setOperationAction(ISD::UADDO, VT, Custom);
1763 setOperationAction(ISD::SSUBO, VT, Custom);
1764 setOperationAction(ISD::USUBO, VT, Custom);
1765 setOperationAction(ISD::SMULO, VT, Custom);
1766 setOperationAction(ISD::UMULO, VT, Custom);
1769 if (!Subtarget->is64Bit()) {
1770 // These libcalls are not available in 32-bit.
1771 setLibcallName(RTLIB::SHL_I128, nullptr);
1772 setLibcallName(RTLIB::SRL_I128, nullptr);
1773 setLibcallName(RTLIB::SRA_I128, nullptr);
1776 // Combine sin / cos into one node or libcall if possible.
1777 if (Subtarget->hasSinCos()) {
1778 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1779 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1780 if (Subtarget->isTargetDarwin()) {
1781 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1782 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1783 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1784 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1788 if (Subtarget->isTargetWin64()) {
1789 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1790 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1791 setOperationAction(ISD::SREM, MVT::i128, Custom);
1792 setOperationAction(ISD::UREM, MVT::i128, Custom);
1793 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1794 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1797 // We have target-specific dag combine patterns for the following nodes:
1798 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1799 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1800 setTargetDAGCombine(ISD::BITCAST);
1801 setTargetDAGCombine(ISD::VSELECT);
1802 setTargetDAGCombine(ISD::SELECT);
1803 setTargetDAGCombine(ISD::SHL);
1804 setTargetDAGCombine(ISD::SRA);
1805 setTargetDAGCombine(ISD::SRL);
1806 setTargetDAGCombine(ISD::OR);
1807 setTargetDAGCombine(ISD::AND);
1808 setTargetDAGCombine(ISD::ADD);
1809 setTargetDAGCombine(ISD::FADD);
1810 setTargetDAGCombine(ISD::FSUB);
1811 setTargetDAGCombine(ISD::FNEG);
1812 setTargetDAGCombine(ISD::FMA);
1813 setTargetDAGCombine(ISD::FMINNUM);
1814 setTargetDAGCombine(ISD::FMAXNUM);
1815 setTargetDAGCombine(ISD::SUB);
1816 setTargetDAGCombine(ISD::LOAD);
1817 setTargetDAGCombine(ISD::MLOAD);
1818 setTargetDAGCombine(ISD::STORE);
1819 setTargetDAGCombine(ISD::MSTORE);
1820 setTargetDAGCombine(ISD::TRUNCATE);
1821 setTargetDAGCombine(ISD::ZERO_EXTEND);
1822 setTargetDAGCombine(ISD::ANY_EXTEND);
1823 setTargetDAGCombine(ISD::SIGN_EXTEND);
1824 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1825 setTargetDAGCombine(ISD::SINT_TO_FP);
1826 setTargetDAGCombine(ISD::UINT_TO_FP);
1827 setTargetDAGCombine(ISD::SETCC);
1828 setTargetDAGCombine(ISD::BUILD_VECTOR);
1829 setTargetDAGCombine(ISD::MUL);
1830 setTargetDAGCombine(ISD::XOR);
1831 setTargetDAGCombine(ISD::MSCATTER);
1832 setTargetDAGCombine(ISD::MGATHER);
1834 computeRegisterProperties(Subtarget->getRegisterInfo());
1836 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1837 MaxStoresPerMemsetOptSize = 8;
1838 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1839 MaxStoresPerMemcpyOptSize = 4;
1840 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1841 MaxStoresPerMemmoveOptSize = 4;
1842 setPrefLoopAlignment(4); // 2^4 bytes.
1844 // A predictable cmov does not hurt on an in-order CPU.
1845 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1846 PredictableSelectIsExpensive = !Subtarget->isAtom();
1847 EnableExtLdPromotion = true;
1848 setPrefFunctionAlignment(4); // 2^4 bytes.
1850 verifyIntrinsicTables();
1853 // This has so far only been implemented for 64-bit MachO.
1854 bool X86TargetLowering::useLoadStackGuardNode() const {
1855 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1858 TargetLoweringBase::LegalizeTypeAction
1859 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1860 if (ExperimentalVectorWideningLegalization &&
1861 VT.getVectorNumElements() != 1 &&
1862 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1863 return TypeWidenVector;
1865 return TargetLoweringBase::getPreferredVectorAction(VT);
1868 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1871 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1873 if (VT.isSimple()) {
1874 MVT VVT = VT.getSimpleVT();
1875 const unsigned NumElts = VVT.getVectorNumElements();
1876 const MVT EltVT = VVT.getVectorElementType();
1877 if (VVT.is512BitVector()) {
1878 if (Subtarget->hasAVX512())
1879 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1880 EltVT == MVT::f32 || EltVT == MVT::f64)
1882 case 8: return MVT::v8i1;
1883 case 16: return MVT::v16i1;
1885 if (Subtarget->hasBWI())
1886 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1888 case 32: return MVT::v32i1;
1889 case 64: return MVT::v64i1;
1893 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1894 if (Subtarget->hasVLX())
1895 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1896 EltVT == MVT::f32 || EltVT == MVT::f64)
1898 case 2: return MVT::v2i1;
1899 case 4: return MVT::v4i1;
1900 case 8: return MVT::v8i1;
1902 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1903 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1905 case 8: return MVT::v8i1;
1906 case 16: return MVT::v16i1;
1907 case 32: return MVT::v32i1;
1912 return VT.changeVectorElementTypeToInteger();
1915 /// Helper for getByValTypeAlignment to determine
1916 /// the desired ByVal argument alignment.
1917 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1920 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1921 if (VTy->getBitWidth() == 128)
1923 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1924 unsigned EltAlign = 0;
1925 getMaxByValAlign(ATy->getElementType(), EltAlign);
1926 if (EltAlign > MaxAlign)
1927 MaxAlign = EltAlign;
1928 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1929 for (auto *EltTy : STy->elements()) {
1930 unsigned EltAlign = 0;
1931 getMaxByValAlign(EltTy, EltAlign);
1932 if (EltAlign > MaxAlign)
1933 MaxAlign = EltAlign;
1940 /// Return the desired alignment for ByVal aggregate
1941 /// function arguments in the caller parameter area. For X86, aggregates
1942 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1943 /// are at 4-byte boundaries.
1944 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1945 const DataLayout &DL) const {
1946 if (Subtarget->is64Bit()) {
1947 // Max of 8 and alignment of type.
1948 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1955 if (Subtarget->hasSSE1())
1956 getMaxByValAlign(Ty, Align);
1960 /// Returns the target specific optimal type for load
1961 /// and store operations as a result of memset, memcpy, and memmove
1962 /// lowering. If DstAlign is zero that means it's safe to destination
1963 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1964 /// means there isn't a need to check it against alignment requirement,
1965 /// probably because the source does not need to be loaded. If 'IsMemset' is
1966 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1967 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1968 /// source is constant so it does not need to be loaded.
1969 /// It returns EVT::Other if the type should be determined using generic
1970 /// target-independent logic.
1972 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1973 unsigned DstAlign, unsigned SrcAlign,
1974 bool IsMemset, bool ZeroMemset,
1976 MachineFunction &MF) const {
1977 const Function *F = MF.getFunction();
1978 if ((!IsMemset || ZeroMemset) &&
1979 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1981 (!Subtarget->isUnalignedMem16Slow() ||
1982 ((DstAlign == 0 || DstAlign >= 16) &&
1983 (SrcAlign == 0 || SrcAlign >= 16)))) {
1985 // FIXME: Check if unaligned 32-byte accesses are slow.
1986 if (Subtarget->hasInt256())
1988 if (Subtarget->hasFp256())
1991 if (Subtarget->hasSSE2())
1993 if (Subtarget->hasSSE1())
1995 } else if (!MemcpyStrSrc && Size >= 8 &&
1996 !Subtarget->is64Bit() &&
1997 Subtarget->hasSSE2()) {
1998 // Do not use f64 to lower memcpy if source is string constant. It's
1999 // better to use i32 to avoid the loads.
2003 // This is a compromise. If we reach here, unaligned accesses may be slow on
2004 // this target. However, creating smaller, aligned accesses could be even
2005 // slower and would certainly be a lot more code.
2006 if (Subtarget->is64Bit() && Size >= 8)
2011 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2013 return X86ScalarSSEf32;
2014 else if (VT == MVT::f64)
2015 return X86ScalarSSEf64;
2020 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2025 switch (VT.getSizeInBits()) {
2027 // 8-byte and under are always assumed to be fast.
2031 *Fast = !Subtarget->isUnalignedMem16Slow();
2034 *Fast = !Subtarget->isUnalignedMem32Slow();
2036 // TODO: What about AVX-512 (512-bit) accesses?
2039 // Misaligned accesses of any size are always allowed.
2043 /// Return the entry encoding for a jump table in the
2044 /// current function. The returned value is a member of the
2045 /// MachineJumpTableInfo::JTEntryKind enum.
2046 unsigned X86TargetLowering::getJumpTableEncoding() const {
2047 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2049 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2050 Subtarget->isPICStyleGOT())
2051 return MachineJumpTableInfo::EK_Custom32;
2053 // Otherwise, use the normal jump table encoding heuristics.
2054 return TargetLowering::getJumpTableEncoding();
2057 bool X86TargetLowering::useSoftFloat() const {
2058 return Subtarget->useSoftFloat();
2062 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2063 const MachineBasicBlock *MBB,
2064 unsigned uid,MCContext &Ctx) const{
2065 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2066 Subtarget->isPICStyleGOT());
2067 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2069 return MCSymbolRefExpr::create(MBB->getSymbol(),
2070 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2073 /// Returns relocation base for the given PIC jumptable.
2074 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2075 SelectionDAG &DAG) const {
2076 if (!Subtarget->is64Bit())
2077 // This doesn't have SDLoc associated with it, but is not really the
2078 // same as a Register.
2079 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2080 getPointerTy(DAG.getDataLayout()));
2084 /// This returns the relocation base for the given PIC jumptable,
2085 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2086 const MCExpr *X86TargetLowering::
2087 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2088 MCContext &Ctx) const {
2089 // X86-64 uses RIP relative addressing based on the jump table label.
2090 if (Subtarget->isPICStyleRIPRel())
2091 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2093 // Otherwise, the reference is relative to the PIC base.
2094 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2097 std::pair<const TargetRegisterClass *, uint8_t>
2098 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2100 const TargetRegisterClass *RRC = nullptr;
2102 switch (VT.SimpleTy) {
2104 return TargetLowering::findRepresentativeClass(TRI, VT);
2105 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2106 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2109 RRC = &X86::VR64RegClass;
2111 case MVT::f32: case MVT::f64:
2112 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2113 case MVT::v4f32: case MVT::v2f64:
2114 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2116 RRC = &X86::VR128RegClass;
2119 return std::make_pair(RRC, Cost);
2122 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2123 unsigned &Offset) const {
2124 if (!Subtarget->isTargetLinux())
2127 if (Subtarget->is64Bit()) {
2128 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2130 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2142 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2143 if (!Subtarget->isTargetAndroid())
2144 return TargetLowering::getSafeStackPointerLocation(IRB);
2146 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2147 // definition of TLS_SLOT_SAFESTACK in
2148 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2149 unsigned AddressSpace, Offset;
2150 if (Subtarget->is64Bit()) {
2151 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2153 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2163 return ConstantExpr::getIntToPtr(
2164 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2165 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2168 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2169 unsigned DestAS) const {
2170 assert(SrcAS != DestAS && "Expected different address spaces!");
2172 return SrcAS < 256 && DestAS < 256;
2175 //===----------------------------------------------------------------------===//
2176 // Return Value Calling Convention Implementation
2177 //===----------------------------------------------------------------------===//
2179 #include "X86GenCallingConv.inc"
2181 bool X86TargetLowering::CanLowerReturn(
2182 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2183 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2184 SmallVector<CCValAssign, 16> RVLocs;
2185 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2186 return CCInfo.CheckReturn(Outs, RetCC_X86);
2189 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2190 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2195 X86TargetLowering::LowerReturn(SDValue Chain,
2196 CallingConv::ID CallConv, bool isVarArg,
2197 const SmallVectorImpl<ISD::OutputArg> &Outs,
2198 const SmallVectorImpl<SDValue> &OutVals,
2199 SDLoc dl, SelectionDAG &DAG) const {
2200 MachineFunction &MF = DAG.getMachineFunction();
2201 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2203 if (CallConv == CallingConv::X86_INTR && !Outs.empty())
2204 report_fatal_error("X86 interrupts may not return any value");
2206 SmallVector<CCValAssign, 16> RVLocs;
2207 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2208 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2211 SmallVector<SDValue, 6> RetOps;
2212 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2213 // Operand #1 = Bytes To Pop
2214 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2217 // Copy the result values into the output registers.
2218 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2219 CCValAssign &VA = RVLocs[i];
2220 assert(VA.isRegLoc() && "Can only return in registers!");
2221 SDValue ValToCopy = OutVals[i];
2222 EVT ValVT = ValToCopy.getValueType();
2224 // Promote values to the appropriate types.
2225 if (VA.getLocInfo() == CCValAssign::SExt)
2226 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2227 else if (VA.getLocInfo() == CCValAssign::ZExt)
2228 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2229 else if (VA.getLocInfo() == CCValAssign::AExt) {
2230 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2231 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2233 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2235 else if (VA.getLocInfo() == CCValAssign::BCvt)
2236 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2238 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2239 "Unexpected FP-extend for return value.");
2241 // If this is x86-64, and we disabled SSE, we can't return FP values,
2242 // or SSE or MMX vectors.
2243 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2244 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2245 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2246 report_fatal_error("SSE register return with SSE disabled");
2248 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2249 // llvm-gcc has never done it right and no one has noticed, so this
2250 // should be OK for now.
2251 if (ValVT == MVT::f64 &&
2252 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2253 report_fatal_error("SSE2 register return with SSE2 disabled");
2255 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2256 // the RET instruction and handled by the FP Stackifier.
2257 if (VA.getLocReg() == X86::FP0 ||
2258 VA.getLocReg() == X86::FP1) {
2259 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2260 // change the value to the FP stack register class.
2261 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2262 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2263 RetOps.push_back(ValToCopy);
2264 // Don't emit a copytoreg.
2268 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2269 // which is returned in RAX / RDX.
2270 if (Subtarget->is64Bit()) {
2271 if (ValVT == MVT::x86mmx) {
2272 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2273 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2274 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2276 // If we don't have SSE2 available, convert to v4f32 so the generated
2277 // register is legal.
2278 if (!Subtarget->hasSSE2())
2279 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2284 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2285 Flag = Chain.getValue(1);
2286 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2289 // All x86 ABIs require that for returning structs by value we copy
2290 // the sret argument into %rax/%eax (depending on ABI) for the return.
2291 // We saved the argument into a virtual register in the entry block,
2292 // so now we copy the value out and into %rax/%eax.
2294 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2295 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2296 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2297 // either case FuncInfo->setSRetReturnReg() will have been called.
2298 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2299 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2300 getPointerTy(MF.getDataLayout()));
2303 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2304 X86::RAX : X86::EAX;
2305 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2306 Flag = Chain.getValue(1);
2308 // RAX/EAX now acts like a return value.
2310 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2313 RetOps[0] = Chain; // Update chain.
2315 // Add the flag if we have it.
2317 RetOps.push_back(Flag);
2319 X86ISD::NodeType opcode = X86ISD::RET_FLAG;
2320 if (CallConv == CallingConv::X86_INTR)
2321 opcode = X86ISD::IRET;
2322 return DAG.getNode(opcode, dl, MVT::Other, RetOps);
2325 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2326 if (N->getNumValues() != 1)
2328 if (!N->hasNUsesOfValue(1, 0))
2331 SDValue TCChain = Chain;
2332 SDNode *Copy = *N->use_begin();
2333 if (Copy->getOpcode() == ISD::CopyToReg) {
2334 // If the copy has a glue operand, we conservatively assume it isn't safe to
2335 // perform a tail call.
2336 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2338 TCChain = Copy->getOperand(0);
2339 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2342 bool HasRet = false;
2343 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2345 if (UI->getOpcode() != X86ISD::RET_FLAG)
2347 // If we are returning more than one value, we can definitely
2348 // not make a tail call see PR19530
2349 if (UI->getNumOperands() > 4)
2351 if (UI->getNumOperands() == 4 &&
2352 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2365 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2366 ISD::NodeType ExtendKind) const {
2368 // TODO: Is this also valid on 32-bit?
2369 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2370 ReturnMVT = MVT::i8;
2372 ReturnMVT = MVT::i32;
2374 EVT MinVT = getRegisterType(Context, ReturnMVT);
2375 return VT.bitsLT(MinVT) ? MinVT : VT;
2378 /// Lower the result values of a call into the
2379 /// appropriate copies out of appropriate physical registers.
2382 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2383 CallingConv::ID CallConv, bool isVarArg,
2384 const SmallVectorImpl<ISD::InputArg> &Ins,
2385 SDLoc dl, SelectionDAG &DAG,
2386 SmallVectorImpl<SDValue> &InVals) const {
2388 // Assign locations to each value returned by this call.
2389 SmallVector<CCValAssign, 16> RVLocs;
2390 bool Is64Bit = Subtarget->is64Bit();
2391 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2393 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2395 // Copy all of the result registers out of their specified physreg.
2396 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2397 CCValAssign &VA = RVLocs[i];
2398 EVT CopyVT = VA.getLocVT();
2400 // If this is x86-64, and we disabled SSE, we can't return FP values
2401 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64 || CopyVT == MVT::f128) &&
2402 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2403 report_fatal_error("SSE register return with SSE disabled");
2406 // If we prefer to use the value in xmm registers, copy it out as f80 and
2407 // use a truncate to move it from fp stack reg to xmm reg.
2408 bool RoundAfterCopy = false;
2409 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2410 isScalarFPTypeInSSEReg(VA.getValVT())) {
2412 RoundAfterCopy = (CopyVT != VA.getLocVT());
2415 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2416 CopyVT, InFlag).getValue(1);
2417 SDValue Val = Chain.getValue(0);
2420 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2421 // This truncation won't change the value.
2422 DAG.getIntPtrConstant(1, dl));
2424 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2425 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2427 InFlag = Chain.getValue(2);
2428 InVals.push_back(Val);
2434 //===----------------------------------------------------------------------===//
2435 // C & StdCall & Fast Calling Convention implementation
2436 //===----------------------------------------------------------------------===//
2437 // StdCall calling convention seems to be standard for many Windows' API
2438 // routines and around. It differs from C calling convention just a little:
2439 // callee should clean up the stack, not caller. Symbols should be also
2440 // decorated in some fancy way :) It doesn't support any vector arguments.
2441 // For info on fast calling convention see Fast Calling Convention (tail call)
2442 // implementation LowerX86_32FastCCCallTo.
2444 /// CallIsStructReturn - Determines whether a call uses struct return
2446 enum StructReturnType {
2451 static StructReturnType
2452 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsMCU) {
2454 return NotStructReturn;
2456 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2457 if (!Flags.isSRet())
2458 return NotStructReturn;
2459 if (Flags.isInReg() || IsMCU)
2460 return RegStructReturn;
2461 return StackStructReturn;
2464 /// Determines whether a function uses struct return semantics.
2465 static StructReturnType
2466 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins, bool IsMCU) {
2468 return NotStructReturn;
2470 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2471 if (!Flags.isSRet())
2472 return NotStructReturn;
2473 if (Flags.isInReg() || IsMCU)
2474 return RegStructReturn;
2475 return StackStructReturn;
2478 /// Make a copy of an aggregate at address specified by "Src" to address
2479 /// "Dst" with size and alignment information specified by the specific
2480 /// parameter attribute. The copy will be passed as a byval function parameter.
2482 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2483 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2485 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2487 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2488 /*isVolatile*/false, /*AlwaysInline=*/true,
2489 /*isTailCall*/false,
2490 MachinePointerInfo(), MachinePointerInfo());
2493 /// Return true if the calling convention is one that we can guarantee TCO for.
2494 static bool canGuaranteeTCO(CallingConv::ID CC) {
2495 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2496 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2499 /// Return true if we might ever do TCO for calls with this calling convention.
2500 static bool mayTailCallThisCC(CallingConv::ID CC) {
2502 // C calling conventions:
2503 case CallingConv::C:
2504 case CallingConv::X86_64_Win64:
2505 case CallingConv::X86_64_SysV:
2506 // Callee pop conventions:
2507 case CallingConv::X86_ThisCall:
2508 case CallingConv::X86_StdCall:
2509 case CallingConv::X86_VectorCall:
2510 case CallingConv::X86_FastCall:
2513 return canGuaranteeTCO(CC);
2517 /// Return true if the function is being made into a tailcall target by
2518 /// changing its ABI.
2519 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2520 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2523 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2525 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2526 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2530 CallingConv::ID CalleeCC = CS.getCallingConv();
2531 if (!mayTailCallThisCC(CalleeCC))
2538 X86TargetLowering::LowerMemArgument(SDValue Chain,
2539 CallingConv::ID CallConv,
2540 const SmallVectorImpl<ISD::InputArg> &Ins,
2541 SDLoc dl, SelectionDAG &DAG,
2542 const CCValAssign &VA,
2543 MachineFrameInfo *MFI,
2545 // Create the nodes corresponding to a load from this parameter slot.
2546 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2547 bool AlwaysUseMutable = shouldGuaranteeTCO(
2548 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2549 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2552 // If value is passed by pointer we have address passed instead of the value
2554 bool ExtendedInMem = VA.isExtInLoc() &&
2555 VA.getValVT().getScalarType() == MVT::i1;
2557 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2558 ValVT = VA.getLocVT();
2560 ValVT = VA.getValVT();
2562 // Calculate SP offset of interrupt parameter, re-arrange the slot normally
2563 // taken by a return address.
2565 if (CallConv == CallingConv::X86_INTR) {
2566 const X86Subtarget& Subtarget =
2567 static_cast<const X86Subtarget&>(DAG.getSubtarget());
2568 // X86 interrupts may take one or two arguments.
2569 // On the stack there will be no return address as in regular call.
2570 // Offset of last argument need to be set to -4/-8 bytes.
2571 // Where offset of the first argument out of two, should be set to 0 bytes.
2572 Offset = (Subtarget.is64Bit() ? 8 : 4) * ((i + 1) % Ins.size() - 1);
2575 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2576 // changed with more analysis.
2577 // In case of tail call optimization mark all arguments mutable. Since they
2578 // could be overwritten by lowering of arguments in case of a tail call.
2579 if (Flags.isByVal()) {
2580 unsigned Bytes = Flags.getByValSize();
2581 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2582 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2583 // Adjust SP offset of interrupt parameter.
2584 if (CallConv == CallingConv::X86_INTR) {
2585 MFI->setObjectOffset(FI, Offset);
2587 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2589 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2590 VA.getLocMemOffset(), isImmutable);
2591 // Adjust SP offset of interrupt parameter.
2592 if (CallConv == CallingConv::X86_INTR) {
2593 MFI->setObjectOffset(FI, Offset);
2596 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2597 SDValue Val = DAG.getLoad(
2598 ValVT, dl, Chain, FIN,
2599 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2601 return ExtendedInMem ?
2602 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2606 // FIXME: Get this from tablegen.
2607 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2608 const X86Subtarget *Subtarget) {
2609 assert(Subtarget->is64Bit());
2611 if (Subtarget->isCallingConvWin64(CallConv)) {
2612 static const MCPhysReg GPR64ArgRegsWin64[] = {
2613 X86::RCX, X86::RDX, X86::R8, X86::R9
2615 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2618 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2619 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2621 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2624 // FIXME: Get this from tablegen.
2625 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2626 CallingConv::ID CallConv,
2627 const X86Subtarget *Subtarget) {
2628 assert(Subtarget->is64Bit());
2629 if (Subtarget->isCallingConvWin64(CallConv)) {
2630 // The XMM registers which might contain var arg parameters are shadowed
2631 // in their paired GPR. So we only need to save the GPR to their home
2633 // TODO: __vectorcall will change this.
2637 const Function *Fn = MF.getFunction();
2638 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2639 bool isSoftFloat = Subtarget->useSoftFloat();
2640 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2641 "SSE register cannot be used when SSE is disabled!");
2642 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2643 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2647 static const MCPhysReg XMMArgRegs64Bit[] = {
2648 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2649 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2651 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2654 SDValue X86TargetLowering::LowerFormalArguments(
2655 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2656 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2657 SmallVectorImpl<SDValue> &InVals) const {
2658 MachineFunction &MF = DAG.getMachineFunction();
2659 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2660 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2662 const Function* Fn = MF.getFunction();
2663 if (Fn->hasExternalLinkage() &&
2664 Subtarget->isTargetCygMing() &&
2665 Fn->getName() == "main")
2666 FuncInfo->setForceFramePointer(true);
2668 MachineFrameInfo *MFI = MF.getFrameInfo();
2669 bool Is64Bit = Subtarget->is64Bit();
2670 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2672 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2673 "Var args not supported with calling convention fastcc, ghc or hipe");
2675 if (CallConv == CallingConv::X86_INTR) {
2676 bool isLegal = Ins.size() == 1 ||
2677 (Ins.size() == 2 && ((Is64Bit && Ins[1].VT == MVT::i64) ||
2678 (!Is64Bit && Ins[1].VT == MVT::i32)));
2680 report_fatal_error("X86 interrupts may take one or two arguments");
2683 // Assign locations to all of the incoming arguments.
2684 SmallVector<CCValAssign, 16> ArgLocs;
2685 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2687 // Allocate shadow area for Win64
2689 CCInfo.AllocateStack(32, 8);
2691 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2693 unsigned LastVal = ~0U;
2695 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2696 CCValAssign &VA = ArgLocs[i];
2697 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2699 assert(VA.getValNo() != LastVal &&
2700 "Don't support value assigned to multiple locs yet");
2702 LastVal = VA.getValNo();
2704 if (VA.isRegLoc()) {
2705 EVT RegVT = VA.getLocVT();
2706 const TargetRegisterClass *RC;
2707 if (RegVT == MVT::i32)
2708 RC = &X86::GR32RegClass;
2709 else if (Is64Bit && RegVT == MVT::i64)
2710 RC = &X86::GR64RegClass;
2711 else if (RegVT == MVT::f32)
2712 RC = &X86::FR32RegClass;
2713 else if (RegVT == MVT::f64)
2714 RC = &X86::FR64RegClass;
2715 else if (RegVT == MVT::f128)
2716 RC = &X86::FR128RegClass;
2717 else if (RegVT.is512BitVector())
2718 RC = &X86::VR512RegClass;
2719 else if (RegVT.is256BitVector())
2720 RC = &X86::VR256RegClass;
2721 else if (RegVT.is128BitVector())
2722 RC = &X86::VR128RegClass;
2723 else if (RegVT == MVT::x86mmx)
2724 RC = &X86::VR64RegClass;
2725 else if (RegVT == MVT::i1)
2726 RC = &X86::VK1RegClass;
2727 else if (RegVT == MVT::v8i1)
2728 RC = &X86::VK8RegClass;
2729 else if (RegVT == MVT::v16i1)
2730 RC = &X86::VK16RegClass;
2731 else if (RegVT == MVT::v32i1)
2732 RC = &X86::VK32RegClass;
2733 else if (RegVT == MVT::v64i1)
2734 RC = &X86::VK64RegClass;
2736 llvm_unreachable("Unknown argument type!");
2738 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2739 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2741 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2742 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2744 if (VA.getLocInfo() == CCValAssign::SExt)
2745 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2746 DAG.getValueType(VA.getValVT()));
2747 else if (VA.getLocInfo() == CCValAssign::ZExt)
2748 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2749 DAG.getValueType(VA.getValVT()));
2750 else if (VA.getLocInfo() == CCValAssign::BCvt)
2751 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2753 if (VA.isExtInLoc()) {
2754 // Handle MMX values passed in XMM regs.
2755 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2756 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2758 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2761 assert(VA.isMemLoc());
2762 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2765 // If value is passed via pointer - do a load.
2766 if (VA.getLocInfo() == CCValAssign::Indirect)
2767 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2768 MachinePointerInfo(), false, false, false, 0);
2770 InVals.push_back(ArgValue);
2773 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2774 // All x86 ABIs require that for returning structs by value we copy the
2775 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2776 // the argument into a virtual register so that we can access it from the
2778 if (Ins[i].Flags.isSRet()) {
2779 unsigned Reg = FuncInfo->getSRetReturnReg();
2781 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2782 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2783 FuncInfo->setSRetReturnReg(Reg);
2785 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2786 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2791 unsigned StackSize = CCInfo.getNextStackOffset();
2792 // Align stack specially for tail calls.
2793 if (shouldGuaranteeTCO(CallConv,
2794 MF.getTarget().Options.GuaranteedTailCallOpt))
2795 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2797 // If the function takes variable number of arguments, make a frame index for
2798 // the start of the first vararg value... for expansion of llvm.va_start. We
2799 // can skip this if there are no va_start calls.
2800 if (MFI->hasVAStart() &&
2801 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2802 CallConv != CallingConv::X86_ThisCall))) {
2803 FuncInfo->setVarArgsFrameIndex(
2804 MFI->CreateFixedObject(1, StackSize, true));
2807 // Figure out if XMM registers are in use.
2808 assert(!(Subtarget->useSoftFloat() &&
2809 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2810 "SSE register cannot be used when SSE is disabled!");
2812 // 64-bit calling conventions support varargs and register parameters, so we
2813 // have to do extra work to spill them in the prologue.
2814 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2815 // Find the first unallocated argument registers.
2816 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2817 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2818 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2819 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2820 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2821 "SSE register cannot be used when SSE is disabled!");
2823 // Gather all the live in physical registers.
2824 SmallVector<SDValue, 6> LiveGPRs;
2825 SmallVector<SDValue, 8> LiveXMMRegs;
2827 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2828 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2830 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2832 if (!ArgXMMs.empty()) {
2833 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2834 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2835 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2836 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2837 LiveXMMRegs.push_back(
2838 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2843 // Get to the caller-allocated home save location. Add 8 to account
2844 // for the return address.
2845 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2846 FuncInfo->setRegSaveFrameIndex(
2847 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2848 // Fixup to set vararg frame on shadow area (4 x i64).
2850 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2852 // For X86-64, if there are vararg parameters that are passed via
2853 // registers, then we must store them to their spots on the stack so
2854 // they may be loaded by deferencing the result of va_next.
2855 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2856 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2857 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2858 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2861 // Store the integer parameter registers.
2862 SmallVector<SDValue, 8> MemOps;
2863 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2864 getPointerTy(DAG.getDataLayout()));
2865 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2866 for (SDValue Val : LiveGPRs) {
2867 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2868 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2870 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2871 MachinePointerInfo::getFixedStack(
2872 DAG.getMachineFunction(),
2873 FuncInfo->getRegSaveFrameIndex(), Offset),
2875 MemOps.push_back(Store);
2879 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2880 // Now store the XMM (fp + vector) parameter registers.
2881 SmallVector<SDValue, 12> SaveXMMOps;
2882 SaveXMMOps.push_back(Chain);
2883 SaveXMMOps.push_back(ALVal);
2884 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2885 FuncInfo->getRegSaveFrameIndex(), dl));
2886 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2887 FuncInfo->getVarArgsFPOffset(), dl));
2888 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2890 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2891 MVT::Other, SaveXMMOps));
2894 if (!MemOps.empty())
2895 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2898 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2899 // Find the largest legal vector type.
2900 MVT VecVT = MVT::Other;
2901 // FIXME: Only some x86_32 calling conventions support AVX512.
2902 if (Subtarget->hasAVX512() &&
2903 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2904 CallConv == CallingConv::Intel_OCL_BI)))
2905 VecVT = MVT::v16f32;
2906 else if (Subtarget->hasAVX())
2908 else if (Subtarget->hasSSE2())
2911 // We forward some GPRs and some vector types.
2912 SmallVector<MVT, 2> RegParmTypes;
2913 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2914 RegParmTypes.push_back(IntVT);
2915 if (VecVT != MVT::Other)
2916 RegParmTypes.push_back(VecVT);
2918 // Compute the set of forwarded registers. The rest are scratch.
2919 SmallVectorImpl<ForwardedRegister> &Forwards =
2920 FuncInfo->getForwardedMustTailRegParms();
2921 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2923 // Conservatively forward AL on x86_64, since it might be used for varargs.
2924 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2925 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2926 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2929 // Copy all forwards from physical to virtual registers.
2930 for (ForwardedRegister &F : Forwards) {
2931 // FIXME: Can we use a less constrained schedule?
2932 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2933 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2934 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2938 // Some CCs need callee pop.
2939 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2940 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2941 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2942 } else if (CallConv == CallingConv::X86_INTR && Ins.size() == 2) {
2943 // X86 interrupts must pop the error code if present
2944 FuncInfo->setBytesToPopOnReturn(Is64Bit ? 8 : 4);
2946 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2947 // If this is an sret function, the return should pop the hidden pointer.
2948 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2949 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2950 argsAreStructReturn(Ins, Subtarget->isTargetMCU()) == StackStructReturn)
2951 FuncInfo->setBytesToPopOnReturn(4);
2955 // RegSaveFrameIndex is X86-64 only.
2956 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2957 if (CallConv == CallingConv::X86_FastCall ||
2958 CallConv == CallingConv::X86_ThisCall)
2959 // fastcc functions can't have varargs.
2960 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2963 FuncInfo->setArgumentStackSize(StackSize);
2965 if (WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo()) {
2966 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2967 if (Personality == EHPersonality::CoreCLR) {
2969 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2970 // that we'd prefer this slot be allocated towards the bottom of the frame
2971 // (i.e. near the stack pointer after allocating the frame). Every
2972 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2973 // offset from the bottom of this and each funclet's frame must be the
2974 // same, so the size of funclets' (mostly empty) frames is dictated by
2975 // how far this slot is from the bottom (since they allocate just enough
2976 // space to accomodate holding this slot at the correct offset).
2977 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2978 EHInfo->PSPSymFrameIdx = PSPSymFI;
2986 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2987 SDValue StackPtr, SDValue Arg,
2988 SDLoc dl, SelectionDAG &DAG,
2989 const CCValAssign &VA,
2990 ISD::ArgFlagsTy Flags) const {
2991 unsigned LocMemOffset = VA.getLocMemOffset();
2992 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2993 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2995 if (Flags.isByVal())
2996 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2998 return DAG.getStore(
2999 Chain, dl, Arg, PtrOff,
3000 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
3004 /// Emit a load of return address if tail call
3005 /// optimization is performed and it is required.
3007 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
3008 SDValue &OutRetAddr, SDValue Chain,
3009 bool IsTailCall, bool Is64Bit,
3010 int FPDiff, SDLoc dl) const {
3011 // Adjust the Return address stack slot.
3012 EVT VT = getPointerTy(DAG.getDataLayout());
3013 OutRetAddr = getReturnAddressFrameIndex(DAG);
3015 // Load the "old" Return address.
3016 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
3017 false, false, false, 0);
3018 return SDValue(OutRetAddr.getNode(), 1);
3021 /// Emit a store of the return address if tail call
3022 /// optimization is performed and it is required (FPDiff!=0).
3023 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
3024 SDValue Chain, SDValue RetAddrFrIdx,
3025 EVT PtrVT, unsigned SlotSize,
3026 int FPDiff, SDLoc dl) {
3027 // Store the return address to the appropriate stack slot.
3028 if (!FPDiff) return Chain;
3029 // Calculate the new stack slot for the return address.
3030 int NewReturnAddrFI =
3031 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
3033 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
3034 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
3035 MachinePointerInfo::getFixedStack(
3036 DAG.getMachineFunction(), NewReturnAddrFI),
3041 /// Returns a vector_shuffle mask for an movs{s|d}, movd
3042 /// operation of specified width.
3043 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
3045 unsigned NumElems = VT.getVectorNumElements();
3046 SmallVector<int, 8> Mask;
3047 Mask.push_back(NumElems);
3048 for (unsigned i = 1; i != NumElems; ++i)
3050 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3054 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3055 SmallVectorImpl<SDValue> &InVals) const {
3056 SelectionDAG &DAG = CLI.DAG;
3058 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3059 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3060 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3061 SDValue Chain = CLI.Chain;
3062 SDValue Callee = CLI.Callee;
3063 CallingConv::ID CallConv = CLI.CallConv;
3064 bool &isTailCall = CLI.IsTailCall;
3065 bool isVarArg = CLI.IsVarArg;
3067 MachineFunction &MF = DAG.getMachineFunction();
3068 bool Is64Bit = Subtarget->is64Bit();
3069 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3070 StructReturnType SR = callIsStructReturn(Outs, Subtarget->isTargetMCU());
3071 bool IsSibcall = false;
3072 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3073 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3075 if (CallConv == CallingConv::X86_INTR)
3076 report_fatal_error("X86 interrupts may not be called directly");
3078 if (Attr.getValueAsString() == "true")
3081 if (Subtarget->isPICStyleGOT() &&
3082 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3083 // If we are using a GOT, disable tail calls to external symbols with
3084 // default visibility. Tail calling such a symbol requires using a GOT
3085 // relocation, which forces early binding of the symbol. This breaks code
3086 // that require lazy function symbol resolution. Using musttail or
3087 // GuaranteedTailCallOpt will override this.
3088 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3089 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3090 G->getGlobal()->hasDefaultVisibility()))
3094 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3096 // Force this to be a tail call. The verifier rules are enough to ensure
3097 // that we can lower this successfully without moving the return address
3100 } else if (isTailCall) {
3101 // Check if it's really possible to do a tail call.
3102 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3103 isVarArg, SR != NotStructReturn,
3104 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3105 Outs, OutVals, Ins, DAG);
3107 // Sibcalls are automatically detected tailcalls which do not require
3109 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3116 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3117 "Var args not supported with calling convention fastcc, ghc or hipe");
3119 // Analyze operands of the call, assigning locations to each operand.
3120 SmallVector<CCValAssign, 16> ArgLocs;
3121 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3123 // Allocate shadow area for Win64
3125 CCInfo.AllocateStack(32, 8);
3127 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3129 // Get a count of how many bytes are to be pushed on the stack.
3130 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3132 // This is a sibcall. The memory operands are available in caller's
3133 // own caller's stack.
3135 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3136 canGuaranteeTCO(CallConv))
3137 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3140 if (isTailCall && !IsSibcall && !IsMustTail) {
3141 // Lower arguments at fp - stackoffset + fpdiff.
3142 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3144 FPDiff = NumBytesCallerPushed - NumBytes;
3146 // Set the delta of movement of the returnaddr stackslot.
3147 // But only set if delta is greater than previous delta.
3148 if (FPDiff < X86Info->getTCReturnAddrDelta())
3149 X86Info->setTCReturnAddrDelta(FPDiff);
3152 unsigned NumBytesToPush = NumBytes;
3153 unsigned NumBytesToPop = NumBytes;
3155 // If we have an inalloca argument, all stack space has already been allocated
3156 // for us and be right at the top of the stack. We don't support multiple
3157 // arguments passed in memory when using inalloca.
3158 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3160 if (!ArgLocs.back().isMemLoc())
3161 report_fatal_error("cannot use inalloca attribute on a register "
3163 if (ArgLocs.back().getLocMemOffset() != 0)
3164 report_fatal_error("any parameter with the inalloca attribute must be "
3165 "the only memory argument");
3169 Chain = DAG.getCALLSEQ_START(
3170 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3172 SDValue RetAddrFrIdx;
3173 // Load return address for tail calls.
3174 if (isTailCall && FPDiff)
3175 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3176 Is64Bit, FPDiff, dl);
3178 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3179 SmallVector<SDValue, 8> MemOpChains;
3182 // Walk the register/memloc assignments, inserting copies/loads. In the case
3183 // of tail call optimization arguments are handle later.
3184 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3185 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3186 // Skip inalloca arguments, they have already been written.
3187 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3188 if (Flags.isInAlloca())
3191 CCValAssign &VA = ArgLocs[i];
3192 EVT RegVT = VA.getLocVT();
3193 SDValue Arg = OutVals[i];
3194 bool isByVal = Flags.isByVal();
3196 // Promote the value if needed.
3197 switch (VA.getLocInfo()) {
3198 default: llvm_unreachable("Unknown loc info!");
3199 case CCValAssign::Full: break;
3200 case CCValAssign::SExt:
3201 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3203 case CCValAssign::ZExt:
3204 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3206 case CCValAssign::AExt:
3207 if (Arg.getValueType().isVector() &&
3208 Arg.getValueType().getVectorElementType() == MVT::i1)
3209 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3210 else if (RegVT.is128BitVector()) {
3211 // Special case: passing MMX values in XMM registers.
3212 Arg = DAG.getBitcast(MVT::i64, Arg);
3213 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3214 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3216 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3218 case CCValAssign::BCvt:
3219 Arg = DAG.getBitcast(RegVT, Arg);
3221 case CCValAssign::Indirect: {
3222 // Store the argument.
3223 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3224 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3225 Chain = DAG.getStore(
3226 Chain, dl, Arg, SpillSlot,
3227 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3234 if (VA.isRegLoc()) {
3235 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3236 if (isVarArg && IsWin64) {
3237 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3238 // shadow reg if callee is a varargs function.
3239 unsigned ShadowReg = 0;
3240 switch (VA.getLocReg()) {
3241 case X86::XMM0: ShadowReg = X86::RCX; break;
3242 case X86::XMM1: ShadowReg = X86::RDX; break;
3243 case X86::XMM2: ShadowReg = X86::R8; break;
3244 case X86::XMM3: ShadowReg = X86::R9; break;
3247 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3249 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3250 assert(VA.isMemLoc());
3251 if (!StackPtr.getNode())
3252 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3253 getPointerTy(DAG.getDataLayout()));
3254 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3255 dl, DAG, VA, Flags));
3259 if (!MemOpChains.empty())
3260 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3262 if (Subtarget->isPICStyleGOT()) {
3263 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3266 RegsToPass.push_back(std::make_pair(
3267 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3268 getPointerTy(DAG.getDataLayout()))));
3270 // If we are tail calling and generating PIC/GOT style code load the
3271 // address of the callee into ECX. The value in ecx is used as target of
3272 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3273 // for tail calls on PIC/GOT architectures. Normally we would just put the
3274 // address of GOT into ebx and then call target@PLT. But for tail calls
3275 // ebx would be restored (since ebx is callee saved) before jumping to the
3278 // Note: The actual moving to ECX is done further down.
3279 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3280 if (G && !G->getGlobal()->hasLocalLinkage() &&
3281 G->getGlobal()->hasDefaultVisibility())
3282 Callee = LowerGlobalAddress(Callee, DAG);
3283 else if (isa<ExternalSymbolSDNode>(Callee))
3284 Callee = LowerExternalSymbol(Callee, DAG);
3288 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3289 // From AMD64 ABI document:
3290 // For calls that may call functions that use varargs or stdargs
3291 // (prototype-less calls or calls to functions containing ellipsis (...) in
3292 // the declaration) %al is used as hidden argument to specify the number
3293 // of SSE registers used. The contents of %al do not need to match exactly
3294 // the number of registers, but must be an ubound on the number of SSE
3295 // registers used and is in the range 0 - 8 inclusive.
3297 // Count the number of XMM registers allocated.
3298 static const MCPhysReg XMMArgRegs[] = {
3299 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3300 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3302 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3303 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3304 && "SSE registers cannot be used when SSE is disabled");
3306 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3307 DAG.getConstant(NumXMMRegs, dl,
3311 if (isVarArg && IsMustTail) {
3312 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3313 for (const auto &F : Forwards) {
3314 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3315 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3319 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3320 // don't need this because the eligibility check rejects calls that require
3321 // shuffling arguments passed in memory.
3322 if (!IsSibcall && isTailCall) {
3323 // Force all the incoming stack arguments to be loaded from the stack
3324 // before any new outgoing arguments are stored to the stack, because the
3325 // outgoing stack slots may alias the incoming argument stack slots, and
3326 // the alias isn't otherwise explicit. This is slightly more conservative
3327 // than necessary, because it means that each store effectively depends
3328 // on every argument instead of just those arguments it would clobber.
3329 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3331 SmallVector<SDValue, 8> MemOpChains2;
3334 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3335 CCValAssign &VA = ArgLocs[i];
3338 assert(VA.isMemLoc());
3339 SDValue Arg = OutVals[i];
3340 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3341 // Skip inalloca arguments. They don't require any work.
3342 if (Flags.isInAlloca())
3344 // Create frame index.
3345 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3346 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3347 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3348 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3350 if (Flags.isByVal()) {
3351 // Copy relative to framepointer.
3352 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3353 if (!StackPtr.getNode())
3354 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3355 getPointerTy(DAG.getDataLayout()));
3356 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3359 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3363 // Store relative to framepointer.
3364 MemOpChains2.push_back(DAG.getStore(
3365 ArgChain, dl, Arg, FIN,
3366 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3371 if (!MemOpChains2.empty())
3372 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3374 // Store the return address to the appropriate stack slot.
3375 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3376 getPointerTy(DAG.getDataLayout()),
3377 RegInfo->getSlotSize(), FPDiff, dl);
3380 // Build a sequence of copy-to-reg nodes chained together with token chain
3381 // and flag operands which copy the outgoing args into registers.
3383 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3384 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3385 RegsToPass[i].second, InFlag);
3386 InFlag = Chain.getValue(1);
3389 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3390 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3391 // In the 64-bit large code model, we have to make all calls
3392 // through a register, since the call instruction's 32-bit
3393 // pc-relative offset may not be large enough to hold the whole
3395 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3396 // If the callee is a GlobalAddress node (quite common, every direct call
3397 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3399 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3401 // We should use extra load for direct calls to dllimported functions in
3403 const GlobalValue *GV = G->getGlobal();
3404 if (!GV->hasDLLImportStorageClass()) {
3405 unsigned char OpFlags = 0;
3406 bool ExtraLoad = false;
3407 unsigned WrapperKind = ISD::DELETED_NODE;
3409 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3410 // external symbols most go through the PLT in PIC mode. If the symbol
3411 // has hidden or protected visibility, or if it is static or local, then
3412 // we don't need to use the PLT - we can directly call it.
3413 if (Subtarget->isTargetELF() &&
3414 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3415 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3416 OpFlags = X86II::MO_PLT;
3417 } else if (Subtarget->isPICStyleStubAny() &&
3418 !GV->isStrongDefinitionForLinker() &&
3419 (!Subtarget->getTargetTriple().isMacOSX() ||
3420 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3421 // PC-relative references to external symbols should go through $stub,
3422 // unless we're building with the leopard linker or later, which
3423 // automatically synthesizes these stubs.
3424 OpFlags = X86II::MO_DARWIN_STUB;
3425 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3426 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3427 // If the function is marked as non-lazy, generate an indirect call
3428 // which loads from the GOT directly. This avoids runtime overhead
3429 // at the cost of eager binding (and one extra byte of encoding).
3430 OpFlags = X86II::MO_GOTPCREL;
3431 WrapperKind = X86ISD::WrapperRIP;
3435 Callee = DAG.getTargetGlobalAddress(
3436 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3438 // Add a wrapper if needed.
3439 if (WrapperKind != ISD::DELETED_NODE)
3440 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3441 getPointerTy(DAG.getDataLayout()), Callee);
3442 // Add extra indirection if needed.
3444 Callee = DAG.getLoad(
3445 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3446 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3449 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3450 unsigned char OpFlags = 0;
3452 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3453 // external symbols should go through the PLT.
3454 if (Subtarget->isTargetELF() &&
3455 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3456 OpFlags = X86II::MO_PLT;
3457 } else if (Subtarget->isPICStyleStubAny() &&
3458 (!Subtarget->getTargetTriple().isMacOSX() ||
3459 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3460 // PC-relative references to external symbols should go through $stub,
3461 // unless we're building with the leopard linker or later, which
3462 // automatically synthesizes these stubs.
3463 OpFlags = X86II::MO_DARWIN_STUB;
3466 Callee = DAG.getTargetExternalSymbol(
3467 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3468 } else if (Subtarget->isTarget64BitILP32() &&
3469 Callee->getValueType(0) == MVT::i32) {
3470 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3471 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3474 // Returns a chain & a flag for retval copy to use.
3475 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3476 SmallVector<SDValue, 8> Ops;
3478 if (!IsSibcall && isTailCall) {
3479 Chain = DAG.getCALLSEQ_END(Chain,
3480 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3481 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3482 InFlag = Chain.getValue(1);
3485 Ops.push_back(Chain);
3486 Ops.push_back(Callee);
3489 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3491 // Add argument registers to the end of the list so that they are known live
3493 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3494 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3495 RegsToPass[i].second.getValueType()));
3497 // Add a register mask operand representing the call-preserved registers.
3498 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3499 assert(Mask && "Missing call preserved mask for calling convention");
3501 // If this is an invoke in a 32-bit function using a funclet-based
3502 // personality, assume the function clobbers all registers. If an exception
3503 // is thrown, the runtime will not restore CSRs.
3504 // FIXME: Model this more precisely so that we can register allocate across
3505 // the normal edge and spill and fill across the exceptional edge.
3506 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3507 const Function *CallerFn = MF.getFunction();
3508 EHPersonality Pers =
3509 CallerFn->hasPersonalityFn()
3510 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3511 : EHPersonality::Unknown;
3512 if (isFuncletEHPersonality(Pers))
3513 Mask = RegInfo->getNoPreservedMask();
3516 Ops.push_back(DAG.getRegisterMask(Mask));
3518 if (InFlag.getNode())
3519 Ops.push_back(InFlag);
3523 //// If this is the first return lowered for this function, add the regs
3524 //// to the liveout set for the function.
3525 // This isn't right, although it's probably harmless on x86; liveouts
3526 // should be computed from returns not tail calls. Consider a void
3527 // function making a tail call to a function returning int.
3528 MF.getFrameInfo()->setHasTailCall();
3529 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3532 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3533 InFlag = Chain.getValue(1);
3535 // Create the CALLSEQ_END node.
3536 unsigned NumBytesForCalleeToPop;
3537 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3538 DAG.getTarget().Options.GuaranteedTailCallOpt))
3539 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3540 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3541 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3542 SR == StackStructReturn)
3543 // If this is a call to a struct-return function, the callee
3544 // pops the hidden struct pointer, so we have to push it back.
3545 // This is common for Darwin/X86, Linux & Mingw32 targets.
3546 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3547 NumBytesForCalleeToPop = 4;
3549 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3551 // Returns a flag for retval copy to use.
3553 Chain = DAG.getCALLSEQ_END(Chain,
3554 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3555 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3558 InFlag = Chain.getValue(1);
3561 // Handle result values, copying them out of physregs into vregs that we
3563 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3564 Ins, dl, DAG, InVals);
3567 //===----------------------------------------------------------------------===//
3568 // Fast Calling Convention (tail call) implementation
3569 //===----------------------------------------------------------------------===//
3571 // Like std call, callee cleans arguments, convention except that ECX is
3572 // reserved for storing the tail called function address. Only 2 registers are
3573 // free for argument passing (inreg). Tail call optimization is performed
3575 // * tailcallopt is enabled
3576 // * caller/callee are fastcc
3577 // On X86_64 architecture with GOT-style position independent code only local
3578 // (within module) calls are supported at the moment.
3579 // To keep the stack aligned according to platform abi the function
3580 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3581 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3582 // If a tail called function callee has more arguments than the caller the
3583 // caller needs to make sure that there is room to move the RETADDR to. This is
3584 // achieved by reserving an area the size of the argument delta right after the
3585 // original RETADDR, but before the saved framepointer or the spilled registers
3586 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3598 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3601 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3602 SelectionDAG& DAG) const {
3603 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3604 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3605 unsigned StackAlignment = TFI.getStackAlignment();
3606 uint64_t AlignMask = StackAlignment - 1;
3607 int64_t Offset = StackSize;
3608 unsigned SlotSize = RegInfo->getSlotSize();
3609 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3610 // Number smaller than 12 so just add the difference.
3611 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3613 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3614 Offset = ((~AlignMask) & Offset) + StackAlignment +
3615 (StackAlignment-SlotSize);
3620 /// Return true if the given stack call argument is already available in the
3621 /// same position (relatively) of the caller's incoming argument stack.
3623 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3624 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3625 const X86InstrInfo *TII) {
3626 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3628 if (Arg.getOpcode() == ISD::CopyFromReg) {
3629 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3630 if (!TargetRegisterInfo::isVirtualRegister(VR))
3632 MachineInstr *Def = MRI->getVRegDef(VR);
3635 if (!Flags.isByVal()) {
3636 if (!TII->isLoadFromStackSlot(Def, FI))
3639 unsigned Opcode = Def->getOpcode();
3640 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3641 Opcode == X86::LEA64_32r) &&
3642 Def->getOperand(1).isFI()) {
3643 FI = Def->getOperand(1).getIndex();
3644 Bytes = Flags.getByValSize();
3648 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3649 if (Flags.isByVal())
3650 // ByVal argument is passed in as a pointer but it's now being
3651 // dereferenced. e.g.
3652 // define @foo(%struct.X* %A) {
3653 // tail call @bar(%struct.X* byval %A)
3656 SDValue Ptr = Ld->getBasePtr();
3657 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3660 FI = FINode->getIndex();
3661 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3662 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3663 FI = FINode->getIndex();
3664 Bytes = Flags.getByValSize();
3668 assert(FI != INT_MAX);
3669 if (!MFI->isFixedObjectIndex(FI))
3671 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3674 /// Check whether the call is eligible for tail call optimization. Targets
3675 /// that want to do tail call optimization should implement this function.
3676 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3677 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3678 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3679 const SmallVectorImpl<ISD::OutputArg> &Outs,
3680 const SmallVectorImpl<SDValue> &OutVals,
3681 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3682 if (!mayTailCallThisCC(CalleeCC))
3685 // If -tailcallopt is specified, make fastcc functions tail-callable.
3686 MachineFunction &MF = DAG.getMachineFunction();
3687 const Function *CallerF = MF.getFunction();
3689 // If the function return type is x86_fp80 and the callee return type is not,
3690 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3691 // perform a tailcall optimization here.
3692 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3695 CallingConv::ID CallerCC = CallerF->getCallingConv();
3696 bool CCMatch = CallerCC == CalleeCC;
3697 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3698 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3700 // Win64 functions have extra shadow space for argument homing. Don't do the
3701 // sibcall if the caller and callee have mismatched expectations for this
3703 if (IsCalleeWin64 != IsCallerWin64)
3706 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3707 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3712 // Look for obvious safe cases to perform tail call optimization that do not
3713 // require ABI changes. This is what gcc calls sibcall.
3715 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3716 // emit a special epilogue.
3717 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3718 if (RegInfo->needsStackRealignment(MF))
3721 // Also avoid sibcall optimization if either caller or callee uses struct
3722 // return semantics.
3723 if (isCalleeStructRet || isCallerStructRet)
3726 // Do not sibcall optimize vararg calls unless all arguments are passed via
3728 if (isVarArg && !Outs.empty()) {
3729 // Optimizing for varargs on Win64 is unlikely to be safe without
3730 // additional testing.
3731 if (IsCalleeWin64 || IsCallerWin64)
3734 SmallVector<CCValAssign, 16> ArgLocs;
3735 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3738 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3739 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3740 if (!ArgLocs[i].isRegLoc())
3744 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3745 // stack. Therefore, if it's not used by the call it is not safe to optimize
3746 // this into a sibcall.
3747 bool Unused = false;
3748 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3755 SmallVector<CCValAssign, 16> RVLocs;
3756 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3758 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3759 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3760 CCValAssign &VA = RVLocs[i];
3761 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3766 // If the calling conventions do not match, then we'd better make sure the
3767 // results are returned in the same way as what the caller expects.
3769 SmallVector<CCValAssign, 16> RVLocs1;
3770 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3772 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3774 SmallVector<CCValAssign, 16> RVLocs2;
3775 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3777 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3779 if (RVLocs1.size() != RVLocs2.size())
3781 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3782 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3784 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3786 if (RVLocs1[i].isRegLoc()) {
3787 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3790 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3796 unsigned StackArgsSize = 0;
3798 // If the callee takes no arguments then go on to check the results of the
3800 if (!Outs.empty()) {
3801 // Check if stack adjustment is needed. For now, do not do this if any
3802 // argument is passed on the stack.
3803 SmallVector<CCValAssign, 16> ArgLocs;
3804 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3807 // Allocate shadow area for Win64
3809 CCInfo.AllocateStack(32, 8);
3811 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3812 StackArgsSize = CCInfo.getNextStackOffset();
3814 if (CCInfo.getNextStackOffset()) {
3815 // Check if the arguments are already laid out in the right way as
3816 // the caller's fixed stack objects.
3817 MachineFrameInfo *MFI = MF.getFrameInfo();
3818 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3819 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3820 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3821 CCValAssign &VA = ArgLocs[i];
3822 SDValue Arg = OutVals[i];
3823 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3824 if (VA.getLocInfo() == CCValAssign::Indirect)
3826 if (!VA.isRegLoc()) {
3827 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3834 // If the tailcall address may be in a register, then make sure it's
3835 // possible to register allocate for it. In 32-bit, the call address can
3836 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3837 // callee-saved registers are restored. These happen to be the same
3838 // registers used to pass 'inreg' arguments so watch out for those.
3839 if (!Subtarget->is64Bit() &&
3840 ((!isa<GlobalAddressSDNode>(Callee) &&
3841 !isa<ExternalSymbolSDNode>(Callee)) ||
3842 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3843 unsigned NumInRegs = 0;
3844 // In PIC we need an extra register to formulate the address computation
3846 unsigned MaxInRegs =
3847 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3849 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3850 CCValAssign &VA = ArgLocs[i];
3853 unsigned Reg = VA.getLocReg();
3856 case X86::EAX: case X86::EDX: case X86::ECX:
3857 if (++NumInRegs == MaxInRegs)
3865 bool CalleeWillPop =
3866 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3867 MF.getTarget().Options.GuaranteedTailCallOpt);
3869 if (unsigned BytesToPop =
3870 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3871 // If we have bytes to pop, the callee must pop them.
3872 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3873 if (!CalleePopMatches)
3875 } else if (CalleeWillPop && StackArgsSize > 0) {
3876 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3884 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3885 const TargetLibraryInfo *libInfo) const {
3886 return X86::createFastISel(funcInfo, libInfo);
3889 //===----------------------------------------------------------------------===//
3890 // Other Lowering Hooks
3891 //===----------------------------------------------------------------------===//
3893 static bool MayFoldLoad(SDValue Op) {
3894 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3897 static bool MayFoldIntoStore(SDValue Op) {
3898 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3901 static bool isTargetShuffle(unsigned Opcode) {
3903 default: return false;
3904 case X86ISD::BLENDI:
3905 case X86ISD::PSHUFB:
3906 case X86ISD::PSHUFD:
3907 case X86ISD::PSHUFHW:
3908 case X86ISD::PSHUFLW:
3910 case X86ISD::PALIGNR:
3911 case X86ISD::MOVLHPS:
3912 case X86ISD::MOVLHPD:
3913 case X86ISD::MOVHLPS:
3914 case X86ISD::MOVLPS:
3915 case X86ISD::MOVLPD:
3916 case X86ISD::MOVSHDUP:
3917 case X86ISD::MOVSLDUP:
3918 case X86ISD::MOVDDUP:
3921 case X86ISD::UNPCKL:
3922 case X86ISD::UNPCKH:
3923 case X86ISD::VPERMILPI:
3924 case X86ISD::VPERM2X128:
3925 case X86ISD::VPERMI:
3926 case X86ISD::VPERMV:
3927 case X86ISD::VPERMV3:
3932 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3933 SDValue V1, unsigned TargetMask,
3934 SelectionDAG &DAG) {
3936 default: llvm_unreachable("Unknown x86 shuffle node");
3937 case X86ISD::PSHUFD:
3938 case X86ISD::PSHUFHW:
3939 case X86ISD::PSHUFLW:
3940 case X86ISD::VPERMILPI:
3941 case X86ISD::VPERMI:
3942 return DAG.getNode(Opc, dl, VT, V1,
3943 DAG.getConstant(TargetMask, dl, MVT::i8));
3947 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3948 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3950 default: llvm_unreachable("Unknown x86 shuffle node");
3951 case X86ISD::MOVLHPS:
3952 case X86ISD::MOVLHPD:
3953 case X86ISD::MOVHLPS:
3954 case X86ISD::MOVLPS:
3955 case X86ISD::MOVLPD:
3958 case X86ISD::UNPCKL:
3959 case X86ISD::UNPCKH:
3960 return DAG.getNode(Opc, dl, VT, V1, V2);
3964 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3965 MachineFunction &MF = DAG.getMachineFunction();
3966 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3967 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3968 int ReturnAddrIndex = FuncInfo->getRAIndex();
3970 if (ReturnAddrIndex == 0) {
3971 // Set up a frame object for the return address.
3972 unsigned SlotSize = RegInfo->getSlotSize();
3973 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3976 FuncInfo->setRAIndex(ReturnAddrIndex);
3979 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3982 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3983 bool hasSymbolicDisplacement) {
3984 // Offset should fit into 32 bit immediate field.
3985 if (!isInt<32>(Offset))
3988 // If we don't have a symbolic displacement - we don't have any extra
3990 if (!hasSymbolicDisplacement)
3993 // FIXME: Some tweaks might be needed for medium code model.
3994 if (M != CodeModel::Small && M != CodeModel::Kernel)
3997 // For small code model we assume that latest object is 16MB before end of 31
3998 // bits boundary. We may also accept pretty large negative constants knowing
3999 // that all objects are in the positive half of address space.
4000 if (M == CodeModel::Small && Offset < 16*1024*1024)
4003 // For kernel code model we know that all object resist in the negative half
4004 // of 32bits address space. We may not accept negative offsets, since they may
4005 // be just off and we may accept pretty large positive ones.
4006 if (M == CodeModel::Kernel && Offset >= 0)
4012 /// Determines whether the callee is required to pop its own arguments.
4013 /// Callee pop is necessary to support tail calls.
4014 bool X86::isCalleePop(CallingConv::ID CallingConv,
4015 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
4016 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
4017 // can guarantee TCO.
4018 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
4021 switch (CallingConv) {
4024 case CallingConv::X86_StdCall:
4025 case CallingConv::X86_FastCall:
4026 case CallingConv::X86_ThisCall:
4027 case CallingConv::X86_VectorCall:
4032 /// \brief Return true if the condition is an unsigned comparison operation.
4033 static bool isX86CCUnsigned(unsigned X86CC) {
4035 default: llvm_unreachable("Invalid integer condition!");
4036 case X86::COND_E: return true;
4037 case X86::COND_G: return false;
4038 case X86::COND_GE: return false;
4039 case X86::COND_L: return false;
4040 case X86::COND_LE: return false;
4041 case X86::COND_NE: return true;
4042 case X86::COND_B: return true;
4043 case X86::COND_A: return true;
4044 case X86::COND_BE: return true;
4045 case X86::COND_AE: return true;
4049 static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
4050 switch (SetCCOpcode) {
4051 default: llvm_unreachable("Invalid integer condition!");
4052 case ISD::SETEQ: return X86::COND_E;
4053 case ISD::SETGT: return X86::COND_G;
4054 case ISD::SETGE: return X86::COND_GE;
4055 case ISD::SETLT: return X86::COND_L;
4056 case ISD::SETLE: return X86::COND_LE;
4057 case ISD::SETNE: return X86::COND_NE;
4058 case ISD::SETULT: return X86::COND_B;
4059 case ISD::SETUGT: return X86::COND_A;
4060 case ISD::SETULE: return X86::COND_BE;
4061 case ISD::SETUGE: return X86::COND_AE;
4065 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
4066 /// condition code, returning the condition code and the LHS/RHS of the
4067 /// comparison to make.
4068 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4069 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4071 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4072 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4073 // X > -1 -> X == 0, jump !sign.
4074 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4075 return X86::COND_NS;
4077 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4078 // X < 0 -> X == 0, jump on sign.
4081 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4083 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4084 return X86::COND_LE;
4088 return TranslateIntegerX86CC(SetCCOpcode);
4091 // First determine if it is required or is profitable to flip the operands.
4093 // If LHS is a foldable load, but RHS is not, flip the condition.
4094 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4095 !ISD::isNON_EXTLoad(RHS.getNode())) {
4096 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4097 std::swap(LHS, RHS);
4100 switch (SetCCOpcode) {
4106 std::swap(LHS, RHS);
4110 // On a floating point condition, the flags are set as follows:
4112 // 0 | 0 | 0 | X > Y
4113 // 0 | 0 | 1 | X < Y
4114 // 1 | 0 | 0 | X == Y
4115 // 1 | 1 | 1 | unordered
4116 switch (SetCCOpcode) {
4117 default: llvm_unreachable("Condcode should be pre-legalized away");
4119 case ISD::SETEQ: return X86::COND_E;
4120 case ISD::SETOLT: // flipped
4122 case ISD::SETGT: return X86::COND_A;
4123 case ISD::SETOLE: // flipped
4125 case ISD::SETGE: return X86::COND_AE;
4126 case ISD::SETUGT: // flipped
4128 case ISD::SETLT: return X86::COND_B;
4129 case ISD::SETUGE: // flipped
4131 case ISD::SETLE: return X86::COND_BE;
4133 case ISD::SETNE: return X86::COND_NE;
4134 case ISD::SETUO: return X86::COND_P;
4135 case ISD::SETO: return X86::COND_NP;
4137 case ISD::SETUNE: return X86::COND_INVALID;
4141 /// Is there a floating point cmov for the specific X86 condition code?
4142 /// Current x86 isa includes the following FP cmov instructions:
4143 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4144 static bool hasFPCMov(unsigned X86CC) {
4160 /// Returns true if the target can instruction select the
4161 /// specified FP immediate natively. If false, the legalizer will
4162 /// materialize the FP immediate as a load from a constant pool.
4163 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4164 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4165 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4171 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4172 ISD::LoadExtType ExtTy,
4174 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4175 // relocation target a movq or addq instruction: don't let the load shrink.
4176 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4177 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4178 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4179 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4183 /// \brief Returns true if it is beneficial to convert a load of a constant
4184 /// to just the constant itself.
4185 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4187 assert(Ty->isIntegerTy());
4189 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4190 if (BitSize == 0 || BitSize > 64)
4195 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4196 unsigned Index) const {
4197 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4200 return (Index == 0 || Index == ResVT.getVectorNumElements());
4203 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4204 // Speculate cttz only if we can directly use TZCNT.
4205 return Subtarget->hasBMI();
4208 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4209 // Speculate ctlz only if we can directly use LZCNT.
4210 return Subtarget->hasLZCNT();
4213 /// Return true if every element in Mask, beginning
4214 /// from position Pos and ending in Pos+Size is undef.
4215 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4216 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4222 /// Return true if Val is undef or if its value falls within the
4223 /// specified range (L, H].
4224 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4225 return (Val < 0) || (Val >= Low && Val < Hi);
4228 /// Val is either less than zero (undef) or equal to the specified value.
4229 static bool isUndefOrEqual(int Val, int CmpVal) {
4230 return (Val < 0 || Val == CmpVal);
4233 /// Return true if every element in Mask, beginning
4234 /// from position Pos and ending in Pos+Size, falls within the specified
4235 /// sequential range (Low, Low+Size]. or is undef.
4236 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4237 unsigned Pos, unsigned Size, int Low) {
4238 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4239 if (!isUndefOrEqual(Mask[i], Low))
4244 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4245 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4246 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4247 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4248 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4251 // The index should be aligned on a vecWidth-bit boundary.
4253 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4255 MVT VT = N->getSimpleValueType(0);
4256 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4257 bool Result = (Index * ElSize) % vecWidth == 0;
4262 /// Return true if the specified INSERT_SUBVECTOR
4263 /// operand specifies a subvector insert that is suitable for input to
4264 /// insertion of 128 or 256-bit subvectors
4265 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4266 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4267 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4269 // The index should be aligned on a vecWidth-bit boundary.
4271 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4273 MVT VT = N->getSimpleValueType(0);
4274 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4275 bool Result = (Index * ElSize) % vecWidth == 0;
4280 bool X86::isVINSERT128Index(SDNode *N) {
4281 return isVINSERTIndex(N, 128);
4284 bool X86::isVINSERT256Index(SDNode *N) {
4285 return isVINSERTIndex(N, 256);
4288 bool X86::isVEXTRACT128Index(SDNode *N) {
4289 return isVEXTRACTIndex(N, 128);
4292 bool X86::isVEXTRACT256Index(SDNode *N) {
4293 return isVEXTRACTIndex(N, 256);
4296 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4297 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4298 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4299 "Illegal extract subvector for VEXTRACT");
4302 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4304 MVT VecVT = N->getOperand(0).getSimpleValueType();
4305 MVT ElVT = VecVT.getVectorElementType();
4307 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4308 return Index / NumElemsPerChunk;
4311 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4312 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4313 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4314 "Illegal insert subvector for VINSERT");
4317 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4319 MVT VecVT = N->getSimpleValueType(0);
4320 MVT ElVT = VecVT.getVectorElementType();
4322 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4323 return Index / NumElemsPerChunk;
4326 /// Return the appropriate immediate to extract the specified
4327 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4328 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4329 return getExtractVEXTRACTImmediate(N, 128);
4332 /// Return the appropriate immediate to extract the specified
4333 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4334 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4335 return getExtractVEXTRACTImmediate(N, 256);
4338 /// Return the appropriate immediate to insert at the specified
4339 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4340 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4341 return getInsertVINSERTImmediate(N, 128);
4344 /// Return the appropriate immediate to insert at the specified
4345 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4346 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4347 return getInsertVINSERTImmediate(N, 256);
4350 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4351 bool X86::isZeroNode(SDValue Elt) {
4352 return isNullConstant(Elt) || isNullFPConstant(Elt);
4355 // Build a vector of constants
4356 // Use an UNDEF node if MaskElt == -1.
4357 // Spilt 64-bit constants in the 32-bit mode.
4358 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4360 SDLoc dl, bool IsMask = false) {
4362 SmallVector<SDValue, 32> Ops;
4365 MVT ConstVecVT = VT;
4366 unsigned NumElts = VT.getVectorNumElements();
4367 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4368 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4369 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4373 MVT EltVT = ConstVecVT.getVectorElementType();
4374 for (unsigned i = 0; i < NumElts; ++i) {
4375 bool IsUndef = Values[i] < 0 && IsMask;
4376 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4377 DAG.getConstant(Values[i], dl, EltVT);
4378 Ops.push_back(OpNode);
4380 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4381 DAG.getConstant(0, dl, EltVT));
4383 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4385 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4389 /// Returns a vector of specified type with all zero elements.
4390 static SDValue getZeroVector(MVT VT, const X86Subtarget *Subtarget,
4391 SelectionDAG &DAG, SDLoc dl) {
4392 assert(VT.isVector() && "Expected a vector type");
4394 // Always build SSE zero vectors as <4 x i32> bitcasted
4395 // to their dest type. This ensures they get CSE'd.
4397 if (VT.is128BitVector()) { // SSE
4398 if (Subtarget->hasSSE2()) { // SSE2
4399 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4400 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4402 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4403 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4405 } else if (VT.is256BitVector()) { // AVX
4406 if (Subtarget->hasInt256()) { // AVX2
4407 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4408 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4409 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4411 // 256-bit logic and arithmetic instructions in AVX are all
4412 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4413 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4414 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4415 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4417 } else if (VT.is512BitVector()) { // AVX-512
4418 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4419 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4420 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4421 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4422 } else if (VT.getVectorElementType() == MVT::i1) {
4424 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4425 && "Unexpected vector type");
4426 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4427 && "Unexpected vector type");
4428 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4429 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4430 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4432 llvm_unreachable("Unexpected vector type");
4434 return DAG.getBitcast(VT, Vec);
4437 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4438 SelectionDAG &DAG, SDLoc dl,
4439 unsigned vectorWidth) {
4440 assert((vectorWidth == 128 || vectorWidth == 256) &&
4441 "Unsupported vector width");
4442 EVT VT = Vec.getValueType();
4443 EVT ElVT = VT.getVectorElementType();
4444 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4445 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4446 VT.getVectorNumElements()/Factor);
4448 // Extract from UNDEF is UNDEF.
4449 if (Vec.getOpcode() == ISD::UNDEF)
4450 return DAG.getUNDEF(ResultVT);
4452 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4453 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4454 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4456 // This is the index of the first element of the vectorWidth-bit chunk
4457 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4458 IdxVal &= ~(ElemsPerChunk - 1);
4460 // If the input is a buildvector just emit a smaller one.
4461 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4462 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4463 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4465 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4466 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4469 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4470 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4471 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4472 /// instructions or a simple subregister reference. Idx is an index in the
4473 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4474 /// lowering EXTRACT_VECTOR_ELT operations easier.
4475 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4476 SelectionDAG &DAG, SDLoc dl) {
4477 assert((Vec.getValueType().is256BitVector() ||
4478 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4479 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4482 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4483 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4484 SelectionDAG &DAG, SDLoc dl) {
4485 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4486 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4489 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4490 unsigned IdxVal, SelectionDAG &DAG,
4491 SDLoc dl, unsigned vectorWidth) {
4492 assert((vectorWidth == 128 || vectorWidth == 256) &&
4493 "Unsupported vector width");
4494 // Inserting UNDEF is Result
4495 if (Vec.getOpcode() == ISD::UNDEF)
4497 EVT VT = Vec.getValueType();
4498 EVT ElVT = VT.getVectorElementType();
4499 EVT ResultVT = Result.getValueType();
4501 // Insert the relevant vectorWidth bits.
4502 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4503 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4505 // This is the index of the first element of the vectorWidth-bit chunk
4506 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4507 IdxVal &= ~(ElemsPerChunk - 1);
4509 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4510 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4513 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4514 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4515 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4516 /// simple superregister reference. Idx is an index in the 128 bits
4517 /// we want. It need not be aligned to a 128-bit boundary. That makes
4518 /// lowering INSERT_VECTOR_ELT operations easier.
4519 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4520 SelectionDAG &DAG, SDLoc dl) {
4521 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4523 // For insertion into the zero index (low half) of a 256-bit vector, it is
4524 // more efficient to generate a blend with immediate instead of an insert*128.
4525 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4526 // extend the subvector to the size of the result vector. Make sure that
4527 // we are not recursing on that node by checking for undef here.
4528 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4529 Result.getOpcode() != ISD::UNDEF) {
4530 EVT ResultVT = Result.getValueType();
4531 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4532 SDValue Undef = DAG.getUNDEF(ResultVT);
4533 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4536 // The blend instruction, and therefore its mask, depend on the data type.
4537 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4538 if (ScalarType.isFloatingPoint()) {
4539 // Choose either vblendps (float) or vblendpd (double).
4540 unsigned ScalarSize = ScalarType.getSizeInBits();
4541 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4542 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4543 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4544 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4547 const X86Subtarget &Subtarget =
4548 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4550 // AVX2 is needed for 256-bit integer blend support.
4551 // Integers must be cast to 32-bit because there is only vpblendd;
4552 // vpblendw can't be used for this because it has a handicapped mask.
4554 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4555 // is still more efficient than using the wrong domain vinsertf128 that
4556 // will be created by InsertSubVector().
4557 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4559 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4560 Result = DAG.getBitcast(CastVT, Result);
4561 Vec256 = DAG.getBitcast(CastVT, Vec256);
4562 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4563 return DAG.getBitcast(ResultVT, Vec256);
4566 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4569 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4570 SelectionDAG &DAG, SDLoc dl) {
4571 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4572 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4575 /// Insert i1-subvector to i1-vector.
4576 static SDValue Insert1BitVector(SDValue Op, SelectionDAG &DAG) {
4579 SDValue Vec = Op.getOperand(0);
4580 SDValue SubVec = Op.getOperand(1);
4581 SDValue Idx = Op.getOperand(2);
4583 if (!isa<ConstantSDNode>(Idx))
4586 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
4587 if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
4590 MVT OpVT = Op.getSimpleValueType();
4591 MVT SubVecVT = SubVec.getSimpleValueType();
4592 unsigned NumElems = OpVT.getVectorNumElements();
4593 unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
4595 assert(IdxVal + SubVecNumElems <= NumElems &&
4596 IdxVal % SubVecVT.getSizeInBits() == 0 &&
4597 "Unexpected index value in INSERT_SUBVECTOR");
4599 // There are 3 possible cases:
4600 // 1. Subvector should be inserted in the lower part (IdxVal == 0)
4601 // 2. Subvector should be inserted in the upper part
4602 // (IdxVal + SubVecNumElems == NumElems)
4603 // 3. Subvector should be inserted in the middle (for example v2i1
4604 // to v16i1, index 2)
4606 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
4607 SDValue Undef = DAG.getUNDEF(OpVT);
4608 SDValue WideSubVec =
4609 DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef, SubVec, ZeroIdx);
4611 return DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4612 DAG.getConstant(IdxVal, dl, MVT::i8));
4614 if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
4615 unsigned ShiftLeft = NumElems - SubVecNumElems;
4616 unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
4617 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, WideSubVec,
4618 DAG.getConstant(ShiftLeft, dl, MVT::i8));
4619 return ShiftRight ? DAG.getNode(X86ISD::VSRLI, dl, OpVT, WideSubVec,
4620 DAG.getConstant(ShiftRight, dl, MVT::i8)) : WideSubVec;
4624 // Zero lower bits of the Vec
4625 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4626 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4627 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4628 // Merge them together
4629 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4632 // Simple case when we put subvector in the upper part
4633 if (IdxVal + SubVecNumElems == NumElems) {
4634 // Zero upper bits of the Vec
4635 WideSubVec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec,
4636 DAG.getConstant(IdxVal, dl, MVT::i8));
4637 SDValue ShiftBits = DAG.getConstant(SubVecNumElems, dl, MVT::i8);
4638 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
4639 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
4640 return DAG.getNode(ISD::OR, dl, OpVT, Vec, WideSubVec);
4642 // Subvector should be inserted in the middle - use shuffle
4643 SmallVector<int, 64> Mask;
4644 for (unsigned i = 0; i < NumElems; ++i)
4645 Mask.push_back(i >= IdxVal && i < IdxVal + SubVecNumElems ?
4647 return DAG.getVectorShuffle(OpVT, dl, WideSubVec, Vec, Mask);
4650 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4651 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4652 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4653 /// large BUILD_VECTORS.
4654 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4655 unsigned NumElems, SelectionDAG &DAG,
4657 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4658 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4661 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4662 unsigned NumElems, SelectionDAG &DAG,
4664 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4665 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4668 /// Returns a vector of specified type with all bits set.
4669 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4670 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4671 /// Then bitcast to their original type, ensuring they get CSE'd.
4672 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4673 SelectionDAG &DAG, SDLoc dl) {
4674 assert(VT.isVector() && "Expected a vector type");
4676 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4678 if (VT.is512BitVector()) {
4679 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4680 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4681 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4682 } else if (VT.is256BitVector()) {
4683 if (Subtarget->hasInt256()) { // AVX2
4684 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4685 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4687 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4688 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4690 } else if (VT.is128BitVector()) {
4691 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4693 llvm_unreachable("Unexpected vector type");
4695 return DAG.getBitcast(VT, Vec);
4698 /// Returns a vector_shuffle node for an unpackl operation.
4699 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4701 unsigned NumElems = VT.getVectorNumElements();
4702 SmallVector<int, 8> Mask;
4703 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4705 Mask.push_back(i + NumElems);
4707 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4710 /// Returns a vector_shuffle node for an unpackh operation.
4711 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4713 unsigned NumElems = VT.getVectorNumElements();
4714 SmallVector<int, 8> Mask;
4715 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4716 Mask.push_back(i + Half);
4717 Mask.push_back(i + NumElems + Half);
4719 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4722 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4723 /// This produces a shuffle where the low element of V2 is swizzled into the
4724 /// zero/undef vector, landing at element Idx.
4725 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4726 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4728 const X86Subtarget *Subtarget,
4729 SelectionDAG &DAG) {
4730 MVT VT = V2.getSimpleValueType();
4732 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4733 unsigned NumElems = VT.getVectorNumElements();
4734 SmallVector<int, 16> MaskVec;
4735 for (unsigned i = 0; i != NumElems; ++i)
4736 // If this is the insertion idx, put the low elt of V2 here.
4737 MaskVec.push_back(i == Idx ? NumElems : i);
4738 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4741 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4742 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4743 /// uses one source. Note that this will set IsUnary for shuffles which use a
4744 /// single input multiple times, and in those cases it will
4745 /// adjust the mask to only have indices within that single input.
4746 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4747 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4748 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4749 unsigned NumElems = VT.getVectorNumElements();
4753 bool IsFakeUnary = false;
4754 switch(N->getOpcode()) {
4755 case X86ISD::BLENDI:
4756 ImmN = N->getOperand(N->getNumOperands()-1);
4757 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4760 ImmN = N->getOperand(N->getNumOperands()-1);
4761 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4762 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4764 case X86ISD::UNPCKH:
4765 DecodeUNPCKHMask(VT, Mask);
4766 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4768 case X86ISD::UNPCKL:
4769 DecodeUNPCKLMask(VT, Mask);
4770 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4772 case X86ISD::MOVHLPS:
4773 DecodeMOVHLPSMask(NumElems, Mask);
4774 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4776 case X86ISD::MOVLHPS:
4777 DecodeMOVLHPSMask(NumElems, Mask);
4778 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4780 case X86ISD::PALIGNR:
4781 ImmN = N->getOperand(N->getNumOperands()-1);
4782 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4784 case X86ISD::PSHUFD:
4785 case X86ISD::VPERMILPI:
4786 ImmN = N->getOperand(N->getNumOperands()-1);
4787 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4790 case X86ISD::PSHUFHW:
4791 ImmN = N->getOperand(N->getNumOperands()-1);
4792 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4795 case X86ISD::PSHUFLW:
4796 ImmN = N->getOperand(N->getNumOperands()-1);
4797 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4800 case X86ISD::PSHUFB: {
4802 SDValue MaskNode = N->getOperand(1);
4803 while (MaskNode->getOpcode() == ISD::BITCAST)
4804 MaskNode = MaskNode->getOperand(0);
4806 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4807 // If we have a build-vector, then things are easy.
4808 MVT VT = MaskNode.getSimpleValueType();
4809 assert(VT.isVector() &&
4810 "Can't produce a non-vector with a build_vector!");
4811 if (!VT.isInteger())
4814 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4816 SmallVector<uint64_t, 32> RawMask;
4817 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4818 SDValue Op = MaskNode->getOperand(i);
4819 if (Op->getOpcode() == ISD::UNDEF) {
4820 RawMask.push_back((uint64_t)SM_SentinelUndef);
4823 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4826 APInt MaskElement = CN->getAPIntValue();
4828 // We now have to decode the element which could be any integer size and
4829 // extract each byte of it.
4830 for (int j = 0; j < NumBytesPerElement; ++j) {
4831 // Note that this is x86 and so always little endian: the low byte is
4832 // the first byte of the mask.
4833 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4834 MaskElement = MaskElement.lshr(8);
4837 DecodePSHUFBMask(RawMask, Mask);
4841 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4845 SDValue Ptr = MaskLoad->getBasePtr();
4846 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4847 Ptr->getOpcode() == X86ISD::WrapperRIP)
4848 Ptr = Ptr->getOperand(0);
4850 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4851 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4854 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4855 DecodePSHUFBMask(C, Mask);
4863 case X86ISD::VPERMI:
4864 ImmN = N->getOperand(N->getNumOperands()-1);
4865 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4870 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4872 case X86ISD::VPERM2X128:
4873 ImmN = N->getOperand(N->getNumOperands()-1);
4874 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4875 if (Mask.empty()) return false;
4876 // Mask only contains negative index if an element is zero.
4877 if (std::any_of(Mask.begin(), Mask.end(),
4878 [](int M){ return M == SM_SentinelZero; }))
4881 case X86ISD::MOVSLDUP:
4882 DecodeMOVSLDUPMask(VT, Mask);
4885 case X86ISD::MOVSHDUP:
4886 DecodeMOVSHDUPMask(VT, Mask);
4889 case X86ISD::MOVDDUP:
4890 DecodeMOVDDUPMask(VT, Mask);
4893 case X86ISD::MOVLHPD:
4894 case X86ISD::MOVLPD:
4895 case X86ISD::MOVLPS:
4896 // Not yet implemented
4898 case X86ISD::VPERMV: {
4900 SDValue MaskNode = N->getOperand(0);
4901 while (MaskNode->getOpcode() == ISD::BITCAST)
4902 MaskNode = MaskNode->getOperand(0);
4904 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4905 SmallVector<uint64_t, 32> RawMask;
4906 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4907 // If we have a build-vector, then things are easy.
4908 assert(MaskNode.getSimpleValueType().isInteger() &&
4909 MaskNode.getSimpleValueType().getVectorNumElements() ==
4910 VT.getVectorNumElements());
4912 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4913 SDValue Op = MaskNode->getOperand(i);
4914 if (Op->getOpcode() == ISD::UNDEF)
4915 RawMask.push_back((uint64_t)SM_SentinelUndef);
4916 else if (isa<ConstantSDNode>(Op)) {
4917 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4918 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4922 DecodeVPERMVMask(RawMask, Mask);
4925 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4926 unsigned NumEltsInMask = MaskNode->getNumOperands();
4927 MaskNode = MaskNode->getOperand(0);
4928 if (auto *CN = dyn_cast<ConstantSDNode>(MaskNode)) {
4929 APInt MaskEltValue = CN->getAPIntValue();
4930 for (unsigned i = 0; i < NumEltsInMask; ++i)
4931 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4932 DecodeVPERMVMask(RawMask, Mask);
4935 // It may be a scalar load
4938 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4942 SDValue Ptr = MaskLoad->getBasePtr();
4943 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4944 Ptr->getOpcode() == X86ISD::WrapperRIP)
4945 Ptr = Ptr->getOperand(0);
4947 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4948 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4951 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4952 DecodeVPERMVMask(C, VT, Mask);
4959 case X86ISD::VPERMV3: {
4961 SDValue MaskNode = N->getOperand(1);
4962 while (MaskNode->getOpcode() == ISD::BITCAST)
4963 MaskNode = MaskNode->getOperand(1);
4965 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4966 // If we have a build-vector, then things are easy.
4967 assert(MaskNode.getSimpleValueType().isInteger() &&
4968 MaskNode.getSimpleValueType().getVectorNumElements() ==
4969 VT.getVectorNumElements());
4971 SmallVector<uint64_t, 32> RawMask;
4972 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4974 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4975 SDValue Op = MaskNode->getOperand(i);
4976 if (Op->getOpcode() == ISD::UNDEF)
4977 RawMask.push_back((uint64_t)SM_SentinelUndef);
4979 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4982 APInt MaskElement = CN->getAPIntValue();
4983 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4986 DecodeVPERMV3Mask(RawMask, Mask);
4990 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4994 SDValue Ptr = MaskLoad->getBasePtr();
4995 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4996 Ptr->getOpcode() == X86ISD::WrapperRIP)
4997 Ptr = Ptr->getOperand(0);
4999 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
5000 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5003 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5004 DecodeVPERMV3Mask(C, VT, Mask);
5011 default: llvm_unreachable("unknown target shuffle node");
5014 // If we have a fake unary shuffle, the shuffle mask is spread across two
5015 // inputs that are actually the same node. Re-map the mask to always point
5016 // into the first input.
5019 if (M >= (int)Mask.size())
5025 /// Returns the scalar element that will make up the ith
5026 /// element of the result of the vector shuffle.
5027 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5030 return SDValue(); // Limit search depth.
5032 SDValue V = SDValue(N, 0);
5033 EVT VT = V.getValueType();
5034 unsigned Opcode = V.getOpcode();
5036 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5037 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5038 int Elt = SV->getMaskElt(Index);
5041 return DAG.getUNDEF(VT.getVectorElementType());
5043 unsigned NumElems = VT.getVectorNumElements();
5044 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5045 : SV->getOperand(1);
5046 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5049 // Recurse into target specific vector shuffles to find scalars.
5050 if (isTargetShuffle(Opcode)) {
5051 MVT ShufVT = V.getSimpleValueType();
5052 unsigned NumElems = ShufVT.getVectorNumElements();
5053 SmallVector<int, 16> ShuffleMask;
5056 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5059 int Elt = ShuffleMask[Index];
5061 return DAG.getUNDEF(ShufVT.getVectorElementType());
5063 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5065 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5069 // Actual nodes that may contain scalar elements
5070 if (Opcode == ISD::BITCAST) {
5071 V = V.getOperand(0);
5072 EVT SrcVT = V.getValueType();
5073 unsigned NumElems = VT.getVectorNumElements();
5075 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5079 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5080 return (Index == 0) ? V.getOperand(0)
5081 : DAG.getUNDEF(VT.getVectorElementType());
5083 if (V.getOpcode() == ISD::BUILD_VECTOR)
5084 return V.getOperand(Index);
5089 /// Custom lower build_vector of v16i8.
5090 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5091 unsigned NumNonZero, unsigned NumZero,
5093 const X86Subtarget* Subtarget,
5094 const TargetLowering &TLI) {
5102 // SSE4.1 - use PINSRB to insert each byte directly.
5103 if (Subtarget->hasSSE41()) {
5104 for (unsigned i = 0; i < 16; ++i) {
5105 bool isNonZero = (NonZeros & (1 << i)) != 0;
5109 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
5111 V = DAG.getUNDEF(MVT::v16i8);
5114 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5115 MVT::v16i8, V, Op.getOperand(i),
5116 DAG.getIntPtrConstant(i, dl));
5123 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5124 for (unsigned i = 0; i < 16; ++i) {
5125 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5126 if (ThisIsNonZero && First) {
5128 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5130 V = DAG.getUNDEF(MVT::v8i16);
5135 SDValue ThisElt, LastElt;
5136 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5137 if (LastIsNonZero) {
5138 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5139 MVT::i16, Op.getOperand(i-1));
5141 if (ThisIsNonZero) {
5142 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5143 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5144 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5146 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5150 if (ThisElt.getNode())
5151 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5152 DAG.getIntPtrConstant(i/2, dl));
5156 return DAG.getBitcast(MVT::v16i8, V);
5159 /// Custom lower build_vector of v8i16.
5160 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5161 unsigned NumNonZero, unsigned NumZero,
5163 const X86Subtarget* Subtarget,
5164 const TargetLowering &TLI) {
5171 for (unsigned i = 0; i < 8; ++i) {
5172 bool isNonZero = (NonZeros & (1 << i)) != 0;
5176 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5178 V = DAG.getUNDEF(MVT::v8i16);
5181 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5182 MVT::v8i16, V, Op.getOperand(i),
5183 DAG.getIntPtrConstant(i, dl));
5190 /// Custom lower build_vector of v4i32 or v4f32.
5191 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5192 const X86Subtarget *Subtarget,
5193 const TargetLowering &TLI) {
5194 // Find all zeroable elements.
5195 std::bitset<4> Zeroable;
5196 for (int i=0; i < 4; ++i) {
5197 SDValue Elt = Op->getOperand(i);
5198 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5200 assert(Zeroable.size() - Zeroable.count() > 1 &&
5201 "We expect at least two non-zero elements!");
5203 // We only know how to deal with build_vector nodes where elements are either
5204 // zeroable or extract_vector_elt with constant index.
5205 SDValue FirstNonZero;
5206 unsigned FirstNonZeroIdx;
5207 for (unsigned i=0; i < 4; ++i) {
5210 SDValue Elt = Op->getOperand(i);
5211 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5212 !isa<ConstantSDNode>(Elt.getOperand(1)))
5214 // Make sure that this node is extracting from a 128-bit vector.
5215 MVT VT = Elt.getOperand(0).getSimpleValueType();
5216 if (!VT.is128BitVector())
5218 if (!FirstNonZero.getNode()) {
5220 FirstNonZeroIdx = i;
5224 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5225 SDValue V1 = FirstNonZero.getOperand(0);
5226 MVT VT = V1.getSimpleValueType();
5228 // See if this build_vector can be lowered as a blend with zero.
5230 unsigned EltMaskIdx, EltIdx;
5232 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5233 if (Zeroable[EltIdx]) {
5234 // The zero vector will be on the right hand side.
5235 Mask[EltIdx] = EltIdx+4;
5239 Elt = Op->getOperand(EltIdx);
5240 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5241 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5242 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5244 Mask[EltIdx] = EltIdx;
5248 // Let the shuffle legalizer deal with blend operations.
5249 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5250 if (V1.getSimpleValueType() != VT)
5251 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5252 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5255 // See if we can lower this build_vector to a INSERTPS.
5256 if (!Subtarget->hasSSE41())
5259 SDValue V2 = Elt.getOperand(0);
5260 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5263 bool CanFold = true;
5264 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5268 SDValue Current = Op->getOperand(i);
5269 SDValue SrcVector = Current->getOperand(0);
5272 CanFold = SrcVector == V1 &&
5273 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5279 assert(V1.getNode() && "Expected at least two non-zero elements!");
5280 if (V1.getSimpleValueType() != MVT::v4f32)
5281 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5282 if (V2.getSimpleValueType() != MVT::v4f32)
5283 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5285 // Ok, we can emit an INSERTPS instruction.
5286 unsigned ZMask = Zeroable.to_ulong();
5288 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5289 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5291 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5292 DAG.getIntPtrConstant(InsertPSMask, DL));
5293 return DAG.getBitcast(VT, Result);
5296 /// Return a vector logical shift node.
5297 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5298 unsigned NumBits, SelectionDAG &DAG,
5299 const TargetLowering &TLI, SDLoc dl) {
5300 assert(VT.is128BitVector() && "Unknown type for VShift");
5301 MVT ShVT = MVT::v2i64;
5302 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5303 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5304 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5305 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5306 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5307 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5311 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5313 // Check if the scalar load can be widened into a vector load. And if
5314 // the address is "base + cst" see if the cst can be "absorbed" into
5315 // the shuffle mask.
5316 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5317 SDValue Ptr = LD->getBasePtr();
5318 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5320 EVT PVT = LD->getValueType(0);
5321 if (PVT != MVT::i32 && PVT != MVT::f32)
5326 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5327 FI = FINode->getIndex();
5329 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5330 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5331 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5332 Offset = Ptr.getConstantOperandVal(1);
5333 Ptr = Ptr.getOperand(0);
5338 // FIXME: 256-bit vector instructions don't require a strict alignment,
5339 // improve this code to support it better.
5340 unsigned RequiredAlign = VT.getSizeInBits()/8;
5341 SDValue Chain = LD->getChain();
5342 // Make sure the stack object alignment is at least 16 or 32.
5343 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5344 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5345 if (MFI->isFixedObjectIndex(FI)) {
5346 // Can't change the alignment. FIXME: It's possible to compute
5347 // the exact stack offset and reference FI + adjust offset instead.
5348 // If someone *really* cares about this. That's the way to implement it.
5351 MFI->setObjectAlignment(FI, RequiredAlign);
5355 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5356 // Ptr + (Offset & ~15).
5359 if ((Offset % RequiredAlign) & 3)
5361 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5364 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5365 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5368 int EltNo = (Offset - StartOffset) >> 2;
5369 unsigned NumElems = VT.getVectorNumElements();
5371 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5372 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5373 LD->getPointerInfo().getWithOffset(StartOffset),
5374 false, false, false, 0);
5376 SmallVector<int, 8> Mask(NumElems, EltNo);
5378 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5384 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5385 /// elements can be replaced by a single large load which has the same value as
5386 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5388 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5390 /// FIXME: we'd also like to handle the case where the last elements are zero
5391 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5392 /// There's even a handy isZeroNode for that purpose.
5393 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5394 SDLoc &DL, SelectionDAG &DAG,
5395 bool isAfterLegalize) {
5396 unsigned NumElems = Elts.size();
5398 LoadSDNode *LDBase = nullptr;
5399 unsigned LastLoadedElt = -1U;
5401 // For each element in the initializer, see if we've found a load or an undef.
5402 // If we don't find an initial load element, or later load elements are
5403 // non-consecutive, bail out.
5404 for (unsigned i = 0; i < NumElems; ++i) {
5405 SDValue Elt = Elts[i];
5406 // Look through a bitcast.
5407 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5408 Elt = Elt.getOperand(0);
5409 if (!Elt.getNode() ||
5410 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5413 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5415 LDBase = cast<LoadSDNode>(Elt.getNode());
5419 if (Elt.getOpcode() == ISD::UNDEF)
5422 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5423 EVT LdVT = Elt.getValueType();
5424 // Each loaded element must be the correct fractional portion of the
5425 // requested vector load.
5426 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5428 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5433 // If we have found an entire vector of loads and undefs, then return a large
5434 // load of the entire vector width starting at the base pointer. If we found
5435 // consecutive loads for the low half, generate a vzext_load node.
5436 if (LastLoadedElt == NumElems - 1) {
5437 assert(LDBase && "Did not find base load for merging consecutive loads");
5438 EVT EltVT = LDBase->getValueType(0);
5439 // Ensure that the input vector size for the merged loads matches the
5440 // cumulative size of the input elements.
5441 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5444 if (isAfterLegalize &&
5445 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5448 SDValue NewLd = SDValue();
5450 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5451 LDBase->getPointerInfo(), LDBase->isVolatile(),
5452 LDBase->isNonTemporal(), LDBase->isInvariant(),
5453 LDBase->getAlignment());
5455 if (LDBase->hasAnyUseOfValue(1)) {
5456 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5458 SDValue(NewLd.getNode(), 1));
5459 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5460 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5461 SDValue(NewLd.getNode(), 1));
5467 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5468 //of a v4i32 / v4f32. It's probably worth generalizing.
5469 EVT EltVT = VT.getVectorElementType();
5470 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5471 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5472 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5473 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5475 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5476 LDBase->getPointerInfo(),
5477 LDBase->getAlignment(),
5478 false/*isVolatile*/, true/*ReadMem*/,
5481 // Make sure the newly-created LOAD is in the same position as LDBase in
5482 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5483 // update uses of LDBase's output chain to use the TokenFactor.
5484 if (LDBase->hasAnyUseOfValue(1)) {
5485 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5486 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5487 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5488 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5489 SDValue(ResNode.getNode(), 1));
5492 return DAG.getBitcast(VT, ResNode);
5497 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5498 /// to generate a splat value for the following cases:
5499 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5500 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5501 /// a scalar load, or a constant.
5502 /// The VBROADCAST node is returned when a pattern is found,
5503 /// or SDValue() otherwise.
5504 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5505 SelectionDAG &DAG) {
5506 // VBROADCAST requires AVX.
5507 // TODO: Splats could be generated for non-AVX CPUs using SSE
5508 // instructions, but there's less potential gain for only 128-bit vectors.
5509 if (!Subtarget->hasAVX())
5512 MVT VT = Op.getSimpleValueType();
5515 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5516 "Unsupported vector type for broadcast.");
5521 switch (Op.getOpcode()) {
5523 // Unknown pattern found.
5526 case ISD::BUILD_VECTOR: {
5527 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5528 BitVector UndefElements;
5529 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5531 // We need a splat of a single value to use broadcast, and it doesn't
5532 // make any sense if the value is only in one element of the vector.
5533 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5537 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5538 Ld.getOpcode() == ISD::ConstantFP);
5540 // Make sure that all of the users of a non-constant load are from the
5541 // BUILD_VECTOR node.
5542 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5547 case ISD::VECTOR_SHUFFLE: {
5548 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5550 // Shuffles must have a splat mask where the first element is
5552 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5555 SDValue Sc = Op.getOperand(0);
5556 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5557 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5559 if (!Subtarget->hasInt256())
5562 // Use the register form of the broadcast instruction available on AVX2.
5563 if (VT.getSizeInBits() >= 256)
5564 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5565 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5568 Ld = Sc.getOperand(0);
5569 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5570 Ld.getOpcode() == ISD::ConstantFP);
5572 // The scalar_to_vector node and the suspected
5573 // load node must have exactly one user.
5574 // Constants may have multiple users.
5576 // AVX-512 has register version of the broadcast
5577 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5578 Ld.getValueType().getSizeInBits() >= 32;
5579 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5586 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5587 bool IsGE256 = (VT.getSizeInBits() >= 256);
5589 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5590 // instruction to save 8 or more bytes of constant pool data.
5591 // TODO: If multiple splats are generated to load the same constant,
5592 // it may be detrimental to overall size. There needs to be a way to detect
5593 // that condition to know if this is truly a size win.
5594 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5596 // Handle broadcasting a single constant scalar from the constant pool
5598 // On Sandybridge (no AVX2), it is still better to load a constant vector
5599 // from the constant pool and not to broadcast it from a scalar.
5600 // But override that restriction when optimizing for size.
5601 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5602 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5603 EVT CVT = Ld.getValueType();
5604 assert(!CVT.isVector() && "Must not broadcast a vector type");
5606 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5607 // For size optimization, also splat v2f64 and v2i64, and for size opt
5608 // with AVX2, also splat i8 and i16.
5609 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5610 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5611 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5612 const Constant *C = nullptr;
5613 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5614 C = CI->getConstantIntValue();
5615 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5616 C = CF->getConstantFPValue();
5618 assert(C && "Invalid constant type");
5620 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5622 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5623 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5625 CVT, dl, DAG.getEntryNode(), CP,
5626 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5627 false, false, Alignment);
5629 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5633 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5635 // Handle AVX2 in-register broadcasts.
5636 if (!IsLoad && Subtarget->hasInt256() &&
5637 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5638 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5640 // The scalar source must be a normal load.
5644 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5645 (Subtarget->hasVLX() && ScalarSize == 64))
5646 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5648 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5649 // double since there is no vbroadcastsd xmm
5650 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5651 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5652 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5655 // Unsupported broadcast.
5659 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5660 /// underlying vector and index.
5662 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5664 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5666 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5667 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5670 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5672 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5674 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5675 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5678 // In this case the vector is the extract_subvector expression and the index
5679 // is 2, as specified by the shuffle.
5680 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5681 SDValue ShuffleVec = SVOp->getOperand(0);
5682 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5683 assert(ShuffleVecVT.getVectorElementType() ==
5684 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5686 int ShuffleIdx = SVOp->getMaskElt(Idx);
5687 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5688 ExtractedFromVec = ShuffleVec;
5694 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5695 MVT VT = Op.getSimpleValueType();
5697 // Skip if insert_vec_elt is not supported.
5698 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5699 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5703 unsigned NumElems = Op.getNumOperands();
5707 SmallVector<unsigned, 4> InsertIndices;
5708 SmallVector<int, 8> Mask(NumElems, -1);
5710 for (unsigned i = 0; i != NumElems; ++i) {
5711 unsigned Opc = Op.getOperand(i).getOpcode();
5713 if (Opc == ISD::UNDEF)
5716 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5717 // Quit if more than 1 elements need inserting.
5718 if (InsertIndices.size() > 1)
5721 InsertIndices.push_back(i);
5725 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5726 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5727 // Quit if non-constant index.
5728 if (!isa<ConstantSDNode>(ExtIdx))
5730 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5732 // Quit if extracted from vector of different type.
5733 if (ExtractedFromVec.getValueType() != VT)
5736 if (!VecIn1.getNode())
5737 VecIn1 = ExtractedFromVec;
5738 else if (VecIn1 != ExtractedFromVec) {
5739 if (!VecIn2.getNode())
5740 VecIn2 = ExtractedFromVec;
5741 else if (VecIn2 != ExtractedFromVec)
5742 // Quit if more than 2 vectors to shuffle
5746 if (ExtractedFromVec == VecIn1)
5748 else if (ExtractedFromVec == VecIn2)
5749 Mask[i] = Idx + NumElems;
5752 if (!VecIn1.getNode())
5755 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5756 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5757 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5758 unsigned Idx = InsertIndices[i];
5759 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5760 DAG.getIntPtrConstant(Idx, DL));
5766 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5767 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5768 Op.getScalarValueSizeInBits() == 1 &&
5769 "Can not convert non-constant vector");
5770 uint64_t Immediate = 0;
5771 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5772 SDValue In = Op.getOperand(idx);
5773 if (In.getOpcode() != ISD::UNDEF)
5774 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5778 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5779 return DAG.getConstant(Immediate, dl, VT);
5781 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5783 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5785 MVT VT = Op.getSimpleValueType();
5786 assert((VT.getVectorElementType() == MVT::i1) &&
5787 "Unexpected type in LowerBUILD_VECTORvXi1!");
5790 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5791 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5792 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5793 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5796 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5797 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5798 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5799 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5802 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5803 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5804 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5805 return DAG.getBitcast(VT, Imm);
5806 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5807 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5808 DAG.getIntPtrConstant(0, dl));
5811 // Vector has one or more non-const elements
5812 uint64_t Immediate = 0;
5813 SmallVector<unsigned, 16> NonConstIdx;
5814 bool IsSplat = true;
5815 bool HasConstElts = false;
5817 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5818 SDValue In = Op.getOperand(idx);
5819 if (In.getOpcode() == ISD::UNDEF)
5821 if (!isa<ConstantSDNode>(In))
5822 NonConstIdx.push_back(idx);
5824 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5825 HasConstElts = true;
5829 else if (In != Op.getOperand(SplatIdx))
5833 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5835 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5836 DAG.getConstant(1, dl, VT),
5837 DAG.getConstant(0, dl, VT));
5839 // insert elements one by one
5843 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5844 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5846 else if (HasConstElts)
5847 Imm = DAG.getConstant(0, dl, VT);
5849 Imm = DAG.getUNDEF(VT);
5850 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5851 DstVec = DAG.getBitcast(VT, Imm);
5853 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5854 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5855 DAG.getIntPtrConstant(0, dl));
5858 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5859 unsigned InsertIdx = NonConstIdx[i];
5860 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5861 Op.getOperand(InsertIdx),
5862 DAG.getIntPtrConstant(InsertIdx, dl));
5867 /// \brief Return true if \p N implements a horizontal binop and return the
5868 /// operands for the horizontal binop into V0 and V1.
5870 /// This is a helper function of LowerToHorizontalOp().
5871 /// This function checks that the build_vector \p N in input implements a
5872 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5873 /// operation to match.
5874 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5875 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5876 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5879 /// This function only analyzes elements of \p N whose indices are
5880 /// in range [BaseIdx, LastIdx).
5881 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5883 unsigned BaseIdx, unsigned LastIdx,
5884 SDValue &V0, SDValue &V1) {
5885 EVT VT = N->getValueType(0);
5887 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5888 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5889 "Invalid Vector in input!");
5891 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5892 bool CanFold = true;
5893 unsigned ExpectedVExtractIdx = BaseIdx;
5894 unsigned NumElts = LastIdx - BaseIdx;
5895 V0 = DAG.getUNDEF(VT);
5896 V1 = DAG.getUNDEF(VT);
5898 // Check if N implements a horizontal binop.
5899 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5900 SDValue Op = N->getOperand(i + BaseIdx);
5903 if (Op->getOpcode() == ISD::UNDEF) {
5904 // Update the expected vector extract index.
5905 if (i * 2 == NumElts)
5906 ExpectedVExtractIdx = BaseIdx;
5907 ExpectedVExtractIdx += 2;
5911 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5916 SDValue Op0 = Op.getOperand(0);
5917 SDValue Op1 = Op.getOperand(1);
5919 // Try to match the following pattern:
5920 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5921 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5922 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5923 Op0.getOperand(0) == Op1.getOperand(0) &&
5924 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5925 isa<ConstantSDNode>(Op1.getOperand(1)));
5929 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5930 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5932 if (i * 2 < NumElts) {
5933 if (V0.getOpcode() == ISD::UNDEF) {
5934 V0 = Op0.getOperand(0);
5935 if (V0.getValueType() != VT)
5939 if (V1.getOpcode() == ISD::UNDEF) {
5940 V1 = Op0.getOperand(0);
5941 if (V1.getValueType() != VT)
5944 if (i * 2 == NumElts)
5945 ExpectedVExtractIdx = BaseIdx;
5948 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5949 if (I0 == ExpectedVExtractIdx)
5950 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5951 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5952 // Try to match the following dag sequence:
5953 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5954 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5958 ExpectedVExtractIdx += 2;
5964 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5965 /// a concat_vector.
5967 /// This is a helper function of LowerToHorizontalOp().
5968 /// This function expects two 256-bit vectors called V0 and V1.
5969 /// At first, each vector is split into two separate 128-bit vectors.
5970 /// Then, the resulting 128-bit vectors are used to implement two
5971 /// horizontal binary operations.
5973 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5975 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5976 /// the two new horizontal binop.
5977 /// When Mode is set, the first horizontal binop dag node would take as input
5978 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5979 /// horizontal binop dag node would take as input the lower 128-bit of V1
5980 /// and the upper 128-bit of V1.
5982 /// HADD V0_LO, V0_HI
5983 /// HADD V1_LO, V1_HI
5985 /// Otherwise, the first horizontal binop dag node takes as input the lower
5986 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5987 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5989 /// HADD V0_LO, V1_LO
5990 /// HADD V0_HI, V1_HI
5992 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5993 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5994 /// the upper 128-bits of the result.
5995 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5996 SDLoc DL, SelectionDAG &DAG,
5997 unsigned X86Opcode, bool Mode,
5998 bool isUndefLO, bool isUndefHI) {
5999 EVT VT = V0.getValueType();
6000 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6001 "Invalid nodes in input!");
6003 unsigned NumElts = VT.getVectorNumElements();
6004 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6005 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6006 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6007 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6008 EVT NewVT = V0_LO.getValueType();
6010 SDValue LO = DAG.getUNDEF(NewVT);
6011 SDValue HI = DAG.getUNDEF(NewVT);
6014 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6015 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6016 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6017 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6018 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6020 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6021 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6022 V1_LO->getOpcode() != ISD::UNDEF))
6023 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6025 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6026 V1_HI->getOpcode() != ISD::UNDEF))
6027 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6030 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6033 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
6035 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
6036 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6037 MVT VT = BV->getSimpleValueType(0);
6038 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
6039 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
6043 unsigned NumElts = VT.getVectorNumElements();
6044 SDValue InVec0 = DAG.getUNDEF(VT);
6045 SDValue InVec1 = DAG.getUNDEF(VT);
6047 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6048 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6050 // Odd-numbered elements in the input build vector are obtained from
6051 // adding two integer/float elements.
6052 // Even-numbered elements in the input build vector are obtained from
6053 // subtracting two integer/float elements.
6054 unsigned ExpectedOpcode = ISD::FSUB;
6055 unsigned NextExpectedOpcode = ISD::FADD;
6056 bool AddFound = false;
6057 bool SubFound = false;
6059 for (unsigned i = 0, e = NumElts; i != e; ++i) {
6060 SDValue Op = BV->getOperand(i);
6062 // Skip 'undef' values.
6063 unsigned Opcode = Op.getOpcode();
6064 if (Opcode == ISD::UNDEF) {
6065 std::swap(ExpectedOpcode, NextExpectedOpcode);
6069 // Early exit if we found an unexpected opcode.
6070 if (Opcode != ExpectedOpcode)
6073 SDValue Op0 = Op.getOperand(0);
6074 SDValue Op1 = Op.getOperand(1);
6076 // Try to match the following pattern:
6077 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6078 // Early exit if we cannot match that sequence.
6079 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6080 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6081 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6082 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6083 Op0.getOperand(1) != Op1.getOperand(1))
6086 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6090 // We found a valid add/sub node. Update the information accordingly.
6096 // Update InVec0 and InVec1.
6097 if (InVec0.getOpcode() == ISD::UNDEF) {
6098 InVec0 = Op0.getOperand(0);
6099 if (InVec0.getSimpleValueType() != VT)
6102 if (InVec1.getOpcode() == ISD::UNDEF) {
6103 InVec1 = Op1.getOperand(0);
6104 if (InVec1.getSimpleValueType() != VT)
6108 // Make sure that operands in input to each add/sub node always
6109 // come from a same pair of vectors.
6110 if (InVec0 != Op0.getOperand(0)) {
6111 if (ExpectedOpcode == ISD::FSUB)
6114 // FADD is commutable. Try to commute the operands
6115 // and then test again.
6116 std::swap(Op0, Op1);
6117 if (InVec0 != Op0.getOperand(0))
6121 if (InVec1 != Op1.getOperand(0))
6124 // Update the pair of expected opcodes.
6125 std::swap(ExpectedOpcode, NextExpectedOpcode);
6128 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6129 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6130 InVec1.getOpcode() != ISD::UNDEF)
6131 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6136 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6137 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6138 const X86Subtarget *Subtarget,
6139 SelectionDAG &DAG) {
6140 MVT VT = BV->getSimpleValueType(0);
6141 unsigned NumElts = VT.getVectorNumElements();
6142 unsigned NumUndefsLO = 0;
6143 unsigned NumUndefsHI = 0;
6144 unsigned Half = NumElts/2;
6146 // Count the number of UNDEF operands in the build_vector in input.
6147 for (unsigned i = 0, e = Half; i != e; ++i)
6148 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6151 for (unsigned i = Half, e = NumElts; i != e; ++i)
6152 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6155 // Early exit if this is either a build_vector of all UNDEFs or all the
6156 // operands but one are UNDEF.
6157 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6161 SDValue InVec0, InVec1;
6162 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6163 // Try to match an SSE3 float HADD/HSUB.
6164 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6165 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6167 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6168 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6169 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6170 // Try to match an SSSE3 integer HADD/HSUB.
6171 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6172 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6174 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6175 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6178 if (!Subtarget->hasAVX())
6181 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6182 // Try to match an AVX horizontal add/sub of packed single/double
6183 // precision floating point values from 256-bit vectors.
6184 SDValue InVec2, InVec3;
6185 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6186 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6187 ((InVec0.getOpcode() == ISD::UNDEF ||
6188 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6189 ((InVec1.getOpcode() == ISD::UNDEF ||
6190 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6191 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6193 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6194 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6195 ((InVec0.getOpcode() == ISD::UNDEF ||
6196 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6197 ((InVec1.getOpcode() == ISD::UNDEF ||
6198 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6199 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6200 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6201 // Try to match an AVX2 horizontal add/sub of signed integers.
6202 SDValue InVec2, InVec3;
6204 bool CanFold = true;
6206 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6207 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6208 ((InVec0.getOpcode() == ISD::UNDEF ||
6209 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6210 ((InVec1.getOpcode() == ISD::UNDEF ||
6211 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6212 X86Opcode = X86ISD::HADD;
6213 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6214 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6215 ((InVec0.getOpcode() == ISD::UNDEF ||
6216 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6217 ((InVec1.getOpcode() == ISD::UNDEF ||
6218 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6219 X86Opcode = X86ISD::HSUB;
6224 // Fold this build_vector into a single horizontal add/sub.
6225 // Do this only if the target has AVX2.
6226 if (Subtarget->hasAVX2())
6227 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6229 // Do not try to expand this build_vector into a pair of horizontal
6230 // add/sub if we can emit a pair of scalar add/sub.
6231 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6234 // Convert this build_vector into a pair of horizontal binop followed by
6236 bool isUndefLO = NumUndefsLO == Half;
6237 bool isUndefHI = NumUndefsHI == Half;
6238 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6239 isUndefLO, isUndefHI);
6243 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6244 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6246 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6247 X86Opcode = X86ISD::HADD;
6248 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6249 X86Opcode = X86ISD::HSUB;
6250 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6251 X86Opcode = X86ISD::FHADD;
6252 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6253 X86Opcode = X86ISD::FHSUB;
6257 // Don't try to expand this build_vector into a pair of horizontal add/sub
6258 // if we can simply emit a pair of scalar add/sub.
6259 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6262 // Convert this build_vector into two horizontal add/sub followed by
6264 bool isUndefLO = NumUndefsLO == Half;
6265 bool isUndefHI = NumUndefsHI == Half;
6266 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6267 isUndefLO, isUndefHI);
6274 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6277 MVT VT = Op.getSimpleValueType();
6278 MVT ExtVT = VT.getVectorElementType();
6279 unsigned NumElems = Op.getNumOperands();
6281 // Generate vectors for predicate vectors.
6282 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6283 return LowerBUILD_VECTORvXi1(Op, DAG);
6285 // Vectors containing all zeros can be matched by pxor and xorps later
6286 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6287 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6288 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6289 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6292 return getZeroVector(VT, Subtarget, DAG, dl);
6295 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6296 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6297 // vpcmpeqd on 256-bit vectors.
6298 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6299 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6302 if (!VT.is512BitVector())
6303 return getOnesVector(VT, Subtarget, DAG, dl);
6306 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6307 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6309 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6310 return HorizontalOp;
6311 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6314 unsigned EVTBits = ExtVT.getSizeInBits();
6316 unsigned NumZero = 0;
6317 unsigned NumNonZero = 0;
6318 uint64_t NonZeros = 0;
6319 bool IsAllConstants = true;
6320 SmallSet<SDValue, 8> Values;
6321 for (unsigned i = 0; i < NumElems; ++i) {
6322 SDValue Elt = Op.getOperand(i);
6323 if (Elt.getOpcode() == ISD::UNDEF)
6326 if (Elt.getOpcode() != ISD::Constant &&
6327 Elt.getOpcode() != ISD::ConstantFP)
6328 IsAllConstants = false;
6329 if (X86::isZeroNode(Elt))
6332 assert(i < sizeof(NonZeros) * 8); // Make sure the shift is within range.
6333 NonZeros |= ((uint64_t)1 << i);
6338 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6339 if (NumNonZero == 0)
6340 return DAG.getUNDEF(VT);
6342 // Special case for single non-zero, non-undef, element.
6343 if (NumNonZero == 1) {
6344 unsigned Idx = countTrailingZeros(NonZeros);
6345 SDValue Item = Op.getOperand(Idx);
6347 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6348 // the value are obviously zero, truncate the value to i32 and do the
6349 // insertion that way. Only do this if the value is non-constant or if the
6350 // value is a constant being inserted into element 0. It is cheaper to do
6351 // a constant pool load than it is to do a movd + shuffle.
6352 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6353 (!IsAllConstants || Idx == 0)) {
6354 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6356 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6357 MVT VecVT = MVT::v4i32;
6359 // Truncate the value (which may itself be a constant) to i32, and
6360 // convert it to a vector with movd (S2V+shuffle to zero extend).
6361 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6362 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6363 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6364 Item, Idx * 2, true, Subtarget, DAG));
6368 // If we have a constant or non-constant insertion into the low element of
6369 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6370 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6371 // depending on what the source datatype is.
6374 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6376 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6377 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6378 if (VT.is512BitVector()) {
6379 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6380 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6381 Item, DAG.getIntPtrConstant(0, dl));
6383 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6384 "Expected an SSE value type!");
6385 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6386 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6387 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6390 // We can't directly insert an i8 or i16 into a vector, so zero extend
6392 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6393 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6394 if (VT.is256BitVector()) {
6395 if (Subtarget->hasAVX()) {
6396 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6397 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6399 // Without AVX, we need to extend to a 128-bit vector and then
6400 // insert into the 256-bit vector.
6401 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6402 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6403 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6406 assert(VT.is128BitVector() && "Expected an SSE value type!");
6407 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6408 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6410 return DAG.getBitcast(VT, Item);
6414 // Is it a vector logical left shift?
6415 if (NumElems == 2 && Idx == 1 &&
6416 X86::isZeroNode(Op.getOperand(0)) &&
6417 !X86::isZeroNode(Op.getOperand(1))) {
6418 unsigned NumBits = VT.getSizeInBits();
6419 return getVShift(true, VT,
6420 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6421 VT, Op.getOperand(1)),
6422 NumBits/2, DAG, *this, dl);
6425 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6428 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6429 // is a non-constant being inserted into an element other than the low one,
6430 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6431 // movd/movss) to move this into the low element, then shuffle it into
6433 if (EVTBits == 32) {
6434 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6435 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6439 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6440 if (Values.size() == 1) {
6441 if (EVTBits == 32) {
6442 // Instead of a shuffle like this:
6443 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6444 // Check if it's possible to issue this instead.
6445 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6446 unsigned Idx = countTrailingZeros(NonZeros);
6447 SDValue Item = Op.getOperand(Idx);
6448 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6449 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6454 // A vector full of immediates; various special cases are already
6455 // handled, so this is best done with a single constant-pool load.
6459 // For AVX-length vectors, see if we can use a vector load to get all of the
6460 // elements, otherwise build the individual 128-bit pieces and use
6461 // shuffles to put them in place.
6462 if (VT.is256BitVector() || VT.is512BitVector()) {
6463 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6465 // Check for a build vector of consecutive loads.
6466 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6469 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6471 // Build both the lower and upper subvector.
6472 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6473 makeArrayRef(&V[0], NumElems/2));
6474 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6475 makeArrayRef(&V[NumElems / 2], NumElems/2));
6477 // Recreate the wider vector with the lower and upper part.
6478 if (VT.is256BitVector())
6479 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6480 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6483 // Let legalizer expand 2-wide build_vectors.
6484 if (EVTBits == 64) {
6485 if (NumNonZero == 1) {
6486 // One half is zero or undef.
6487 unsigned Idx = countTrailingZeros(NonZeros);
6488 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6489 Op.getOperand(Idx));
6490 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6495 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6496 if (EVTBits == 8 && NumElems == 16)
6497 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros, NumNonZero, NumZero,
6498 DAG, Subtarget, *this))
6501 if (EVTBits == 16 && NumElems == 8)
6502 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros, NumNonZero, NumZero,
6503 DAG, Subtarget, *this))
6506 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6507 if (EVTBits == 32 && NumElems == 4)
6508 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6511 // If element VT is == 32 bits, turn it into a number of shuffles.
6512 SmallVector<SDValue, 8> V(NumElems);
6513 if (NumElems == 4 && NumZero > 0) {
6514 for (unsigned i = 0; i < 4; ++i) {
6515 bool isZero = !(NonZeros & (1ULL << i));
6517 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6519 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6522 for (unsigned i = 0; i < 2; ++i) {
6523 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6526 V[i] = V[i*2]; // Must be a zero vector.
6529 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6532 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6535 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6540 bool Reverse1 = (NonZeros & 0x3) == 2;
6541 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6545 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6546 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6548 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6551 if (Values.size() > 1 && VT.is128BitVector()) {
6552 // Check for a build vector of consecutive loads.
6553 for (unsigned i = 0; i < NumElems; ++i)
6554 V[i] = Op.getOperand(i);
6556 // Check for elements which are consecutive loads.
6557 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6560 // Check for a build vector from mostly shuffle plus few inserting.
6561 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6564 // For SSE 4.1, use insertps to put the high elements into the low element.
6565 if (Subtarget->hasSSE41()) {
6567 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6568 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6570 Result = DAG.getUNDEF(VT);
6572 for (unsigned i = 1; i < NumElems; ++i) {
6573 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6574 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6575 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6580 // Otherwise, expand into a number of unpckl*, start by extending each of
6581 // our (non-undef) elements to the full vector width with the element in the
6582 // bottom slot of the vector (which generates no code for SSE).
6583 for (unsigned i = 0; i < NumElems; ++i) {
6584 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6585 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6587 V[i] = DAG.getUNDEF(VT);
6590 // Next, we iteratively mix elements, e.g. for v4f32:
6591 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6592 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6593 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6594 unsigned EltStride = NumElems >> 1;
6595 while (EltStride != 0) {
6596 for (unsigned i = 0; i < EltStride; ++i) {
6597 // If V[i+EltStride] is undef and this is the first round of mixing,
6598 // then it is safe to just drop this shuffle: V[i] is already in the
6599 // right place, the one element (since it's the first round) being
6600 // inserted as undef can be dropped. This isn't safe for successive
6601 // rounds because they will permute elements within both vectors.
6602 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6603 EltStride == NumElems/2)
6606 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6615 // 256-bit AVX can use the vinsertf128 instruction
6616 // to create 256-bit vectors from two other 128-bit ones.
6617 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6619 MVT ResVT = Op.getSimpleValueType();
6621 assert((ResVT.is256BitVector() ||
6622 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6624 SDValue V1 = Op.getOperand(0);
6625 SDValue V2 = Op.getOperand(1);
6626 unsigned NumElems = ResVT.getVectorNumElements();
6627 if (ResVT.is256BitVector())
6628 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6630 if (Op.getNumOperands() == 4) {
6631 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6632 ResVT.getVectorNumElements()/2);
6633 SDValue V3 = Op.getOperand(2);
6634 SDValue V4 = Op.getOperand(3);
6635 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6636 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6638 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6641 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6642 const X86Subtarget *Subtarget,
6643 SelectionDAG & DAG) {
6645 MVT ResVT = Op.getSimpleValueType();
6646 unsigned NumOfOperands = Op.getNumOperands();
6648 assert(isPowerOf2_32(NumOfOperands) &&
6649 "Unexpected number of operands in CONCAT_VECTORS");
6651 SDValue Undef = DAG.getUNDEF(ResVT);
6652 if (NumOfOperands > 2) {
6653 // Specialize the cases when all, or all but one, of the operands are undef.
6654 unsigned NumOfDefinedOps = 0;
6656 for (unsigned i = 0; i < NumOfOperands; i++)
6657 if (!Op.getOperand(i).isUndef()) {
6661 if (NumOfDefinedOps == 0)
6663 if (NumOfDefinedOps == 1) {
6664 unsigned SubVecNumElts =
6665 Op.getOperand(OpIdx).getValueType().getVectorNumElements();
6666 SDValue IdxVal = DAG.getIntPtrConstant(SubVecNumElts * OpIdx, dl);
6667 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef,
6668 Op.getOperand(OpIdx), IdxVal);
6671 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6672 ResVT.getVectorNumElements()/2);
6673 SmallVector<SDValue, 2> Ops;
6674 for (unsigned i = 0; i < NumOfOperands/2; i++)
6675 Ops.push_back(Op.getOperand(i));
6676 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6678 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6679 Ops.push_back(Op.getOperand(i));
6680 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6681 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6685 SDValue V1 = Op.getOperand(0);
6686 SDValue V2 = Op.getOperand(1);
6687 unsigned NumElems = ResVT.getVectorNumElements();
6688 assert(V1.getValueType() == V2.getValueType() &&
6689 V1.getValueType().getVectorNumElements() == NumElems/2 &&
6690 "Unexpected operands in CONCAT_VECTORS");
6692 if (ResVT.getSizeInBits() >= 16)
6693 return Op; // The operation is legal with KUNPCK
6695 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6696 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6697 SDValue ZeroVec = getZeroVector(ResVT, Subtarget, DAG, dl);
6698 if (IsZeroV1 && IsZeroV2)
6701 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6703 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6705 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V1, ZeroIdx);
6707 SDValue IdxVal = DAG.getIntPtrConstant(NumElems/2, dl);
6709 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, IdxVal);
6712 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, ZeroVec, V2, IdxVal);
6714 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6715 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, V1, V2, IdxVal);
6718 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6719 const X86Subtarget *Subtarget,
6720 SelectionDAG &DAG) {
6721 MVT VT = Op.getSimpleValueType();
6722 if (VT.getVectorElementType() == MVT::i1)
6723 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6725 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6726 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6727 Op.getNumOperands() == 4)));
6729 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6730 // from two other 128-bit ones.
6732 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6733 return LowerAVXCONCAT_VECTORS(Op, DAG);
6736 //===----------------------------------------------------------------------===//
6737 // Vector shuffle lowering
6739 // This is an experimental code path for lowering vector shuffles on x86. It is
6740 // designed to handle arbitrary vector shuffles and blends, gracefully
6741 // degrading performance as necessary. It works hard to recognize idiomatic
6742 // shuffles and lower them to optimal instruction patterns without leaving
6743 // a framework that allows reasonably efficient handling of all vector shuffle
6745 //===----------------------------------------------------------------------===//
6747 /// \brief Tiny helper function to identify a no-op mask.
6749 /// This is a somewhat boring predicate function. It checks whether the mask
6750 /// array input, which is assumed to be a single-input shuffle mask of the kind
6751 /// used by the X86 shuffle instructions (not a fully general
6752 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6753 /// in-place shuffle are 'no-op's.
6754 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6755 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6756 if (Mask[i] != -1 && Mask[i] != i)
6761 /// \brief Helper function to classify a mask as a single-input mask.
6763 /// This isn't a generic single-input test because in the vector shuffle
6764 /// lowering we canonicalize single inputs to be the first input operand. This
6765 /// means we can more quickly test for a single input by only checking whether
6766 /// an input from the second operand exists. We also assume that the size of
6767 /// mask corresponds to the size of the input vectors which isn't true in the
6768 /// fully general case.
6769 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6771 if (M >= (int)Mask.size())
6776 /// \brief Test whether there are elements crossing 128-bit lanes in this
6779 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6780 /// and we routinely test for these.
6781 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6782 int LaneSize = 128 / VT.getScalarSizeInBits();
6783 int Size = Mask.size();
6784 for (int i = 0; i < Size; ++i)
6785 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6790 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6792 /// This checks a shuffle mask to see if it is performing the same
6793 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6794 /// that it is also not lane-crossing. It may however involve a blend from the
6795 /// same lane of a second vector.
6797 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6798 /// non-trivial to compute in the face of undef lanes. The representation is
6799 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6800 /// entries from both V1 and V2 inputs to the wider mask.
6802 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6803 SmallVectorImpl<int> &RepeatedMask) {
6804 int LaneSize = 128 / VT.getScalarSizeInBits();
6805 RepeatedMask.resize(LaneSize, -1);
6806 int Size = Mask.size();
6807 for (int i = 0; i < Size; ++i) {
6810 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6811 // This entry crosses lanes, so there is no way to model this shuffle.
6814 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6815 if (RepeatedMask[i % LaneSize] == -1)
6816 // This is the first non-undef entry in this slot of a 128-bit lane.
6817 RepeatedMask[i % LaneSize] =
6818 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6819 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6820 // Found a mismatch with the repeated mask.
6826 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6829 /// This is a fast way to test a shuffle mask against a fixed pattern:
6831 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6833 /// It returns true if the mask is exactly as wide as the argument list, and
6834 /// each element of the mask is either -1 (signifying undef) or the value given
6835 /// in the argument.
6836 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6837 ArrayRef<int> ExpectedMask) {
6838 if (Mask.size() != ExpectedMask.size())
6841 int Size = Mask.size();
6843 // If the values are build vectors, we can look through them to find
6844 // equivalent inputs that make the shuffles equivalent.
6845 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6846 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6848 for (int i = 0; i < Size; ++i)
6849 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6850 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6851 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6852 if (!MaskBV || !ExpectedBV ||
6853 MaskBV->getOperand(Mask[i] % Size) !=
6854 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6861 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6863 /// This helper function produces an 8-bit shuffle immediate corresponding to
6864 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6865 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6868 /// NB: We rely heavily on "undef" masks preserving the input lane.
6869 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6870 SelectionDAG &DAG) {
6871 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6872 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6873 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6874 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6875 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6878 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6879 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6880 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6881 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6882 return DAG.getConstant(Imm, DL, MVT::i8);
6885 /// \brief Compute whether each element of a shuffle is zeroable.
6887 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6888 /// Either it is an undef element in the shuffle mask, the element of the input
6889 /// referenced is undef, or the element of the input referenced is known to be
6890 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6891 /// as many lanes with this technique as possible to simplify the remaining
6893 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6894 SDValue V1, SDValue V2) {
6895 SmallBitVector Zeroable(Mask.size(), false);
6897 while (V1.getOpcode() == ISD::BITCAST)
6898 V1 = V1->getOperand(0);
6899 while (V2.getOpcode() == ISD::BITCAST)
6900 V2 = V2->getOperand(0);
6902 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6903 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6905 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6907 // Handle the easy cases.
6908 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6913 // If this is an index into a build_vector node (which has the same number
6914 // of elements), dig out the input value and use it.
6915 SDValue V = M < Size ? V1 : V2;
6916 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6919 SDValue Input = V.getOperand(M % Size);
6920 // The UNDEF opcode check really should be dead code here, but not quite
6921 // worth asserting on (it isn't invalid, just unexpected).
6922 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6929 // X86 has dedicated unpack instructions that can handle specific blend
6930 // operations: UNPCKH and UNPCKL.
6931 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6932 SDValue V1, SDValue V2,
6933 SelectionDAG &DAG) {
6934 int NumElts = VT.getVectorNumElements();
6935 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6936 SmallVector<int, 8> Unpckl;
6937 SmallVector<int, 8> Unpckh;
6939 for (int i = 0; i < NumElts; ++i) {
6940 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6941 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6942 int HiPos = LoPos + NumEltsInLane / 2;
6943 Unpckl.push_back(LoPos);
6944 Unpckh.push_back(HiPos);
6947 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6948 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6949 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6950 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6952 // Commute and try again.
6953 ShuffleVectorSDNode::commuteMask(Unpckl);
6954 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6955 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6957 ShuffleVectorSDNode::commuteMask(Unpckh);
6958 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6959 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6964 /// \brief Try to emit a bitmask instruction for a shuffle.
6966 /// This handles cases where we can model a blend exactly as a bitmask due to
6967 /// one of the inputs being zeroable.
6968 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6969 SDValue V2, ArrayRef<int> Mask,
6970 SelectionDAG &DAG) {
6971 MVT EltVT = VT.getVectorElementType();
6972 int NumEltBits = EltVT.getSizeInBits();
6973 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6974 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6975 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6977 if (EltVT.isFloatingPoint()) {
6978 Zero = DAG.getBitcast(EltVT, Zero);
6979 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6981 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6982 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6984 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6987 if (Mask[i] % Size != i)
6988 return SDValue(); // Not a blend.
6990 V = Mask[i] < Size ? V1 : V2;
6991 else if (V != (Mask[i] < Size ? V1 : V2))
6992 return SDValue(); // Can only let one input through the mask.
6994 VMaskOps[i] = AllOnes;
6997 return SDValue(); // No non-zeroable elements!
6999 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
7000 V = DAG.getNode(VT.isFloatingPoint()
7001 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
7006 /// \brief Try to emit a blend instruction for a shuffle using bit math.
7008 /// This is used as a fallback approach when first class blend instructions are
7009 /// unavailable. Currently it is only suitable for integer vectors, but could
7010 /// be generalized for floating point vectors if desirable.
7011 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
7012 SDValue V2, ArrayRef<int> Mask,
7013 SelectionDAG &DAG) {
7014 assert(VT.isInteger() && "Only supports integer vector types!");
7015 MVT EltVT = VT.getVectorElementType();
7016 int NumEltBits = EltVT.getSizeInBits();
7017 SDValue Zero = DAG.getConstant(0, DL, EltVT);
7018 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
7020 SmallVector<SDValue, 16> MaskOps;
7021 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7022 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
7023 return SDValue(); // Shuffled input!
7024 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
7027 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
7028 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
7029 // We have to cast V2 around.
7030 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
7031 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
7032 DAG.getBitcast(MaskVT, V1Mask),
7033 DAG.getBitcast(MaskVT, V2)));
7034 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
7037 /// \brief Try to emit a blend instruction for a shuffle.
7039 /// This doesn't do any checks for the availability of instructions for blending
7040 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7041 /// be matched in the backend with the type given. What it does check for is
7042 /// that the shuffle mask is a blend, or convertible into a blend with zero.
7043 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7044 SDValue V2, ArrayRef<int> Original,
7045 const X86Subtarget *Subtarget,
7046 SelectionDAG &DAG) {
7047 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7048 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7049 SmallVector<int, 8> Mask(Original.begin(), Original.end());
7050 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7051 bool ForceV1Zero = false, ForceV2Zero = false;
7053 // Attempt to generate the binary blend mask. If an input is zero then
7054 // we can use any lane.
7055 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
7056 unsigned BlendMask = 0;
7057 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7063 if (M == i + Size) {
7064 BlendMask |= 1u << i;
7075 BlendMask |= 1u << i;
7080 return SDValue(); // Shuffled input!
7083 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
7085 V1 = getZeroVector(VT, Subtarget, DAG, DL);
7087 V2 = getZeroVector(VT, Subtarget, DAG, DL);
7089 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
7090 unsigned ScaledMask = 0;
7091 for (int i = 0; i != Size; ++i)
7092 if (BlendMask & (1u << i))
7093 for (int j = 0; j != Scale; ++j)
7094 ScaledMask |= 1u << (i * Scale + j);
7098 switch (VT.SimpleTy) {
7103 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7104 DAG.getConstant(BlendMask, DL, MVT::i8));
7108 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7112 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7113 // that instruction.
7114 if (Subtarget->hasAVX2()) {
7115 // Scale the blend by the number of 32-bit dwords per element.
7116 int Scale = VT.getScalarSizeInBits() / 32;
7117 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7118 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7119 V1 = DAG.getBitcast(BlendVT, V1);
7120 V2 = DAG.getBitcast(BlendVT, V2);
7121 return DAG.getBitcast(
7122 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7123 DAG.getConstant(BlendMask, DL, MVT::i8)));
7127 // For integer shuffles we need to expand the mask and cast the inputs to
7128 // v8i16s prior to blending.
7129 int Scale = 8 / VT.getVectorNumElements();
7130 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
7131 V1 = DAG.getBitcast(MVT::v8i16, V1);
7132 V2 = DAG.getBitcast(MVT::v8i16, V2);
7133 return DAG.getBitcast(VT,
7134 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7135 DAG.getConstant(BlendMask, DL, MVT::i8)));
7139 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7140 SmallVector<int, 8> RepeatedMask;
7141 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7142 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7143 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7145 for (int i = 0; i < 8; ++i)
7146 if (RepeatedMask[i] >= 16)
7147 BlendMask |= 1u << i;
7148 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7149 DAG.getConstant(BlendMask, DL, MVT::i8));
7155 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
7156 "256-bit byte-blends require AVX2 support!");
7158 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
7159 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
7162 // Scale the blend by the number of bytes per element.
7163 int Scale = VT.getScalarSizeInBits() / 8;
7165 // This form of blend is always done on bytes. Compute the byte vector
7167 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7169 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7170 // mix of LLVM's code generator and the x86 backend. We tell the code
7171 // generator that boolean values in the elements of an x86 vector register
7172 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7173 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7174 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7175 // of the element (the remaining are ignored) and 0 in that high bit would
7176 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7177 // the LLVM model for boolean values in vector elements gets the relevant
7178 // bit set, it is set backwards and over constrained relative to x86's
7180 SmallVector<SDValue, 32> VSELECTMask;
7181 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7182 for (int j = 0; j < Scale; ++j)
7183 VSELECTMask.push_back(
7184 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7185 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7188 V1 = DAG.getBitcast(BlendVT, V1);
7189 V2 = DAG.getBitcast(BlendVT, V2);
7190 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7191 DAG.getNode(ISD::BUILD_VECTOR, DL,
7192 BlendVT, VSELECTMask),
7197 llvm_unreachable("Not a supported integer vector type!");
7201 /// \brief Try to lower as a blend of elements from two inputs followed by
7202 /// a single-input permutation.
7204 /// This matches the pattern where we can blend elements from two inputs and
7205 /// then reduce the shuffle to a single-input permutation.
7206 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7209 SelectionDAG &DAG) {
7210 // We build up the blend mask while checking whether a blend is a viable way
7211 // to reduce the shuffle.
7212 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7213 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7215 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7219 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7221 if (BlendMask[Mask[i] % Size] == -1)
7222 BlendMask[Mask[i] % Size] = Mask[i];
7223 else if (BlendMask[Mask[i] % Size] != Mask[i])
7224 return SDValue(); // Can't blend in the needed input!
7226 PermuteMask[i] = Mask[i] % Size;
7229 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7230 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7233 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7234 /// blends and permutes.
7236 /// This matches the extremely common pattern for handling combined
7237 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7238 /// operations. It will try to pick the best arrangement of shuffles and
7240 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7244 SelectionDAG &DAG) {
7245 // Shuffle the input elements into the desired positions in V1 and V2 and
7246 // blend them together.
7247 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7248 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7249 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7250 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7251 if (Mask[i] >= 0 && Mask[i] < Size) {
7252 V1Mask[i] = Mask[i];
7254 } else if (Mask[i] >= Size) {
7255 V2Mask[i] = Mask[i] - Size;
7256 BlendMask[i] = i + Size;
7259 // Try to lower with the simpler initial blend strategy unless one of the
7260 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7261 // shuffle may be able to fold with a load or other benefit. However, when
7262 // we'll have to do 2x as many shuffles in order to achieve this, blending
7263 // first is a better strategy.
7264 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7265 if (SDValue BlendPerm =
7266 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7269 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7270 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7271 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7274 /// \brief Try to lower a vector shuffle as a byte rotation.
7276 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7277 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7278 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7279 /// try to generically lower a vector shuffle through such an pattern. It
7280 /// does not check for the profitability of lowering either as PALIGNR or
7281 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7282 /// This matches shuffle vectors that look like:
7284 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7286 /// Essentially it concatenates V1 and V2, shifts right by some number of
7287 /// elements, and takes the low elements as the result. Note that while this is
7288 /// specified as a *right shift* because x86 is little-endian, it is a *left
7289 /// rotate* of the vector lanes.
7290 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7293 const X86Subtarget *Subtarget,
7294 SelectionDAG &DAG) {
7295 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7297 int NumElts = Mask.size();
7298 int NumLanes = VT.getSizeInBits() / 128;
7299 int NumLaneElts = NumElts / NumLanes;
7301 // We need to detect various ways of spelling a rotation:
7302 // [11, 12, 13, 14, 15, 0, 1, 2]
7303 // [-1, 12, 13, 14, -1, -1, 1, -1]
7304 // [-1, -1, -1, -1, -1, -1, 1, 2]
7305 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7306 // [-1, 4, 5, 6, -1, -1, 9, -1]
7307 // [-1, 4, 5, 6, -1, -1, -1, -1]
7310 for (int l = 0; l < NumElts; l += NumLaneElts) {
7311 for (int i = 0; i < NumLaneElts; ++i) {
7312 if (Mask[l + i] == -1)
7314 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7316 // Get the mod-Size index and lane correct it.
7317 int LaneIdx = (Mask[l + i] % NumElts) - l;
7318 // Make sure it was in this lane.
7319 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7322 // Determine where a rotated vector would have started.
7323 int StartIdx = i - LaneIdx;
7325 // The identity rotation isn't interesting, stop.
7328 // If we found the tail of a vector the rotation must be the missing
7329 // front. If we found the head of a vector, it must be how much of the
7331 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7334 Rotation = CandidateRotation;
7335 else if (Rotation != CandidateRotation)
7336 // The rotations don't match, so we can't match this mask.
7339 // Compute which value this mask is pointing at.
7340 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7342 // Compute which of the two target values this index should be assigned
7343 // to. This reflects whether the high elements are remaining or the low
7344 // elements are remaining.
7345 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7347 // Either set up this value if we've not encountered it before, or check
7348 // that it remains consistent.
7351 else if (TargetV != MaskV)
7352 // This may be a rotation, but it pulls from the inputs in some
7353 // unsupported interleaving.
7358 // Check that we successfully analyzed the mask, and normalize the results.
7359 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7360 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7366 // The actual rotate instruction rotates bytes, so we need to scale the
7367 // rotation based on how many bytes are in the vector lane.
7368 int Scale = 16 / NumLaneElts;
7370 // SSSE3 targets can use the palignr instruction.
7371 if (Subtarget->hasSSSE3()) {
7372 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7373 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7374 Lo = DAG.getBitcast(AlignVT, Lo);
7375 Hi = DAG.getBitcast(AlignVT, Hi);
7377 return DAG.getBitcast(
7378 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7379 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7382 assert(VT.is128BitVector() &&
7383 "Rotate-based lowering only supports 128-bit lowering!");
7384 assert(Mask.size() <= 16 &&
7385 "Can shuffle at most 16 bytes in a 128-bit vector!");
7387 // Default SSE2 implementation
7388 int LoByteShift = 16 - Rotation * Scale;
7389 int HiByteShift = Rotation * Scale;
7391 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7392 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7393 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7395 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7396 DAG.getConstant(LoByteShift, DL, MVT::i8));
7397 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7398 DAG.getConstant(HiByteShift, DL, MVT::i8));
7399 return DAG.getBitcast(VT,
7400 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7403 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7405 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7406 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7407 /// matches elements from one of the input vectors shuffled to the left or
7408 /// right with zeroable elements 'shifted in'. It handles both the strictly
7409 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7412 /// PSHL : (little-endian) left bit shift.
7413 /// [ zz, 0, zz, 2 ]
7414 /// [ -1, 4, zz, -1 ]
7415 /// PSRL : (little-endian) right bit shift.
7417 /// [ -1, -1, 7, zz]
7418 /// PSLLDQ : (little-endian) left byte shift
7419 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7420 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7421 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7422 /// PSRLDQ : (little-endian) right byte shift
7423 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7424 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7425 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7426 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7427 SDValue V2, ArrayRef<int> Mask,
7428 SelectionDAG &DAG) {
7429 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7431 int Size = Mask.size();
7432 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7434 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7435 for (int i = 0; i < Size; i += Scale)
7436 for (int j = 0; j < Shift; ++j)
7437 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7443 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7444 for (int i = 0; i != Size; i += Scale) {
7445 unsigned Pos = Left ? i + Shift : i;
7446 unsigned Low = Left ? i : i + Shift;
7447 unsigned Len = Scale - Shift;
7448 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7449 Low + (V == V1 ? 0 : Size)))
7453 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7454 bool ByteShift = ShiftEltBits > 64;
7455 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7456 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7457 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7459 // Normalize the scale for byte shifts to still produce an i64 element
7461 Scale = ByteShift ? Scale / 2 : Scale;
7463 // We need to round trip through the appropriate type for the shift.
7464 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7465 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7466 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7467 "Illegal integer vector type");
7468 V = DAG.getBitcast(ShiftVT, V);
7470 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7471 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7472 return DAG.getBitcast(VT, V);
7475 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7476 // keep doubling the size of the integer elements up to that. We can
7477 // then shift the elements of the integer vector by whole multiples of
7478 // their width within the elements of the larger integer vector. Test each
7479 // multiple to see if we can find a match with the moved element indices
7480 // and that the shifted in elements are all zeroable.
7481 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7482 for (int Shift = 1; Shift != Scale; ++Shift)
7483 for (bool Left : {true, false})
7484 if (CheckZeros(Shift, Scale, Left))
7485 for (SDValue V : {V1, V2})
7486 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7493 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7494 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7495 SDValue V2, ArrayRef<int> Mask,
7496 SelectionDAG &DAG) {
7497 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7498 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7500 int Size = Mask.size();
7501 int HalfSize = Size / 2;
7502 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7504 // Upper half must be undefined.
7505 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7508 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7509 // Remainder of lower half result is zero and upper half is all undef.
7510 auto LowerAsEXTRQ = [&]() {
7511 // Determine the extraction length from the part of the
7512 // lower half that isn't zeroable.
7514 for (; Len > 0; --Len)
7515 if (!Zeroable[Len - 1])
7517 assert(Len > 0 && "Zeroable shuffle mask");
7519 // Attempt to match first Len sequential elements from the lower half.
7522 for (int i = 0; i != Len; ++i) {
7526 SDValue &V = (M < Size ? V1 : V2);
7529 // The extracted elements must start at a valid index and all mask
7530 // elements must be in the lower half.
7531 if (i > M || M >= HalfSize)
7534 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7545 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7546 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7547 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7548 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7549 DAG.getConstant(BitLen, DL, MVT::i8),
7550 DAG.getConstant(BitIdx, DL, MVT::i8));
7553 if (SDValue ExtrQ = LowerAsEXTRQ())
7556 // INSERTQ: Extract lowest Len elements from lower half of second source and
7557 // insert over first source, starting at Idx.
7558 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7559 auto LowerAsInsertQ = [&]() {
7560 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7563 // Attempt to match first source from mask before insertion point.
7564 if (isUndefInRange(Mask, 0, Idx)) {
7566 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7568 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7574 // Extend the extraction length looking to match both the insertion of
7575 // the second source and the remaining elements of the first.
7576 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7581 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7583 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7589 // Match the remaining elements of the lower half.
7590 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7592 } else if ((!Base || (Base == V1)) &&
7593 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7595 } else if ((!Base || (Base == V2)) &&
7596 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7603 // We may not have a base (first source) - this can safely be undefined.
7605 Base = DAG.getUNDEF(VT);
7607 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7608 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7609 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7610 DAG.getConstant(BitLen, DL, MVT::i8),
7611 DAG.getConstant(BitIdx, DL, MVT::i8));
7618 if (SDValue InsertQ = LowerAsInsertQ())
7624 /// \brief Lower a vector shuffle as a zero or any extension.
7626 /// Given a specific number of elements, element bit width, and extension
7627 /// stride, produce either a zero or any extension based on the available
7628 /// features of the subtarget. The extended elements are consecutive and
7629 /// begin and can start from an offseted element index in the input; to
7630 /// avoid excess shuffling the offset must either being in the bottom lane
7631 /// or at the start of a higher lane. All extended elements must be from
7633 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7634 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7635 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7636 assert(Scale > 1 && "Need a scale to extend.");
7637 int EltBits = VT.getScalarSizeInBits();
7638 int NumElements = VT.getVectorNumElements();
7639 int NumEltsPerLane = 128 / EltBits;
7640 int OffsetLane = Offset / NumEltsPerLane;
7641 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7642 "Only 8, 16, and 32 bit elements can be extended.");
7643 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7644 assert(0 <= Offset && "Extension offset must be positive.");
7645 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7646 "Extension offset must be in the first lane or start an upper lane.");
7648 // Check that an index is in same lane as the base offset.
7649 auto SafeOffset = [&](int Idx) {
7650 return OffsetLane == (Idx / NumEltsPerLane);
7653 // Shift along an input so that the offset base moves to the first element.
7654 auto ShuffleOffset = [&](SDValue V) {
7658 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7659 for (int i = 0; i * Scale < NumElements; ++i) {
7660 int SrcIdx = i + Offset;
7661 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7663 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7666 // Found a valid zext mask! Try various lowering strategies based on the
7667 // input type and available ISA extensions.
7668 if (Subtarget->hasSSE41()) {
7669 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7670 // PUNPCK will catch this in a later shuffle match.
7671 if (Offset && Scale == 2 && VT.is128BitVector())
7673 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7674 NumElements / Scale);
7675 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7676 return DAG.getBitcast(VT, InputV);
7679 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7681 // For any extends we can cheat for larger element sizes and use shuffle
7682 // instructions that can fold with a load and/or copy.
7683 if (AnyExt && EltBits == 32) {
7684 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7686 return DAG.getBitcast(
7687 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7688 DAG.getBitcast(MVT::v4i32, InputV),
7689 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7691 if (AnyExt && EltBits == 16 && Scale > 2) {
7692 int PSHUFDMask[4] = {Offset / 2, -1,
7693 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7694 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7695 DAG.getBitcast(MVT::v4i32, InputV),
7696 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7697 int PSHUFWMask[4] = {1, -1, -1, -1};
7698 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7699 return DAG.getBitcast(
7700 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7701 DAG.getBitcast(MVT::v8i16, InputV),
7702 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7705 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7707 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7708 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7709 assert(VT.is128BitVector() && "Unexpected vector width!");
7711 int LoIdx = Offset * EltBits;
7712 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7713 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7714 DAG.getConstant(EltBits, DL, MVT::i8),
7715 DAG.getConstant(LoIdx, DL, MVT::i8)));
7717 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7718 !SafeOffset(Offset + 1))
7719 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7721 int HiIdx = (Offset + 1) * EltBits;
7722 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7723 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7724 DAG.getConstant(EltBits, DL, MVT::i8),
7725 DAG.getConstant(HiIdx, DL, MVT::i8)));
7726 return DAG.getNode(ISD::BITCAST, DL, VT,
7727 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7730 // If this would require more than 2 unpack instructions to expand, use
7731 // pshufb when available. We can only use more than 2 unpack instructions
7732 // when zero extending i8 elements which also makes it easier to use pshufb.
7733 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7734 assert(NumElements == 16 && "Unexpected byte vector width!");
7735 SDValue PSHUFBMask[16];
7736 for (int i = 0; i < 16; ++i) {
7737 int Idx = Offset + (i / Scale);
7738 PSHUFBMask[i] = DAG.getConstant(
7739 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7741 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7742 return DAG.getBitcast(VT,
7743 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7744 DAG.getNode(ISD::BUILD_VECTOR, DL,
7745 MVT::v16i8, PSHUFBMask)));
7748 // If we are extending from an offset, ensure we start on a boundary that
7749 // we can unpack from.
7750 int AlignToUnpack = Offset % (NumElements / Scale);
7751 if (AlignToUnpack) {
7752 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7753 for (int i = AlignToUnpack; i < NumElements; ++i)
7754 ShMask[i - AlignToUnpack] = i;
7755 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7756 Offset -= AlignToUnpack;
7759 // Otherwise emit a sequence of unpacks.
7761 unsigned UnpackLoHi = X86ISD::UNPCKL;
7762 if (Offset >= (NumElements / 2)) {
7763 UnpackLoHi = X86ISD::UNPCKH;
7764 Offset -= (NumElements / 2);
7767 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7768 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7769 : getZeroVector(InputVT, Subtarget, DAG, DL);
7770 InputV = DAG.getBitcast(InputVT, InputV);
7771 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7775 } while (Scale > 1);
7776 return DAG.getBitcast(VT, InputV);
7779 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7781 /// This routine will try to do everything in its power to cleverly lower
7782 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7783 /// check for the profitability of this lowering, it tries to aggressively
7784 /// match this pattern. It will use all of the micro-architectural details it
7785 /// can to emit an efficient lowering. It handles both blends with all-zero
7786 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7787 /// masking out later).
7789 /// The reason we have dedicated lowering for zext-style shuffles is that they
7790 /// are both incredibly common and often quite performance sensitive.
7791 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7792 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7793 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7794 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7796 int Bits = VT.getSizeInBits();
7797 int NumLanes = Bits / 128;
7798 int NumElements = VT.getVectorNumElements();
7799 int NumEltsPerLane = NumElements / NumLanes;
7800 assert(VT.getScalarSizeInBits() <= 32 &&
7801 "Exceeds 32-bit integer zero extension limit");
7802 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7804 // Define a helper function to check a particular ext-scale and lower to it if
7806 auto Lower = [&](int Scale) -> SDValue {
7811 for (int i = 0; i < NumElements; ++i) {
7814 continue; // Valid anywhere but doesn't tell us anything.
7815 if (i % Scale != 0) {
7816 // Each of the extended elements need to be zeroable.
7820 // We no longer are in the anyext case.
7825 // Each of the base elements needs to be consecutive indices into the
7826 // same input vector.
7827 SDValue V = M < NumElements ? V1 : V2;
7828 M = M % NumElements;
7831 Offset = M - (i / Scale);
7832 } else if (InputV != V)
7833 return SDValue(); // Flip-flopping inputs.
7835 // Offset must start in the lowest 128-bit lane or at the start of an
7837 // FIXME: Is it ever worth allowing a negative base offset?
7838 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7839 (Offset % NumEltsPerLane) == 0))
7842 // If we are offsetting, all referenced entries must come from the same
7844 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7847 if ((M % NumElements) != (Offset + (i / Scale)))
7848 return SDValue(); // Non-consecutive strided elements.
7852 // If we fail to find an input, we have a zero-shuffle which should always
7853 // have already been handled.
7854 // FIXME: Maybe handle this here in case during blending we end up with one?
7858 // If we are offsetting, don't extend if we only match a single input, we
7859 // can always do better by using a basic PSHUF or PUNPCK.
7860 if (Offset != 0 && Matches < 2)
7863 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7864 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7867 // The widest scale possible for extending is to a 64-bit integer.
7868 assert(Bits % 64 == 0 &&
7869 "The number of bits in a vector must be divisible by 64 on x86!");
7870 int NumExtElements = Bits / 64;
7872 // Each iteration, try extending the elements half as much, but into twice as
7874 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7875 assert(NumElements % NumExtElements == 0 &&
7876 "The input vector size must be divisible by the extended size.");
7877 if (SDValue V = Lower(NumElements / NumExtElements))
7881 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7885 // Returns one of the source operands if the shuffle can be reduced to a
7886 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7887 auto CanZExtLowHalf = [&]() {
7888 for (int i = NumElements / 2; i != NumElements; ++i)
7891 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7893 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7898 if (SDValue V = CanZExtLowHalf()) {
7899 V = DAG.getBitcast(MVT::v2i64, V);
7900 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7901 return DAG.getBitcast(VT, V);
7904 // No viable ext lowering found.
7908 /// \brief Try to get a scalar value for a specific element of a vector.
7910 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7911 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7912 SelectionDAG &DAG) {
7913 MVT VT = V.getSimpleValueType();
7914 MVT EltVT = VT.getVectorElementType();
7915 while (V.getOpcode() == ISD::BITCAST)
7916 V = V.getOperand(0);
7917 // If the bitcasts shift the element size, we can't extract an equivalent
7919 MVT NewVT = V.getSimpleValueType();
7920 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7923 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7924 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7925 // Ensure the scalar operand is the same size as the destination.
7926 // FIXME: Add support for scalar truncation where possible.
7927 SDValue S = V.getOperand(Idx);
7928 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7929 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7935 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7937 /// This is particularly important because the set of instructions varies
7938 /// significantly based on whether the operand is a load or not.
7939 static bool isShuffleFoldableLoad(SDValue V) {
7940 while (V.getOpcode() == ISD::BITCAST)
7941 V = V.getOperand(0);
7943 return ISD::isNON_EXTLoad(V.getNode());
7946 /// \brief Try to lower insertion of a single element into a zero vector.
7948 /// This is a common pattern that we have especially efficient patterns to lower
7949 /// across all subtarget feature sets.
7950 static SDValue lowerVectorShuffleAsElementInsertion(
7951 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7952 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7953 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7955 MVT EltVT = VT.getVectorElementType();
7957 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7958 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7960 bool IsV1Zeroable = true;
7961 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7962 if (i != V2Index && !Zeroable[i]) {
7963 IsV1Zeroable = false;
7967 // Check for a single input from a SCALAR_TO_VECTOR node.
7968 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7969 // all the smarts here sunk into that routine. However, the current
7970 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7971 // vector shuffle lowering is dead.
7972 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7974 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7975 // We need to zext the scalar if it is smaller than an i32.
7976 V2S = DAG.getBitcast(EltVT, V2S);
7977 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7978 // Using zext to expand a narrow element won't work for non-zero
7983 // Zero-extend directly to i32.
7985 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7987 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7988 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7989 EltVT == MVT::i16) {
7990 // Either not inserting from the low element of the input or the input
7991 // element size is too small to use VZEXT_MOVL to clear the high bits.
7995 if (!IsV1Zeroable) {
7996 // If V1 can't be treated as a zero vector we have fewer options to lower
7997 // this. We can't support integer vectors or non-zero targets cheaply, and
7998 // the V1 elements can't be permuted in any way.
7999 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8000 if (!VT.isFloatingPoint() || V2Index != 0)
8002 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8003 V1Mask[V2Index] = -1;
8004 if (!isNoopShuffleMask(V1Mask))
8006 // This is essentially a special case blend operation, but if we have
8007 // general purpose blend operations, they are always faster. Bail and let
8008 // the rest of the lowering handle these as blends.
8009 if (Subtarget->hasSSE41())
8012 // Otherwise, use MOVSD or MOVSS.
8013 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8014 "Only two types of floating point element types to handle!");
8015 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8019 // This lowering only works for the low element with floating point vectors.
8020 if (VT.isFloatingPoint() && V2Index != 0)
8023 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8025 V2 = DAG.getBitcast(VT, V2);
8028 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8029 // the desired position. Otherwise it is more efficient to do a vector
8030 // shift left. We know that we can do a vector shift left because all
8031 // the inputs are zero.
8032 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8033 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8034 V2Shuffle[V2Index] = 0;
8035 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8037 V2 = DAG.getBitcast(MVT::v2i64, V2);
8039 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8040 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
8041 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
8042 DAG.getDataLayout(), VT)));
8043 V2 = DAG.getBitcast(VT, V2);
8049 /// \brief Try to lower broadcast of a single - truncated - integer element,
8050 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
8052 /// This assumes we have AVX2.
8053 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
8055 const X86Subtarget *Subtarget,
8056 SelectionDAG &DAG) {
8057 assert(Subtarget->hasAVX2() &&
8058 "We can only lower integer broadcasts with AVX2!");
8060 EVT EltVT = VT.getVectorElementType();
8061 EVT V0VT = V0.getValueType();
8063 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
8064 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
8066 EVT V0EltVT = V0VT.getVectorElementType();
8067 if (!V0EltVT.isInteger())
8070 const unsigned EltSize = EltVT.getSizeInBits();
8071 const unsigned V0EltSize = V0EltVT.getSizeInBits();
8073 // This is only a truncation if the original element type is larger.
8074 if (V0EltSize <= EltSize)
8077 assert(((V0EltSize % EltSize) == 0) &&
8078 "Scalar type sizes must all be powers of 2 on x86!");
8080 const unsigned V0Opc = V0.getOpcode();
8081 const unsigned Scale = V0EltSize / EltSize;
8082 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
8084 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
8085 V0Opc != ISD::BUILD_VECTOR)
8088 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
8090 // If we're extracting non-least-significant bits, shift so we can truncate.
8091 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
8092 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
8093 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
8094 if (const int OffsetIdx = BroadcastIdx % Scale)
8095 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
8096 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
8098 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
8099 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
8102 /// \brief Try to lower broadcast of a single element.
8104 /// For convenience, this code also bundles all of the subtarget feature set
8105 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8106 /// a convenient way to factor it out.
8107 /// FIXME: This is very similar to LowerVectorBroadcast - can we merge them?
8108 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
8110 const X86Subtarget *Subtarget,
8111 SelectionDAG &DAG) {
8112 if (!Subtarget->hasAVX())
8114 if (VT.isInteger() && !Subtarget->hasAVX2())
8117 // Check that the mask is a broadcast.
8118 int BroadcastIdx = -1;
8120 if (M >= 0 && BroadcastIdx == -1)
8122 else if (M >= 0 && M != BroadcastIdx)
8125 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8126 "a sorted mask where the broadcast "
8129 // Go up the chain of (vector) values to find a scalar load that we can
8130 // combine with the broadcast.
8132 switch (V.getOpcode()) {
8133 case ISD::CONCAT_VECTORS: {
8134 int OperandSize = Mask.size() / V.getNumOperands();
8135 V = V.getOperand(BroadcastIdx / OperandSize);
8136 BroadcastIdx %= OperandSize;
8140 case ISD::INSERT_SUBVECTOR: {
8141 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8142 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8146 int BeginIdx = (int)ConstantIdx->getZExtValue();
8148 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
8149 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8150 BroadcastIdx -= BeginIdx;
8161 // Check if this is a broadcast of a scalar. We special case lowering
8162 // for scalars so that we can more effectively fold with loads.
8163 // First, look through bitcast: if the original value has a larger element
8164 // type than the shuffle, the broadcast element is in essence truncated.
8165 // Make that explicit to ease folding.
8166 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
8167 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
8168 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
8169 return TruncBroadcast;
8171 // Also check the simpler case, where we can directly reuse the scalar.
8172 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8173 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8174 V = V.getOperand(BroadcastIdx);
8176 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8177 // Only AVX2 has register broadcasts.
8178 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8180 } else if (MayFoldLoad(V) && !cast<LoadSDNode>(V)->isVolatile()) {
8181 // If we are broadcasting a load that is only used by the shuffle
8182 // then we can reduce the vector load to the broadcasted scalar load.
8183 LoadSDNode *Ld = cast<LoadSDNode>(V);
8184 SDValue BaseAddr = Ld->getOperand(1);
8185 EVT AddrVT = BaseAddr.getValueType();
8186 EVT SVT = VT.getScalarType();
8187 unsigned Offset = BroadcastIdx * SVT.getStoreSize();
8188 SDValue NewAddr = DAG.getNode(
8189 ISD::ADD, DL, AddrVT, BaseAddr,
8190 DAG.getConstant(Offset, DL, AddrVT));
8191 V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
8192 DAG.getMachineFunction().getMachineMemOperand(
8193 Ld->getMemOperand(), Offset, SVT.getStoreSize()));
8194 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8195 // We can't broadcast from a vector register without AVX2, and we can only
8196 // broadcast from the zero-element of a vector register.
8200 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8203 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8204 // INSERTPS when the V1 elements are already in the correct locations
8205 // because otherwise we can just always use two SHUFPS instructions which
8206 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8207 // perform INSERTPS if a single V1 element is out of place and all V2
8208 // elements are zeroable.
8209 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8211 SelectionDAG &DAG) {
8212 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8213 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8214 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8215 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8217 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8220 int V1DstIndex = -1;
8221 int V2DstIndex = -1;
8222 bool V1UsedInPlace = false;
8224 for (int i = 0; i < 4; ++i) {
8225 // Synthesize a zero mask from the zeroable elements (includes undefs).
8231 // Flag if we use any V1 inputs in place.
8233 V1UsedInPlace = true;
8237 // We can only insert a single non-zeroable element.
8238 if (V1DstIndex != -1 || V2DstIndex != -1)
8242 // V1 input out of place for insertion.
8245 // V2 input for insertion.
8250 // Don't bother if we have no (non-zeroable) element for insertion.
8251 if (V1DstIndex == -1 && V2DstIndex == -1)
8254 // Determine element insertion src/dst indices. The src index is from the
8255 // start of the inserted vector, not the start of the concatenated vector.
8256 unsigned V2SrcIndex = 0;
8257 if (V1DstIndex != -1) {
8258 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8259 // and don't use the original V2 at all.
8260 V2SrcIndex = Mask[V1DstIndex];
8261 V2DstIndex = V1DstIndex;
8264 V2SrcIndex = Mask[V2DstIndex] - 4;
8267 // If no V1 inputs are used in place, then the result is created only from
8268 // the zero mask and the V2 insertion - so remove V1 dependency.
8270 V1 = DAG.getUNDEF(MVT::v4f32);
8272 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8273 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8275 // Insert the V2 element into the desired position.
8277 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8278 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8281 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8282 /// UNPCK instruction.
8284 /// This specifically targets cases where we end up with alternating between
8285 /// the two inputs, and so can permute them into something that feeds a single
8286 /// UNPCK instruction. Note that this routine only targets integer vectors
8287 /// because for floating point vectors we have a generalized SHUFPS lowering
8288 /// strategy that handles everything that doesn't *exactly* match an unpack,
8289 /// making this clever lowering unnecessary.
8290 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8291 SDValue V1, SDValue V2,
8293 SelectionDAG &DAG) {
8294 assert(!VT.isFloatingPoint() &&
8295 "This routine only supports integer vectors.");
8296 assert(!isSingleInputShuffleMask(Mask) &&
8297 "This routine should only be used when blending two inputs.");
8298 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8300 int Size = Mask.size();
8302 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8303 return M >= 0 && M % Size < Size / 2;
8305 int NumHiInputs = std::count_if(
8306 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8308 bool UnpackLo = NumLoInputs >= NumHiInputs;
8310 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8311 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8312 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8314 for (int i = 0; i < Size; ++i) {
8318 // Each element of the unpack contains Scale elements from this mask.
8319 int UnpackIdx = i / Scale;
8321 // We only handle the case where V1 feeds the first slots of the unpack.
8322 // We rely on canonicalization to ensure this is the case.
8323 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8326 // Setup the mask for this input. The indexing is tricky as we have to
8327 // handle the unpack stride.
8328 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8329 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8333 // If we will have to shuffle both inputs to use the unpack, check whether
8334 // we can just unpack first and shuffle the result. If so, skip this unpack.
8335 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8336 !isNoopShuffleMask(V2Mask))
8339 // Shuffle the inputs into place.
8340 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8341 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8343 // Cast the inputs to the type we will use to unpack them.
8344 V1 = DAG.getBitcast(UnpackVT, V1);
8345 V2 = DAG.getBitcast(UnpackVT, V2);
8347 // Unpack the inputs and cast the result back to the desired type.
8348 return DAG.getBitcast(
8349 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8353 // We try each unpack from the largest to the smallest to try and find one
8354 // that fits this mask.
8355 int OrigNumElements = VT.getVectorNumElements();
8356 int OrigScalarSize = VT.getScalarSizeInBits();
8357 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8358 int Scale = ScalarSize / OrigScalarSize;
8359 int NumElements = OrigNumElements / Scale;
8360 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8361 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8365 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8367 if (NumLoInputs == 0 || NumHiInputs == 0) {
8368 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8369 "We have to have *some* inputs!");
8370 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8372 // FIXME: We could consider the total complexity of the permute of each
8373 // possible unpacking. Or at the least we should consider how many
8374 // half-crossings are created.
8375 // FIXME: We could consider commuting the unpacks.
8377 SmallVector<int, 32> PermMask;
8378 PermMask.assign(Size, -1);
8379 for (int i = 0; i < Size; ++i) {
8383 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8386 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8388 return DAG.getVectorShuffle(
8389 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8391 DAG.getUNDEF(VT), PermMask);
8397 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8399 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8400 /// support for floating point shuffles but not integer shuffles. These
8401 /// instructions will incur a domain crossing penalty on some chips though so
8402 /// it is better to avoid lowering through this for integer vectors where
8404 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8405 const X86Subtarget *Subtarget,
8406 SelectionDAG &DAG) {
8408 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8409 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8410 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8411 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8412 ArrayRef<int> Mask = SVOp->getMask();
8413 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8415 if (isSingleInputShuffleMask(Mask)) {
8416 // Use low duplicate instructions for masks that match their pattern.
8417 if (Subtarget->hasSSE3())
8418 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8419 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8421 // Straight shuffle of a single input vector. Simulate this by using the
8422 // single input as both of the "inputs" to this instruction..
8423 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8425 if (Subtarget->hasAVX()) {
8426 // If we have AVX, we can use VPERMILPS which will allow folding a load
8427 // into the shuffle.
8428 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8429 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8432 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8433 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8435 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8436 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8438 // If we have a single input, insert that into V1 if we can do so cheaply.
8439 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8440 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8441 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8443 // Try inverting the insertion since for v2 masks it is easy to do and we
8444 // can't reliably sort the mask one way or the other.
8445 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8446 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8447 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8448 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8452 // Try to use one of the special instruction patterns to handle two common
8453 // blend patterns if a zero-blend above didn't work.
8454 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8455 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8456 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8457 // We can either use a special instruction to load over the low double or
8458 // to move just the low double.
8460 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8462 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8464 if (Subtarget->hasSSE41())
8465 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8469 // Use dedicated unpack instructions for masks that match their pattern.
8471 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8474 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8475 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8476 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8479 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8481 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8482 /// the integer unit to minimize domain crossing penalties. However, for blends
8483 /// it falls back to the floating point shuffle operation with appropriate bit
8485 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8486 const X86Subtarget *Subtarget,
8487 SelectionDAG &DAG) {
8489 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8490 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8491 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8492 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8493 ArrayRef<int> Mask = SVOp->getMask();
8494 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8496 if (isSingleInputShuffleMask(Mask)) {
8497 // Check for being able to broadcast a single element.
8498 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8499 Mask, Subtarget, DAG))
8502 // Straight shuffle of a single input vector. For everything from SSE2
8503 // onward this has a single fast instruction with no scary immediates.
8504 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8505 V1 = DAG.getBitcast(MVT::v4i32, V1);
8506 int WidenedMask[4] = {
8507 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8508 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8509 return DAG.getBitcast(
8511 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8512 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8514 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8515 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8516 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8517 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8519 // If we have a blend of two PACKUS operations an the blend aligns with the
8520 // low and half halves, we can just merge the PACKUS operations. This is
8521 // particularly important as it lets us merge shuffles that this routine itself
8523 auto GetPackNode = [](SDValue V) {
8524 while (V.getOpcode() == ISD::BITCAST)
8525 V = V.getOperand(0);
8527 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8529 if (SDValue V1Pack = GetPackNode(V1))
8530 if (SDValue V2Pack = GetPackNode(V2))
8531 return DAG.getBitcast(MVT::v2i64,
8532 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8533 Mask[0] == 0 ? V1Pack.getOperand(0)
8534 : V1Pack.getOperand(1),
8535 Mask[1] == 2 ? V2Pack.getOperand(0)
8536 : V2Pack.getOperand(1)));
8538 // Try to use shift instructions.
8540 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8543 // When loading a scalar and then shuffling it into a vector we can often do
8544 // the insertion cheaply.
8545 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8546 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8548 // Try inverting the insertion since for v2 masks it is easy to do and we
8549 // can't reliably sort the mask one way or the other.
8550 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8551 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8552 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8555 // We have different paths for blend lowering, but they all must use the
8556 // *exact* same predicate.
8557 bool IsBlendSupported = Subtarget->hasSSE41();
8558 if (IsBlendSupported)
8559 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8563 // Use dedicated unpack instructions for masks that match their pattern.
8565 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8568 // Try to use byte rotation instructions.
8569 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8570 if (Subtarget->hasSSSE3())
8571 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8572 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8575 // If we have direct support for blends, we should lower by decomposing into
8576 // a permute. That will be faster than the domain cross.
8577 if (IsBlendSupported)
8578 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8581 // We implement this with SHUFPD which is pretty lame because it will likely
8582 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8583 // However, all the alternatives are still more cycles and newer chips don't
8584 // have this problem. It would be really nice if x86 had better shuffles here.
8585 V1 = DAG.getBitcast(MVT::v2f64, V1);
8586 V2 = DAG.getBitcast(MVT::v2f64, V2);
8587 return DAG.getBitcast(MVT::v2i64,
8588 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8591 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8593 /// This is used to disable more specialized lowerings when the shufps lowering
8594 /// will happen to be efficient.
8595 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8596 // This routine only handles 128-bit shufps.
8597 assert(Mask.size() == 4 && "Unsupported mask size!");
8599 // To lower with a single SHUFPS we need to have the low half and high half
8600 // each requiring a single input.
8601 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8603 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8609 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8611 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8612 /// It makes no assumptions about whether this is the *best* lowering, it simply
8614 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8615 ArrayRef<int> Mask, SDValue V1,
8616 SDValue V2, SelectionDAG &DAG) {
8617 SDValue LowV = V1, HighV = V2;
8618 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8621 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8623 if (NumV2Elements == 1) {
8625 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8628 // Compute the index adjacent to V2Index and in the same half by toggling
8630 int V2AdjIndex = V2Index ^ 1;
8632 if (Mask[V2AdjIndex] == -1) {
8633 // Handles all the cases where we have a single V2 element and an undef.
8634 // This will only ever happen in the high lanes because we commute the
8635 // vector otherwise.
8637 std::swap(LowV, HighV);
8638 NewMask[V2Index] -= 4;
8640 // Handle the case where the V2 element ends up adjacent to a V1 element.
8641 // To make this work, blend them together as the first step.
8642 int V1Index = V2AdjIndex;
8643 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8644 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8645 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8647 // Now proceed to reconstruct the final blend as we have the necessary
8648 // high or low half formed.
8655 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8656 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8658 } else if (NumV2Elements == 2) {
8659 if (Mask[0] < 4 && Mask[1] < 4) {
8660 // Handle the easy case where we have V1 in the low lanes and V2 in the
8664 } else if (Mask[2] < 4 && Mask[3] < 4) {
8665 // We also handle the reversed case because this utility may get called
8666 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8667 // arrange things in the right direction.
8673 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8674 // trying to place elements directly, just blend them and set up the final
8675 // shuffle to place them.
8677 // The first two blend mask elements are for V1, the second two are for
8679 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8680 Mask[2] < 4 ? Mask[2] : Mask[3],
8681 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8682 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8683 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8684 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8686 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8689 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8690 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8691 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8692 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8695 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8696 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8699 /// \brief Lower 4-lane 32-bit floating point shuffles.
8701 /// Uses instructions exclusively from the floating point unit to minimize
8702 /// domain crossing penalties, as these are sufficient to implement all v4f32
8704 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8705 const X86Subtarget *Subtarget,
8706 SelectionDAG &DAG) {
8708 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8709 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8710 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8711 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8712 ArrayRef<int> Mask = SVOp->getMask();
8713 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8716 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8718 if (NumV2Elements == 0) {
8719 // Check for being able to broadcast a single element.
8720 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8721 Mask, Subtarget, DAG))
8724 // Use even/odd duplicate instructions for masks that match their pattern.
8725 if (Subtarget->hasSSE3()) {
8726 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8727 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8728 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8729 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8732 if (Subtarget->hasAVX()) {
8733 // If we have AVX, we can use VPERMILPS which will allow folding a load
8734 // into the shuffle.
8735 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8736 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8739 // Otherwise, use a straight shuffle of a single input vector. We pass the
8740 // input vector to both operands to simulate this with a SHUFPS.
8741 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8742 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8745 // There are special ways we can lower some single-element blends. However, we
8746 // have custom ways we can lower more complex single-element blends below that
8747 // we defer to if both this and BLENDPS fail to match, so restrict this to
8748 // when the V2 input is targeting element 0 of the mask -- that is the fast
8750 if (NumV2Elements == 1 && Mask[0] >= 4)
8751 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8752 Mask, Subtarget, DAG))
8755 if (Subtarget->hasSSE41()) {
8756 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8760 // Use INSERTPS if we can complete the shuffle efficiently.
8761 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8764 if (!isSingleSHUFPSMask(Mask))
8765 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8766 DL, MVT::v4f32, V1, V2, Mask, DAG))
8770 // Use dedicated unpack instructions for masks that match their pattern.
8772 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8775 // Otherwise fall back to a SHUFPS lowering strategy.
8776 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8779 /// \brief Lower 4-lane i32 vector shuffles.
8781 /// We try to handle these with integer-domain shuffles where we can, but for
8782 /// blends we use the floating point domain blend instructions.
8783 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8784 const X86Subtarget *Subtarget,
8785 SelectionDAG &DAG) {
8787 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8788 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8789 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8790 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8791 ArrayRef<int> Mask = SVOp->getMask();
8792 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8794 // Whenever we can lower this as a zext, that instruction is strictly faster
8795 // than any alternative. It also allows us to fold memory operands into the
8796 // shuffle in many cases.
8797 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8798 Mask, Subtarget, DAG))
8802 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8804 if (NumV2Elements == 0) {
8805 // Check for being able to broadcast a single element.
8806 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8807 Mask, Subtarget, DAG))
8810 // Straight shuffle of a single input vector. For everything from SSE2
8811 // onward this has a single fast instruction with no scary immediates.
8812 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8813 // but we aren't actually going to use the UNPCK instruction because doing
8814 // so prevents folding a load into this instruction or making a copy.
8815 const int UnpackLoMask[] = {0, 0, 1, 1};
8816 const int UnpackHiMask[] = {2, 2, 3, 3};
8817 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8818 Mask = UnpackLoMask;
8819 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8820 Mask = UnpackHiMask;
8822 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8823 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8826 // Try to use shift instructions.
8828 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8831 // There are special ways we can lower some single-element blends.
8832 if (NumV2Elements == 1)
8833 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8834 Mask, Subtarget, DAG))
8837 // We have different paths for blend lowering, but they all must use the
8838 // *exact* same predicate.
8839 bool IsBlendSupported = Subtarget->hasSSE41();
8840 if (IsBlendSupported)
8841 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8845 if (SDValue Masked =
8846 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8849 // Use dedicated unpack instructions for masks that match their pattern.
8851 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8854 // Try to use byte rotation instructions.
8855 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8856 if (Subtarget->hasSSSE3())
8857 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8858 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8861 // If we have direct support for blends, we should lower by decomposing into
8862 // a permute. That will be faster than the domain cross.
8863 if (IsBlendSupported)
8864 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8867 // Try to lower by permuting the inputs into an unpack instruction.
8868 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8872 // We implement this with SHUFPS because it can blend from two vectors.
8873 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8874 // up the inputs, bypassing domain shift penalties that we would encur if we
8875 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8877 return DAG.getBitcast(
8879 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8880 DAG.getBitcast(MVT::v4f32, V2), Mask));
8883 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8884 /// shuffle lowering, and the most complex part.
8886 /// The lowering strategy is to try to form pairs of input lanes which are
8887 /// targeted at the same half of the final vector, and then use a dword shuffle
8888 /// to place them onto the right half, and finally unpack the paired lanes into
8889 /// their final position.
8891 /// The exact breakdown of how to form these dword pairs and align them on the
8892 /// correct sides is really tricky. See the comments within the function for
8893 /// more of the details.
8895 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8896 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8897 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8898 /// vector, form the analogous 128-bit 8-element Mask.
8899 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8900 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8901 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8902 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8903 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8905 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8906 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8907 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8909 SmallVector<int, 4> LoInputs;
8910 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8911 [](int M) { return M >= 0; });
8912 std::sort(LoInputs.begin(), LoInputs.end());
8913 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8914 SmallVector<int, 4> HiInputs;
8915 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8916 [](int M) { return M >= 0; });
8917 std::sort(HiInputs.begin(), HiInputs.end());
8918 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8920 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8921 int NumHToL = LoInputs.size() - NumLToL;
8923 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8924 int NumHToH = HiInputs.size() - NumLToH;
8925 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8926 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8927 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8928 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8930 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8931 // such inputs we can swap two of the dwords across the half mark and end up
8932 // with <=2 inputs to each half in each half. Once there, we can fall through
8933 // to the generic code below. For example:
8935 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8936 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8938 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8939 // and an existing 2-into-2 on the other half. In this case we may have to
8940 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8941 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8942 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8943 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8944 // half than the one we target for fixing) will be fixed when we re-enter this
8945 // path. We will also combine away any sequence of PSHUFD instructions that
8946 // result into a single instruction. Here is an example of the tricky case:
8948 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8949 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8951 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8953 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8954 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8956 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8957 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8959 // The result is fine to be handled by the generic logic.
8960 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8961 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8962 int AOffset, int BOffset) {
8963 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8964 "Must call this with A having 3 or 1 inputs from the A half.");
8965 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8966 "Must call this with B having 1 or 3 inputs from the B half.");
8967 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8968 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8970 bool ThreeAInputs = AToAInputs.size() == 3;
8972 // Compute the index of dword with only one word among the three inputs in
8973 // a half by taking the sum of the half with three inputs and subtracting
8974 // the sum of the actual three inputs. The difference is the remaining
8977 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8978 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8979 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8980 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8981 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8982 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8983 int TripleNonInputIdx =
8984 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8985 TripleDWord = TripleNonInputIdx / 2;
8987 // We use xor with one to compute the adjacent DWord to whichever one the
8989 OneInputDWord = (OneInput / 2) ^ 1;
8991 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8992 // and BToA inputs. If there is also such a problem with the BToB and AToB
8993 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8994 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8995 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8996 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8997 // Compute how many inputs will be flipped by swapping these DWords. We
8999 // to balance this to ensure we don't form a 3-1 shuffle in the other
9001 int NumFlippedAToBInputs =
9002 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
9003 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
9004 int NumFlippedBToBInputs =
9005 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
9006 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
9007 if ((NumFlippedAToBInputs == 1 &&
9008 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
9009 (NumFlippedBToBInputs == 1 &&
9010 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
9011 // We choose whether to fix the A half or B half based on whether that
9012 // half has zero flipped inputs. At zero, we may not be able to fix it
9013 // with that half. We also bias towards fixing the B half because that
9014 // will more commonly be the high half, and we have to bias one way.
9015 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
9016 ArrayRef<int> Inputs) {
9017 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
9018 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
9019 PinnedIdx ^ 1) != Inputs.end();
9020 // Determine whether the free index is in the flipped dword or the
9021 // unflipped dword based on where the pinned index is. We use this bit
9022 // in an xor to conditionally select the adjacent dword.
9023 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
9024 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9025 FixFreeIdx) != Inputs.end();
9026 if (IsFixIdxInput == IsFixFreeIdxInput)
9028 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
9029 FixFreeIdx) != Inputs.end();
9030 assert(IsFixIdxInput != IsFixFreeIdxInput &&
9031 "We need to be changing the number of flipped inputs!");
9032 int PSHUFHalfMask[] = {0, 1, 2, 3};
9033 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
9034 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
9036 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
9039 if (M != -1 && M == FixIdx)
9041 else if (M != -1 && M == FixFreeIdx)
9044 if (NumFlippedBToBInputs != 0) {
9046 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9047 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
9049 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
9050 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
9051 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9056 int PSHUFDMask[] = {0, 1, 2, 3};
9057 PSHUFDMask[ADWord] = BDWord;
9058 PSHUFDMask[BDWord] = ADWord;
9061 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9062 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9064 // Adjust the mask to match the new locations of A and B.
9066 if (M != -1 && M/2 == ADWord)
9067 M = 2 * BDWord + M % 2;
9068 else if (M != -1 && M/2 == BDWord)
9069 M = 2 * ADWord + M % 2;
9071 // Recurse back into this routine to re-compute state now that this isn't
9072 // a 3 and 1 problem.
9073 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
9076 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9077 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9078 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9079 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9081 // At this point there are at most two inputs to the low and high halves from
9082 // each half. That means the inputs can always be grouped into dwords and
9083 // those dwords can then be moved to the correct half with a dword shuffle.
9084 // We use at most one low and one high word shuffle to collect these paired
9085 // inputs into dwords, and finally a dword shuffle to place them.
9086 int PSHUFLMask[4] = {-1, -1, -1, -1};
9087 int PSHUFHMask[4] = {-1, -1, -1, -1};
9088 int PSHUFDMask[4] = {-1, -1, -1, -1};
9090 // First fix the masks for all the inputs that are staying in their
9091 // original halves. This will then dictate the targets of the cross-half
9093 auto fixInPlaceInputs =
9094 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9095 MutableArrayRef<int> SourceHalfMask,
9096 MutableArrayRef<int> HalfMask, int HalfOffset) {
9097 if (InPlaceInputs.empty())
9099 if (InPlaceInputs.size() == 1) {
9100 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9101 InPlaceInputs[0] - HalfOffset;
9102 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9105 if (IncomingInputs.empty()) {
9106 // Just fix all of the in place inputs.
9107 for (int Input : InPlaceInputs) {
9108 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9109 PSHUFDMask[Input / 2] = Input / 2;
9114 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9115 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9116 InPlaceInputs[0] - HalfOffset;
9117 // Put the second input next to the first so that they are packed into
9118 // a dword. We find the adjacent index by toggling the low bit.
9119 int AdjIndex = InPlaceInputs[0] ^ 1;
9120 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9121 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9122 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9124 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9125 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9127 // Now gather the cross-half inputs and place them into a free dword of
9128 // their target half.
9129 // FIXME: This operation could almost certainly be simplified dramatically to
9130 // look more like the 3-1 fixing operation.
9131 auto moveInputsToRightHalf = [&PSHUFDMask](
9132 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9133 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9134 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9136 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9137 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9139 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9141 int LowWord = Word & ~1;
9142 int HighWord = Word | 1;
9143 return isWordClobbered(SourceHalfMask, LowWord) ||
9144 isWordClobbered(SourceHalfMask, HighWord);
9147 if (IncomingInputs.empty())
9150 if (ExistingInputs.empty()) {
9151 // Map any dwords with inputs from them into the right half.
9152 for (int Input : IncomingInputs) {
9153 // If the source half mask maps over the inputs, turn those into
9154 // swaps and use the swapped lane.
9155 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9156 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9157 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9158 Input - SourceOffset;
9159 // We have to swap the uses in our half mask in one sweep.
9160 for (int &M : HalfMask)
9161 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9163 else if (M == Input)
9164 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9166 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9167 Input - SourceOffset &&
9168 "Previous placement doesn't match!");
9170 // Note that this correctly re-maps both when we do a swap and when
9171 // we observe the other side of the swap above. We rely on that to
9172 // avoid swapping the members of the input list directly.
9173 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9176 // Map the input's dword into the correct half.
9177 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9178 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9180 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9182 "Previous placement doesn't match!");
9185 // And just directly shift any other-half mask elements to be same-half
9186 // as we will have mirrored the dword containing the element into the
9187 // same position within that half.
9188 for (int &M : HalfMask)
9189 if (M >= SourceOffset && M < SourceOffset + 4) {
9190 M = M - SourceOffset + DestOffset;
9191 assert(M >= 0 && "This should never wrap below zero!");
9196 // Ensure we have the input in a viable dword of its current half. This
9197 // is particularly tricky because the original position may be clobbered
9198 // by inputs being moved and *staying* in that half.
9199 if (IncomingInputs.size() == 1) {
9200 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9201 int InputFixed = std::find(std::begin(SourceHalfMask),
9202 std::end(SourceHalfMask), -1) -
9203 std::begin(SourceHalfMask) + SourceOffset;
9204 SourceHalfMask[InputFixed - SourceOffset] =
9205 IncomingInputs[0] - SourceOffset;
9206 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9208 IncomingInputs[0] = InputFixed;
9210 } else if (IncomingInputs.size() == 2) {
9211 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9212 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9213 // We have two non-adjacent or clobbered inputs we need to extract from
9214 // the source half. To do this, we need to map them into some adjacent
9215 // dword slot in the source mask.
9216 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9217 IncomingInputs[1] - SourceOffset};
9219 // If there is a free slot in the source half mask adjacent to one of
9220 // the inputs, place the other input in it. We use (Index XOR 1) to
9221 // compute an adjacent index.
9222 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9223 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9224 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9225 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9226 InputsFixed[1] = InputsFixed[0] ^ 1;
9227 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9228 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9229 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9230 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9231 InputsFixed[0] = InputsFixed[1] ^ 1;
9232 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9233 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9234 // The two inputs are in the same DWord but it is clobbered and the
9235 // adjacent DWord isn't used at all. Move both inputs to the free
9237 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9238 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9239 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9240 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9242 // The only way we hit this point is if there is no clobbering
9243 // (because there are no off-half inputs to this half) and there is no
9244 // free slot adjacent to one of the inputs. In this case, we have to
9245 // swap an input with a non-input.
9246 for (int i = 0; i < 4; ++i)
9247 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9248 "We can't handle any clobbers here!");
9249 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9250 "Cannot have adjacent inputs here!");
9252 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9253 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9255 // We also have to update the final source mask in this case because
9256 // it may need to undo the above swap.
9257 for (int &M : FinalSourceHalfMask)
9258 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9259 M = InputsFixed[1] + SourceOffset;
9260 else if (M == InputsFixed[1] + SourceOffset)
9261 M = (InputsFixed[0] ^ 1) + SourceOffset;
9263 InputsFixed[1] = InputsFixed[0] ^ 1;
9266 // Point everything at the fixed inputs.
9267 for (int &M : HalfMask)
9268 if (M == IncomingInputs[0])
9269 M = InputsFixed[0] + SourceOffset;
9270 else if (M == IncomingInputs[1])
9271 M = InputsFixed[1] + SourceOffset;
9273 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9274 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9277 llvm_unreachable("Unhandled input size!");
9280 // Now hoist the DWord down to the right half.
9281 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9282 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9283 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9284 for (int &M : HalfMask)
9285 for (int Input : IncomingInputs)
9287 M = FreeDWord * 2 + Input % 2;
9289 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9290 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9291 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9292 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9294 // Now enact all the shuffles we've computed to move the inputs into their
9296 if (!isNoopShuffleMask(PSHUFLMask))
9297 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9298 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9299 if (!isNoopShuffleMask(PSHUFHMask))
9300 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9301 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9302 if (!isNoopShuffleMask(PSHUFDMask))
9305 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9306 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9308 // At this point, each half should contain all its inputs, and we can then
9309 // just shuffle them into their final position.
9310 assert(std::count_if(LoMask.begin(), LoMask.end(),
9311 [](int M) { return M >= 4; }) == 0 &&
9312 "Failed to lift all the high half inputs to the low mask!");
9313 assert(std::count_if(HiMask.begin(), HiMask.end(),
9314 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9315 "Failed to lift all the low half inputs to the high mask!");
9317 // Do a half shuffle for the low mask.
9318 if (!isNoopShuffleMask(LoMask))
9319 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9320 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9322 // Do a half shuffle with the high mask after shifting its values down.
9323 for (int &M : HiMask)
9326 if (!isNoopShuffleMask(HiMask))
9327 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9328 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9333 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9334 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9335 SDValue V2, ArrayRef<int> Mask,
9336 SelectionDAG &DAG, bool &V1InUse,
9338 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9344 int Size = Mask.size();
9345 int Scale = 16 / Size;
9346 for (int i = 0; i < 16; ++i) {
9347 if (Mask[i / Scale] == -1) {
9348 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9350 const int ZeroMask = 0x80;
9351 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9353 int V2Idx = Mask[i / Scale] < Size
9355 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9356 if (Zeroable[i / Scale])
9357 V1Idx = V2Idx = ZeroMask;
9358 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9359 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9360 V1InUse |= (ZeroMask != V1Idx);
9361 V2InUse |= (ZeroMask != V2Idx);
9366 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9367 DAG.getBitcast(MVT::v16i8, V1),
9368 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9370 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9371 DAG.getBitcast(MVT::v16i8, V2),
9372 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9374 // If we need shuffled inputs from both, blend the two.
9376 if (V1InUse && V2InUse)
9377 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9379 V = V1InUse ? V1 : V2;
9381 // Cast the result back to the correct type.
9382 return DAG.getBitcast(VT, V);
9385 /// \brief Generic lowering of 8-lane i16 shuffles.
9387 /// This handles both single-input shuffles and combined shuffle/blends with
9388 /// two inputs. The single input shuffles are immediately delegated to
9389 /// a dedicated lowering routine.
9391 /// The blends are lowered in one of three fundamental ways. If there are few
9392 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9393 /// of the input is significantly cheaper when lowered as an interleaving of
9394 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9395 /// halves of the inputs separately (making them have relatively few inputs)
9396 /// and then concatenate them.
9397 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9398 const X86Subtarget *Subtarget,
9399 SelectionDAG &DAG) {
9401 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9402 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9403 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9404 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9405 ArrayRef<int> OrigMask = SVOp->getMask();
9406 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9407 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9408 MutableArrayRef<int> Mask(MaskStorage);
9410 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9412 // Whenever we can lower this as a zext, that instruction is strictly faster
9413 // than any alternative.
9414 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9415 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9418 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9420 auto isV2 = [](int M) { return M >= 8; };
9422 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9424 if (NumV2Inputs == 0) {
9425 // Check for being able to broadcast a single element.
9426 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9427 Mask, Subtarget, DAG))
9430 // Try to use shift instructions.
9432 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9435 // Use dedicated unpack instructions for masks that match their pattern.
9437 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9440 // Try to use byte rotation instructions.
9441 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9442 Mask, Subtarget, DAG))
9445 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9449 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9450 "All single-input shuffles should be canonicalized to be V1-input "
9453 // Try to use shift instructions.
9455 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9458 // See if we can use SSE4A Extraction / Insertion.
9459 if (Subtarget->hasSSE4A())
9460 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9463 // There are special ways we can lower some single-element blends.
9464 if (NumV2Inputs == 1)
9465 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9466 Mask, Subtarget, DAG))
9469 // We have different paths for blend lowering, but they all must use the
9470 // *exact* same predicate.
9471 bool IsBlendSupported = Subtarget->hasSSE41();
9472 if (IsBlendSupported)
9473 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9477 if (SDValue Masked =
9478 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9481 // Use dedicated unpack instructions for masks that match their pattern.
9483 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9486 // Try to use byte rotation instructions.
9487 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9488 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9491 if (SDValue BitBlend =
9492 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9495 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9499 // If we can't directly blend but can use PSHUFB, that will be better as it
9500 // can both shuffle and set up the inefficient blend.
9501 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9502 bool V1InUse, V2InUse;
9503 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9507 // We can always bit-blend if we have to so the fallback strategy is to
9508 // decompose into single-input permutes and blends.
9509 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9513 /// \brief Check whether a compaction lowering can be done by dropping even
9514 /// elements and compute how many times even elements must be dropped.
9516 /// This handles shuffles which take every Nth element where N is a power of
9517 /// two. Example shuffle masks:
9519 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9520 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9521 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9522 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9523 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9524 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9526 /// Any of these lanes can of course be undef.
9528 /// This routine only supports N <= 3.
9529 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9532 /// \returns N above, or the number of times even elements must be dropped if
9533 /// there is such a number. Otherwise returns zero.
9534 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9535 // Figure out whether we're looping over two inputs or just one.
9536 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9538 // The modulus for the shuffle vector entries is based on whether this is
9539 // a single input or not.
9540 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9541 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9542 "We should only be called with masks with a power-of-2 size!");
9544 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9546 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9547 // and 2^3 simultaneously. This is because we may have ambiguity with
9548 // partially undef inputs.
9549 bool ViableForN[3] = {true, true, true};
9551 for (int i = 0, e = Mask.size(); i < e; ++i) {
9552 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9557 bool IsAnyViable = false;
9558 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9559 if (ViableForN[j]) {
9562 // The shuffle mask must be equal to (i * 2^N) % M.
9563 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9566 ViableForN[j] = false;
9568 // Early exit if we exhaust the possible powers of two.
9573 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9577 // Return 0 as there is no viable power of two.
9581 /// \brief Generic lowering of v16i8 shuffles.
9583 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9584 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9585 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9586 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9588 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9589 const X86Subtarget *Subtarget,
9590 SelectionDAG &DAG) {
9592 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9593 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9594 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9595 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9596 ArrayRef<int> Mask = SVOp->getMask();
9597 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9599 // Try to use shift instructions.
9601 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9604 // Try to use byte rotation instructions.
9605 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9606 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9609 // Try to use a zext lowering.
9610 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9611 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9614 // See if we can use SSE4A Extraction / Insertion.
9615 if (Subtarget->hasSSE4A())
9616 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9620 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9622 // For single-input shuffles, there are some nicer lowering tricks we can use.
9623 if (NumV2Elements == 0) {
9624 // Check for being able to broadcast a single element.
9625 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9626 Mask, Subtarget, DAG))
9629 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9630 // Notably, this handles splat and partial-splat shuffles more efficiently.
9631 // However, it only makes sense if the pre-duplication shuffle simplifies
9632 // things significantly. Currently, this means we need to be able to
9633 // express the pre-duplication shuffle as an i16 shuffle.
9635 // FIXME: We should check for other patterns which can be widened into an
9636 // i16 shuffle as well.
9637 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9638 for (int i = 0; i < 16; i += 2)
9639 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9644 auto tryToWidenViaDuplication = [&]() -> SDValue {
9645 if (!canWidenViaDuplication(Mask))
9647 SmallVector<int, 4> LoInputs;
9648 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9649 [](int M) { return M >= 0 && M < 8; });
9650 std::sort(LoInputs.begin(), LoInputs.end());
9651 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9653 SmallVector<int, 4> HiInputs;
9654 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9655 [](int M) { return M >= 8; });
9656 std::sort(HiInputs.begin(), HiInputs.end());
9657 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9660 bool TargetLo = LoInputs.size() >= HiInputs.size();
9661 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9662 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9664 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9665 SmallDenseMap<int, int, 8> LaneMap;
9666 for (int I : InPlaceInputs) {
9667 PreDupI16Shuffle[I/2] = I/2;
9670 int j = TargetLo ? 0 : 4, je = j + 4;
9671 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9672 // Check if j is already a shuffle of this input. This happens when
9673 // there are two adjacent bytes after we move the low one.
9674 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9675 // If we haven't yet mapped the input, search for a slot into which
9677 while (j < je && PreDupI16Shuffle[j] != -1)
9681 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9684 // Map this input with the i16 shuffle.
9685 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9688 // Update the lane map based on the mapping we ended up with.
9689 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9691 V1 = DAG.getBitcast(
9693 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9694 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9696 // Unpack the bytes to form the i16s that will be shuffled into place.
9697 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9698 MVT::v16i8, V1, V1);
9700 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9701 for (int i = 0; i < 16; ++i)
9702 if (Mask[i] != -1) {
9703 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9704 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9705 if (PostDupI16Shuffle[i / 2] == -1)
9706 PostDupI16Shuffle[i / 2] = MappedMask;
9708 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9709 "Conflicting entrties in the original shuffle!");
9711 return DAG.getBitcast(
9713 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9714 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9716 if (SDValue V = tryToWidenViaDuplication())
9720 if (SDValue Masked =
9721 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9724 // Use dedicated unpack instructions for masks that match their pattern.
9726 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9729 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9730 // with PSHUFB. It is important to do this before we attempt to generate any
9731 // blends but after all of the single-input lowerings. If the single input
9732 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9733 // want to preserve that and we can DAG combine any longer sequences into
9734 // a PSHUFB in the end. But once we start blending from multiple inputs,
9735 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9736 // and there are *very* few patterns that would actually be faster than the
9737 // PSHUFB approach because of its ability to zero lanes.
9739 // FIXME: The only exceptions to the above are blends which are exact
9740 // interleavings with direct instructions supporting them. We currently don't
9741 // handle those well here.
9742 if (Subtarget->hasSSSE3()) {
9743 bool V1InUse = false;
9744 bool V2InUse = false;
9746 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9747 DAG, V1InUse, V2InUse);
9749 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9750 // do so. This avoids using them to handle blends-with-zero which is
9751 // important as a single pshufb is significantly faster for that.
9752 if (V1InUse && V2InUse) {
9753 if (Subtarget->hasSSE41())
9754 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9755 Mask, Subtarget, DAG))
9758 // We can use an unpack to do the blending rather than an or in some
9759 // cases. Even though the or may be (very minorly) more efficient, we
9760 // preference this lowering because there are common cases where part of
9761 // the complexity of the shuffles goes away when we do the final blend as
9763 // FIXME: It might be worth trying to detect if the unpack-feeding
9764 // shuffles will both be pshufb, in which case we shouldn't bother with
9766 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9767 DL, MVT::v16i8, V1, V2, Mask, DAG))
9774 // There are special ways we can lower some single-element blends.
9775 if (NumV2Elements == 1)
9776 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9777 Mask, Subtarget, DAG))
9780 if (SDValue BitBlend =
9781 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9784 // Check whether a compaction lowering can be done. This handles shuffles
9785 // which take every Nth element for some even N. See the helper function for
9788 // We special case these as they can be particularly efficiently handled with
9789 // the PACKUSB instruction on x86 and they show up in common patterns of
9790 // rearranging bytes to truncate wide elements.
9791 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9792 // NumEvenDrops is the power of two stride of the elements. Another way of
9793 // thinking about it is that we need to drop the even elements this many
9794 // times to get the original input.
9795 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9797 // First we need to zero all the dropped bytes.
9798 assert(NumEvenDrops <= 3 &&
9799 "No support for dropping even elements more than 3 times.");
9800 // We use the mask type to pick which bytes are preserved based on how many
9801 // elements are dropped.
9802 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9803 SDValue ByteClearMask = DAG.getBitcast(
9804 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9805 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9807 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9809 // Now pack things back together.
9810 V1 = DAG.getBitcast(MVT::v8i16, V1);
9811 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9812 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9813 for (int i = 1; i < NumEvenDrops; ++i) {
9814 Result = DAG.getBitcast(MVT::v8i16, Result);
9815 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9821 // Handle multi-input cases by blending single-input shuffles.
9822 if (NumV2Elements > 0)
9823 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9826 // The fallback path for single-input shuffles widens this into two v8i16
9827 // vectors with unpacks, shuffles those, and then pulls them back together
9831 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9832 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9833 for (int i = 0; i < 16; ++i)
9835 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9837 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9839 SDValue VLoHalf, VHiHalf;
9840 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9841 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9843 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9844 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9845 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9846 [](int M) { return M >= 0 && M % 2 == 1; })) {
9847 // Use a mask to drop the high bytes.
9848 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9849 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9850 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9852 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9853 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9855 // Squash the masks to point directly into VLoHalf.
9856 for (int &M : LoBlendMask)
9859 for (int &M : HiBlendMask)
9863 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9864 // VHiHalf so that we can blend them as i16s.
9865 VLoHalf = DAG.getBitcast(
9866 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9867 VHiHalf = DAG.getBitcast(
9868 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9871 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9872 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9874 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9877 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9879 /// This routine breaks down the specific type of 128-bit shuffle and
9880 /// dispatches to the lowering routines accordingly.
9881 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9882 MVT VT, const X86Subtarget *Subtarget,
9883 SelectionDAG &DAG) {
9884 switch (VT.SimpleTy) {
9886 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9888 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9890 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9892 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9894 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9896 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9899 llvm_unreachable("Unimplemented!");
9903 /// \brief Helper function to test whether a shuffle mask could be
9904 /// simplified by widening the elements being shuffled.
9906 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9907 /// leaves it in an unspecified state.
9909 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9910 /// shuffle masks. The latter have the special property of a '-2' representing
9911 /// a zero-ed lane of a vector.
9912 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9913 SmallVectorImpl<int> &WidenedMask) {
9914 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9915 // If both elements are undef, its trivial.
9916 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9917 WidenedMask.push_back(SM_SentinelUndef);
9921 // Check for an undef mask and a mask value properly aligned to fit with
9922 // a pair of values. If we find such a case, use the non-undef mask's value.
9923 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9924 WidenedMask.push_back(Mask[i + 1] / 2);
9927 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9928 WidenedMask.push_back(Mask[i] / 2);
9932 // When zeroing, we need to spread the zeroing across both lanes to widen.
9933 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9934 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9935 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9936 WidenedMask.push_back(SM_SentinelZero);
9942 // Finally check if the two mask values are adjacent and aligned with
9944 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9945 WidenedMask.push_back(Mask[i] / 2);
9949 // Otherwise we can't safely widen the elements used in this shuffle.
9952 assert(WidenedMask.size() == Mask.size() / 2 &&
9953 "Incorrect size of mask after widening the elements!");
9958 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9960 /// This routine just extracts two subvectors, shuffles them independently, and
9961 /// then concatenates them back together. This should work effectively with all
9962 /// AVX vector shuffle types.
9963 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9964 SDValue V2, ArrayRef<int> Mask,
9965 SelectionDAG &DAG) {
9966 assert(VT.getSizeInBits() >= 256 &&
9967 "Only for 256-bit or wider vector shuffles!");
9968 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9969 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9971 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9972 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9974 int NumElements = VT.getVectorNumElements();
9975 int SplitNumElements = NumElements / 2;
9976 MVT ScalarVT = VT.getVectorElementType();
9977 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9979 // Rather than splitting build-vectors, just build two narrower build
9980 // vectors. This helps shuffling with splats and zeros.
9981 auto SplitVector = [&](SDValue V) {
9982 while (V.getOpcode() == ISD::BITCAST)
9983 V = V->getOperand(0);
9985 MVT OrigVT = V.getSimpleValueType();
9986 int OrigNumElements = OrigVT.getVectorNumElements();
9987 int OrigSplitNumElements = OrigNumElements / 2;
9988 MVT OrigScalarVT = OrigVT.getVectorElementType();
9989 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9993 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9995 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9996 DAG.getIntPtrConstant(0, DL));
9997 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9998 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
10001 SmallVector<SDValue, 16> LoOps, HiOps;
10002 for (int i = 0; i < OrigSplitNumElements; ++i) {
10003 LoOps.push_back(BV->getOperand(i));
10004 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
10006 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
10007 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
10009 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
10010 DAG.getBitcast(SplitVT, HiV));
10013 SDValue LoV1, HiV1, LoV2, HiV2;
10014 std::tie(LoV1, HiV1) = SplitVector(V1);
10015 std::tie(LoV2, HiV2) = SplitVector(V2);
10017 // Now create two 4-way blends of these half-width vectors.
10018 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
10019 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
10020 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
10021 for (int i = 0; i < SplitNumElements; ++i) {
10022 int M = HalfMask[i];
10023 if (M >= NumElements) {
10024 if (M >= NumElements + SplitNumElements)
10028 V2BlendMask.push_back(M - NumElements);
10029 V1BlendMask.push_back(-1);
10030 BlendMask.push_back(SplitNumElements + i);
10031 } else if (M >= 0) {
10032 if (M >= SplitNumElements)
10036 V2BlendMask.push_back(-1);
10037 V1BlendMask.push_back(M);
10038 BlendMask.push_back(i);
10040 V2BlendMask.push_back(-1);
10041 V1BlendMask.push_back(-1);
10042 BlendMask.push_back(-1);
10046 // Because the lowering happens after all combining takes place, we need to
10047 // manually combine these blend masks as much as possible so that we create
10048 // a minimal number of high-level vector shuffle nodes.
10050 // First try just blending the halves of V1 or V2.
10051 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10052 return DAG.getUNDEF(SplitVT);
10053 if (!UseLoV2 && !UseHiV2)
10054 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10055 if (!UseLoV1 && !UseHiV1)
10056 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10058 SDValue V1Blend, V2Blend;
10059 if (UseLoV1 && UseHiV1) {
10061 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10063 // We only use half of V1 so map the usage down into the final blend mask.
10064 V1Blend = UseLoV1 ? LoV1 : HiV1;
10065 for (int i = 0; i < SplitNumElements; ++i)
10066 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10067 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10069 if (UseLoV2 && UseHiV2) {
10071 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10073 // We only use half of V2 so map the usage down into the final blend mask.
10074 V2Blend = UseLoV2 ? LoV2 : HiV2;
10075 for (int i = 0; i < SplitNumElements; ++i)
10076 if (BlendMask[i] >= SplitNumElements)
10077 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10079 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10081 SDValue Lo = HalfBlend(LoMask);
10082 SDValue Hi = HalfBlend(HiMask);
10083 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10086 /// \brief Either split a vector in halves or decompose the shuffles and the
10089 /// This is provided as a good fallback for many lowerings of non-single-input
10090 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10091 /// between splitting the shuffle into 128-bit components and stitching those
10092 /// back together vs. extracting the single-input shuffles and blending those
10094 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10095 SDValue V2, ArrayRef<int> Mask,
10096 SelectionDAG &DAG) {
10097 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10098 "lower single-input shuffles as it "
10099 "could then recurse on itself.");
10100 int Size = Mask.size();
10102 // If this can be modeled as a broadcast of two elements followed by a blend,
10103 // prefer that lowering. This is especially important because broadcasts can
10104 // often fold with memory operands.
10105 auto DoBothBroadcast = [&] {
10106 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10109 if (V2BroadcastIdx == -1)
10110 V2BroadcastIdx = M - Size;
10111 else if (M - Size != V2BroadcastIdx)
10113 } else if (M >= 0) {
10114 if (V1BroadcastIdx == -1)
10115 V1BroadcastIdx = M;
10116 else if (M != V1BroadcastIdx)
10121 if (DoBothBroadcast())
10122 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10125 // If the inputs all stem from a single 128-bit lane of each input, then we
10126 // split them rather than blending because the split will decompose to
10127 // unusually few instructions.
10128 int LaneCount = VT.getSizeInBits() / 128;
10129 int LaneSize = Size / LaneCount;
10130 SmallBitVector LaneInputs[2];
10131 LaneInputs[0].resize(LaneCount, false);
10132 LaneInputs[1].resize(LaneCount, false);
10133 for (int i = 0; i < Size; ++i)
10135 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10136 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10137 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10139 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10140 // that the decomposed single-input shuffles don't end up here.
10141 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10144 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10145 /// a permutation and blend of those lanes.
10147 /// This essentially blends the out-of-lane inputs to each lane into the lane
10148 /// from a permuted copy of the vector. This lowering strategy results in four
10149 /// instructions in the worst case for a single-input cross lane shuffle which
10150 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10151 /// of. Special cases for each particular shuffle pattern should be handled
10152 /// prior to trying this lowering.
10153 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10154 SDValue V1, SDValue V2,
10155 ArrayRef<int> Mask,
10156 SelectionDAG &DAG) {
10157 // FIXME: This should probably be generalized for 512-bit vectors as well.
10158 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
10159 int LaneSize = Mask.size() / 2;
10161 // If there are only inputs from one 128-bit lane, splitting will in fact be
10162 // less expensive. The flags track whether the given lane contains an element
10163 // that crosses to another lane.
10164 bool LaneCrossing[2] = {false, false};
10165 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10166 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10167 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10168 if (!LaneCrossing[0] || !LaneCrossing[1])
10169 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10171 if (isSingleInputShuffleMask(Mask)) {
10172 SmallVector<int, 32> FlippedBlendMask;
10173 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10174 FlippedBlendMask.push_back(
10175 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10177 : Mask[i] % LaneSize +
10178 (i / LaneSize) * LaneSize + Size));
10180 // Flip the vector, and blend the results which should now be in-lane. The
10181 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10182 // 5 for the high source. The value 3 selects the high half of source 2 and
10183 // the value 2 selects the low half of source 2. We only use source 2 to
10184 // allow folding it into a memory operand.
10185 unsigned PERMMask = 3 | 2 << 4;
10186 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10187 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10188 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10191 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10192 // will be handled by the above logic and a blend of the results, much like
10193 // other patterns in AVX.
10194 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10197 /// \brief Handle lowering 2-lane 128-bit shuffles.
10198 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10199 SDValue V2, ArrayRef<int> Mask,
10200 const X86Subtarget *Subtarget,
10201 SelectionDAG &DAG) {
10202 // TODO: If minimizing size and one of the inputs is a zero vector and the
10203 // the zero vector has only one use, we could use a VPERM2X128 to save the
10204 // instruction bytes needed to explicitly generate the zero vector.
10206 // Blends are faster and handle all the non-lane-crossing cases.
10207 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10211 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10212 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10214 // If either input operand is a zero vector, use VPERM2X128 because its mask
10215 // allows us to replace the zero input with an implicit zero.
10216 if (!IsV1Zero && !IsV2Zero) {
10217 // Check for patterns which can be matched with a single insert of a 128-bit
10219 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10220 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10221 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10222 VT.getVectorNumElements() / 2);
10223 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10224 DAG.getIntPtrConstant(0, DL));
10225 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10226 OnlyUsesV1 ? V1 : V2,
10227 DAG.getIntPtrConstant(0, DL));
10228 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10232 // Otherwise form a 128-bit permutation. After accounting for undefs,
10233 // convert the 64-bit shuffle mask selection values into 128-bit
10234 // selection bits by dividing the indexes by 2 and shifting into positions
10235 // defined by a vperm2*128 instruction's immediate control byte.
10237 // The immediate permute control byte looks like this:
10238 // [1:0] - select 128 bits from sources for low half of destination
10240 // [3] - zero low half of destination
10241 // [5:4] - select 128 bits from sources for high half of destination
10243 // [7] - zero high half of destination
10245 int MaskLO = Mask[0];
10246 if (MaskLO == SM_SentinelUndef)
10247 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10249 int MaskHI = Mask[2];
10250 if (MaskHI == SM_SentinelUndef)
10251 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10253 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10255 // If either input is a zero vector, replace it with an undef input.
10256 // Shuffle mask values < 4 are selecting elements of V1.
10257 // Shuffle mask values >= 4 are selecting elements of V2.
10258 // Adjust each half of the permute mask by clearing the half that was
10259 // selecting the zero vector and setting the zero mask bit.
10261 V1 = DAG.getUNDEF(VT);
10263 PermMask = (PermMask & 0xf0) | 0x08;
10265 PermMask = (PermMask & 0x0f) | 0x80;
10268 V2 = DAG.getUNDEF(VT);
10270 PermMask = (PermMask & 0xf0) | 0x08;
10272 PermMask = (PermMask & 0x0f) | 0x80;
10275 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10276 DAG.getConstant(PermMask, DL, MVT::i8));
10279 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10280 /// shuffling each lane.
10282 /// This will only succeed when the result of fixing the 128-bit lanes results
10283 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10284 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10285 /// the lane crosses early and then use simpler shuffles within each lane.
10287 /// FIXME: It might be worthwhile at some point to support this without
10288 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10289 /// in x86 only floating point has interesting non-repeating shuffles, and even
10290 /// those are still *marginally* more expensive.
10291 static SDValue lowerVectorShuffleByMerging128BitLanes(
10292 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10293 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10294 assert(!isSingleInputShuffleMask(Mask) &&
10295 "This is only useful with multiple inputs.");
10297 int Size = Mask.size();
10298 int LaneSize = 128 / VT.getScalarSizeInBits();
10299 int NumLanes = Size / LaneSize;
10300 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10302 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10303 // check whether the in-128-bit lane shuffles share a repeating pattern.
10304 SmallVector<int, 4> Lanes;
10305 Lanes.resize(NumLanes, -1);
10306 SmallVector<int, 4> InLaneMask;
10307 InLaneMask.resize(LaneSize, -1);
10308 for (int i = 0; i < Size; ++i) {
10312 int j = i / LaneSize;
10314 if (Lanes[j] < 0) {
10315 // First entry we've seen for this lane.
10316 Lanes[j] = Mask[i] / LaneSize;
10317 } else if (Lanes[j] != Mask[i] / LaneSize) {
10318 // This doesn't match the lane selected previously!
10322 // Check that within each lane we have a consistent shuffle mask.
10323 int k = i % LaneSize;
10324 if (InLaneMask[k] < 0) {
10325 InLaneMask[k] = Mask[i] % LaneSize;
10326 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10327 // This doesn't fit a repeating in-lane mask.
10332 // First shuffle the lanes into place.
10333 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10334 VT.getSizeInBits() / 64);
10335 SmallVector<int, 8> LaneMask;
10336 LaneMask.resize(NumLanes * 2, -1);
10337 for (int i = 0; i < NumLanes; ++i)
10338 if (Lanes[i] >= 0) {
10339 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10340 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10343 V1 = DAG.getBitcast(LaneVT, V1);
10344 V2 = DAG.getBitcast(LaneVT, V2);
10345 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10347 // Cast it back to the type we actually want.
10348 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10350 // Now do a simple shuffle that isn't lane crossing.
10351 SmallVector<int, 8> NewMask;
10352 NewMask.resize(Size, -1);
10353 for (int i = 0; i < Size; ++i)
10355 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10356 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10357 "Must not introduce lane crosses at this point!");
10359 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10362 /// Lower shuffles where an entire half of a 256-bit vector is UNDEF.
10363 /// This allows for fast cases such as subvector extraction/insertion
10364 /// or shuffling smaller vector types which can lower more efficiently.
10365 static SDValue lowerVectorShuffleWithUndefHalf(SDLoc DL, MVT VT, SDValue V1,
10366 SDValue V2, ArrayRef<int> Mask,
10367 const X86Subtarget *Subtarget,
10368 SelectionDAG &DAG) {
10369 assert(VT.getSizeInBits() == 256 && "Expected 256-bit vector");
10371 unsigned NumElts = VT.getVectorNumElements();
10372 unsigned HalfNumElts = NumElts / 2;
10373 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(), HalfNumElts);
10375 bool UndefLower = isUndefInRange(Mask, 0, HalfNumElts);
10376 bool UndefUpper = isUndefInRange(Mask, HalfNumElts, HalfNumElts);
10377 if (!UndefLower && !UndefUpper)
10380 // Upper half is undef and lower half is whole upper subvector.
10381 // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
10383 isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) {
10384 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10385 DAG.getIntPtrConstant(HalfNumElts, DL));
10386 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10387 DAG.getIntPtrConstant(0, DL));
10390 // Lower half is undef and upper half is whole lower subvector.
10391 // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
10393 isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) {
10394 SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
10395 DAG.getIntPtrConstant(0, DL));
10396 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
10397 DAG.getIntPtrConstant(HalfNumElts, DL));
10400 // AVX2 supports efficient immediate 64-bit element cross-lane shuffles.
10401 if (UndefLower && Subtarget->hasAVX2() &&
10402 (VT == MVT::v4f64 || VT == MVT::v4i64))
10405 // If the shuffle only uses the lower halves of the input operands,
10406 // then extract them and perform the 'half' shuffle at half width.
10407 // e.g. vector_shuffle <X, X, X, X, u, u, u, u> or <X, X, u, u>
10408 int HalfIdx1 = -1, HalfIdx2 = -1;
10409 SmallVector<int, 8> HalfMask;
10410 unsigned Offset = UndefLower ? HalfNumElts : 0;
10411 for (unsigned i = 0; i != HalfNumElts; ++i) {
10412 int M = Mask[i + Offset];
10414 HalfMask.push_back(M);
10418 // Determine which of the 4 half vectors this element is from.
10419 // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2.
10420 int HalfIdx = M / HalfNumElts;
10422 // Only shuffle using the lower halves of the inputs.
10423 // TODO: Investigate usefulness of shuffling with upper halves.
10424 if (HalfIdx != 0 && HalfIdx != 2)
10427 // Determine the element index into its half vector source.
10428 int HalfElt = M % HalfNumElts;
10430 // We can shuffle with up to 2 half vectors, set the new 'half'
10431 // shuffle mask accordingly.
10432 if (-1 == HalfIdx1 || HalfIdx1 == HalfIdx) {
10433 HalfMask.push_back(HalfElt);
10434 HalfIdx1 = HalfIdx;
10437 if (-1 == HalfIdx2 || HalfIdx2 == HalfIdx) {
10438 HalfMask.push_back(HalfElt + HalfNumElts);
10439 HalfIdx2 = HalfIdx;
10443 // Too many half vectors referenced.
10446 assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length");
10448 auto GetHalfVector = [&](int HalfIdx) {
10450 return DAG.getUNDEF(HalfVT);
10451 SDValue V = (HalfIdx < 2 ? V1 : V2);
10452 HalfIdx = (HalfIdx % 2) * HalfNumElts;
10453 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V,
10454 DAG.getIntPtrConstant(HalfIdx, DL));
10457 SDValue Half1 = GetHalfVector(HalfIdx1);
10458 SDValue Half2 = GetHalfVector(HalfIdx2);
10459 SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask);
10460 return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V,
10461 DAG.getIntPtrConstant(Offset, DL));
10464 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10467 /// This returns true if the elements from a particular input are already in the
10468 /// slot required by the given mask and require no permutation.
10469 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10470 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10471 int Size = Mask.size();
10472 for (int i = 0; i < Size; ++i)
10473 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10479 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10480 ArrayRef<int> Mask, SDValue V1,
10481 SDValue V2, SelectionDAG &DAG) {
10483 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10484 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10485 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10486 int NumElts = VT.getVectorNumElements();
10487 bool ShufpdMask = true;
10488 bool CommutableMask = true;
10489 unsigned Immediate = 0;
10490 for (int i = 0; i < NumElts; ++i) {
10493 int Val = (i & 6) + NumElts * (i & 1);
10494 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10495 if (Mask[i] < Val || Mask[i] > Val + 1)
10496 ShufpdMask = false;
10497 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10498 CommutableMask = false;
10499 Immediate |= (Mask[i] % 2) << i;
10502 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10503 DAG.getConstant(Immediate, DL, MVT::i8));
10504 if (CommutableMask)
10505 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10506 DAG.getConstant(Immediate, DL, MVT::i8));
10510 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10512 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10513 /// isn't available.
10514 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10515 const X86Subtarget *Subtarget,
10516 SelectionDAG &DAG) {
10518 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10519 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10520 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10521 ArrayRef<int> Mask = SVOp->getMask();
10522 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10524 SmallVector<int, 4> WidenedMask;
10525 if (canWidenShuffleElements(Mask, WidenedMask))
10526 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10529 if (isSingleInputShuffleMask(Mask)) {
10530 // Check for being able to broadcast a single element.
10531 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10532 Mask, Subtarget, DAG))
10535 // Use low duplicate instructions for masks that match their pattern.
10536 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10537 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10539 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10540 // Non-half-crossing single input shuffles can be lowerid with an
10541 // interleaved permutation.
10542 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10543 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10544 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10545 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10548 // With AVX2 we have direct support for this permutation.
10549 if (Subtarget->hasAVX2())
10550 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10551 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10553 // Otherwise, fall back.
10554 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10558 // Use dedicated unpack instructions for masks that match their pattern.
10560 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10563 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10567 // Check if the blend happens to exactly fit that of SHUFPD.
10569 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10572 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10573 // shuffle. However, if we have AVX2 and either inputs are already in place,
10574 // we will be able to shuffle even across lanes the other input in a single
10575 // instruction so skip this pattern.
10576 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10577 isShuffleMaskInputInPlace(1, Mask))))
10578 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10579 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10582 // If we have AVX2 then we always want to lower with a blend because an v4 we
10583 // can fully permute the elements.
10584 if (Subtarget->hasAVX2())
10585 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10588 // Otherwise fall back on generic lowering.
10589 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10592 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10594 /// This routine is only called when we have AVX2 and thus a reasonable
10595 /// instruction set for v4i64 shuffling..
10596 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10597 const X86Subtarget *Subtarget,
10598 SelectionDAG &DAG) {
10600 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10601 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10602 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10603 ArrayRef<int> Mask = SVOp->getMask();
10604 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10605 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10607 SmallVector<int, 4> WidenedMask;
10608 if (canWidenShuffleElements(Mask, WidenedMask))
10609 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10612 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10616 // Check for being able to broadcast a single element.
10617 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10618 Mask, Subtarget, DAG))
10621 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10622 // use lower latency instructions that will operate on both 128-bit lanes.
10623 SmallVector<int, 2> RepeatedMask;
10624 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10625 if (isSingleInputShuffleMask(Mask)) {
10626 int PSHUFDMask[] = {-1, -1, -1, -1};
10627 for (int i = 0; i < 2; ++i)
10628 if (RepeatedMask[i] >= 0) {
10629 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10630 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10632 return DAG.getBitcast(
10634 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10635 DAG.getBitcast(MVT::v8i32, V1),
10636 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10640 // AVX2 provides a direct instruction for permuting a single input across
10642 if (isSingleInputShuffleMask(Mask))
10643 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10644 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10646 // Try to use shift instructions.
10647 if (SDValue Shift =
10648 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10651 // Use dedicated unpack instructions for masks that match their pattern.
10653 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10656 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10657 // shuffle. However, if we have AVX2 and either inputs are already in place,
10658 // we will be able to shuffle even across lanes the other input in a single
10659 // instruction so skip this pattern.
10660 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10661 isShuffleMaskInputInPlace(1, Mask))))
10662 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10663 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10666 // Otherwise fall back on generic blend lowering.
10667 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10671 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10673 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10674 /// isn't available.
10675 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10676 const X86Subtarget *Subtarget,
10677 SelectionDAG &DAG) {
10679 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10680 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10681 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10682 ArrayRef<int> Mask = SVOp->getMask();
10683 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10685 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10689 // Check for being able to broadcast a single element.
10690 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10691 Mask, Subtarget, DAG))
10694 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10695 // options to efficiently lower the shuffle.
10696 SmallVector<int, 4> RepeatedMask;
10697 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10698 assert(RepeatedMask.size() == 4 &&
10699 "Repeated masks must be half the mask width!");
10701 // Use even/odd duplicate instructions for masks that match their pattern.
10702 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10703 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10704 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10705 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10707 if (isSingleInputShuffleMask(Mask))
10708 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10709 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10711 // Use dedicated unpack instructions for masks that match their pattern.
10713 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10716 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10717 // have already handled any direct blends. We also need to squash the
10718 // repeated mask into a simulated v4f32 mask.
10719 for (int i = 0; i < 4; ++i)
10720 if (RepeatedMask[i] >= 8)
10721 RepeatedMask[i] -= 4;
10722 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10725 // If we have a single input shuffle with different shuffle patterns in the
10726 // two 128-bit lanes use the variable mask to VPERMILPS.
10727 if (isSingleInputShuffleMask(Mask)) {
10728 SDValue VPermMask[8];
10729 for (int i = 0; i < 8; ++i)
10730 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10731 : DAG.getConstant(Mask[i], DL, MVT::i32);
10732 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10733 return DAG.getNode(
10734 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10735 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10737 if (Subtarget->hasAVX2())
10738 return DAG.getNode(
10739 X86ISD::VPERMV, DL, MVT::v8f32,
10740 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10742 // Otherwise, fall back.
10743 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10747 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10749 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10750 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10753 // If we have AVX2 then we always want to lower with a blend because at v8 we
10754 // can fully permute the elements.
10755 if (Subtarget->hasAVX2())
10756 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10759 // Otherwise fall back on generic lowering.
10760 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10763 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10765 /// This routine is only called when we have AVX2 and thus a reasonable
10766 /// instruction set for v8i32 shuffling..
10767 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10768 const X86Subtarget *Subtarget,
10769 SelectionDAG &DAG) {
10771 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10772 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10773 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10774 ArrayRef<int> Mask = SVOp->getMask();
10775 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10776 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10778 // Whenever we can lower this as a zext, that instruction is strictly faster
10779 // than any alternative. It also allows us to fold memory operands into the
10780 // shuffle in many cases.
10781 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10782 Mask, Subtarget, DAG))
10785 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10789 // Check for being able to broadcast a single element.
10790 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10791 Mask, Subtarget, DAG))
10794 // If the shuffle mask is repeated in each 128-bit lane we can use more
10795 // efficient instructions that mirror the shuffles across the two 128-bit
10797 SmallVector<int, 4> RepeatedMask;
10798 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10799 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10800 if (isSingleInputShuffleMask(Mask))
10801 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10802 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10804 // Use dedicated unpack instructions for masks that match their pattern.
10806 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10810 // Try to use shift instructions.
10811 if (SDValue Shift =
10812 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10815 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10816 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10819 // If the shuffle patterns aren't repeated but it is a single input, directly
10820 // generate a cross-lane VPERMD instruction.
10821 if (isSingleInputShuffleMask(Mask)) {
10822 SDValue VPermMask[8];
10823 for (int i = 0; i < 8; ++i)
10824 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10825 : DAG.getConstant(Mask[i], DL, MVT::i32);
10826 return DAG.getNode(
10827 X86ISD::VPERMV, DL, MVT::v8i32,
10828 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10831 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10833 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10834 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10837 // Otherwise fall back on generic blend lowering.
10838 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10842 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10844 /// This routine is only called when we have AVX2 and thus a reasonable
10845 /// instruction set for v16i16 shuffling..
10846 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10847 const X86Subtarget *Subtarget,
10848 SelectionDAG &DAG) {
10850 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10851 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10852 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10853 ArrayRef<int> Mask = SVOp->getMask();
10854 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10855 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10857 // Whenever we can lower this as a zext, that instruction is strictly faster
10858 // than any alternative. It also allows us to fold memory operands into the
10859 // shuffle in many cases.
10860 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10861 Mask, Subtarget, DAG))
10864 // Check for being able to broadcast a single element.
10865 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10866 Mask, Subtarget, DAG))
10869 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10873 // Use dedicated unpack instructions for masks that match their pattern.
10875 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10878 // Try to use shift instructions.
10879 if (SDValue Shift =
10880 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10883 // Try to use byte rotation instructions.
10884 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10885 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10888 if (isSingleInputShuffleMask(Mask)) {
10889 // There are no generalized cross-lane shuffle operations available on i16
10891 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10892 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10895 SmallVector<int, 8> RepeatedMask;
10896 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10897 // As this is a single-input shuffle, the repeated mask should be
10898 // a strictly valid v8i16 mask that we can pass through to the v8i16
10899 // lowering to handle even the v16 case.
10900 return lowerV8I16GeneralSingleInputVectorShuffle(
10901 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10904 SDValue PSHUFBMask[32];
10905 for (int i = 0; i < 16; ++i) {
10906 if (Mask[i] == -1) {
10907 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10911 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10912 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10913 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10914 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10916 return DAG.getBitcast(MVT::v16i16,
10917 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10918 DAG.getBitcast(MVT::v32i8, V1),
10919 DAG.getNode(ISD::BUILD_VECTOR, DL,
10920 MVT::v32i8, PSHUFBMask)));
10923 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10925 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10926 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10929 // Otherwise fall back on generic lowering.
10930 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10933 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10935 /// This routine is only called when we have AVX2 and thus a reasonable
10936 /// instruction set for v32i8 shuffling..
10937 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10938 const X86Subtarget *Subtarget,
10939 SelectionDAG &DAG) {
10941 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10942 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10943 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10944 ArrayRef<int> Mask = SVOp->getMask();
10945 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10946 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10948 // Whenever we can lower this as a zext, that instruction is strictly faster
10949 // than any alternative. It also allows us to fold memory operands into the
10950 // shuffle in many cases.
10951 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10952 Mask, Subtarget, DAG))
10955 // Check for being able to broadcast a single element.
10956 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10957 Mask, Subtarget, DAG))
10960 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10964 // Use dedicated unpack instructions for masks that match their pattern.
10966 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10969 // Try to use shift instructions.
10970 if (SDValue Shift =
10971 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10974 // Try to use byte rotation instructions.
10975 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10976 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10979 if (isSingleInputShuffleMask(Mask)) {
10980 // There are no generalized cross-lane shuffle operations available on i8
10982 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10983 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10986 SDValue PSHUFBMask[32];
10987 for (int i = 0; i < 32; ++i)
10990 ? DAG.getUNDEF(MVT::i8)
10991 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10994 return DAG.getNode(
10995 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10996 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10999 // Try to simplify this by merging 128-bit lanes to enable a lane-based
11001 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
11002 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
11005 // Otherwise fall back on generic lowering.
11006 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
11009 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
11011 /// This routine either breaks down the specific type of a 256-bit x86 vector
11012 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
11013 /// together based on the available instructions.
11014 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11015 MVT VT, const X86Subtarget *Subtarget,
11016 SelectionDAG &DAG) {
11018 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11019 ArrayRef<int> Mask = SVOp->getMask();
11021 // If we have a single input to the zero element, insert that into V1 if we
11022 // can do so cheaply.
11023 int NumElts = VT.getVectorNumElements();
11024 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
11025 return M >= NumElts;
11028 if (NumV2Elements == 1 && Mask[0] >= NumElts)
11029 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
11030 DL, VT, V1, V2, Mask, Subtarget, DAG))
11033 // Handle special cases where the lower or upper half is UNDEF.
11035 lowerVectorShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG))
11038 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
11039 // can check for those subtargets here and avoid much of the subtarget
11040 // querying in the per-vector-type lowering routines. With AVX1 we have
11041 // essentially *zero* ability to manipulate a 256-bit vector with integer
11042 // types. Since we'll use floating point types there eventually, just
11043 // immediately cast everything to a float and operate entirely in that domain.
11044 if (VT.isInteger() && !Subtarget->hasAVX2()) {
11045 int ElementBits = VT.getScalarSizeInBits();
11046 if (ElementBits < 32)
11047 // No floating point type available, decompose into 128-bit vectors.
11048 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11050 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
11051 VT.getVectorNumElements());
11052 V1 = DAG.getBitcast(FpVT, V1);
11053 V2 = DAG.getBitcast(FpVT, V2);
11054 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
11057 switch (VT.SimpleTy) {
11059 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11061 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11063 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11065 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11067 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11069 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11072 llvm_unreachable("Not a valid 256-bit x86 vector type!");
11076 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
11077 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
11078 ArrayRef<int> Mask,
11079 SDValue V1, SDValue V2,
11080 SelectionDAG &DAG) {
11081 assert(VT.getScalarSizeInBits() == 64 &&
11082 "Unexpected element type size for 128bit shuffle.");
11084 // To handle 256 bit vector requires VLX and most probably
11085 // function lowerV2X128VectorShuffle() is better solution.
11086 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
11088 SmallVector<int, 4> WidenedMask;
11089 if (!canWidenShuffleElements(Mask, WidenedMask))
11092 // Form a 128-bit permutation.
11093 // Convert the 64-bit shuffle mask selection values into 128-bit selection
11094 // bits defined by a vshuf64x2 instruction's immediate control byte.
11095 unsigned PermMask = 0, Imm = 0;
11096 unsigned ControlBitsNum = WidenedMask.size() / 2;
11098 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
11099 if (WidenedMask[i] == SM_SentinelZero)
11102 // Use first element in place of undef mask.
11103 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
11104 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
11107 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
11108 DAG.getConstant(PermMask, DL, MVT::i8));
11111 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
11112 ArrayRef<int> Mask, SDValue V1,
11113 SDValue V2, SelectionDAG &DAG) {
11115 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
11117 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
11118 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
11120 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
11121 if (isSingleInputShuffleMask(Mask))
11122 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
11124 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
11127 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
11128 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11129 const X86Subtarget *Subtarget,
11130 SelectionDAG &DAG) {
11132 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11133 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
11134 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11135 ArrayRef<int> Mask = SVOp->getMask();
11136 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11138 if (SDValue Shuf128 =
11139 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
11142 if (SDValue Unpck =
11143 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
11146 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
11149 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
11150 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11151 const X86Subtarget *Subtarget,
11152 SelectionDAG &DAG) {
11154 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11155 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11156 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11157 ArrayRef<int> Mask = SVOp->getMask();
11158 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11160 if (SDValue Unpck =
11161 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
11164 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
11167 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11168 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11169 const X86Subtarget *Subtarget,
11170 SelectionDAG &DAG) {
11172 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11173 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11174 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11175 ArrayRef<int> Mask = SVOp->getMask();
11176 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11178 if (SDValue Shuf128 =
11179 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
11182 if (SDValue Unpck =
11183 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
11186 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
11189 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11190 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11191 const X86Subtarget *Subtarget,
11192 SelectionDAG &DAG) {
11194 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11195 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11196 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11197 ArrayRef<int> Mask = SVOp->getMask();
11198 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11200 if (SDValue Unpck =
11201 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
11204 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
11207 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11208 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11209 const X86Subtarget *Subtarget,
11210 SelectionDAG &DAG) {
11212 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11213 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11214 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11215 ArrayRef<int> Mask = SVOp->getMask();
11216 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11217 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11219 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
11222 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11223 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11224 const X86Subtarget *Subtarget,
11225 SelectionDAG &DAG) {
11227 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11228 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11229 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11230 ArrayRef<int> Mask = SVOp->getMask();
11231 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11232 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11234 // FIXME: Implement direct support for this type!
11235 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11238 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11240 /// This routine either breaks down the specific type of a 512-bit x86 vector
11241 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11242 /// together based on the available instructions.
11243 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11244 MVT VT, const X86Subtarget *Subtarget,
11245 SelectionDAG &DAG) {
11247 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11248 ArrayRef<int> Mask = SVOp->getMask();
11249 assert(Subtarget->hasAVX512() &&
11250 "Cannot lower 512-bit vectors w/ basic ISA!");
11252 // Check for being able to broadcast a single element.
11253 if (SDValue Broadcast =
11254 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
11257 // Dispatch to each element type for lowering. If we don't have supprot for
11258 // specific element type shuffles at 512 bits, immediately split them and
11259 // lower them. Each lowering routine of a given type is allowed to assume that
11260 // the requisite ISA extensions for that element type are available.
11261 switch (VT.SimpleTy) {
11263 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11265 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11267 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11269 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11271 if (Subtarget->hasBWI())
11272 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11275 if (Subtarget->hasBWI())
11276 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11280 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11283 // Otherwise fall back on splitting.
11284 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11287 // Lower vXi1 vector shuffles.
11288 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11289 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11290 // vector, shuffle and then truncate it back.
11291 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11292 MVT VT, const X86Subtarget *Subtarget,
11293 SelectionDAG &DAG) {
11295 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11296 ArrayRef<int> Mask = SVOp->getMask();
11297 assert(Subtarget->hasAVX512() &&
11298 "Cannot lower 512-bit vectors w/o basic ISA!");
11300 switch (VT.SimpleTy) {
11302 llvm_unreachable("Expected a vector of i1 elements");
11304 ExtVT = MVT::v2i64;
11307 ExtVT = MVT::v4i32;
11310 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11313 ExtVT = MVT::v16i32;
11316 ExtVT = MVT::v32i16;
11319 ExtVT = MVT::v64i8;
11323 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11324 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11325 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11326 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11328 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11331 V2 = DAG.getUNDEF(ExtVT);
11332 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11333 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11334 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11335 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11337 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11338 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11339 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11341 /// \brief Top-level lowering for x86 vector shuffles.
11343 /// This handles decomposition, canonicalization, and lowering of all x86
11344 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11345 /// above in helper routines. The canonicalization attempts to widen shuffles
11346 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11347 /// s.t. only one of the two inputs needs to be tested, etc.
11348 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11349 SelectionDAG &DAG) {
11350 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11351 ArrayRef<int> Mask = SVOp->getMask();
11352 SDValue V1 = Op.getOperand(0);
11353 SDValue V2 = Op.getOperand(1);
11354 MVT VT = Op.getSimpleValueType();
11355 int NumElements = VT.getVectorNumElements();
11357 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11359 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11360 "Can't lower MMX shuffles");
11362 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11363 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11364 if (V1IsUndef && V2IsUndef)
11365 return DAG.getUNDEF(VT);
11367 // When we create a shuffle node we put the UNDEF node to second operand,
11368 // but in some cases the first operand may be transformed to UNDEF.
11369 // In this case we should just commute the node.
11371 return DAG.getCommutedVectorShuffle(*SVOp);
11373 // Check for non-undef masks pointing at an undef vector and make the masks
11374 // undef as well. This makes it easier to match the shuffle based solely on
11378 if (M >= NumElements) {
11379 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11380 for (int &M : NewMask)
11381 if (M >= NumElements)
11383 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11386 // We actually see shuffles that are entirely re-arrangements of a set of
11387 // zero inputs. This mostly happens while decomposing complex shuffles into
11388 // simple ones. Directly lower these as a buildvector of zeros.
11389 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11390 if (Zeroable.all())
11391 return getZeroVector(VT, Subtarget, DAG, dl);
11393 // Try to collapse shuffles into using a vector type with fewer elements but
11394 // wider element types. We cap this to not form integers or floating point
11395 // elements wider than 64 bits, but it might be interesting to form i128
11396 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11397 SmallVector<int, 16> WidenedMask;
11398 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11399 canWidenShuffleElements(Mask, WidenedMask)) {
11400 MVT NewEltVT = VT.isFloatingPoint()
11401 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11402 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11403 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11404 // Make sure that the new vector type is legal. For example, v2f64 isn't
11406 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11407 V1 = DAG.getBitcast(NewVT, V1);
11408 V2 = DAG.getBitcast(NewVT, V2);
11409 return DAG.getBitcast(
11410 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11414 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11415 for (int M : SVOp->getMask())
11417 ++NumUndefElements;
11418 else if (M < NumElements)
11423 // Commute the shuffle as needed such that more elements come from V1 than
11424 // V2. This allows us to match the shuffle pattern strictly on how many
11425 // elements come from V1 without handling the symmetric cases.
11426 if (NumV2Elements > NumV1Elements)
11427 return DAG.getCommutedVectorShuffle(*SVOp);
11429 // When the number of V1 and V2 elements are the same, try to minimize the
11430 // number of uses of V2 in the low half of the vector. When that is tied,
11431 // ensure that the sum of indices for V1 is equal to or lower than the sum
11432 // indices for V2. When those are equal, try to ensure that the number of odd
11433 // indices for V1 is lower than the number of odd indices for V2.
11434 if (NumV1Elements == NumV2Elements) {
11435 int LowV1Elements = 0, LowV2Elements = 0;
11436 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11437 if (M >= NumElements)
11441 if (LowV2Elements > LowV1Elements) {
11442 return DAG.getCommutedVectorShuffle(*SVOp);
11443 } else if (LowV2Elements == LowV1Elements) {
11444 int SumV1Indices = 0, SumV2Indices = 0;
11445 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11446 if (SVOp->getMask()[i] >= NumElements)
11448 else if (SVOp->getMask()[i] >= 0)
11450 if (SumV2Indices < SumV1Indices) {
11451 return DAG.getCommutedVectorShuffle(*SVOp);
11452 } else if (SumV2Indices == SumV1Indices) {
11453 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11454 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11455 if (SVOp->getMask()[i] >= NumElements)
11456 NumV2OddIndices += i % 2;
11457 else if (SVOp->getMask()[i] >= 0)
11458 NumV1OddIndices += i % 2;
11459 if (NumV2OddIndices < NumV1OddIndices)
11460 return DAG.getCommutedVectorShuffle(*SVOp);
11465 // For each vector width, delegate to a specialized lowering routine.
11466 if (VT.is128BitVector())
11467 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11469 if (VT.is256BitVector())
11470 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11472 if (VT.is512BitVector())
11473 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11476 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11477 llvm_unreachable("Unimplemented!");
11480 // This function assumes its argument is a BUILD_VECTOR of constants or
11481 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11483 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11484 unsigned &MaskValue) {
11486 unsigned NumElems = BuildVector->getNumOperands();
11488 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11489 // We don't handle the >2 lanes case right now.
11490 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11494 unsigned NumElemsInLane = NumElems / NumLanes;
11496 // Blend for v16i16 should be symmetric for the both lanes.
11497 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11498 SDValue EltCond = BuildVector->getOperand(i);
11499 SDValue SndLaneEltCond =
11500 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11502 int Lane1Cond = -1, Lane2Cond = -1;
11503 if (isa<ConstantSDNode>(EltCond))
11504 Lane1Cond = !isNullConstant(EltCond);
11505 if (isa<ConstantSDNode>(SndLaneEltCond))
11506 Lane2Cond = !isNullConstant(SndLaneEltCond);
11508 unsigned LaneMask = 0;
11509 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11510 // Lane1Cond != 0, means we want the first argument.
11511 // Lane1Cond == 0, means we want the second argument.
11512 // The encoding of this argument is 0 for the first argument, 1
11513 // for the second. Therefore, invert the condition.
11514 LaneMask = !Lane1Cond << i;
11515 else if (Lane1Cond < 0)
11516 LaneMask = !Lane2Cond << i;
11520 MaskValue |= LaneMask;
11522 MaskValue |= LaneMask << NumElemsInLane;
11527 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11528 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11529 const X86Subtarget *Subtarget,
11530 SelectionDAG &DAG) {
11531 SDValue Cond = Op.getOperand(0);
11532 SDValue LHS = Op.getOperand(1);
11533 SDValue RHS = Op.getOperand(2);
11535 MVT VT = Op.getSimpleValueType();
11537 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11539 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11541 // Only non-legal VSELECTs reach this lowering, convert those into generic
11542 // shuffles and re-use the shuffle lowering path for blends.
11543 SmallVector<int, 32> Mask;
11544 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11545 SDValue CondElt = CondBV->getOperand(i);
11547 isa<ConstantSDNode>(CondElt) ? i + (isNullConstant(CondElt) ? Size : 0)
11550 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11553 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11554 // A vselect where all conditions and data are constants can be optimized into
11555 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11556 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11557 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11558 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11561 // Try to lower this to a blend-style vector shuffle. This can handle all
11562 // constant condition cases.
11563 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11566 // Variable blends are only legal from SSE4.1 onward.
11567 if (!Subtarget->hasSSE41())
11570 // Only some types will be legal on some subtargets. If we can emit a legal
11571 // VSELECT-matching blend, return Op, and but if we need to expand, return
11573 switch (Op.getSimpleValueType().SimpleTy) {
11575 // Most of the vector types have blends past SSE4.1.
11579 // The byte blends for AVX vectors were introduced only in AVX2.
11580 if (Subtarget->hasAVX2())
11587 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11588 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11591 // FIXME: We should custom lower this by fixing the condition and using i8
11597 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11598 MVT VT = Op.getSimpleValueType();
11601 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11604 if (VT.getSizeInBits() == 8) {
11605 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11606 Op.getOperand(0), Op.getOperand(1));
11607 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11608 DAG.getValueType(VT));
11609 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11612 if (VT.getSizeInBits() == 16) {
11613 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11614 if (isNullConstant(Op.getOperand(1)))
11615 return DAG.getNode(
11616 ISD::TRUNCATE, dl, MVT::i16,
11617 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11618 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11619 Op.getOperand(1)));
11620 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11621 Op.getOperand(0), Op.getOperand(1));
11622 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11623 DAG.getValueType(VT));
11624 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11627 if (VT == MVT::f32) {
11628 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11629 // the result back to FR32 register. It's only worth matching if the
11630 // result has a single use which is a store or a bitcast to i32. And in
11631 // the case of a store, it's not worth it if the index is a constant 0,
11632 // because a MOVSSmr can be used instead, which is smaller and faster.
11633 if (!Op.hasOneUse())
11635 SDNode *User = *Op.getNode()->use_begin();
11636 if ((User->getOpcode() != ISD::STORE ||
11637 isNullConstant(Op.getOperand(1))) &&
11638 (User->getOpcode() != ISD::BITCAST ||
11639 User->getValueType(0) != MVT::i32))
11641 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11642 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11644 return DAG.getBitcast(MVT::f32, Extract);
11647 if (VT == MVT::i32 || VT == MVT::i64) {
11648 // ExtractPS/pextrq works with constant index.
11649 if (isa<ConstantSDNode>(Op.getOperand(1)))
11655 /// Extract one bit from mask vector, like v16i1 or v8i1.
11656 /// AVX-512 feature.
11658 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11659 SDValue Vec = Op.getOperand(0);
11661 MVT VecVT = Vec.getSimpleValueType();
11662 SDValue Idx = Op.getOperand(1);
11663 MVT EltVT = Op.getSimpleValueType();
11665 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11666 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11667 "Unexpected vector type in ExtractBitFromMaskVector");
11669 // variable index can't be handled in mask registers,
11670 // extend vector to VR512
11671 if (!isa<ConstantSDNode>(Idx)) {
11672 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11673 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11674 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11675 ExtVT.getVectorElementType(), Ext, Idx);
11676 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11679 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11680 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11681 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11682 rc = getRegClassFor(MVT::v16i1);
11683 unsigned MaxSift = rc->getSize()*8 - 1;
11684 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11685 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11686 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11687 DAG.getConstant(MaxSift, dl, MVT::i8));
11688 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11689 DAG.getIntPtrConstant(0, dl));
11693 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11694 SelectionDAG &DAG) const {
11696 SDValue Vec = Op.getOperand(0);
11697 MVT VecVT = Vec.getSimpleValueType();
11698 SDValue Idx = Op.getOperand(1);
11700 if (Op.getSimpleValueType() == MVT::i1)
11701 return ExtractBitFromMaskVector(Op, DAG);
11703 if (!isa<ConstantSDNode>(Idx)) {
11704 if (VecVT.is512BitVector() ||
11705 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11706 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11709 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11710 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11711 MaskEltVT.getSizeInBits());
11713 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11714 auto PtrVT = getPointerTy(DAG.getDataLayout());
11715 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11716 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11717 DAG.getConstant(0, dl, PtrVT));
11718 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11719 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11720 DAG.getConstant(0, dl, PtrVT));
11725 // If this is a 256-bit vector result, first extract the 128-bit vector and
11726 // then extract the element from the 128-bit vector.
11727 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11729 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11730 // Get the 128-bit vector.
11731 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11732 MVT EltVT = VecVT.getVectorElementType();
11734 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11735 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11737 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11738 // this can be done with a mask.
11739 IdxVal &= ElemsPerChunk - 1;
11740 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11741 DAG.getConstant(IdxVal, dl, MVT::i32));
11744 assert(VecVT.is128BitVector() && "Unexpected vector length");
11746 if (Subtarget->hasSSE41())
11747 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11750 MVT VT = Op.getSimpleValueType();
11751 // TODO: handle v16i8.
11752 if (VT.getSizeInBits() == 16) {
11753 SDValue Vec = Op.getOperand(0);
11754 if (isNullConstant(Op.getOperand(1)))
11755 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11756 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11757 DAG.getBitcast(MVT::v4i32, Vec),
11758 Op.getOperand(1)));
11759 // Transform it so it match pextrw which produces a 32-bit result.
11760 MVT EltVT = MVT::i32;
11761 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11762 Op.getOperand(0), Op.getOperand(1));
11763 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11764 DAG.getValueType(VT));
11765 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11768 if (VT.getSizeInBits() == 32) {
11769 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11773 // SHUFPS the element to the lowest double word, then movss.
11774 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11775 MVT VVT = Op.getOperand(0).getSimpleValueType();
11776 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11777 DAG.getUNDEF(VVT), Mask);
11778 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11779 DAG.getIntPtrConstant(0, dl));
11782 if (VT.getSizeInBits() == 64) {
11783 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11784 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11785 // to match extract_elt for f64.
11786 if (isNullConstant(Op.getOperand(1)))
11789 // UNPCKHPD the element to the lowest double word, then movsd.
11790 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11791 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11792 int Mask[2] = { 1, -1 };
11793 MVT VVT = Op.getOperand(0).getSimpleValueType();
11794 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11795 DAG.getUNDEF(VVT), Mask);
11796 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11797 DAG.getIntPtrConstant(0, dl));
11803 /// Insert one bit to mask vector, like v16i1 or v8i1.
11804 /// AVX-512 feature.
11806 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11808 SDValue Vec = Op.getOperand(0);
11809 SDValue Elt = Op.getOperand(1);
11810 SDValue Idx = Op.getOperand(2);
11811 MVT VecVT = Vec.getSimpleValueType();
11813 if (!isa<ConstantSDNode>(Idx)) {
11814 // Non constant index. Extend source and destination,
11815 // insert element and then truncate the result.
11816 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11817 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11818 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11819 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11820 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11821 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11824 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11825 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11827 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11828 DAG.getConstant(IdxVal, dl, MVT::i8));
11829 if (Vec.getOpcode() == ISD::UNDEF)
11831 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11834 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11835 SelectionDAG &DAG) const {
11836 MVT VT = Op.getSimpleValueType();
11837 MVT EltVT = VT.getVectorElementType();
11839 if (EltVT == MVT::i1)
11840 return InsertBitToMaskVector(Op, DAG);
11843 SDValue N0 = Op.getOperand(0);
11844 SDValue N1 = Op.getOperand(1);
11845 SDValue N2 = Op.getOperand(2);
11846 if (!isa<ConstantSDNode>(N2))
11848 auto *N2C = cast<ConstantSDNode>(N2);
11849 unsigned IdxVal = N2C->getZExtValue();
11851 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11852 // into that, and then insert the subvector back into the result.
11853 if (VT.is256BitVector() || VT.is512BitVector()) {
11854 // With a 256-bit vector, we can insert into the zero element efficiently
11855 // using a blend if we have AVX or AVX2 and the right data type.
11856 if (VT.is256BitVector() && IdxVal == 0) {
11857 // TODO: It is worthwhile to cast integer to floating point and back
11858 // and incur a domain crossing penalty if that's what we'll end up
11859 // doing anyway after extracting to a 128-bit vector.
11860 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11861 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11862 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11863 N2 = DAG.getIntPtrConstant(1, dl);
11864 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11868 // Get the desired 128-bit vector chunk.
11869 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11871 // Insert the element into the desired chunk.
11872 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11873 assert(isPowerOf2_32(NumEltsIn128));
11874 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11875 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11877 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11878 DAG.getConstant(IdxIn128, dl, MVT::i32));
11880 // Insert the changed part back into the bigger vector
11881 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11883 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11885 if (Subtarget->hasSSE41()) {
11886 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11888 if (VT == MVT::v8i16) {
11889 Opc = X86ISD::PINSRW;
11891 assert(VT == MVT::v16i8);
11892 Opc = X86ISD::PINSRB;
11895 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11897 if (N1.getValueType() != MVT::i32)
11898 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11899 if (N2.getValueType() != MVT::i32)
11900 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11901 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11904 if (EltVT == MVT::f32) {
11905 // Bits [7:6] of the constant are the source select. This will always be
11906 // zero here. The DAG Combiner may combine an extract_elt index into
11907 // these bits. For example (insert (extract, 3), 2) could be matched by
11908 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11909 // Bits [5:4] of the constant are the destination select. This is the
11910 // value of the incoming immediate.
11911 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11912 // combine either bitwise AND or insert of float 0.0 to set these bits.
11914 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11915 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11916 // If this is an insertion of 32-bits into the low 32-bits of
11917 // a vector, we prefer to generate a blend with immediate rather
11918 // than an insertps. Blends are simpler operations in hardware and so
11919 // will always have equal or better performance than insertps.
11920 // But if optimizing for size and there's a load folding opportunity,
11921 // generate insertps because blendps does not have a 32-bit memory
11923 N2 = DAG.getIntPtrConstant(1, dl);
11924 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11925 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11927 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11928 // Create this as a scalar to vector..
11929 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11930 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11933 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11934 // PINSR* works with constant index.
11939 if (EltVT == MVT::i8)
11942 if (EltVT.getSizeInBits() == 16) {
11943 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11944 // as its second argument.
11945 if (N1.getValueType() != MVT::i32)
11946 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11947 if (N2.getValueType() != MVT::i32)
11948 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11949 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11954 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11956 MVT OpVT = Op.getSimpleValueType();
11958 // If this is a 256-bit vector result, first insert into a 128-bit
11959 // vector and then insert into the 256-bit vector.
11960 if (!OpVT.is128BitVector()) {
11961 // Insert into a 128-bit vector.
11962 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11963 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11964 OpVT.getVectorNumElements() / SizeFactor);
11966 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11968 // Insert the 128-bit vector.
11969 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11972 if (OpVT == MVT::v1i64 &&
11973 Op.getOperand(0).getValueType() == MVT::i64)
11974 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11976 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11977 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11978 return DAG.getBitcast(
11979 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11982 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11983 // a simple subregister reference or explicit instructions to grab
11984 // upper bits of a vector.
11985 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11986 SelectionDAG &DAG) {
11988 SDValue In = Op.getOperand(0);
11989 SDValue Idx = Op.getOperand(1);
11990 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11991 MVT ResVT = Op.getSimpleValueType();
11992 MVT InVT = In.getSimpleValueType();
11994 if (Subtarget->hasFp256()) {
11995 if (ResVT.is128BitVector() &&
11996 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11997 isa<ConstantSDNode>(Idx)) {
11998 return Extract128BitVector(In, IdxVal, DAG, dl);
12000 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
12001 isa<ConstantSDNode>(Idx)) {
12002 return Extract256BitVector(In, IdxVal, DAG, dl);
12008 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
12009 // simple superregister reference or explicit instructions to insert
12010 // the upper bits of a vector.
12011 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
12012 SelectionDAG &DAG) {
12013 if (!Subtarget->hasAVX())
12017 SDValue Vec = Op.getOperand(0);
12018 SDValue SubVec = Op.getOperand(1);
12019 SDValue Idx = Op.getOperand(2);
12021 if (!isa<ConstantSDNode>(Idx))
12024 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12025 MVT OpVT = Op.getSimpleValueType();
12026 MVT SubVecVT = SubVec.getSimpleValueType();
12028 // Fold two 16-byte subvector loads into one 32-byte load:
12029 // (insert_subvector (insert_subvector undef, (load addr), 0),
12030 // (load addr + 16), Elts/2)
12032 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
12033 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
12034 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
12035 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
12036 if (Idx2 && Idx2->getZExtValue() == 0) {
12037 SDValue SubVec2 = Vec.getOperand(1);
12038 // If needed, look through a bitcast to get to the load.
12039 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
12040 SubVec2 = SubVec2.getOperand(0);
12042 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
12044 unsigned Alignment = FirstLd->getAlignment();
12045 unsigned AS = FirstLd->getAddressSpace();
12046 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
12047 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
12048 OpVT, AS, Alignment, &Fast) && Fast) {
12049 SDValue Ops[] = { SubVec2, SubVec };
12050 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
12057 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
12058 SubVecVT.is128BitVector())
12059 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
12061 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
12062 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
12064 if (OpVT.getVectorElementType() == MVT::i1)
12065 return Insert1BitVector(Op, DAG);
12070 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
12071 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
12072 // one of the above mentioned nodes. It has to be wrapped because otherwise
12073 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
12074 // be used to form addressing mode. These wrapped nodes will be selected
12077 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
12078 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
12080 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12081 // global base reg.
12082 unsigned char OpFlag = 0;
12083 unsigned WrapperKind = X86ISD::Wrapper;
12084 CodeModel::Model M = DAG.getTarget().getCodeModel();
12086 if (Subtarget->isPICStyleRIPRel() &&
12087 (M == CodeModel::Small || M == CodeModel::Kernel))
12088 WrapperKind = X86ISD::WrapperRIP;
12089 else if (Subtarget->isPICStyleGOT())
12090 OpFlag = X86II::MO_GOTOFF;
12091 else if (Subtarget->isPICStyleStubPIC())
12092 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12094 auto PtrVT = getPointerTy(DAG.getDataLayout());
12095 SDValue Result = DAG.getTargetConstantPool(
12096 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
12098 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12099 // With PIC, the address is actually $g + Offset.
12102 DAG.getNode(ISD::ADD, DL, PtrVT,
12103 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12109 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
12110 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
12112 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12113 // global base reg.
12114 unsigned char OpFlag = 0;
12115 unsigned WrapperKind = X86ISD::Wrapper;
12116 CodeModel::Model M = DAG.getTarget().getCodeModel();
12118 if (Subtarget->isPICStyleRIPRel() &&
12119 (M == CodeModel::Small || M == CodeModel::Kernel))
12120 WrapperKind = X86ISD::WrapperRIP;
12121 else if (Subtarget->isPICStyleGOT())
12122 OpFlag = X86II::MO_GOTOFF;
12123 else if (Subtarget->isPICStyleStubPIC())
12124 OpFlag = X86II::MO_PIC_BASE_OFFSET;
12126 auto PtrVT = getPointerTy(DAG.getDataLayout());
12127 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
12129 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12131 // With PIC, the address is actually $g + Offset.
12134 DAG.getNode(ISD::ADD, DL, PtrVT,
12135 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12141 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
12142 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
12144 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12145 // global base reg.
12146 unsigned char OpFlag = 0;
12147 unsigned WrapperKind = X86ISD::Wrapper;
12148 CodeModel::Model M = DAG.getTarget().getCodeModel();
12150 if (Subtarget->isPICStyleRIPRel() &&
12151 (M == CodeModel::Small || M == CodeModel::Kernel)) {
12152 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
12153 OpFlag = X86II::MO_GOTPCREL;
12154 WrapperKind = X86ISD::WrapperRIP;
12155 } else if (Subtarget->isPICStyleGOT()) {
12156 OpFlag = X86II::MO_GOT;
12157 } else if (Subtarget->isPICStyleStubPIC()) {
12158 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
12159 } else if (Subtarget->isPICStyleStubNoDynamic()) {
12160 OpFlag = X86II::MO_DARWIN_NONLAZY;
12163 auto PtrVT = getPointerTy(DAG.getDataLayout());
12164 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
12167 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12169 // With PIC, the address is actually $g + Offset.
12170 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
12171 !Subtarget->is64Bit()) {
12173 DAG.getNode(ISD::ADD, DL, PtrVT,
12174 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
12177 // For symbols that require a load from a stub to get the address, emit the
12179 if (isGlobalStubReference(OpFlag))
12180 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
12181 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12182 false, false, false, 0);
12188 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
12189 // Create the TargetBlockAddressAddress node.
12190 unsigned char OpFlags =
12191 Subtarget->ClassifyBlockAddressReference();
12192 CodeModel::Model M = DAG.getTarget().getCodeModel();
12193 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
12194 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
12196 auto PtrVT = getPointerTy(DAG.getDataLayout());
12197 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
12199 if (Subtarget->isPICStyleRIPRel() &&
12200 (M == CodeModel::Small || M == CodeModel::Kernel))
12201 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12203 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12205 // With PIC, the address is actually $g + Offset.
12206 if (isGlobalRelativeToPICBase(OpFlags)) {
12207 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12208 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12215 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
12216 int64_t Offset, SelectionDAG &DAG) const {
12217 // Create the TargetGlobalAddress node, folding in the constant
12218 // offset if it is legal.
12219 unsigned char OpFlags =
12220 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
12221 CodeModel::Model M = DAG.getTarget().getCodeModel();
12222 auto PtrVT = getPointerTy(DAG.getDataLayout());
12224 if (OpFlags == X86II::MO_NO_FLAG &&
12225 X86::isOffsetSuitableForCodeModel(Offset, M)) {
12226 // A direct static reference to a global.
12227 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
12230 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
12233 if (Subtarget->isPICStyleRIPRel() &&
12234 (M == CodeModel::Small || M == CodeModel::Kernel))
12235 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
12237 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12239 // With PIC, the address is actually $g + Offset.
12240 if (isGlobalRelativeToPICBase(OpFlags)) {
12241 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12242 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12245 // For globals that require a load from a stub to get the address, emit the
12247 if (isGlobalStubReference(OpFlags))
12248 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12249 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12250 false, false, false, 0);
12252 // If there was a non-zero offset that we didn't fold, create an explicit
12253 // addition for it.
12255 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12256 DAG.getConstant(Offset, dl, PtrVT));
12262 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12263 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12264 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12265 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12269 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12270 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12271 unsigned char OperandFlags, bool LocalDynamic = false) {
12272 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12273 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12275 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12276 GA->getValueType(0),
12280 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12284 SDValue Ops[] = { Chain, TGA, *InFlag };
12285 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12287 SDValue Ops[] = { Chain, TGA };
12288 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12291 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12292 MFI->setAdjustsStack(true);
12293 MFI->setHasCalls(true);
12295 SDValue Flag = Chain.getValue(1);
12296 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12299 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12301 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12304 SDLoc dl(GA); // ? function entry point might be better
12305 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12306 DAG.getNode(X86ISD::GlobalBaseReg,
12307 SDLoc(), PtrVT), InFlag);
12308 InFlag = Chain.getValue(1);
12310 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12313 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12315 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12317 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12318 X86::RAX, X86II::MO_TLSGD);
12321 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12327 // Get the start address of the TLS block for this module.
12328 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12329 .getInfo<X86MachineFunctionInfo>();
12330 MFI->incNumLocalDynamicTLSAccesses();
12334 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12335 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12338 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12339 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12340 InFlag = Chain.getValue(1);
12341 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12342 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12345 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12349 unsigned char OperandFlags = X86II::MO_DTPOFF;
12350 unsigned WrapperKind = X86ISD::Wrapper;
12351 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12352 GA->getValueType(0),
12353 GA->getOffset(), OperandFlags);
12354 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12356 // Add x@dtpoff with the base.
12357 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12360 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12361 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12362 const EVT PtrVT, TLSModel::Model model,
12363 bool is64Bit, bool isPIC) {
12366 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12367 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12368 is64Bit ? 257 : 256));
12370 SDValue ThreadPointer =
12371 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12372 MachinePointerInfo(Ptr), false, false, false, 0);
12374 unsigned char OperandFlags = 0;
12375 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12377 unsigned WrapperKind = X86ISD::Wrapper;
12378 if (model == TLSModel::LocalExec) {
12379 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12380 } else if (model == TLSModel::InitialExec) {
12382 OperandFlags = X86II::MO_GOTTPOFF;
12383 WrapperKind = X86ISD::WrapperRIP;
12385 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12388 llvm_unreachable("Unexpected model");
12391 // emit "addl x@ntpoff,%eax" (local exec)
12392 // or "addl x@indntpoff,%eax" (initial exec)
12393 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12395 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12396 GA->getOffset(), OperandFlags);
12397 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12399 if (model == TLSModel::InitialExec) {
12400 if (isPIC && !is64Bit) {
12401 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12402 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12406 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12407 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12408 false, false, false, 0);
12411 // The address of the thread local variable is the add of the thread
12412 // pointer with the offset of the variable.
12413 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12417 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12419 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12421 // Cygwin uses emutls.
12422 // FIXME: It may be EmulatedTLS-generic also for X86-Android.
12423 if (Subtarget->isTargetWindowsCygwin())
12424 return LowerToTLSEmulatedModel(GA, DAG);
12426 const GlobalValue *GV = GA->getGlobal();
12427 auto PtrVT = getPointerTy(DAG.getDataLayout());
12429 if (Subtarget->isTargetELF()) {
12430 if (DAG.getTarget().Options.EmulatedTLS)
12431 return LowerToTLSEmulatedModel(GA, DAG);
12432 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12434 case TLSModel::GeneralDynamic:
12435 if (Subtarget->is64Bit())
12436 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12437 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12438 case TLSModel::LocalDynamic:
12439 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12440 Subtarget->is64Bit());
12441 case TLSModel::InitialExec:
12442 case TLSModel::LocalExec:
12443 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12444 DAG.getTarget().getRelocationModel() ==
12447 llvm_unreachable("Unknown TLS model.");
12450 if (Subtarget->isTargetDarwin()) {
12451 // Darwin only has one model of TLS. Lower to that.
12452 unsigned char OpFlag = 0;
12453 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12454 X86ISD::WrapperRIP : X86ISD::Wrapper;
12456 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12457 // global base reg.
12458 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12459 !Subtarget->is64Bit();
12461 OpFlag = X86II::MO_TLVP_PIC_BASE;
12463 OpFlag = X86II::MO_TLVP;
12465 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12466 GA->getValueType(0),
12467 GA->getOffset(), OpFlag);
12468 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12470 // With PIC32, the address is actually $g + Offset.
12472 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12473 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12476 // Lowering the machine isd will make sure everything is in the right
12478 SDValue Chain = DAG.getEntryNode();
12479 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12480 SDValue Args[] = { Chain, Offset };
12481 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12483 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12484 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12485 MFI->setAdjustsStack(true);
12487 // And our return value (tls address) is in the standard call return value
12489 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12490 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12493 if (Subtarget->isTargetKnownWindowsMSVC() ||
12494 Subtarget->isTargetWindowsGNU()) {
12495 // Just use the implicit TLS architecture
12496 // Need to generate someting similar to:
12497 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12499 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12500 // mov rcx, qword [rdx+rcx*8]
12501 // mov eax, .tls$:tlsvar
12502 // [rax+rcx] contains the address
12503 // Windows 64bit: gs:0x58
12504 // Windows 32bit: fs:__tls_array
12507 SDValue Chain = DAG.getEntryNode();
12509 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12510 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12511 // use its literal value of 0x2C.
12512 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12513 ? Type::getInt8PtrTy(*DAG.getContext(),
12515 : Type::getInt32PtrTy(*DAG.getContext(),
12518 SDValue TlsArray = Subtarget->is64Bit()
12519 ? DAG.getIntPtrConstant(0x58, dl)
12520 : (Subtarget->isTargetWindowsGNU()
12521 ? DAG.getIntPtrConstant(0x2C, dl)
12522 : DAG.getExternalSymbol("_tls_array", PtrVT));
12524 SDValue ThreadPointer =
12525 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12529 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12530 res = ThreadPointer;
12532 // Load the _tls_index variable
12533 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12534 if (Subtarget->is64Bit())
12535 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12536 MachinePointerInfo(), MVT::i32, false, false,
12539 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12542 auto &DL = DAG.getDataLayout();
12544 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12545 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12547 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12550 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12553 // Get the offset of start of .tls section
12554 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12555 GA->getValueType(0),
12556 GA->getOffset(), X86II::MO_SECREL);
12557 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12559 // The address of the thread local variable is the add of the thread
12560 // pointer with the offset of the variable.
12561 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12564 llvm_unreachable("TLS not implemented for this target.");
12567 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12568 /// and take a 2 x i32 value to shift plus a shift amount.
12569 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12570 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12571 MVT VT = Op.getSimpleValueType();
12572 unsigned VTBits = VT.getSizeInBits();
12574 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12575 SDValue ShOpLo = Op.getOperand(0);
12576 SDValue ShOpHi = Op.getOperand(1);
12577 SDValue ShAmt = Op.getOperand(2);
12578 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12579 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12581 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12582 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12583 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12584 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12585 : DAG.getConstant(0, dl, VT);
12587 SDValue Tmp2, Tmp3;
12588 if (Op.getOpcode() == ISD::SHL_PARTS) {
12589 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12590 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12592 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12593 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12596 // If the shift amount is larger or equal than the width of a part we can't
12597 // rely on the results of shld/shrd. Insert a test and select the appropriate
12598 // values for large shift amounts.
12599 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12600 DAG.getConstant(VTBits, dl, MVT::i8));
12601 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12602 AndNode, DAG.getConstant(0, dl, MVT::i8));
12605 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12606 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12607 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12609 if (Op.getOpcode() == ISD::SHL_PARTS) {
12610 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12611 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12613 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12614 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12617 SDValue Ops[2] = { Lo, Hi };
12618 return DAG.getMergeValues(Ops, dl);
12621 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12622 SelectionDAG &DAG) const {
12623 SDValue Src = Op.getOperand(0);
12624 MVT SrcVT = Src.getSimpleValueType();
12625 MVT VT = Op.getSimpleValueType();
12628 if (SrcVT.isVector()) {
12629 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12630 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12631 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12632 DAG.getUNDEF(SrcVT)));
12634 if (SrcVT.getVectorElementType() == MVT::i1) {
12635 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12636 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12637 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12642 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12643 "Unknown SINT_TO_FP to lower!");
12645 // These are really Legal; return the operand so the caller accepts it as
12647 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12649 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12650 Subtarget->is64Bit()) {
12654 unsigned Size = SrcVT.getSizeInBits()/8;
12655 MachineFunction &MF = DAG.getMachineFunction();
12656 auto PtrVT = getPointerTy(MF.getDataLayout());
12657 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12658 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12659 SDValue Chain = DAG.getStore(
12660 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12661 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12663 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12666 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12668 SelectionDAG &DAG) const {
12672 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12674 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12676 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12678 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12680 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12681 MachineMemOperand *MMO;
12683 int SSFI = FI->getIndex();
12684 MMO = DAG.getMachineFunction().getMachineMemOperand(
12685 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12686 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12688 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12689 StackSlot = StackSlot.getOperand(1);
12691 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12692 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12694 Tys, Ops, SrcVT, MMO);
12697 Chain = Result.getValue(1);
12698 SDValue InFlag = Result.getValue(2);
12700 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12701 // shouldn't be necessary except that RFP cannot be live across
12702 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12703 MachineFunction &MF = DAG.getMachineFunction();
12704 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12705 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12706 auto PtrVT = getPointerTy(MF.getDataLayout());
12707 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12708 Tys = DAG.getVTList(MVT::Other);
12710 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12712 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12713 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12714 MachineMemOperand::MOStore, SSFISize, SSFISize);
12716 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12717 Ops, Op.getValueType(), MMO);
12718 Result = DAG.getLoad(
12719 Op.getValueType(), DL, Chain, StackSlot,
12720 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12721 false, false, false, 0);
12727 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12728 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12729 SelectionDAG &DAG) const {
12730 // This algorithm is not obvious. Here it is what we're trying to output:
12733 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12734 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12736 haddpd %xmm0, %xmm0
12738 pshufd $0x4e, %xmm0, %xmm1
12744 LLVMContext *Context = DAG.getContext();
12746 // Build some magic constants.
12747 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12748 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12749 auto PtrVT = getPointerTy(DAG.getDataLayout());
12750 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12752 SmallVector<Constant*,2> CV1;
12754 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12755 APInt(64, 0x4330000000000000ULL))));
12757 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12758 APInt(64, 0x4530000000000000ULL))));
12759 Constant *C1 = ConstantVector::get(CV1);
12760 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12762 // Load the 64-bit value into an XMM register.
12763 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12766 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12767 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12768 false, false, false, 16);
12770 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12773 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12774 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12775 false, false, false, 16);
12776 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12777 // TODO: Are there any fast-math-flags to propagate here?
12778 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12781 if (Subtarget->hasSSE3()) {
12782 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12783 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12785 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12786 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12788 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12789 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12792 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12793 DAG.getIntPtrConstant(0, dl));
12796 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12797 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12798 SelectionDAG &DAG) const {
12800 // FP constant to bias correct the final result.
12801 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12804 // Load the 32-bit value into an XMM register.
12805 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12808 // Zero out the upper parts of the register.
12809 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12811 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12812 DAG.getBitcast(MVT::v2f64, Load),
12813 DAG.getIntPtrConstant(0, dl));
12815 // Or the load with the bias.
12816 SDValue Or = DAG.getNode(
12817 ISD::OR, dl, MVT::v2i64,
12818 DAG.getBitcast(MVT::v2i64,
12819 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12820 DAG.getBitcast(MVT::v2i64,
12821 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12823 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12824 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12826 // Subtract the bias.
12827 // TODO: Are there any fast-math-flags to propagate here?
12828 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12830 // Handle final rounding.
12831 MVT DestVT = Op.getSimpleValueType();
12833 if (DestVT.bitsLT(MVT::f64))
12834 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12835 DAG.getIntPtrConstant(0, dl));
12836 if (DestVT.bitsGT(MVT::f64))
12837 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12839 // Handle final rounding.
12843 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12844 const X86Subtarget &Subtarget) {
12845 // The algorithm is the following:
12846 // #ifdef __SSE4_1__
12847 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12848 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12849 // (uint4) 0x53000000, 0xaa);
12851 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12852 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12854 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12855 // return (float4) lo + fhi;
12857 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12858 // reassociate the two FADDs, and if we do that, the algorithm fails
12859 // spectacularly (PR24512).
12860 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12861 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12862 // there's also the MachineCombiner reassociations happening on Machine IR.
12863 if (DAG.getTarget().Options.UnsafeFPMath)
12867 SDValue V = Op->getOperand(0);
12868 MVT VecIntVT = V.getSimpleValueType();
12869 bool Is128 = VecIntVT == MVT::v4i32;
12870 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12871 // If we convert to something else than the supported type, e.g., to v4f64,
12873 if (VecFloatVT != Op->getSimpleValueType(0))
12876 unsigned NumElts = VecIntVT.getVectorNumElements();
12877 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12878 "Unsupported custom type");
12879 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12881 // In the #idef/#else code, we have in common:
12882 // - The vector of constants:
12888 // Create the splat vector for 0x4b000000.
12889 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12890 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12891 CstLow, CstLow, CstLow, CstLow};
12892 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12893 makeArrayRef(&CstLowArray[0], NumElts));
12894 // Create the splat vector for 0x53000000.
12895 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12896 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12897 CstHigh, CstHigh, CstHigh, CstHigh};
12898 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12899 makeArrayRef(&CstHighArray[0], NumElts));
12901 // Create the right shift.
12902 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12903 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12904 CstShift, CstShift, CstShift, CstShift};
12905 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12906 makeArrayRef(&CstShiftArray[0], NumElts));
12907 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12910 if (Subtarget.hasSSE41()) {
12911 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12912 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12913 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12914 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12915 // Low will be bitcasted right away, so do not bother bitcasting back to its
12917 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12918 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12919 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12920 // (uint4) 0x53000000, 0xaa);
12921 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12922 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12923 // High will be bitcasted right away, so do not bother bitcasting back to
12924 // its original type.
12925 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12926 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12928 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12929 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12930 CstMask, CstMask, CstMask);
12931 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12932 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12933 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12935 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12936 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12939 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12940 SDValue CstFAdd = DAG.getConstantFP(
12941 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12942 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12943 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12944 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12945 makeArrayRef(&CstFAddArray[0], NumElts));
12947 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12948 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12949 // TODO: Are there any fast-math-flags to propagate here?
12951 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12952 // return (float4) lo + fhi;
12953 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12954 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12957 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12958 SelectionDAG &DAG) const {
12959 SDValue N0 = Op.getOperand(0);
12960 MVT SVT = N0.getSimpleValueType();
12963 switch (SVT.SimpleTy) {
12965 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12970 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12971 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12972 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12976 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12979 assert(Subtarget->hasAVX512());
12980 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12981 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12985 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12986 SelectionDAG &DAG) const {
12987 SDValue N0 = Op.getOperand(0);
12989 auto PtrVT = getPointerTy(DAG.getDataLayout());
12991 if (Op.getSimpleValueType().isVector())
12992 return lowerUINT_TO_FP_vec(Op, DAG);
12994 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12995 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12996 // the optimization here.
12997 if (DAG.SignBitIsZero(N0))
12998 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
13000 MVT SrcVT = N0.getSimpleValueType();
13001 MVT DstVT = Op.getSimpleValueType();
13003 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
13004 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
13005 // Conversions from unsigned i32 to f32/f64 are legal,
13006 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
13010 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
13011 return LowerUINT_TO_FP_i64(Op, DAG);
13012 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
13013 return LowerUINT_TO_FP_i32(Op, DAG);
13014 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
13017 // Make a 64-bit buffer, and use it to build an FILD.
13018 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
13019 if (SrcVT == MVT::i32) {
13020 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
13021 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
13022 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13023 StackSlot, MachinePointerInfo(),
13025 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
13026 OffsetSlot, MachinePointerInfo(),
13028 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
13032 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
13033 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13034 StackSlot, MachinePointerInfo(),
13036 // For i64 source, we need to add the appropriate power of 2 if the input
13037 // was negative. This is the same as the optimization in
13038 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
13039 // we must be careful to do the computation in x87 extended precision, not
13040 // in SSE. (The generic code can't know it's OK to do this, or how to.)
13041 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
13042 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
13043 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
13044 MachineMemOperand::MOLoad, 8, 8);
13046 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
13047 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
13048 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
13051 APInt FF(32, 0x5F800000ULL);
13053 // Check whether the sign bit is set.
13054 SDValue SignSet = DAG.getSetCC(
13055 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
13056 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
13058 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
13059 SDValue FudgePtr = DAG.getConstantPool(
13060 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
13062 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
13063 SDValue Zero = DAG.getIntPtrConstant(0, dl);
13064 SDValue Four = DAG.getIntPtrConstant(4, dl);
13065 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
13067 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
13069 // Load the value out, extending it from f32 to f80.
13070 // FIXME: Avoid the extend by constructing the right constant pool?
13071 SDValue Fudge = DAG.getExtLoad(
13072 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
13073 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
13074 false, false, false, 4);
13075 // Extend everything to 80 bits to force it to be done on x87.
13076 // TODO: Are there any fast-math-flags to propagate here?
13077 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
13078 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
13079 DAG.getIntPtrConstant(0, dl));
13082 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
13083 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
13084 // just return an <SDValue(), SDValue()> pair.
13085 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
13086 // to i16, i32 or i64, and we lower it to a legal sequence.
13087 // If lowered to the final integer result we return a <result, SDValue()> pair.
13088 // Otherwise we lower it to a sequence ending with a FIST, return a
13089 // <FIST, StackSlot> pair, and the caller is responsible for loading
13090 // the final integer result from StackSlot.
13091 std::pair<SDValue,SDValue>
13092 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
13093 bool IsSigned, bool IsReplace) const {
13096 EVT DstTy = Op.getValueType();
13097 EVT TheVT = Op.getOperand(0).getValueType();
13098 auto PtrVT = getPointerTy(DAG.getDataLayout());
13100 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
13101 // f16 must be promoted before using the lowering in this routine.
13102 // fp128 does not use this lowering.
13103 return std::make_pair(SDValue(), SDValue());
13106 // If using FIST to compute an unsigned i64, we'll need some fixup
13107 // to handle values above the maximum signed i64. A FIST is always
13108 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
13109 bool UnsignedFixup = !IsSigned &&
13110 DstTy == MVT::i64 &&
13111 (!Subtarget->is64Bit() ||
13112 !isScalarFPTypeInSSEReg(TheVT));
13114 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
13115 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
13116 // The low 32 bits of the fist result will have the correct uint32 result.
13117 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
13121 assert(DstTy.getSimpleVT() <= MVT::i64 &&
13122 DstTy.getSimpleVT() >= MVT::i16 &&
13123 "Unknown FP_TO_INT to lower!");
13125 // These are really Legal.
13126 if (DstTy == MVT::i32 &&
13127 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13128 return std::make_pair(SDValue(), SDValue());
13129 if (Subtarget->is64Bit() &&
13130 DstTy == MVT::i64 &&
13131 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
13132 return std::make_pair(SDValue(), SDValue());
13134 // We lower FP->int64 into FISTP64 followed by a load from a temporary
13136 MachineFunction &MF = DAG.getMachineFunction();
13137 unsigned MemSize = DstTy.getSizeInBits()/8;
13138 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13139 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13142 switch (DstTy.getSimpleVT().SimpleTy) {
13143 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
13144 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
13145 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
13146 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
13149 SDValue Chain = DAG.getEntryNode();
13150 SDValue Value = Op.getOperand(0);
13151 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
13153 if (UnsignedFixup) {
13155 // Conversion to unsigned i64 is implemented with a select,
13156 // depending on whether the source value fits in the range
13157 // of a signed i64. Let Thresh be the FP equivalent of
13158 // 0x8000000000000000ULL.
13160 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
13161 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
13162 // Fist-to-mem64 FistSrc
13163 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
13164 // to XOR'ing the high 32 bits with Adjust.
13166 // Being a power of 2, Thresh is exactly representable in all FP formats.
13167 // For X87 we'd like to use the smallest FP type for this constant, but
13168 // for DAG type consistency we have to match the FP operand type.
13170 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
13171 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
13172 bool LosesInfo = false;
13173 if (TheVT == MVT::f64)
13174 // The rounding mode is irrelevant as the conversion should be exact.
13175 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
13177 else if (TheVT == MVT::f80)
13178 Status = Thresh.convert(APFloat::x87DoubleExtended,
13179 APFloat::rmNearestTiesToEven, &LosesInfo);
13181 assert(Status == APFloat::opOK && !LosesInfo &&
13182 "FP conversion should have been exact");
13184 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
13186 SDValue Cmp = DAG.getSetCC(DL,
13187 getSetCCResultType(DAG.getDataLayout(),
13188 *DAG.getContext(), TheVT),
13189 Value, ThreshVal, ISD::SETLT);
13190 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
13191 DAG.getConstant(0, DL, MVT::i32),
13192 DAG.getConstant(0x80000000, DL, MVT::i32));
13193 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
13194 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
13195 *DAG.getContext(), TheVT),
13196 Value, ThreshVal, ISD::SETLT);
13197 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
13200 // FIXME This causes a redundant load/store if the SSE-class value is already
13201 // in memory, such as if it is on the callstack.
13202 if (isScalarFPTypeInSSEReg(TheVT)) {
13203 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
13204 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
13205 MachinePointerInfo::getFixedStack(MF, SSFI), false,
13207 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
13209 Chain, StackSlot, DAG.getValueType(TheVT)
13212 MachineMemOperand *MMO =
13213 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13214 MachineMemOperand::MOLoad, MemSize, MemSize);
13215 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
13216 Chain = Value.getValue(1);
13217 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
13218 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
13221 MachineMemOperand *MMO =
13222 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
13223 MachineMemOperand::MOStore, MemSize, MemSize);
13225 if (UnsignedFixup) {
13227 // Insert the FIST, load its result as two i32's,
13228 // and XOR the high i32 with Adjust.
13230 SDValue FistOps[] = { Chain, Value, StackSlot };
13231 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13232 FistOps, DstTy, MMO);
13234 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
13235 MachinePointerInfo(),
13236 false, false, false, 0);
13237 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
13238 DAG.getConstant(4, DL, PtrVT));
13240 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
13241 MachinePointerInfo(),
13242 false, false, false, 0);
13243 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13245 if (Subtarget->is64Bit()) {
13246 // Join High32 and Low32 into a 64-bit result.
13247 // (High32 << 32) | Low32
13248 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13249 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13250 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13251 DAG.getConstant(32, DL, MVT::i8));
13252 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13253 return std::make_pair(Result, SDValue());
13256 SDValue ResultOps[] = { Low32, High32 };
13258 SDValue pair = IsReplace
13259 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13260 : DAG.getMergeValues(ResultOps, DL);
13261 return std::make_pair(pair, SDValue());
13263 // Build the FP_TO_INT*_IN_MEM
13264 SDValue Ops[] = { Chain, Value, StackSlot };
13265 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13267 return std::make_pair(FIST, StackSlot);
13271 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13272 const X86Subtarget *Subtarget) {
13273 MVT VT = Op->getSimpleValueType(0);
13274 SDValue In = Op->getOperand(0);
13275 MVT InVT = In.getSimpleValueType();
13278 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13279 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13281 // Optimize vectors in AVX mode:
13284 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13285 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13286 // Concat upper and lower parts.
13289 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13290 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13291 // Concat upper and lower parts.
13294 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13295 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13296 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13299 if (Subtarget->hasInt256())
13300 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13302 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13303 SDValue Undef = DAG.getUNDEF(InVT);
13304 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13305 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13306 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13308 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13309 VT.getVectorNumElements()/2);
13311 OpLo = DAG.getBitcast(HVT, OpLo);
13312 OpHi = DAG.getBitcast(HVT, OpHi);
13314 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13317 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13318 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13319 MVT VT = Op->getSimpleValueType(0);
13320 SDValue In = Op->getOperand(0);
13321 MVT InVT = In.getSimpleValueType();
13323 unsigned int NumElts = VT.getVectorNumElements();
13324 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13327 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13328 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13330 assert(InVT.getVectorElementType() == MVT::i1);
13331 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13333 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13335 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13337 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13338 if (VT.is512BitVector())
13340 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13343 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13344 SelectionDAG &DAG) {
13345 if (Subtarget->hasFp256())
13346 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13352 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13353 SelectionDAG &DAG) {
13355 MVT VT = Op.getSimpleValueType();
13356 SDValue In = Op.getOperand(0);
13357 MVT SVT = In.getSimpleValueType();
13359 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13360 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13362 if (Subtarget->hasFp256())
13363 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13366 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13367 VT.getVectorNumElements() != SVT.getVectorNumElements());
13371 static SDValue LowerTruncateVecI1(SDValue Op, SelectionDAG &DAG,
13372 const X86Subtarget *Subtarget) {
13375 MVT VT = Op.getSimpleValueType();
13376 SDValue In = Op.getOperand(0);
13377 MVT InVT = In.getSimpleValueType();
13379 assert(VT.getVectorElementType() == MVT::i1 && "Unexected vector type.");
13381 // Shift LSB to MSB and use VPMOVB2M - SKX.
13382 unsigned ShiftInx = InVT.getScalarSizeInBits() - 1;
13383 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13384 Subtarget->hasBWI()) || // legal, will go to VPMOVB2M, VPMOVW2M
13385 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13386 InVT.getScalarSizeInBits() <= 16 && Subtarget->hasBWI() &&
13387 Subtarget->hasVLX())) { // legal, will go to VPMOVB2M, VPMOVW2M
13388 // Shift packed bytes not supported natively, bitcast to dword
13389 MVT ExtVT = MVT::getVectorVT(MVT::i16, InVT.getSizeInBits()/16);
13390 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, ExtVT,
13391 DAG.getBitcast(ExtVT, In),
13392 DAG.getConstant(ShiftInx, DL, ExtVT));
13393 ShiftNode = DAG.getBitcast(InVT, ShiftNode);
13394 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13396 if ((InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13397 Subtarget->hasDQI()) || // legal, will go to VPMOVD2M, VPMOVQ2M
13398 ((InVT.is256BitVector() || InVT.is128BitVector()) &&
13399 InVT.getScalarSizeInBits() >= 32 && Subtarget->hasDQI() &&
13400 Subtarget->hasVLX())) { // legal, will go to VPMOVD2M, VPMOVQ2M
13402 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13403 DAG.getConstant(ShiftInx, DL, InVT));
13404 return DAG.getNode(X86ISD::CVT2MASK, DL, VT, ShiftNode);
13407 // Shift LSB to MSB, extend if necessary and use TESTM.
13408 unsigned NumElts = InVT.getVectorNumElements();
13409 if (InVT.getSizeInBits() < 512 &&
13410 (InVT.getScalarType() == MVT::i8 || InVT.getScalarType() == MVT::i16 ||
13411 !Subtarget->hasVLX())) {
13412 assert((NumElts == 8 || NumElts == 16) && "Unexected vector type.");
13414 // TESTD/Q should be used (if BW supported we use CVT2MASK above),
13415 // so vector should be extended to packed dword/qword.
13416 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(512/NumElts), NumElts);
13417 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13419 ShiftInx = InVT.getScalarSizeInBits() - 1;
13422 SDValue ShiftNode = DAG.getNode(ISD::SHL, DL, InVT, In,
13423 DAG.getConstant(ShiftInx, DL, InVT));
13424 return DAG.getNode(X86ISD::TESTM, DL, VT, ShiftNode, ShiftNode);
13427 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13429 MVT VT = Op.getSimpleValueType();
13430 SDValue In = Op.getOperand(0);
13431 MVT InVT = In.getSimpleValueType();
13433 if (VT == MVT::i1) {
13434 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13435 "Invalid scalar TRUNCATE operation");
13436 if (InVT.getSizeInBits() >= 32)
13438 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13439 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13441 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13442 "Invalid TRUNCATE operation");
13444 if (VT.getVectorElementType() == MVT::i1)
13445 return LowerTruncateVecI1(Op, DAG, Subtarget);
13447 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13448 if (Subtarget->hasAVX512()) {
13449 // word to byte only under BWI
13450 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13451 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13452 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13453 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13455 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13456 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13457 if (Subtarget->hasInt256()) {
13458 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13459 In = DAG.getBitcast(MVT::v8i32, In);
13460 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13462 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13463 DAG.getIntPtrConstant(0, DL));
13466 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13467 DAG.getIntPtrConstant(0, DL));
13468 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13469 DAG.getIntPtrConstant(2, DL));
13470 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13471 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13472 static const int ShufMask[] = {0, 2, 4, 6};
13473 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13476 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13477 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13478 if (Subtarget->hasInt256()) {
13479 In = DAG.getBitcast(MVT::v32i8, In);
13481 SmallVector<SDValue,32> pshufbMask;
13482 for (unsigned i = 0; i < 2; ++i) {
13483 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13484 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13485 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13486 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13487 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13488 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13489 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13490 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13491 for (unsigned j = 0; j < 8; ++j)
13492 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13494 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13495 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13496 In = DAG.getBitcast(MVT::v4i64, In);
13498 static const int ShufMask[] = {0, 2, -1, -1};
13499 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13501 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13502 DAG.getIntPtrConstant(0, DL));
13503 return DAG.getBitcast(VT, In);
13506 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13507 DAG.getIntPtrConstant(0, DL));
13509 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13510 DAG.getIntPtrConstant(4, DL));
13512 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13513 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13515 // The PSHUFB mask:
13516 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13517 -1, -1, -1, -1, -1, -1, -1, -1};
13519 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13520 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13521 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13523 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13524 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13526 // The MOVLHPS Mask:
13527 static const int ShufMask2[] = {0, 1, 4, 5};
13528 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13529 return DAG.getBitcast(MVT::v8i16, res);
13532 // Handle truncation of V256 to V128 using shuffles.
13533 if (!VT.is128BitVector() || !InVT.is256BitVector())
13536 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13538 unsigned NumElems = VT.getVectorNumElements();
13539 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13541 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13542 // Prepare truncation shuffle mask
13543 for (unsigned i = 0; i != NumElems; ++i)
13544 MaskVec[i] = i * 2;
13545 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13546 DAG.getUNDEF(NVT), &MaskVec[0]);
13547 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13548 DAG.getIntPtrConstant(0, DL));
13551 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13552 SelectionDAG &DAG) const {
13553 assert(!Op.getSimpleValueType().isVector());
13555 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13556 /*IsSigned=*/ true, /*IsReplace=*/ false);
13557 SDValue FIST = Vals.first, StackSlot = Vals.second;
13558 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13559 if (!FIST.getNode())
13562 if (StackSlot.getNode())
13563 // Load the result.
13564 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13565 FIST, StackSlot, MachinePointerInfo(),
13566 false, false, false, 0);
13568 // The node is the result.
13572 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13573 SelectionDAG &DAG) const {
13574 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13575 /*IsSigned=*/ false, /*IsReplace=*/ false);
13576 SDValue FIST = Vals.first, StackSlot = Vals.second;
13577 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13578 if (!FIST.getNode())
13581 if (StackSlot.getNode())
13582 // Load the result.
13583 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13584 FIST, StackSlot, MachinePointerInfo(),
13585 false, false, false, 0);
13587 // The node is the result.
13591 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13593 MVT VT = Op.getSimpleValueType();
13594 SDValue In = Op.getOperand(0);
13595 MVT SVT = In.getSimpleValueType();
13597 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13599 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13600 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13601 In, DAG.getUNDEF(SVT)));
13604 /// The only differences between FABS and FNEG are the mask and the logic op.
13605 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13606 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13607 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13608 "Wrong opcode for lowering FABS or FNEG.");
13610 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13612 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13613 // into an FNABS. We'll lower the FABS after that if it is still in use.
13615 for (SDNode *User : Op->uses())
13616 if (User->getOpcode() == ISD::FNEG)
13620 MVT VT = Op.getSimpleValueType();
13622 bool IsF128 = (VT == MVT::f128);
13624 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13625 // decide if we should generate a 16-byte constant mask when we only need 4 or
13626 // 8 bytes for the scalar case.
13632 if (VT.isVector()) {
13634 EltVT = VT.getVectorElementType();
13635 NumElts = VT.getVectorNumElements();
13636 } else if (IsF128) {
13637 // SSE instructions are used for optimized f128 logical operations.
13638 LogicVT = MVT::f128;
13642 // There are no scalar bitwise logical SSE/AVX instructions, so we
13643 // generate a 16-byte vector constant and logic op even for the scalar case.
13644 // Using a 16-byte mask allows folding the load of the mask with
13645 // the logic op, so it can save (~4 bytes) on code size.
13646 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13648 NumElts = (VT == MVT::f64) ? 2 : 4;
13651 unsigned EltBits = EltVT.getSizeInBits();
13652 LLVMContext *Context = DAG.getContext();
13653 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13655 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13656 Constant *C = ConstantInt::get(*Context, MaskElt);
13657 C = ConstantVector::getSplat(NumElts, C);
13658 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13659 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13660 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13662 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13663 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13664 false, false, false, Alignment);
13666 SDValue Op0 = Op.getOperand(0);
13667 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13669 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13670 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13672 if (VT.isVector() || IsF128)
13673 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13675 // For the scalar case extend to a 128-bit vector, perform the logic op,
13676 // and extract the scalar result back out.
13677 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13678 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13679 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13680 DAG.getIntPtrConstant(0, dl));
13683 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13684 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13685 LLVMContext *Context = DAG.getContext();
13686 SDValue Op0 = Op.getOperand(0);
13687 SDValue Op1 = Op.getOperand(1);
13689 MVT VT = Op.getSimpleValueType();
13690 MVT SrcVT = Op1.getSimpleValueType();
13691 bool IsF128 = (VT == MVT::f128);
13693 // If second operand is smaller, extend it first.
13694 if (SrcVT.bitsLT(VT)) {
13695 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13698 // And if it is bigger, shrink it first.
13699 if (SrcVT.bitsGT(VT)) {
13700 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13704 // At this point the operands and the result should have the same
13705 // type, and that won't be f80 since that is not custom lowered.
13706 assert((VT == MVT::f64 || VT == MVT::f32 || IsF128) &&
13707 "Unexpected type in LowerFCOPYSIGN");
13709 const fltSemantics &Sem =
13710 VT == MVT::f64 ? APFloat::IEEEdouble :
13711 (IsF128 ? APFloat::IEEEquad : APFloat::IEEEsingle);
13712 const unsigned SizeInBits = VT.getSizeInBits();
13714 SmallVector<Constant *, 4> CV(
13715 VT == MVT::f64 ? 2 : (IsF128 ? 1 : 4),
13716 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13718 // First, clear all bits but the sign bit from the second operand (sign).
13719 CV[0] = ConstantFP::get(*Context,
13720 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13721 Constant *C = ConstantVector::get(CV);
13722 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13723 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13725 // Perform all logic operations as 16-byte vectors because there are no
13726 // scalar FP logic instructions in SSE. This allows load folding of the
13727 // constants into the logic instructions.
13728 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : (IsF128 ? MVT::f128 : MVT::v4f32);
13730 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13731 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13732 false, false, false, 16);
13734 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13735 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13737 // Next, clear the sign bit from the first operand (magnitude).
13738 // If it's a constant, we can clear it here.
13739 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13740 APFloat APF = Op0CN->getValueAPF();
13741 // If the magnitude is a positive zero, the sign bit alone is enough.
13742 if (APF.isPosZero())
13743 return IsF128 ? SignBit :
13744 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13745 DAG.getIntPtrConstant(0, dl));
13747 CV[0] = ConstantFP::get(*Context, APF);
13749 CV[0] = ConstantFP::get(
13751 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13753 C = ConstantVector::get(CV);
13754 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13756 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13757 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13758 false, false, false, 16);
13759 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13760 if (!isa<ConstantFPSDNode>(Op0)) {
13762 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13763 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13765 // OR the magnitude value with the sign bit.
13766 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13767 return IsF128 ? Val :
13768 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13769 DAG.getIntPtrConstant(0, dl));
13772 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13773 SDValue N0 = Op.getOperand(0);
13775 MVT VT = Op.getSimpleValueType();
13777 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13778 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13779 DAG.getConstant(1, dl, VT));
13780 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13783 // Check whether an OR'd tree is PTEST-able.
13784 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13785 SelectionDAG &DAG) {
13786 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13788 if (!Subtarget->hasSSE41())
13791 if (!Op->hasOneUse())
13794 SDNode *N = Op.getNode();
13797 SmallVector<SDValue, 8> Opnds;
13798 DenseMap<SDValue, unsigned> VecInMap;
13799 SmallVector<SDValue, 8> VecIns;
13800 EVT VT = MVT::Other;
13802 // Recognize a special case where a vector is casted into wide integer to
13804 Opnds.push_back(N->getOperand(0));
13805 Opnds.push_back(N->getOperand(1));
13807 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13808 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13809 // BFS traverse all OR'd operands.
13810 if (I->getOpcode() == ISD::OR) {
13811 Opnds.push_back(I->getOperand(0));
13812 Opnds.push_back(I->getOperand(1));
13813 // Re-evaluate the number of nodes to be traversed.
13814 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13818 // Quit if a non-EXTRACT_VECTOR_ELT
13819 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13822 // Quit if without a constant index.
13823 SDValue Idx = I->getOperand(1);
13824 if (!isa<ConstantSDNode>(Idx))
13827 SDValue ExtractedFromVec = I->getOperand(0);
13828 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13829 if (M == VecInMap.end()) {
13830 VT = ExtractedFromVec.getValueType();
13831 // Quit if not 128/256-bit vector.
13832 if (!VT.is128BitVector() && !VT.is256BitVector())
13834 // Quit if not the same type.
13835 if (VecInMap.begin() != VecInMap.end() &&
13836 VT != VecInMap.begin()->first.getValueType())
13838 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13839 VecIns.push_back(ExtractedFromVec);
13841 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13844 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13845 "Not extracted from 128-/256-bit vector.");
13847 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13849 for (DenseMap<SDValue, unsigned>::const_iterator
13850 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13851 // Quit if not all elements are used.
13852 if (I->second != FullMask)
13856 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13858 // Cast all vectors into TestVT for PTEST.
13859 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13860 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13862 // If more than one full vectors are evaluated, OR them first before PTEST.
13863 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13864 // Each iteration will OR 2 nodes and append the result until there is only
13865 // 1 node left, i.e. the final OR'd value of all vectors.
13866 SDValue LHS = VecIns[Slot];
13867 SDValue RHS = VecIns[Slot + 1];
13868 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13871 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13872 VecIns.back(), VecIns.back());
13875 /// \brief return true if \c Op has a use that doesn't just read flags.
13876 static bool hasNonFlagsUse(SDValue Op) {
13877 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13879 SDNode *User = *UI;
13880 unsigned UOpNo = UI.getOperandNo();
13881 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13882 // Look pass truncate.
13883 UOpNo = User->use_begin().getOperandNo();
13884 User = *User->use_begin();
13887 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13888 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13894 /// Emit nodes that will be selected as "test Op0,Op0", or something
13896 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13897 SelectionDAG &DAG) const {
13898 if (Op.getValueType() == MVT::i1) {
13899 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13900 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13901 DAG.getConstant(0, dl, MVT::i8));
13903 // CF and OF aren't always set the way we want. Determine which
13904 // of these we need.
13905 bool NeedCF = false;
13906 bool NeedOF = false;
13909 case X86::COND_A: case X86::COND_AE:
13910 case X86::COND_B: case X86::COND_BE:
13913 case X86::COND_G: case X86::COND_GE:
13914 case X86::COND_L: case X86::COND_LE:
13915 case X86::COND_O: case X86::COND_NO: {
13916 // Check if we really need to set the
13917 // Overflow flag. If NoSignedWrap is present
13918 // that is not actually needed.
13919 switch (Op->getOpcode()) {
13924 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13925 if (BinNode->Flags.hasNoSignedWrap())
13935 // See if we can use the EFLAGS value from the operand instead of
13936 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13937 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13938 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13939 // Emit a CMP with 0, which is the TEST pattern.
13940 //if (Op.getValueType() == MVT::i1)
13941 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13942 // DAG.getConstant(0, MVT::i1));
13943 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13944 DAG.getConstant(0, dl, Op.getValueType()));
13946 unsigned Opcode = 0;
13947 unsigned NumOperands = 0;
13949 // Truncate operations may prevent the merge of the SETCC instruction
13950 // and the arithmetic instruction before it. Attempt to truncate the operands
13951 // of the arithmetic instruction and use a reduced bit-width instruction.
13952 bool NeedTruncation = false;
13953 SDValue ArithOp = Op;
13954 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13955 SDValue Arith = Op->getOperand(0);
13956 // Both the trunc and the arithmetic op need to have one user each.
13957 if (Arith->hasOneUse())
13958 switch (Arith.getOpcode()) {
13965 NeedTruncation = true;
13971 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13972 // which may be the result of a CAST. We use the variable 'Op', which is the
13973 // non-casted variable when we check for possible users.
13974 switch (ArithOp.getOpcode()) {
13976 // Due to an isel shortcoming, be conservative if this add is likely to be
13977 // selected as part of a load-modify-store instruction. When the root node
13978 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13979 // uses of other nodes in the match, such as the ADD in this case. This
13980 // leads to the ADD being left around and reselected, with the result being
13981 // two adds in the output. Alas, even if none our users are stores, that
13982 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13983 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13984 // climbing the DAG back to the root, and it doesn't seem to be worth the
13986 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13987 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13988 if (UI->getOpcode() != ISD::CopyToReg &&
13989 UI->getOpcode() != ISD::SETCC &&
13990 UI->getOpcode() != ISD::STORE)
13993 if (ConstantSDNode *C =
13994 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13995 // An add of one will be selected as an INC.
13996 if (C->isOne() && !Subtarget->slowIncDec()) {
13997 Opcode = X86ISD::INC;
14002 // An add of negative one (subtract of one) will be selected as a DEC.
14003 if (C->isAllOnesValue() && !Subtarget->slowIncDec()) {
14004 Opcode = X86ISD::DEC;
14010 // Otherwise use a regular EFLAGS-setting add.
14011 Opcode = X86ISD::ADD;
14016 // If we have a constant logical shift that's only used in a comparison
14017 // against zero turn it into an equivalent AND. This allows turning it into
14018 // a TEST instruction later.
14019 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
14020 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
14021 EVT VT = Op.getValueType();
14022 unsigned BitWidth = VT.getSizeInBits();
14023 unsigned ShAmt = Op->getConstantOperandVal(1);
14024 if (ShAmt >= BitWidth) // Avoid undefined shifts.
14026 APInt Mask = ArithOp.getOpcode() == ISD::SRL
14027 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
14028 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
14029 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
14031 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
14032 DAG.getConstant(Mask, dl, VT));
14033 DAG.ReplaceAllUsesWith(Op, New);
14039 // If the primary and result isn't used, don't bother using X86ISD::AND,
14040 // because a TEST instruction will be better.
14041 if (!hasNonFlagsUse(Op))
14047 // Due to the ISEL shortcoming noted above, be conservative if this op is
14048 // likely to be selected as part of a load-modify-store instruction.
14049 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14050 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14051 if (UI->getOpcode() == ISD::STORE)
14054 // Otherwise use a regular EFLAGS-setting instruction.
14055 switch (ArithOp.getOpcode()) {
14056 default: llvm_unreachable("unexpected operator!");
14057 case ISD::SUB: Opcode = X86ISD::SUB; break;
14058 case ISD::XOR: Opcode = X86ISD::XOR; break;
14059 case ISD::AND: Opcode = X86ISD::AND; break;
14061 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
14062 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
14063 if (EFLAGS.getNode())
14066 Opcode = X86ISD::OR;
14080 return SDValue(Op.getNode(), 1);
14086 // If we found that truncation is beneficial, perform the truncation and
14088 if (NeedTruncation) {
14089 EVT VT = Op.getValueType();
14090 SDValue WideVal = Op->getOperand(0);
14091 EVT WideVT = WideVal.getValueType();
14092 unsigned ConvertedOp = 0;
14093 // Use a target machine opcode to prevent further DAGCombine
14094 // optimizations that may separate the arithmetic operations
14095 // from the setcc node.
14096 switch (WideVal.getOpcode()) {
14098 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
14099 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
14100 case ISD::AND: ConvertedOp = X86ISD::AND; break;
14101 case ISD::OR: ConvertedOp = X86ISD::OR; break;
14102 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
14106 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14107 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
14108 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
14109 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
14110 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
14116 // Emit a CMP with 0, which is the TEST pattern.
14117 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14118 DAG.getConstant(0, dl, Op.getValueType()));
14120 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14121 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
14123 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
14124 DAG.ReplaceAllUsesWith(Op, New);
14125 return SDValue(New.getNode(), 1);
14128 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
14130 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
14131 SDLoc dl, SelectionDAG &DAG) const {
14132 if (isNullConstant(Op1))
14133 return EmitTest(Op0, X86CC, dl, DAG);
14135 assert(!(isa<ConstantSDNode>(Op1) && Op0.getValueType() == MVT::i1) &&
14136 "Unexpected comparison operation for MVT::i1 operands");
14138 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
14139 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
14140 // Do the comparison at i32 if it's smaller, besides the Atom case.
14141 // This avoids subregister aliasing issues. Keep the smaller reference
14142 // if we're optimizing for size, however, as that'll allow better folding
14143 // of memory operations.
14144 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
14145 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
14146 !Subtarget->isAtom()) {
14147 unsigned ExtendOp =
14148 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
14149 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
14150 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
14152 // Use SUB instead of CMP to enable CSE between SUB and CMP.
14153 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
14154 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
14156 return SDValue(Sub.getNode(), 1);
14158 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
14161 /// Convert a comparison if required by the subtarget.
14162 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
14163 SelectionDAG &DAG) const {
14164 // If the subtarget does not support the FUCOMI instruction, floating-point
14165 // comparisons have to be converted.
14166 if (Subtarget->hasCMov() ||
14167 Cmp.getOpcode() != X86ISD::CMP ||
14168 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
14169 !Cmp.getOperand(1).getValueType().isFloatingPoint())
14172 // The instruction selector will select an FUCOM instruction instead of
14173 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
14174 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
14175 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
14177 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
14178 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
14179 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
14180 DAG.getConstant(8, dl, MVT::i8));
14181 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
14183 // Some 64-bit targets lack SAHF support, but they do support FCOMI.
14184 assert(Subtarget->hasLAHFSAHF() && "Target doesn't support SAHF or FCOMI?");
14185 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
14188 /// The minimum architected relative accuracy is 2^-12. We need one
14189 /// Newton-Raphson step to have a good float result (24 bits of precision).
14190 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
14191 DAGCombinerInfo &DCI,
14192 unsigned &RefinementSteps,
14193 bool &UseOneConstNR) const {
14194 EVT VT = Op.getValueType();
14195 const char *RecipOp;
14197 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
14198 // TODO: Add support for AVX512 (v16f32).
14199 // It is likely not profitable to do this for f64 because a double-precision
14200 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
14201 // instructions: convert to single, rsqrtss, convert back to double, refine
14202 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
14203 // along with FMA, this could be a throughput win.
14204 if (VT == MVT::f32 && Subtarget->hasSSE1())
14206 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14207 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14208 RecipOp = "vec-sqrtf";
14212 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14213 if (!Recips.isEnabled(RecipOp))
14216 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14217 UseOneConstNR = false;
14218 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
14221 /// The minimum architected relative accuracy is 2^-12. We need one
14222 /// Newton-Raphson step to have a good float result (24 bits of precision).
14223 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
14224 DAGCombinerInfo &DCI,
14225 unsigned &RefinementSteps) const {
14226 EVT VT = Op.getValueType();
14227 const char *RecipOp;
14229 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
14230 // TODO: Add support for AVX512 (v16f32).
14231 // It is likely not profitable to do this for f64 because a double-precision
14232 // reciprocal estimate with refinement on x86 prior to FMA requires
14233 // 15 instructions: convert to single, rcpss, convert back to double, refine
14234 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
14235 // along with FMA, this could be a throughput win.
14236 if (VT == MVT::f32 && Subtarget->hasSSE1())
14238 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
14239 (VT == MVT::v8f32 && Subtarget->hasAVX()))
14240 RecipOp = "vec-divf";
14244 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
14245 if (!Recips.isEnabled(RecipOp))
14248 RefinementSteps = Recips.getRefinementSteps(RecipOp);
14249 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
14252 /// If we have at least two divisions that use the same divisor, convert to
14253 /// multplication by a reciprocal. This may need to be adjusted for a given
14254 /// CPU if a division's cost is not at least twice the cost of a multiplication.
14255 /// This is because we still need one division to calculate the reciprocal and
14256 /// then we need two multiplies by that reciprocal as replacements for the
14257 /// original divisions.
14258 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
14262 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
14263 /// if it's possible.
14264 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
14265 SDLoc dl, SelectionDAG &DAG) const {
14266 SDValue Op0 = And.getOperand(0);
14267 SDValue Op1 = And.getOperand(1);
14268 if (Op0.getOpcode() == ISD::TRUNCATE)
14269 Op0 = Op0.getOperand(0);
14270 if (Op1.getOpcode() == ISD::TRUNCATE)
14271 Op1 = Op1.getOperand(0);
14274 if (Op1.getOpcode() == ISD::SHL)
14275 std::swap(Op0, Op1);
14276 if (Op0.getOpcode() == ISD::SHL) {
14277 if (isOneConstant(Op0.getOperand(0))) {
14278 // If we looked past a truncate, check that it's only truncating away
14280 unsigned BitWidth = Op0.getValueSizeInBits();
14281 unsigned AndBitWidth = And.getValueSizeInBits();
14282 if (BitWidth > AndBitWidth) {
14284 DAG.computeKnownBits(Op0, Zeros, Ones);
14285 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14289 RHS = Op0.getOperand(1);
14291 } else if (Op1.getOpcode() == ISD::Constant) {
14292 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14293 uint64_t AndRHSVal = AndRHS->getZExtValue();
14294 SDValue AndLHS = Op0;
14296 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14297 LHS = AndLHS.getOperand(0);
14298 RHS = AndLHS.getOperand(1);
14301 // Use BT if the immediate can't be encoded in a TEST instruction.
14302 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14304 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14308 if (LHS.getNode()) {
14309 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14310 // instruction. Since the shift amount is in-range-or-undefined, we know
14311 // that doing a bittest on the i32 value is ok. We extend to i32 because
14312 // the encoding for the i16 version is larger than the i32 version.
14313 // Also promote i16 to i32 for performance / code size reason.
14314 if (LHS.getValueType() == MVT::i8 ||
14315 LHS.getValueType() == MVT::i16)
14316 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14318 // If the operand types disagree, extend the shift amount to match. Since
14319 // BT ignores high bits (like shifts) we can use anyextend.
14320 if (LHS.getValueType() != RHS.getValueType())
14321 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14323 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14324 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14325 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14326 DAG.getConstant(Cond, dl, MVT::i8), BT);
14332 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14334 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14339 // SSE Condition code mapping:
14348 switch (SetCCOpcode) {
14349 default: llvm_unreachable("Unexpected SETCC condition");
14351 case ISD::SETEQ: SSECC = 0; break;
14353 case ISD::SETGT: Swap = true; // Fallthrough
14355 case ISD::SETOLT: SSECC = 1; break;
14357 case ISD::SETGE: Swap = true; // Fallthrough
14359 case ISD::SETOLE: SSECC = 2; break;
14360 case ISD::SETUO: SSECC = 3; break;
14362 case ISD::SETNE: SSECC = 4; break;
14363 case ISD::SETULE: Swap = true; // Fallthrough
14364 case ISD::SETUGE: SSECC = 5; break;
14365 case ISD::SETULT: Swap = true; // Fallthrough
14366 case ISD::SETUGT: SSECC = 6; break;
14367 case ISD::SETO: SSECC = 7; break;
14369 case ISD::SETONE: SSECC = 8; break;
14372 std::swap(Op0, Op1);
14377 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14378 // ones, and then concatenate the result back.
14379 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14380 MVT VT = Op.getSimpleValueType();
14382 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14383 "Unsupported value type for operation");
14385 unsigned NumElems = VT.getVectorNumElements();
14387 SDValue CC = Op.getOperand(2);
14389 // Extract the LHS vectors
14390 SDValue LHS = Op.getOperand(0);
14391 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14392 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14394 // Extract the RHS vectors
14395 SDValue RHS = Op.getOperand(1);
14396 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14397 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14399 // Issue the operation on the smaller types and concatenate the result back
14400 MVT EltVT = VT.getVectorElementType();
14401 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14402 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14403 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14404 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14407 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14408 SDValue Op0 = Op.getOperand(0);
14409 SDValue Op1 = Op.getOperand(1);
14410 SDValue CC = Op.getOperand(2);
14411 MVT VT = Op.getSimpleValueType();
14414 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14415 "Unexpected type for boolean compare operation");
14416 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14417 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14418 DAG.getConstant(-1, dl, VT));
14419 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14420 DAG.getConstant(-1, dl, VT));
14421 switch (SetCCOpcode) {
14422 default: llvm_unreachable("Unexpected SETCC condition");
14424 // (x == y) -> ~(x ^ y)
14425 return DAG.getNode(ISD::XOR, dl, VT,
14426 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14427 DAG.getConstant(-1, dl, VT));
14429 // (x != y) -> (x ^ y)
14430 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14433 // (x > y) -> (x & ~y)
14434 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14437 // (x < y) -> (~x & y)
14438 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14441 // (x <= y) -> (~x | y)
14442 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14445 // (x >=y) -> (x | ~y)
14446 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14450 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14451 const X86Subtarget *Subtarget) {
14452 SDValue Op0 = Op.getOperand(0);
14453 SDValue Op1 = Op.getOperand(1);
14454 SDValue CC = Op.getOperand(2);
14455 MVT VT = Op.getSimpleValueType();
14458 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14459 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14460 "Cannot set masked compare for this operation");
14462 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14464 bool Unsigned = false;
14467 switch (SetCCOpcode) {
14468 default: llvm_unreachable("Unexpected SETCC condition");
14469 case ISD::SETNE: SSECC = 4; break;
14470 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14471 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14472 case ISD::SETLT: Swap = true; //fall-through
14473 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14474 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14475 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14476 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14477 case ISD::SETULE: Unsigned = true; //fall-through
14478 case ISD::SETLE: SSECC = 2; break;
14482 std::swap(Op0, Op1);
14484 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14485 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14486 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14487 DAG.getConstant(SSECC, dl, MVT::i8));
14490 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14491 /// operand \p Op1. If non-trivial (for example because it's not constant)
14492 /// return an empty value.
14493 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14495 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14499 MVT VT = Op1.getSimpleValueType();
14500 MVT EVT = VT.getVectorElementType();
14501 unsigned n = VT.getVectorNumElements();
14502 SmallVector<SDValue, 8> ULTOp1;
14504 for (unsigned i = 0; i < n; ++i) {
14505 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14506 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14509 // Avoid underflow.
14510 APInt Val = Elt->getAPIntValue();
14514 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14517 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14520 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14521 SelectionDAG &DAG) {
14522 SDValue Op0 = Op.getOperand(0);
14523 SDValue Op1 = Op.getOperand(1);
14524 SDValue CC = Op.getOperand(2);
14525 MVT VT = Op.getSimpleValueType();
14526 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14527 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14532 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14533 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14536 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14537 unsigned Opc = X86ISD::CMPP;
14538 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14539 assert(VT.getVectorNumElements() <= 16);
14540 Opc = X86ISD::CMPM;
14542 // In the two special cases we can't handle, emit two comparisons.
14545 unsigned CombineOpc;
14546 if (SetCCOpcode == ISD::SETUEQ) {
14547 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14549 assert(SetCCOpcode == ISD::SETONE);
14550 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14553 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14554 DAG.getConstant(CC0, dl, MVT::i8));
14555 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14556 DAG.getConstant(CC1, dl, MVT::i8));
14557 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14559 // Handle all other FP comparisons here.
14560 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14561 DAG.getConstant(SSECC, dl, MVT::i8));
14564 MVT VTOp0 = Op0.getSimpleValueType();
14565 assert(VTOp0 == Op1.getSimpleValueType() &&
14566 "Expected operands with same type!");
14567 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14568 "Invalid number of packed elements for source and destination!");
14570 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14571 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14572 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14573 // legalizer firstly checks if the first operand in input to the setcc has
14574 // a legal type. If so, then it promotes the return type to that same type.
14575 // Otherwise, the return type is promoted to the 'next legal type' which,
14576 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14578 // We reach this code only if the following two conditions are met:
14579 // 1. Both return type and operand type have been promoted to wider types
14580 // by the type legalizer.
14581 // 2. The original operand type has been promoted to a 256-bit vector.
14583 // Note that condition 2. only applies for AVX targets.
14584 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14585 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14588 // The non-AVX512 code below works under the assumption that source and
14589 // destination types are the same.
14590 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14591 "Value types for source and destination must be the same!");
14593 // Break 256-bit integer vector compare into smaller ones.
14594 if (VT.is256BitVector() && !Subtarget->hasInt256())
14595 return Lower256IntVSETCC(Op, DAG);
14597 MVT OpVT = Op1.getSimpleValueType();
14598 if (OpVT.getVectorElementType() == MVT::i1)
14599 return LowerBoolVSETCC_AVX512(Op, DAG);
14601 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14602 if (Subtarget->hasAVX512()) {
14603 if (Op1.getSimpleValueType().is512BitVector() ||
14604 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14605 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14606 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14608 // In AVX-512 architecture setcc returns mask with i1 elements,
14609 // But there is no compare instruction for i8 and i16 elements in KNL.
14610 // We are not talking about 512-bit operands in this case, these
14611 // types are illegal.
14613 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14614 OpVT.getVectorElementType().getSizeInBits() >= 8))
14615 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14616 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14619 // Lower using XOP integer comparisons.
14620 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14621 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14622 // Translate compare code to XOP PCOM compare mode.
14623 unsigned CmpMode = 0;
14624 switch (SetCCOpcode) {
14625 default: llvm_unreachable("Unexpected SETCC condition");
14627 case ISD::SETLT: CmpMode = 0x00; break;
14629 case ISD::SETLE: CmpMode = 0x01; break;
14631 case ISD::SETGT: CmpMode = 0x02; break;
14633 case ISD::SETGE: CmpMode = 0x03; break;
14634 case ISD::SETEQ: CmpMode = 0x04; break;
14635 case ISD::SETNE: CmpMode = 0x05; break;
14638 // Are we comparing unsigned or signed integers?
14639 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14640 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14642 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14643 DAG.getConstant(CmpMode, dl, MVT::i8));
14646 // We are handling one of the integer comparisons here. Since SSE only has
14647 // GT and EQ comparisons for integer, swapping operands and multiple
14648 // operations may be required for some comparisons.
14650 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14651 bool Subus = false;
14653 switch (SetCCOpcode) {
14654 default: llvm_unreachable("Unexpected SETCC condition");
14655 case ISD::SETNE: Invert = true;
14656 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14657 case ISD::SETLT: Swap = true;
14658 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14659 case ISD::SETGE: Swap = true;
14660 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14661 Invert = true; break;
14662 case ISD::SETULT: Swap = true;
14663 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14664 FlipSigns = true; break;
14665 case ISD::SETUGE: Swap = true;
14666 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14667 FlipSigns = true; Invert = true; break;
14670 // Special case: Use min/max operations for SETULE/SETUGE
14671 MVT VET = VT.getVectorElementType();
14673 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14674 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14677 switch (SetCCOpcode) {
14679 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14680 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14683 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14686 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14687 if (!MinMax && hasSubus) {
14688 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14690 // t = psubus Op0, Op1
14691 // pcmpeq t, <0..0>
14692 switch (SetCCOpcode) {
14694 case ISD::SETULT: {
14695 // If the comparison is against a constant we can turn this into a
14696 // setule. With psubus, setule does not require a swap. This is
14697 // beneficial because the constant in the register is no longer
14698 // destructed as the destination so it can be hoisted out of a loop.
14699 // Only do this pre-AVX since vpcmp* is no longer destructive.
14700 if (Subtarget->hasAVX())
14702 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14703 if (ULEOp1.getNode()) {
14705 Subus = true; Invert = false; Swap = false;
14709 // Psubus is better than flip-sign because it requires no inversion.
14710 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14711 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14715 Opc = X86ISD::SUBUS;
14721 std::swap(Op0, Op1);
14723 // Check that the operation in question is available (most are plain SSE2,
14724 // but PCMPGTQ and PCMPEQQ have different requirements).
14725 if (VT == MVT::v2i64) {
14726 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14727 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14729 // First cast everything to the right type.
14730 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14731 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14733 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14734 // bits of the inputs before performing those operations. The lower
14735 // compare is always unsigned.
14738 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14740 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14741 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14742 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14743 Sign, Zero, Sign, Zero);
14745 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14746 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14748 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14749 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14750 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14752 // Create masks for only the low parts/high parts of the 64 bit integers.
14753 static const int MaskHi[] = { 1, 1, 3, 3 };
14754 static const int MaskLo[] = { 0, 0, 2, 2 };
14755 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14756 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14757 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14759 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14760 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14763 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14765 return DAG.getBitcast(VT, Result);
14768 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14769 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14770 // pcmpeqd + pshufd + pand.
14771 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14773 // First cast everything to the right type.
14774 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14775 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14778 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14780 // Make sure the lower and upper halves are both all-ones.
14781 static const int Mask[] = { 1, 0, 3, 2 };
14782 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14783 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14786 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14788 return DAG.getBitcast(VT, Result);
14792 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14793 // bits of the inputs before performing those operations.
14795 MVT EltVT = VT.getVectorElementType();
14796 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14798 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14799 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14802 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14804 // If the logical-not of the result is required, perform that now.
14806 Result = DAG.getNOT(dl, Result, VT);
14809 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14812 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14813 getZeroVector(VT, Subtarget, DAG, dl));
14818 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14820 MVT VT = Op.getSimpleValueType();
14822 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14824 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14825 && "SetCC type must be 8-bit or 1-bit integer");
14826 SDValue Op0 = Op.getOperand(0);
14827 SDValue Op1 = Op.getOperand(1);
14829 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14831 // Optimize to BT if possible.
14832 // Lower (X & (1 << N)) == 0 to BT(X, N).
14833 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14834 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14835 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14836 isNullConstant(Op1) &&
14837 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14838 if (SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG)) {
14840 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14845 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14847 if ((isOneConstant(Op1) || isNullConstant(Op1)) &&
14848 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14850 // If the input is a setcc, then reuse the input setcc or use a new one with
14851 // the inverted condition.
14852 if (Op0.getOpcode() == X86ISD::SETCC) {
14853 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14854 bool Invert = (CC == ISD::SETNE) ^ isNullConstant(Op1);
14858 CCode = X86::GetOppositeBranchCondition(CCode);
14859 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14860 DAG.getConstant(CCode, dl, MVT::i8),
14861 Op0.getOperand(1));
14863 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14867 if ((Op0.getValueType() == MVT::i1) && isOneConstant(Op1) &&
14868 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14870 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14871 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14874 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14875 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14876 if (X86CC == X86::COND_INVALID)
14879 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14880 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14881 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14882 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14884 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14888 SDValue X86TargetLowering::LowerSETCCE(SDValue Op, SelectionDAG &DAG) const {
14889 SDValue LHS = Op.getOperand(0);
14890 SDValue RHS = Op.getOperand(1);
14891 SDValue Carry = Op.getOperand(2);
14892 SDValue Cond = Op.getOperand(3);
14895 assert(LHS.getSimpleValueType().isInteger() && "SETCCE is integer only.");
14896 X86::CondCode CC = TranslateIntegerX86CC(cast<CondCodeSDNode>(Cond)->get());
14898 assert(Carry.getOpcode() != ISD::CARRY_FALSE);
14899 SDVTList VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14900 SDValue Cmp = DAG.getNode(X86ISD::SBB, DL, VTs, LHS, RHS, Carry);
14901 return DAG.getNode(X86ISD::SETCC, DL, Op.getValueType(),
14902 DAG.getConstant(CC, DL, MVT::i8), Cmp.getValue(1));
14905 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14906 static bool isX86LogicalCmp(SDValue Op) {
14907 unsigned Opc = Op.getNode()->getOpcode();
14908 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14909 Opc == X86ISD::SAHF)
14911 if (Op.getResNo() == 1 &&
14912 (Opc == X86ISD::ADD ||
14913 Opc == X86ISD::SUB ||
14914 Opc == X86ISD::ADC ||
14915 Opc == X86ISD::SBB ||
14916 Opc == X86ISD::SMUL ||
14917 Opc == X86ISD::UMUL ||
14918 Opc == X86ISD::INC ||
14919 Opc == X86ISD::DEC ||
14920 Opc == X86ISD::OR ||
14921 Opc == X86ISD::XOR ||
14922 Opc == X86ISD::AND))
14925 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14931 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14932 if (V.getOpcode() != ISD::TRUNCATE)
14935 SDValue VOp0 = V.getOperand(0);
14936 unsigned InBits = VOp0.getValueSizeInBits();
14937 unsigned Bits = V.getValueSizeInBits();
14938 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14941 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14942 bool addTest = true;
14943 SDValue Cond = Op.getOperand(0);
14944 SDValue Op1 = Op.getOperand(1);
14945 SDValue Op2 = Op.getOperand(2);
14947 MVT VT = Op1.getSimpleValueType();
14950 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14951 // are available or VBLENDV if AVX is available.
14952 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14953 if (Cond.getOpcode() == ISD::SETCC &&
14954 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14955 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14956 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14957 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14958 int SSECC = translateX86FSETCC(
14959 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14962 if (Subtarget->hasAVX512()) {
14963 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14964 DAG.getConstant(SSECC, DL, MVT::i8));
14965 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14968 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14969 DAG.getConstant(SSECC, DL, MVT::i8));
14971 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14972 // of 3 logic instructions for size savings and potentially speed.
14973 // Unfortunately, there is no scalar form of VBLENDV.
14975 // If either operand is a constant, don't try this. We can expect to
14976 // optimize away at least one of the logic instructions later in that
14977 // case, so that sequence would be faster than a variable blend.
14979 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14980 // uses XMM0 as the selection register. That may need just as many
14981 // instructions as the AND/ANDN/OR sequence due to register moves, so
14984 if (Subtarget->hasAVX() &&
14985 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14987 // Convert to vectors, do a VSELECT, and convert back to scalar.
14988 // All of the conversions should be optimized away.
14990 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14991 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14992 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14993 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14995 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14996 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14998 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
15000 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
15001 VSel, DAG.getIntPtrConstant(0, DL));
15003 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
15004 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
15005 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
15009 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
15011 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
15012 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
15013 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
15014 Op1Scalar = Op1.getOperand(0);
15016 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
15017 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
15018 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
15019 Op2Scalar = Op2.getOperand(0);
15020 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
15021 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
15022 Op1Scalar.getValueType(),
15023 Cond, Op1Scalar, Op2Scalar);
15024 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
15025 return DAG.getBitcast(VT, newSelect);
15026 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
15027 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
15028 DAG.getIntPtrConstant(0, DL));
15032 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
15033 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
15034 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15035 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
15036 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
15037 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
15038 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
15040 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
15043 if (Cond.getOpcode() == ISD::SETCC) {
15044 SDValue NewCond = LowerSETCC(Cond, DAG);
15045 if (NewCond.getNode())
15049 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
15050 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
15051 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
15052 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
15053 if (Cond.getOpcode() == X86ISD::SETCC &&
15054 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
15055 isNullConstant(Cond.getOperand(1).getOperand(1))) {
15056 SDValue Cmp = Cond.getOperand(1);
15058 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
15060 if ((isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15061 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
15062 SDValue Y = isAllOnesConstant(Op2) ? Op1 : Op2;
15064 SDValue CmpOp0 = Cmp.getOperand(0);
15065 // Apply further optimizations for special cases
15066 // (select (x != 0), -1, 0) -> neg & sbb
15067 // (select (x == 0), 0, -1) -> neg & sbb
15068 if (isNullConstant(Y) &&
15069 (isAllOnesConstant(Op1) == (CondCode == X86::COND_NE))) {
15070 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
15071 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
15072 DAG.getConstant(0, DL,
15073 CmpOp0.getValueType()),
15075 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15076 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15077 SDValue(Neg.getNode(), 1));
15081 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
15082 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
15083 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15085 SDValue Res = // Res = 0 or -1.
15086 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15087 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
15089 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_E))
15090 Res = DAG.getNOT(DL, Res, Res.getValueType());
15092 if (!isNullConstant(Op2))
15093 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
15098 // Look past (and (setcc_carry (cmp ...)), 1).
15099 if (Cond.getOpcode() == ISD::AND &&
15100 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15101 isOneConstant(Cond.getOperand(1)))
15102 Cond = Cond.getOperand(0);
15104 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15105 // setting operand in place of the X86ISD::SETCC.
15106 unsigned CondOpcode = Cond.getOpcode();
15107 if (CondOpcode == X86ISD::SETCC ||
15108 CondOpcode == X86ISD::SETCC_CARRY) {
15109 CC = Cond.getOperand(0);
15111 SDValue Cmp = Cond.getOperand(1);
15112 unsigned Opc = Cmp.getOpcode();
15113 MVT VT = Op.getSimpleValueType();
15115 bool IllegalFPCMov = false;
15116 if (VT.isFloatingPoint() && !VT.isVector() &&
15117 !isScalarFPTypeInSSEReg(VT)) // FPStack?
15118 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
15120 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
15121 Opc == X86ISD::BT) { // FIXME
15125 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15126 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15127 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15128 Cond.getOperand(0).getValueType() != MVT::i8)) {
15129 SDValue LHS = Cond.getOperand(0);
15130 SDValue RHS = Cond.getOperand(1);
15131 unsigned X86Opcode;
15134 switch (CondOpcode) {
15135 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15136 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15137 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15138 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15139 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15140 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15141 default: llvm_unreachable("unexpected overflowing operator");
15143 if (CondOpcode == ISD::UMULO)
15144 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15147 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15149 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
15151 if (CondOpcode == ISD::UMULO)
15152 Cond = X86Op.getValue(2);
15154 Cond = X86Op.getValue(1);
15156 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
15161 // Look past the truncate if the high bits are known zero.
15162 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15163 Cond = Cond.getOperand(0);
15165 // We know the result of AND is compared against zero. Try to match
15167 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15168 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG)) {
15169 CC = NewSetCC.getOperand(0);
15170 Cond = NewSetCC.getOperand(1);
15177 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
15178 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
15181 // a < b ? -1 : 0 -> RES = ~setcc_carry
15182 // a < b ? 0 : -1 -> RES = setcc_carry
15183 // a >= b ? -1 : 0 -> RES = setcc_carry
15184 // a >= b ? 0 : -1 -> RES = ~setcc_carry
15185 if (Cond.getOpcode() == X86ISD::SUB) {
15186 Cond = ConvertCmpIfNecessary(Cond, DAG);
15187 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
15189 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
15190 (isAllOnesConstant(Op1) || isAllOnesConstant(Op2)) &&
15191 (isNullConstant(Op1) || isNullConstant(Op2))) {
15192 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15193 DAG.getConstant(X86::COND_B, DL, MVT::i8),
15195 if (isAllOnesConstant(Op1) != (CondCode == X86::COND_B))
15196 return DAG.getNOT(DL, Res, Res.getValueType());
15201 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15202 // widen the cmov and push the truncate through. This avoids introducing a new
15203 // branch during isel and doesn't add any extensions.
15204 if (Op.getValueType() == MVT::i8 &&
15205 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15206 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15207 if (T1.getValueType() == T2.getValueType() &&
15208 // Blacklist CopyFromReg to avoid partial register stalls.
15209 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15210 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15211 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15212 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15216 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15217 // condition is true.
15218 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15219 SDValue Ops[] = { Op2, Op1, CC, Cond };
15220 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15223 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
15224 const X86Subtarget *Subtarget,
15225 SelectionDAG &DAG) {
15226 MVT VT = Op->getSimpleValueType(0);
15227 SDValue In = Op->getOperand(0);
15228 MVT InVT = In.getSimpleValueType();
15229 MVT VTElt = VT.getVectorElementType();
15230 MVT InVTElt = InVT.getVectorElementType();
15234 if ((InVTElt == MVT::i1) &&
15235 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15236 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15238 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15239 VTElt.getSizeInBits() <= 16)) ||
15241 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15242 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15244 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15245 VTElt.getSizeInBits() >= 32))))
15246 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15248 unsigned int NumElts = VT.getVectorNumElements();
15250 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
15253 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15254 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15255 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15256 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15259 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15260 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
15262 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
15265 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
15267 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
15268 if (VT.is512BitVector())
15270 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
15273 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
15274 const X86Subtarget *Subtarget,
15275 SelectionDAG &DAG) {
15276 SDValue In = Op->getOperand(0);
15277 MVT VT = Op->getSimpleValueType(0);
15278 MVT InVT = In.getSimpleValueType();
15279 assert(VT.getSizeInBits() == InVT.getSizeInBits());
15281 MVT InSVT = InVT.getVectorElementType();
15282 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
15284 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
15286 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15291 // SSE41 targets can use the pmovsx* instructions directly.
15292 if (Subtarget->hasSSE41())
15293 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15295 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15299 // As SRAI is only available on i16/i32 types, we expand only up to i32
15300 // and handle i64 separately.
15301 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
15302 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15303 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15304 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15305 Curr = DAG.getBitcast(CurrVT, Curr);
15308 SDValue SignExt = Curr;
15309 if (CurrVT != InVT) {
15310 unsigned SignExtShift =
15311 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15312 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15313 DAG.getConstant(SignExtShift, dl, MVT::i8));
15319 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15320 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15321 DAG.getConstant(31, dl, MVT::i8));
15322 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15323 return DAG.getBitcast(VT, Ext);
15329 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15330 SelectionDAG &DAG) {
15331 MVT VT = Op->getSimpleValueType(0);
15332 SDValue In = Op->getOperand(0);
15333 MVT InVT = In.getSimpleValueType();
15336 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15337 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15339 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15340 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15341 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15344 if (Subtarget->hasInt256())
15345 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15347 // Optimize vectors in AVX mode
15348 // Sign extend v8i16 to v8i32 and
15351 // Divide input vector into two parts
15352 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15353 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15354 // concat the vectors to original VT
15356 unsigned NumElems = InVT.getVectorNumElements();
15357 SDValue Undef = DAG.getUNDEF(InVT);
15359 SmallVector<int,8> ShufMask1(NumElems, -1);
15360 for (unsigned i = 0; i != NumElems/2; ++i)
15363 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15365 SmallVector<int,8> ShufMask2(NumElems, -1);
15366 for (unsigned i = 0; i != NumElems/2; ++i)
15367 ShufMask2[i] = i + NumElems/2;
15369 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15371 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15372 VT.getVectorNumElements()/2);
15374 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15375 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15377 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15380 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15381 // may emit an illegal shuffle but the expansion is still better than scalar
15382 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15383 // we'll emit a shuffle and a arithmetic shift.
15384 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15385 // TODO: It is possible to support ZExt by zeroing the undef values during
15386 // the shuffle phase or after the shuffle.
15387 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15388 SelectionDAG &DAG) {
15389 MVT RegVT = Op.getSimpleValueType();
15390 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15391 assert(RegVT.isInteger() &&
15392 "We only custom lower integer vector sext loads.");
15394 // Nothing useful we can do without SSE2 shuffles.
15395 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15397 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15399 EVT MemVT = Ld->getMemoryVT();
15400 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15401 unsigned RegSz = RegVT.getSizeInBits();
15403 ISD::LoadExtType Ext = Ld->getExtensionType();
15405 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15406 && "Only anyext and sext are currently implemented.");
15407 assert(MemVT != RegVT && "Cannot extend to the same type");
15408 assert(MemVT.isVector() && "Must load a vector from memory");
15410 unsigned NumElems = RegVT.getVectorNumElements();
15411 unsigned MemSz = MemVT.getSizeInBits();
15412 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15414 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15415 // The only way in which we have a legal 256-bit vector result but not the
15416 // integer 256-bit operations needed to directly lower a sextload is if we
15417 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15418 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15419 // correctly legalized. We do this late to allow the canonical form of
15420 // sextload to persist throughout the rest of the DAG combiner -- it wants
15421 // to fold together any extensions it can, and so will fuse a sign_extend
15422 // of an sextload into a sextload targeting a wider value.
15424 if (MemSz == 128) {
15425 // Just switch this to a normal load.
15426 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15427 "it must be a legal 128-bit vector "
15429 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15430 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15431 Ld->isInvariant(), Ld->getAlignment());
15433 assert(MemSz < 128 &&
15434 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15435 // Do an sext load to a 128-bit vector type. We want to use the same
15436 // number of elements, but elements half as wide. This will end up being
15437 // recursively lowered by this routine, but will succeed as we definitely
15438 // have all the necessary features if we're using AVX1.
15440 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15441 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15443 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15444 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15445 Ld->isNonTemporal(), Ld->isInvariant(),
15446 Ld->getAlignment());
15449 // Replace chain users with the new chain.
15450 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15451 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15453 // Finally, do a normal sign-extend to the desired register.
15454 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15457 // All sizes must be a power of two.
15458 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15459 "Non-power-of-two elements are not custom lowered!");
15461 // Attempt to load the original value using scalar loads.
15462 // Find the largest scalar type that divides the total loaded size.
15463 MVT SclrLoadTy = MVT::i8;
15464 for (MVT Tp : MVT::integer_valuetypes()) {
15465 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15470 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15471 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15473 SclrLoadTy = MVT::f64;
15475 // Calculate the number of scalar loads that we need to perform
15476 // in order to load our vector from memory.
15477 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15479 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15480 "Can only lower sext loads with a single scalar load!");
15482 unsigned loadRegZize = RegSz;
15483 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15486 // Represent our vector as a sequence of elements which are the
15487 // largest scalar that we can load.
15488 EVT LoadUnitVecVT = EVT::getVectorVT(
15489 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15491 // Represent the data using the same element type that is stored in
15492 // memory. In practice, we ''widen'' MemVT.
15494 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15495 loadRegZize / MemVT.getScalarSizeInBits());
15497 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15498 "Invalid vector type");
15500 // We can't shuffle using an illegal type.
15501 assert(TLI.isTypeLegal(WideVecVT) &&
15502 "We only lower types that form legal widened vector types");
15504 SmallVector<SDValue, 8> Chains;
15505 SDValue Ptr = Ld->getBasePtr();
15506 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15507 TLI.getPointerTy(DAG.getDataLayout()));
15508 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15510 for (unsigned i = 0; i < NumLoads; ++i) {
15511 // Perform a single load.
15512 SDValue ScalarLoad =
15513 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15514 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15515 Ld->getAlignment());
15516 Chains.push_back(ScalarLoad.getValue(1));
15517 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15518 // another round of DAGCombining.
15520 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15522 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15523 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15525 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15528 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15530 // Bitcast the loaded value to a vector of the original element type, in
15531 // the size of the target vector type.
15532 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15533 unsigned SizeRatio = RegSz / MemSz;
15535 if (Ext == ISD::SEXTLOAD) {
15536 // If we have SSE4.1, we can directly emit a VSEXT node.
15537 if (Subtarget->hasSSE41()) {
15538 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15539 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15543 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15545 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15546 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15548 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15549 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15553 // Redistribute the loaded elements into the different locations.
15554 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15555 for (unsigned i = 0; i != NumElems; ++i)
15556 ShuffleVec[i * SizeRatio] = i;
15558 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15559 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15561 // Bitcast to the requested type.
15562 Shuff = DAG.getBitcast(RegVT, Shuff);
15563 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15567 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15568 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15569 // from the AND / OR.
15570 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15571 Opc = Op.getOpcode();
15572 if (Opc != ISD::OR && Opc != ISD::AND)
15574 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15575 Op.getOperand(0).hasOneUse() &&
15576 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15577 Op.getOperand(1).hasOneUse());
15580 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15581 // 1 and that the SETCC node has a single use.
15582 static bool isXor1OfSetCC(SDValue Op) {
15583 if (Op.getOpcode() != ISD::XOR)
15585 if (isOneConstant(Op.getOperand(1)))
15586 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15587 Op.getOperand(0).hasOneUse();
15591 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15592 bool addTest = true;
15593 SDValue Chain = Op.getOperand(0);
15594 SDValue Cond = Op.getOperand(1);
15595 SDValue Dest = Op.getOperand(2);
15598 bool Inverted = false;
15600 if (Cond.getOpcode() == ISD::SETCC) {
15601 // Check for setcc([su]{add,sub,mul}o == 0).
15602 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15603 isNullConstant(Cond.getOperand(1)) &&
15604 Cond.getOperand(0).getResNo() == 1 &&
15605 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15606 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15607 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15608 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15609 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15610 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15612 Cond = Cond.getOperand(0);
15614 SDValue NewCond = LowerSETCC(Cond, DAG);
15615 if (NewCond.getNode())
15620 // FIXME: LowerXALUO doesn't handle these!!
15621 else if (Cond.getOpcode() == X86ISD::ADD ||
15622 Cond.getOpcode() == X86ISD::SUB ||
15623 Cond.getOpcode() == X86ISD::SMUL ||
15624 Cond.getOpcode() == X86ISD::UMUL)
15625 Cond = LowerXALUO(Cond, DAG);
15628 // Look pass (and (setcc_carry (cmp ...)), 1).
15629 if (Cond.getOpcode() == ISD::AND &&
15630 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY &&
15631 isOneConstant(Cond.getOperand(1)))
15632 Cond = Cond.getOperand(0);
15634 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15635 // setting operand in place of the X86ISD::SETCC.
15636 unsigned CondOpcode = Cond.getOpcode();
15637 if (CondOpcode == X86ISD::SETCC ||
15638 CondOpcode == X86ISD::SETCC_CARRY) {
15639 CC = Cond.getOperand(0);
15641 SDValue Cmp = Cond.getOperand(1);
15642 unsigned Opc = Cmp.getOpcode();
15643 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15644 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15648 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15652 // These can only come from an arithmetic instruction with overflow,
15653 // e.g. SADDO, UADDO.
15654 Cond = Cond.getNode()->getOperand(1);
15660 CondOpcode = Cond.getOpcode();
15661 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15662 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15663 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15664 Cond.getOperand(0).getValueType() != MVT::i8)) {
15665 SDValue LHS = Cond.getOperand(0);
15666 SDValue RHS = Cond.getOperand(1);
15667 unsigned X86Opcode;
15670 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15671 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15673 switch (CondOpcode) {
15674 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15676 if (isOneConstant(RHS)) {
15677 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15680 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15681 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15683 if (isOneConstant(RHS)) {
15684 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15687 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15688 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15689 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15690 default: llvm_unreachable("unexpected overflowing operator");
15693 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15694 if (CondOpcode == ISD::UMULO)
15695 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15698 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15700 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15702 if (CondOpcode == ISD::UMULO)
15703 Cond = X86Op.getValue(2);
15705 Cond = X86Op.getValue(1);
15707 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15711 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15712 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15713 if (CondOpc == ISD::OR) {
15714 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15715 // two branches instead of an explicit OR instruction with a
15717 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15718 isX86LogicalCmp(Cmp)) {
15719 CC = Cond.getOperand(0).getOperand(0);
15720 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15721 Chain, Dest, CC, Cmp);
15722 CC = Cond.getOperand(1).getOperand(0);
15726 } else { // ISD::AND
15727 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15728 // two branches instead of an explicit AND instruction with a
15729 // separate test. However, we only do this if this block doesn't
15730 // have a fall-through edge, because this requires an explicit
15731 // jmp when the condition is false.
15732 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15733 isX86LogicalCmp(Cmp) &&
15734 Op.getNode()->hasOneUse()) {
15735 X86::CondCode CCode =
15736 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15737 CCode = X86::GetOppositeBranchCondition(CCode);
15738 CC = DAG.getConstant(CCode, dl, MVT::i8);
15739 SDNode *User = *Op.getNode()->use_begin();
15740 // Look for an unconditional branch following this conditional branch.
15741 // We need this because we need to reverse the successors in order
15742 // to implement FCMP_OEQ.
15743 if (User->getOpcode() == ISD::BR) {
15744 SDValue FalseBB = User->getOperand(1);
15746 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15747 assert(NewBR == User);
15751 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15752 Chain, Dest, CC, Cmp);
15753 X86::CondCode CCode =
15754 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15755 CCode = X86::GetOppositeBranchCondition(CCode);
15756 CC = DAG.getConstant(CCode, dl, MVT::i8);
15762 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15763 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15764 // It should be transformed during dag combiner except when the condition
15765 // is set by a arithmetics with overflow node.
15766 X86::CondCode CCode =
15767 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15768 CCode = X86::GetOppositeBranchCondition(CCode);
15769 CC = DAG.getConstant(CCode, dl, MVT::i8);
15770 Cond = Cond.getOperand(0).getOperand(1);
15772 } else if (Cond.getOpcode() == ISD::SETCC &&
15773 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15774 // For FCMP_OEQ, we can emit
15775 // two branches instead of an explicit AND instruction with a
15776 // separate test. However, we only do this if this block doesn't
15777 // have a fall-through edge, because this requires an explicit
15778 // jmp when the condition is false.
15779 if (Op.getNode()->hasOneUse()) {
15780 SDNode *User = *Op.getNode()->use_begin();
15781 // Look for an unconditional branch following this conditional branch.
15782 // We need this because we need to reverse the successors in order
15783 // to implement FCMP_OEQ.
15784 if (User->getOpcode() == ISD::BR) {
15785 SDValue FalseBB = User->getOperand(1);
15787 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15788 assert(NewBR == User);
15792 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15793 Cond.getOperand(0), Cond.getOperand(1));
15794 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15795 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15796 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15797 Chain, Dest, CC, Cmp);
15798 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15803 } else if (Cond.getOpcode() == ISD::SETCC &&
15804 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15805 // For FCMP_UNE, we can emit
15806 // two branches instead of an explicit AND instruction with a
15807 // separate test. However, we only do this if this block doesn't
15808 // have a fall-through edge, because this requires an explicit
15809 // jmp when the condition is false.
15810 if (Op.getNode()->hasOneUse()) {
15811 SDNode *User = *Op.getNode()->use_begin();
15812 // Look for an unconditional branch following this conditional branch.
15813 // We need this because we need to reverse the successors in order
15814 // to implement FCMP_UNE.
15815 if (User->getOpcode() == ISD::BR) {
15816 SDValue FalseBB = User->getOperand(1);
15818 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15819 assert(NewBR == User);
15822 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15823 Cond.getOperand(0), Cond.getOperand(1));
15824 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15825 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15826 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15827 Chain, Dest, CC, Cmp);
15828 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15838 // Look pass the truncate if the high bits are known zero.
15839 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15840 Cond = Cond.getOperand(0);
15842 // We know the result of AND is compared against zero. Try to match
15844 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15845 if (SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG)) {
15846 CC = NewSetCC.getOperand(0);
15847 Cond = NewSetCC.getOperand(1);
15854 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15855 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15856 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15858 Cond = ConvertCmpIfNecessary(Cond, DAG);
15859 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15860 Chain, Dest, CC, Cond);
15863 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15864 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15865 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15866 // that the guard pages used by the OS virtual memory manager are allocated in
15867 // correct sequence.
15869 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15870 SelectionDAG &DAG) const {
15871 MachineFunction &MF = DAG.getMachineFunction();
15872 bool SplitStack = MF.shouldSplitStack();
15873 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15878 SDNode *Node = Op.getNode();
15879 SDValue Chain = Op.getOperand(0);
15880 SDValue Size = Op.getOperand(1);
15881 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15882 EVT VT = Node->getValueType(0);
15884 // Chain the dynamic stack allocation so that it doesn't modify the stack
15885 // pointer when other instructions are using the stack.
15886 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
15888 bool Is64Bit = Subtarget->is64Bit();
15889 MVT SPTy = getPointerTy(DAG.getDataLayout());
15893 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15894 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15895 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15896 " not tell us which reg is the stack pointer!");
15897 EVT VT = Node->getValueType(0);
15898 SDValue Tmp3 = Node->getOperand(2);
15900 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15901 Chain = SP.getValue(1);
15902 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15903 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15904 unsigned StackAlign = TFI.getStackAlignment();
15905 Result = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15906 if (Align > StackAlign)
15907 Result = DAG.getNode(ISD::AND, dl, VT, Result,
15908 DAG.getConstant(-(uint64_t)Align, dl, VT));
15909 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Result); // Output chain
15910 } else if (SplitStack) {
15911 MachineRegisterInfo &MRI = MF.getRegInfo();
15914 // The 64 bit implementation of segmented stacks needs to clobber both r10
15915 // r11. This makes it impossible to use it along with nested parameters.
15916 const Function *F = MF.getFunction();
15918 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15920 if (I->hasNestAttr())
15921 report_fatal_error("Cannot use segmented stacks with functions that "
15922 "have nested arguments.");
15925 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15926 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15927 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15928 Result = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15929 DAG.getRegister(Vreg, SPTy));
15932 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15934 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15935 Flag = Chain.getValue(1);
15936 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15938 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15940 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15941 unsigned SPReg = RegInfo->getStackRegister();
15942 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15943 Chain = SP.getValue(1);
15946 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15947 DAG.getConstant(-(uint64_t)Align, dl, VT));
15948 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15954 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15955 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
15957 SDValue Ops[2] = {Result, Chain};
15958 return DAG.getMergeValues(Ops, dl);
15961 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15962 MachineFunction &MF = DAG.getMachineFunction();
15963 auto PtrVT = getPointerTy(MF.getDataLayout());
15964 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15966 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15969 if (!Subtarget->is64Bit() ||
15970 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15971 // vastart just stores the address of the VarArgsFrameIndex slot into the
15972 // memory location argument.
15973 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15974 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15975 MachinePointerInfo(SV), false, false, 0);
15979 // gp_offset (0 - 6 * 8)
15980 // fp_offset (48 - 48 + 8 * 16)
15981 // overflow_arg_area (point to parameters coming in memory).
15983 SmallVector<SDValue, 8> MemOps;
15984 SDValue FIN = Op.getOperand(1);
15986 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15987 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15989 FIN, MachinePointerInfo(SV), false, false, 0);
15990 MemOps.push_back(Store);
15993 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15994 Store = DAG.getStore(Op.getOperand(0), DL,
15995 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15997 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15998 MemOps.push_back(Store);
16000 // Store ptr to overflow_arg_area
16001 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
16002 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
16003 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
16004 MachinePointerInfo(SV, 8),
16006 MemOps.push_back(Store);
16008 // Store ptr to reg_save_area.
16009 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
16010 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
16011 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
16012 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
16013 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
16014 MemOps.push_back(Store);
16015 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
16018 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
16019 assert(Subtarget->is64Bit() &&
16020 "LowerVAARG only handles 64-bit va_arg!");
16021 assert(Op.getNode()->getNumOperands() == 4);
16023 MachineFunction &MF = DAG.getMachineFunction();
16024 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
16025 // The Win64 ABI uses char* instead of a structure.
16026 return DAG.expandVAArg(Op.getNode());
16028 SDValue Chain = Op.getOperand(0);
16029 SDValue SrcPtr = Op.getOperand(1);
16030 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16031 unsigned Align = Op.getConstantOperandVal(3);
16034 EVT ArgVT = Op.getNode()->getValueType(0);
16035 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16036 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
16039 // Decide which area this value should be read from.
16040 // TODO: Implement the AMD64 ABI in its entirety. This simple
16041 // selection mechanism works only for the basic types.
16042 if (ArgVT == MVT::f80) {
16043 llvm_unreachable("va_arg for f80 not yet implemented");
16044 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
16045 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
16046 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
16047 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
16049 llvm_unreachable("Unhandled argument type in LowerVAARG");
16052 if (ArgMode == 2) {
16053 // Sanity Check: Make sure using fp_offset makes sense.
16054 assert(!Subtarget->useSoftFloat() &&
16055 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
16056 Subtarget->hasSSE1());
16059 // Insert VAARG_64 node into the DAG
16060 // VAARG_64 returns two values: Variable Argument Address, Chain
16061 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
16062 DAG.getConstant(ArgMode, dl, MVT::i8),
16063 DAG.getConstant(Align, dl, MVT::i32)};
16064 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
16065 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
16066 VTs, InstOps, MVT::i64,
16067 MachinePointerInfo(SV),
16069 /*Volatile=*/false,
16071 /*WriteMem=*/true);
16072 Chain = VAARG.getValue(1);
16074 // Load the next argument and return it
16075 return DAG.getLoad(ArgVT, dl,
16078 MachinePointerInfo(),
16079 false, false, false, 0);
16082 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
16083 SelectionDAG &DAG) {
16084 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
16085 // where a va_list is still an i8*.
16086 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
16087 if (Subtarget->isCallingConvWin64(
16088 DAG.getMachineFunction().getFunction()->getCallingConv()))
16089 // Probably a Win64 va_copy.
16090 return DAG.expandVACopy(Op.getNode());
16092 SDValue Chain = Op.getOperand(0);
16093 SDValue DstPtr = Op.getOperand(1);
16094 SDValue SrcPtr = Op.getOperand(2);
16095 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
16096 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16099 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
16100 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
16102 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
16105 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
16106 // amount is a constant. Takes immediate version of shift as input.
16107 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
16108 SDValue SrcOp, uint64_t ShiftAmt,
16109 SelectionDAG &DAG) {
16110 MVT ElementType = VT.getVectorElementType();
16112 // Fold this packed shift into its first operand if ShiftAmt is 0.
16116 // Check for ShiftAmt >= element width
16117 if (ShiftAmt >= ElementType.getSizeInBits()) {
16118 if (Opc == X86ISD::VSRAI)
16119 ShiftAmt = ElementType.getSizeInBits() - 1;
16121 return DAG.getConstant(0, dl, VT);
16124 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
16125 && "Unknown target vector shift-by-constant node");
16127 // Fold this packed vector shift into a build vector if SrcOp is a
16128 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
16129 if (VT == SrcOp.getSimpleValueType() &&
16130 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
16131 SmallVector<SDValue, 8> Elts;
16132 unsigned NumElts = SrcOp->getNumOperands();
16133 ConstantSDNode *ND;
16136 default: llvm_unreachable(nullptr);
16137 case X86ISD::VSHLI:
16138 for (unsigned i=0; i!=NumElts; ++i) {
16139 SDValue CurrentOp = SrcOp->getOperand(i);
16140 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16141 Elts.push_back(CurrentOp);
16144 ND = cast<ConstantSDNode>(CurrentOp);
16145 const APInt &C = ND->getAPIntValue();
16146 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
16149 case X86ISD::VSRLI:
16150 for (unsigned i=0; i!=NumElts; ++i) {
16151 SDValue CurrentOp = SrcOp->getOperand(i);
16152 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16153 Elts.push_back(CurrentOp);
16156 ND = cast<ConstantSDNode>(CurrentOp);
16157 const APInt &C = ND->getAPIntValue();
16158 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
16161 case X86ISD::VSRAI:
16162 for (unsigned i=0; i!=NumElts; ++i) {
16163 SDValue CurrentOp = SrcOp->getOperand(i);
16164 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16165 Elts.push_back(CurrentOp);
16168 ND = cast<ConstantSDNode>(CurrentOp);
16169 const APInt &C = ND->getAPIntValue();
16170 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
16175 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16178 return DAG.getNode(Opc, dl, VT, SrcOp,
16179 DAG.getConstant(ShiftAmt, dl, MVT::i8));
16182 // getTargetVShiftNode - Handle vector element shifts where the shift amount
16183 // may or may not be a constant. Takes immediate version of shift as input.
16184 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
16185 SDValue SrcOp, SDValue ShAmt,
16186 SelectionDAG &DAG) {
16187 MVT SVT = ShAmt.getSimpleValueType();
16188 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
16190 // Catch shift-by-constant.
16191 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16192 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16193 CShAmt->getZExtValue(), DAG);
16195 // Change opcode to non-immediate version
16197 default: llvm_unreachable("Unknown target vector shift node");
16198 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16199 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16200 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16203 const X86Subtarget &Subtarget =
16204 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16205 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16206 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16207 // Let the shuffle legalizer expand this shift amount node.
16208 SDValue Op0 = ShAmt.getOperand(0);
16209 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16210 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16212 // Need to build a vector containing shift amount.
16213 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16214 SmallVector<SDValue, 4> ShOps;
16215 ShOps.push_back(ShAmt);
16216 if (SVT == MVT::i32) {
16217 ShOps.push_back(DAG.getConstant(0, dl, SVT));
16218 ShOps.push_back(DAG.getUNDEF(SVT));
16220 ShOps.push_back(DAG.getUNDEF(SVT));
16222 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16223 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16226 // The return type has to be a 128-bit type with the same element
16227 // type as the input type.
16228 MVT EltVT = VT.getVectorElementType();
16229 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16231 ShAmt = DAG.getBitcast(ShVT, ShAmt);
16232 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16235 /// \brief Return Mask with the necessary casting or extending
16236 /// for \p Mask according to \p MaskVT when lowering masking intrinsics
16237 static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
16238 const X86Subtarget *Subtarget,
16239 SelectionDAG &DAG, SDLoc dl) {
16241 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
16242 // Mask should be extended
16243 Mask = DAG.getNode(ISD::ANY_EXTEND, dl,
16244 MVT::getIntegerVT(MaskVT.getSizeInBits()), Mask);
16247 if (Mask.getSimpleValueType() == MVT::i64 && Subtarget->is32Bit()) {
16248 if (MaskVT == MVT::v64i1) {
16249 assert(Subtarget->hasBWI() && "Expected AVX512BW target!");
16250 // In case 32bit mode, bitcast i64 is illegal, extend/split it.
16252 Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16253 DAG.getConstant(0, dl, MVT::i32));
16254 Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Mask,
16255 DAG.getConstant(1, dl, MVT::i32));
16257 Lo = DAG.getBitcast(MVT::v32i1, Lo);
16258 Hi = DAG.getBitcast(MVT::v32i1, Hi);
16260 return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
16262 // MaskVT require < 64bit. Truncate mask (should succeed in any case),
16264 MVT TruncVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
16265 return DAG.getBitcast(MaskVT,
16266 DAG.getNode(ISD::TRUNCATE, dl, TruncVT, Mask));
16270 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16271 Mask.getSimpleValueType().getSizeInBits());
16272 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16273 // are extracted by EXTRACT_SUBVECTOR.
16274 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16275 DAG.getBitcast(BitcastVT, Mask),
16276 DAG.getIntPtrConstant(0, dl));
16280 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16281 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16282 /// necessary casting or extending for \p Mask when lowering masking intrinsics
16283 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16284 SDValue PreservedSrc,
16285 const X86Subtarget *Subtarget,
16286 SelectionDAG &DAG) {
16287 MVT VT = Op.getSimpleValueType();
16288 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16289 unsigned OpcodeSelect = ISD::VSELECT;
16292 if (isAllOnesConstant(Mask))
16295 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16297 switch (Op.getOpcode()) {
16299 case X86ISD::PCMPEQM:
16300 case X86ISD::PCMPGTM:
16302 case X86ISD::CMPMU:
16303 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16304 case X86ISD::VFPCLASS:
16305 case X86ISD::VFPCLASSS:
16306 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16307 case X86ISD::VTRUNC:
16308 case X86ISD::VTRUNCS:
16309 case X86ISD::VTRUNCUS:
16310 // We can't use ISD::VSELECT here because it is not always "Legal"
16311 // for the destination type. For example vpmovqb require only AVX512
16312 // and vselect that can operate on byte element type require BWI
16313 OpcodeSelect = X86ISD::SELECT;
16316 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16317 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16318 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16321 /// \brief Creates an SDNode for a predicated scalar operation.
16322 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16323 /// The mask is coming as MVT::i8 and it should be truncated
16324 /// to MVT::i1 while lowering masking intrinsics.
16325 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16326 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16327 /// for a scalar instruction.
16328 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16329 SDValue PreservedSrc,
16330 const X86Subtarget *Subtarget,
16331 SelectionDAG &DAG) {
16332 if (isAllOnesConstant(Mask))
16335 MVT VT = Op.getSimpleValueType();
16337 // The mask should be of type MVT::i1
16338 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16340 if (Op.getOpcode() == X86ISD::FSETCC)
16341 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16342 if (Op.getOpcode() == X86ISD::VFPCLASS ||
16343 Op.getOpcode() == X86ISD::VFPCLASSS)
16344 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16346 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16347 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16348 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16351 static int getSEHRegistrationNodeSize(const Function *Fn) {
16352 if (!Fn->hasPersonalityFn())
16353 report_fatal_error(
16354 "querying registration node size for function without personality");
16355 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16356 // WinEHStatePass for the full struct definition.
16357 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16358 case EHPersonality::MSVC_X86SEH: return 24;
16359 case EHPersonality::MSVC_CXX: return 16;
16362 report_fatal_error(
16363 "can only recover FP for 32-bit MSVC EH personality functions");
16366 /// When the MSVC runtime transfers control to us, either to an outlined
16367 /// function or when returning to a parent frame after catching an exception, we
16368 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16369 /// Here's the math:
16370 /// RegNodeBase = EntryEBP - RegNodeSize
16371 /// ParentFP = RegNodeBase - ParentFrameOffset
16372 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16373 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16374 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16375 SDValue EntryEBP) {
16376 MachineFunction &MF = DAG.getMachineFunction();
16379 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16380 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16382 // It's possible that the parent function no longer has a personality function
16383 // if the exceptional code was optimized away, in which case we just return
16384 // the incoming EBP.
16385 if (!Fn->hasPersonalityFn())
16388 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16389 // registration, or the .set_setframe offset.
16390 MCSymbol *OffsetSym =
16391 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16392 GlobalValue::getRealLinkageName(Fn->getName()));
16393 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16394 SDValue ParentFrameOffset =
16395 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16397 // Return EntryEBP + ParentFrameOffset for x64. This adjusts from RSP after
16398 // prologue to RBP in the parent function.
16399 const X86Subtarget &Subtarget =
16400 static_cast<const X86Subtarget &>(DAG.getSubtarget());
16401 if (Subtarget.is64Bit())
16402 return DAG.getNode(ISD::ADD, dl, PtrVT, EntryEBP, ParentFrameOffset);
16404 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16405 // RegNodeBase = EntryEBP - RegNodeSize
16406 // ParentFP = RegNodeBase - ParentFrameOffset
16407 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16408 DAG.getConstant(RegNodeSize, dl, PtrVT));
16409 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, ParentFrameOffset);
16412 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16413 SelectionDAG &DAG) {
16415 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16416 MVT VT = Op.getSimpleValueType();
16417 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16419 switch(IntrData->Type) {
16420 case INTR_TYPE_1OP:
16421 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16422 case INTR_TYPE_2OP:
16423 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16425 case INTR_TYPE_2OP_IMM8:
16426 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16427 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16428 case INTR_TYPE_3OP:
16429 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16430 Op.getOperand(2), Op.getOperand(3));
16431 case INTR_TYPE_4OP:
16432 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16433 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16434 case INTR_TYPE_1OP_MASK_RM: {
16435 SDValue Src = Op.getOperand(1);
16436 SDValue PassThru = Op.getOperand(2);
16437 SDValue Mask = Op.getOperand(3);
16438 SDValue RoundingMode;
16439 // We allways add rounding mode to the Node.
16440 // If the rounding mode is not specified, we add the
16441 // "current direction" mode.
16442 if (Op.getNumOperands() == 4)
16444 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16446 RoundingMode = Op.getOperand(4);
16447 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16448 if (IntrWithRoundingModeOpcode != 0)
16449 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16450 X86::STATIC_ROUNDING::CUR_DIRECTION)
16451 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16452 dl, Op.getValueType(), Src, RoundingMode),
16453 Mask, PassThru, Subtarget, DAG);
16454 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16456 Mask, PassThru, Subtarget, DAG);
16458 case INTR_TYPE_1OP_MASK: {
16459 SDValue Src = Op.getOperand(1);
16460 SDValue PassThru = Op.getOperand(2);
16461 SDValue Mask = Op.getOperand(3);
16462 // We add rounding mode to the Node when
16463 // - RM Opcode is specified and
16464 // - RM is not "current direction".
16465 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16466 if (IntrWithRoundingModeOpcode != 0) {
16467 SDValue Rnd = Op.getOperand(4);
16468 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16469 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16470 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16471 dl, Op.getValueType(),
16473 Mask, PassThru, Subtarget, DAG);
16476 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16477 Mask, PassThru, Subtarget, DAG);
16479 case INTR_TYPE_SCALAR_MASK: {
16480 SDValue Src1 = Op.getOperand(1);
16481 SDValue Src2 = Op.getOperand(2);
16482 SDValue passThru = Op.getOperand(3);
16483 SDValue Mask = Op.getOperand(4);
16484 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16485 Mask, passThru, Subtarget, DAG);
16487 case INTR_TYPE_SCALAR_MASK_RM: {
16488 SDValue Src1 = Op.getOperand(1);
16489 SDValue Src2 = Op.getOperand(2);
16490 SDValue Src0 = Op.getOperand(3);
16491 SDValue Mask = Op.getOperand(4);
16492 // There are 2 kinds of intrinsics in this group:
16493 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16494 // (2) With rounding mode and sae - 7 operands.
16495 if (Op.getNumOperands() == 6) {
16496 SDValue Sae = Op.getOperand(5);
16497 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16498 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16500 Mask, Src0, Subtarget, DAG);
16502 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16503 SDValue RoundingMode = Op.getOperand(5);
16504 SDValue Sae = Op.getOperand(6);
16505 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16506 RoundingMode, Sae),
16507 Mask, Src0, Subtarget, DAG);
16509 case INTR_TYPE_2OP_MASK:
16510 case INTR_TYPE_2OP_IMM8_MASK: {
16511 SDValue Src1 = Op.getOperand(1);
16512 SDValue Src2 = Op.getOperand(2);
16513 SDValue PassThru = Op.getOperand(3);
16514 SDValue Mask = Op.getOperand(4);
16516 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16517 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16519 // We specify 2 possible opcodes for intrinsics with rounding modes.
16520 // First, we check if the intrinsic may have non-default rounding mode,
16521 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16522 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16523 if (IntrWithRoundingModeOpcode != 0) {
16524 SDValue Rnd = Op.getOperand(5);
16525 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16526 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16527 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16528 dl, Op.getValueType(),
16530 Mask, PassThru, Subtarget, DAG);
16533 // TODO: Intrinsics should have fast-math-flags to propagate.
16534 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16535 Mask, PassThru, Subtarget, DAG);
16537 case INTR_TYPE_2OP_MASK_RM: {
16538 SDValue Src1 = Op.getOperand(1);
16539 SDValue Src2 = Op.getOperand(2);
16540 SDValue PassThru = Op.getOperand(3);
16541 SDValue Mask = Op.getOperand(4);
16542 // We specify 2 possible modes for intrinsics, with/without rounding
16544 // First, we check if the intrinsic have rounding mode (6 operands),
16545 // if not, we set rounding mode to "current".
16547 if (Op.getNumOperands() == 6)
16548 Rnd = Op.getOperand(5);
16550 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16551 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16553 Mask, PassThru, Subtarget, DAG);
16555 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16556 SDValue Src1 = Op.getOperand(1);
16557 SDValue Src2 = Op.getOperand(2);
16558 SDValue Src3 = Op.getOperand(3);
16559 SDValue PassThru = Op.getOperand(4);
16560 SDValue Mask = Op.getOperand(5);
16561 SDValue Sae = Op.getOperand(6);
16563 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16565 Mask, PassThru, Subtarget, DAG);
16567 case INTR_TYPE_3OP_MASK_RM: {
16568 SDValue Src1 = Op.getOperand(1);
16569 SDValue Src2 = Op.getOperand(2);
16570 SDValue Imm = Op.getOperand(3);
16571 SDValue PassThru = Op.getOperand(4);
16572 SDValue Mask = Op.getOperand(5);
16573 // We specify 2 possible modes for intrinsics, with/without rounding
16575 // First, we check if the intrinsic have rounding mode (7 operands),
16576 // if not, we set rounding mode to "current".
16578 if (Op.getNumOperands() == 7)
16579 Rnd = Op.getOperand(6);
16581 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16582 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16583 Src1, Src2, Imm, Rnd),
16584 Mask, PassThru, Subtarget, DAG);
16586 case INTR_TYPE_3OP_IMM8_MASK:
16587 case INTR_TYPE_3OP_MASK:
16588 case INSERT_SUBVEC: {
16589 SDValue Src1 = Op.getOperand(1);
16590 SDValue Src2 = Op.getOperand(2);
16591 SDValue Src3 = Op.getOperand(3);
16592 SDValue PassThru = Op.getOperand(4);
16593 SDValue Mask = Op.getOperand(5);
16595 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16596 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16597 else if (IntrData->Type == INSERT_SUBVEC) {
16598 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16599 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16600 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16601 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16602 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16605 // We specify 2 possible opcodes for intrinsics with rounding modes.
16606 // First, we check if the intrinsic may have non-default rounding mode,
16607 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16608 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16609 if (IntrWithRoundingModeOpcode != 0) {
16610 SDValue Rnd = Op.getOperand(6);
16611 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16612 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16613 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16614 dl, Op.getValueType(),
16615 Src1, Src2, Src3, Rnd),
16616 Mask, PassThru, Subtarget, DAG);
16619 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16621 Mask, PassThru, Subtarget, DAG);
16623 case VPERM_3OP_MASKZ:
16624 case VPERM_3OP_MASK:{
16625 // Src2 is the PassThru
16626 SDValue Src1 = Op.getOperand(1);
16627 SDValue Src2 = Op.getOperand(2);
16628 SDValue Src3 = Op.getOperand(3);
16629 SDValue Mask = Op.getOperand(4);
16630 MVT VT = Op.getSimpleValueType();
16631 SDValue PassThru = SDValue();
16633 // set PassThru element
16634 if (IntrData->Type == VPERM_3OP_MASKZ)
16635 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16637 PassThru = DAG.getBitcast(VT, Src2);
16639 // Swap Src1 and Src2 in the node creation
16640 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16641 dl, Op.getValueType(),
16643 Mask, PassThru, Subtarget, DAG);
16647 case FMA_OP_MASK: {
16648 SDValue Src1 = Op.getOperand(1);
16649 SDValue Src2 = Op.getOperand(2);
16650 SDValue Src3 = Op.getOperand(3);
16651 SDValue Mask = Op.getOperand(4);
16652 MVT VT = Op.getSimpleValueType();
16653 SDValue PassThru = SDValue();
16655 // set PassThru element
16656 if (IntrData->Type == FMA_OP_MASKZ)
16657 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16658 else if (IntrData->Type == FMA_OP_MASK3)
16663 // We specify 2 possible opcodes for intrinsics with rounding modes.
16664 // First, we check if the intrinsic may have non-default rounding mode,
16665 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16666 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16667 if (IntrWithRoundingModeOpcode != 0) {
16668 SDValue Rnd = Op.getOperand(5);
16669 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16670 X86::STATIC_ROUNDING::CUR_DIRECTION)
16671 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16672 dl, Op.getValueType(),
16673 Src1, Src2, Src3, Rnd),
16674 Mask, PassThru, Subtarget, DAG);
16676 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16677 dl, Op.getValueType(),
16679 Mask, PassThru, Subtarget, DAG);
16681 case TERLOG_OP_MASK:
16682 case TERLOG_OP_MASKZ: {
16683 SDValue Src1 = Op.getOperand(1);
16684 SDValue Src2 = Op.getOperand(2);
16685 SDValue Src3 = Op.getOperand(3);
16686 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16687 SDValue Mask = Op.getOperand(5);
16688 MVT VT = Op.getSimpleValueType();
16689 SDValue PassThru = Src1;
16690 // Set PassThru element.
16691 if (IntrData->Type == TERLOG_OP_MASKZ)
16692 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16694 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16695 Src1, Src2, Src3, Src4),
16696 Mask, PassThru, Subtarget, DAG);
16699 // FPclass intrinsics with mask
16700 SDValue Src1 = Op.getOperand(1);
16701 MVT VT = Src1.getSimpleValueType();
16702 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16703 SDValue Imm = Op.getOperand(2);
16704 SDValue Mask = Op.getOperand(3);
16705 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16706 Mask.getSimpleValueType().getSizeInBits());
16707 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16708 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16709 DAG.getTargetConstant(0, dl, MaskVT),
16711 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16712 DAG.getUNDEF(BitcastVT), FPclassMask,
16713 DAG.getIntPtrConstant(0, dl));
16714 return DAG.getBitcast(Op.getValueType(), Res);
16717 SDValue Src1 = Op.getOperand(1);
16718 SDValue Imm = Op.getOperand(2);
16719 SDValue Mask = Op.getOperand(3);
16720 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16721 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16722 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16723 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16726 case CMP_MASK_CC: {
16727 // Comparison intrinsics with masks.
16728 // Example of transformation:
16729 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16730 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16732 // (v8i1 (insert_subvector undef,
16733 // (v2i1 (and (PCMPEQM %a, %b),
16734 // (extract_subvector
16735 // (v8i1 (bitcast %mask)), 0))), 0))))
16736 MVT VT = Op.getOperand(1).getSimpleValueType();
16737 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16738 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16739 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16740 Mask.getSimpleValueType().getSizeInBits());
16742 if (IntrData->Type == CMP_MASK_CC) {
16743 SDValue CC = Op.getOperand(3);
16744 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16745 // We specify 2 possible opcodes for intrinsics with rounding modes.
16746 // First, we check if the intrinsic may have non-default rounding mode,
16747 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16748 if (IntrData->Opc1 != 0) {
16749 SDValue Rnd = Op.getOperand(5);
16750 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16751 X86::STATIC_ROUNDING::CUR_DIRECTION)
16752 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16753 Op.getOperand(2), CC, Rnd);
16755 //default rounding mode
16757 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16758 Op.getOperand(2), CC);
16761 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16762 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16765 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16766 DAG.getTargetConstant(0, dl,
16769 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16770 DAG.getUNDEF(BitcastVT), CmpMask,
16771 DAG.getIntPtrConstant(0, dl));
16772 return DAG.getBitcast(Op.getValueType(), Res);
16774 case CMP_MASK_SCALAR_CC: {
16775 SDValue Src1 = Op.getOperand(1);
16776 SDValue Src2 = Op.getOperand(2);
16777 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16778 SDValue Mask = Op.getOperand(4);
16781 if (IntrData->Opc1 != 0) {
16782 SDValue Rnd = Op.getOperand(5);
16783 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16784 X86::STATIC_ROUNDING::CUR_DIRECTION)
16785 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16787 //default rounding mode
16789 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16791 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16792 DAG.getTargetConstant(0, dl,
16796 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16797 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16798 DAG.getValueType(MVT::i1));
16800 case COMI: { // Comparison intrinsics
16801 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16802 SDValue LHS = Op.getOperand(1);
16803 SDValue RHS = Op.getOperand(2);
16804 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16805 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16806 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16807 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16808 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16809 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16811 case COMI_RM: { // Comparison intrinsics with Sae
16812 SDValue LHS = Op.getOperand(1);
16813 SDValue RHS = Op.getOperand(2);
16814 SDValue CC = Op.getOperand(3);
16815 SDValue Sae = Op.getOperand(4);
16816 auto ComiType = TranslateX86ConstCondToX86CC(CC);
16817 // choose between ordered and unordered (comi/ucomi)
16818 unsigned comiOp = std::get<0>(ComiType) ? IntrData->Opc0 : IntrData->Opc1;
16820 if (cast<ConstantSDNode>(Sae)->getZExtValue() !=
16821 X86::STATIC_ROUNDING::CUR_DIRECTION)
16822 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS, Sae);
16824 Cond = DAG.getNode(comiOp, dl, MVT::i32, LHS, RHS);
16825 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16826 DAG.getConstant(std::get<1>(ComiType), dl, MVT::i8), Cond);
16827 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16830 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16831 Op.getOperand(1), Op.getOperand(2), DAG);
16833 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16834 Op.getSimpleValueType(),
16836 Op.getOperand(2), DAG),
16837 Op.getOperand(4), Op.getOperand(3), Subtarget,
16839 case COMPRESS_EXPAND_IN_REG: {
16840 SDValue Mask = Op.getOperand(3);
16841 SDValue DataToCompress = Op.getOperand(1);
16842 SDValue PassThru = Op.getOperand(2);
16843 if (isAllOnesConstant(Mask)) // return data as is
16844 return Op.getOperand(1);
16846 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16848 Mask, PassThru, Subtarget, DAG);
16851 SDValue Mask = Op.getOperand(1);
16852 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16853 Mask.getSimpleValueType().getSizeInBits());
16854 Mask = DAG.getBitcast(MaskVT, Mask);
16855 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Mask);
16858 SDValue Mask = Op.getOperand(3);
16859 MVT VT = Op.getSimpleValueType();
16860 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16861 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16862 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16866 MVT VT = Op.getSimpleValueType();
16867 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits()/2);
16869 SDValue Src1 = getMaskNode(Op.getOperand(1), MaskVT, Subtarget, DAG, dl);
16870 SDValue Src2 = getMaskNode(Op.getOperand(2), MaskVT, Subtarget, DAG, dl);
16871 // Arguments should be swapped.
16872 SDValue Res = DAG.getNode(IntrData->Opc0, dl,
16873 MVT::getVectorVT(MVT::i1, VT.getSizeInBits()),
16875 return DAG.getBitcast(VT, Res);
16877 case CONVERT_TO_MASK: {
16878 MVT SrcVT = Op.getOperand(1).getSimpleValueType();
16879 MVT MaskVT = MVT::getVectorVT(MVT::i1, SrcVT.getVectorNumElements());
16880 MVT BitcastVT = MVT::getVectorVT(MVT::i1, VT.getSizeInBits());
16882 SDValue CvtMask = DAG.getNode(IntrData->Opc0, dl, MaskVT,
16884 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16885 DAG.getUNDEF(BitcastVT), CvtMask,
16886 DAG.getIntPtrConstant(0, dl));
16887 return DAG.getBitcast(Op.getValueType(), Res);
16889 case CONVERT_MASK_TO_VEC: {
16890 SDValue Mask = Op.getOperand(1);
16891 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16892 SDValue VMask = getMaskNode(Mask, MaskVT, Subtarget, DAG, dl);
16893 return DAG.getNode(IntrData->Opc0, dl, VT, VMask);
16895 case BRCST_SUBVEC_TO_VEC: {
16896 SDValue Src = Op.getOperand(1);
16897 SDValue Passthru = Op.getOperand(2);
16898 SDValue Mask = Op.getOperand(3);
16899 EVT resVT = Passthru.getValueType();
16900 SDValue subVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, resVT,
16901 DAG.getUNDEF(resVT), Src,
16902 DAG.getIntPtrConstant(0, dl));
16904 if (Src.getSimpleValueType().is256BitVector() && resVT.is512BitVector())
16905 immVal = DAG.getConstant(0x44, dl, MVT::i8);
16907 immVal = DAG.getConstant(0, dl, MVT::i8);
16908 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16909 subVec, subVec, immVal),
16910 Mask, Passthru, Subtarget, DAG);
16918 default: return SDValue(); // Don't custom lower most intrinsics.
16920 case Intrinsic::x86_avx2_permd:
16921 case Intrinsic::x86_avx2_permps:
16922 // Operands intentionally swapped. Mask is last operand to intrinsic,
16923 // but second operand for node/instruction.
16924 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16925 Op.getOperand(2), Op.getOperand(1));
16927 // ptest and testp intrinsics. The intrinsic these come from are designed to
16928 // return an integer value, not just an instruction so lower it to the ptest
16929 // or testp pattern and a setcc for the result.
16930 case Intrinsic::x86_sse41_ptestz:
16931 case Intrinsic::x86_sse41_ptestc:
16932 case Intrinsic::x86_sse41_ptestnzc:
16933 case Intrinsic::x86_avx_ptestz_256:
16934 case Intrinsic::x86_avx_ptestc_256:
16935 case Intrinsic::x86_avx_ptestnzc_256:
16936 case Intrinsic::x86_avx_vtestz_ps:
16937 case Intrinsic::x86_avx_vtestc_ps:
16938 case Intrinsic::x86_avx_vtestnzc_ps:
16939 case Intrinsic::x86_avx_vtestz_pd:
16940 case Intrinsic::x86_avx_vtestc_pd:
16941 case Intrinsic::x86_avx_vtestnzc_pd:
16942 case Intrinsic::x86_avx_vtestz_ps_256:
16943 case Intrinsic::x86_avx_vtestc_ps_256:
16944 case Intrinsic::x86_avx_vtestnzc_ps_256:
16945 case Intrinsic::x86_avx_vtestz_pd_256:
16946 case Intrinsic::x86_avx_vtestc_pd_256:
16947 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16948 bool IsTestPacked = false;
16951 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16952 case Intrinsic::x86_avx_vtestz_ps:
16953 case Intrinsic::x86_avx_vtestz_pd:
16954 case Intrinsic::x86_avx_vtestz_ps_256:
16955 case Intrinsic::x86_avx_vtestz_pd_256:
16956 IsTestPacked = true; // Fallthrough
16957 case Intrinsic::x86_sse41_ptestz:
16958 case Intrinsic::x86_avx_ptestz_256:
16960 X86CC = X86::COND_E;
16962 case Intrinsic::x86_avx_vtestc_ps:
16963 case Intrinsic::x86_avx_vtestc_pd:
16964 case Intrinsic::x86_avx_vtestc_ps_256:
16965 case Intrinsic::x86_avx_vtestc_pd_256:
16966 IsTestPacked = true; // Fallthrough
16967 case Intrinsic::x86_sse41_ptestc:
16968 case Intrinsic::x86_avx_ptestc_256:
16970 X86CC = X86::COND_B;
16972 case Intrinsic::x86_avx_vtestnzc_ps:
16973 case Intrinsic::x86_avx_vtestnzc_pd:
16974 case Intrinsic::x86_avx_vtestnzc_ps_256:
16975 case Intrinsic::x86_avx_vtestnzc_pd_256:
16976 IsTestPacked = true; // Fallthrough
16977 case Intrinsic::x86_sse41_ptestnzc:
16978 case Intrinsic::x86_avx_ptestnzc_256:
16980 X86CC = X86::COND_A;
16984 SDValue LHS = Op.getOperand(1);
16985 SDValue RHS = Op.getOperand(2);
16986 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16987 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16988 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16989 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16990 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16992 case Intrinsic::x86_avx512_kortestz_w:
16993 case Intrinsic::x86_avx512_kortestc_w: {
16994 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16995 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16996 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16997 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16998 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16999 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17000 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17003 case Intrinsic::x86_sse42_pcmpistria128:
17004 case Intrinsic::x86_sse42_pcmpestria128:
17005 case Intrinsic::x86_sse42_pcmpistric128:
17006 case Intrinsic::x86_sse42_pcmpestric128:
17007 case Intrinsic::x86_sse42_pcmpistrio128:
17008 case Intrinsic::x86_sse42_pcmpestrio128:
17009 case Intrinsic::x86_sse42_pcmpistris128:
17010 case Intrinsic::x86_sse42_pcmpestris128:
17011 case Intrinsic::x86_sse42_pcmpistriz128:
17012 case Intrinsic::x86_sse42_pcmpestriz128: {
17016 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17017 case Intrinsic::x86_sse42_pcmpistria128:
17018 Opcode = X86ISD::PCMPISTRI;
17019 X86CC = X86::COND_A;
17021 case Intrinsic::x86_sse42_pcmpestria128:
17022 Opcode = X86ISD::PCMPESTRI;
17023 X86CC = X86::COND_A;
17025 case Intrinsic::x86_sse42_pcmpistric128:
17026 Opcode = X86ISD::PCMPISTRI;
17027 X86CC = X86::COND_B;
17029 case Intrinsic::x86_sse42_pcmpestric128:
17030 Opcode = X86ISD::PCMPESTRI;
17031 X86CC = X86::COND_B;
17033 case Intrinsic::x86_sse42_pcmpistrio128:
17034 Opcode = X86ISD::PCMPISTRI;
17035 X86CC = X86::COND_O;
17037 case Intrinsic::x86_sse42_pcmpestrio128:
17038 Opcode = X86ISD::PCMPESTRI;
17039 X86CC = X86::COND_O;
17041 case Intrinsic::x86_sse42_pcmpistris128:
17042 Opcode = X86ISD::PCMPISTRI;
17043 X86CC = X86::COND_S;
17045 case Intrinsic::x86_sse42_pcmpestris128:
17046 Opcode = X86ISD::PCMPESTRI;
17047 X86CC = X86::COND_S;
17049 case Intrinsic::x86_sse42_pcmpistriz128:
17050 Opcode = X86ISD::PCMPISTRI;
17051 X86CC = X86::COND_E;
17053 case Intrinsic::x86_sse42_pcmpestriz128:
17054 Opcode = X86ISD::PCMPESTRI;
17055 X86CC = X86::COND_E;
17058 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17059 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17060 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17061 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17062 DAG.getConstant(X86CC, dl, MVT::i8),
17063 SDValue(PCMP.getNode(), 1));
17064 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17067 case Intrinsic::x86_sse42_pcmpistri128:
17068 case Intrinsic::x86_sse42_pcmpestri128: {
17070 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17071 Opcode = X86ISD::PCMPISTRI;
17073 Opcode = X86ISD::PCMPESTRI;
17075 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17076 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17077 return DAG.getNode(Opcode, dl, VTs, NewOps);
17080 case Intrinsic::x86_seh_lsda: {
17081 // Compute the symbol for the LSDA. We know it'll get emitted later.
17082 MachineFunction &MF = DAG.getMachineFunction();
17083 SDValue Op1 = Op.getOperand(1);
17084 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
17085 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
17086 GlobalValue::getRealLinkageName(Fn->getName()));
17088 // Generate a simple absolute symbol reference. This intrinsic is only
17089 // supported on 32-bit Windows, which isn't PIC.
17090 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
17091 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
17094 case Intrinsic::x86_seh_recoverfp: {
17095 SDValue FnOp = Op.getOperand(1);
17096 SDValue IncomingFPOp = Op.getOperand(2);
17097 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
17098 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
17100 report_fatal_error(
17101 "llvm.x86.seh.recoverfp must take a function as the first argument");
17102 return recoverFramePointer(DAG, Fn, IncomingFPOp);
17105 case Intrinsic::localaddress: {
17106 // Returns one of the stack, base, or frame pointer registers, depending on
17107 // which is used to reference local variables.
17108 MachineFunction &MF = DAG.getMachineFunction();
17109 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17111 if (RegInfo->hasBasePointer(MF))
17112 Reg = RegInfo->getBaseRegister();
17113 else // This function handles the SP or FP case.
17114 Reg = RegInfo->getPtrSizedFrameRegister(MF);
17115 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
17120 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17121 SDValue Src, SDValue Mask, SDValue Base,
17122 SDValue Index, SDValue ScaleOp, SDValue Chain,
17123 const X86Subtarget * Subtarget) {
17125 auto *C = cast<ConstantSDNode>(ScaleOp);
17126 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17127 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17128 Index.getSimpleValueType().getVectorNumElements());
17130 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17132 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17134 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17135 Mask.getSimpleValueType().getSizeInBits());
17137 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17138 // are extracted by EXTRACT_SUBVECTOR.
17139 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17140 DAG.getBitcast(BitcastVT, Mask),
17141 DAG.getIntPtrConstant(0, dl));
17143 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17144 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17145 SDValue Segment = DAG.getRegister(0, MVT::i32);
17146 if (Src.getOpcode() == ISD::UNDEF)
17147 Src = getZeroVector(Op.getSimpleValueType(), Subtarget, DAG, dl);
17148 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17149 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17150 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17151 return DAG.getMergeValues(RetOps, dl);
17154 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17155 SDValue Src, SDValue Mask, SDValue Base,
17156 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17158 auto *C = cast<ConstantSDNode>(ScaleOp);
17159 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17160 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17161 SDValue Segment = DAG.getRegister(0, MVT::i32);
17162 MVT MaskVT = MVT::getVectorVT(MVT::i1,
17163 Index.getSimpleValueType().getVectorNumElements());
17165 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17167 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17169 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17170 Mask.getSimpleValueType().getSizeInBits());
17172 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17173 // are extracted by EXTRACT_SUBVECTOR.
17174 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17175 DAG.getBitcast(BitcastVT, Mask),
17176 DAG.getIntPtrConstant(0, dl));
17178 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17179 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17180 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17181 return SDValue(Res, 1);
17184 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17185 SDValue Mask, SDValue Base, SDValue Index,
17186 SDValue ScaleOp, SDValue Chain) {
17188 auto *C = cast<ConstantSDNode>(ScaleOp);
17189 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
17190 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
17191 SDValue Segment = DAG.getRegister(0, MVT::i32);
17193 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17195 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17197 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
17199 MaskInReg = DAG.getBitcast(MaskVT, Mask);
17200 //SDVTList VTs = DAG.getVTList(MVT::Other);
17201 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17202 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17203 return SDValue(Res, 0);
17206 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17207 // read performance monitor counters (x86_rdpmc).
17208 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17209 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17210 SmallVectorImpl<SDValue> &Results) {
17211 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17212 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17215 // The ECX register is used to select the index of the performance counter
17217 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17219 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17221 // Reads the content of a 64-bit performance counter and returns it in the
17222 // registers EDX:EAX.
17223 if (Subtarget->is64Bit()) {
17224 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17225 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17228 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17229 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17232 Chain = HI.getValue(1);
17234 if (Subtarget->is64Bit()) {
17235 // The EAX register is loaded with the low-order 32 bits. The EDX register
17236 // is loaded with the supported high-order bits of the counter.
17237 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17238 DAG.getConstant(32, DL, MVT::i8));
17239 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17240 Results.push_back(Chain);
17244 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17245 SDValue Ops[] = { LO, HI };
17246 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17247 Results.push_back(Pair);
17248 Results.push_back(Chain);
17251 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17252 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17253 // also used to custom lower READCYCLECOUNTER nodes.
17254 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17255 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17256 SmallVectorImpl<SDValue> &Results) {
17257 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17258 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17261 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17262 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17263 // and the EAX register is loaded with the low-order 32 bits.
17264 if (Subtarget->is64Bit()) {
17265 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17266 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17269 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17270 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17273 SDValue Chain = HI.getValue(1);
17275 if (Opcode == X86ISD::RDTSCP_DAG) {
17276 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17278 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17279 // the ECX register. Add 'ecx' explicitly to the chain.
17280 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17282 // Explicitly store the content of ECX at the location passed in input
17283 // to the 'rdtscp' intrinsic.
17284 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17285 MachinePointerInfo(), false, false, 0);
17288 if (Subtarget->is64Bit()) {
17289 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17290 // the EAX register is loaded with the low-order 32 bits.
17291 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17292 DAG.getConstant(32, DL, MVT::i8));
17293 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17294 Results.push_back(Chain);
17298 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17299 SDValue Ops[] = { LO, HI };
17300 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17301 Results.push_back(Pair);
17302 Results.push_back(Chain);
17305 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17306 SelectionDAG &DAG) {
17307 SmallVector<SDValue, 2> Results;
17309 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17311 return DAG.getMergeValues(Results, DL);
17314 static SDValue MarkEHRegistrationNode(SDValue Op, SelectionDAG &DAG) {
17315 MachineFunction &MF = DAG.getMachineFunction();
17316 SDValue Chain = Op.getOperand(0);
17317 SDValue RegNode = Op.getOperand(2);
17318 WinEHFuncInfo *EHInfo = MF.getWinEHFuncInfo();
17320 report_fatal_error("EH registrations only live in functions using WinEH");
17322 // Cast the operand to an alloca, and remember the frame index.
17323 auto *FINode = dyn_cast<FrameIndexSDNode>(RegNode);
17325 report_fatal_error("llvm.x86.seh.ehregnode expects a static alloca");
17326 EHInfo->EHRegNodeFrameIndex = FINode->getIndex();
17328 // Return the chain operand without making any DAG nodes.
17332 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
17333 /// return truncate Store/MaskedStore Node
17334 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17338 SDValue Mask = Op.getOperand(4);
17339 SDValue DataToTruncate = Op.getOperand(3);
17340 SDValue Addr = Op.getOperand(2);
17341 SDValue Chain = Op.getOperand(0);
17343 MVT VT = DataToTruncate.getSimpleValueType();
17344 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
17346 if (isAllOnesConstant(Mask)) // return just a truncate store
17347 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17348 MachinePointerInfo(), SVT, false, false,
17349 SVT.getScalarSizeInBits()/8);
17351 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
17352 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
17353 Mask.getSimpleValueType().getSizeInBits());
17354 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17355 // are extracted by EXTRACT_SUBVECTOR.
17356 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17357 DAG.getBitcast(BitcastVT, Mask),
17358 DAG.getIntPtrConstant(0, dl));
17360 MachineMemOperand *MMO = DAG.getMachineFunction().
17361 getMachineMemOperand(MachinePointerInfo(),
17362 MachineMemOperand::MOStore, SVT.getStoreSize(),
17363 SVT.getScalarSizeInBits()/8);
17365 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17366 VMask, SVT, MMO, true);
17369 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17370 SelectionDAG &DAG) {
17371 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17373 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17375 if (IntNo == llvm::Intrinsic::x86_seh_ehregnode)
17376 return MarkEHRegistrationNode(Op, DAG);
17381 switch(IntrData->Type) {
17382 default: llvm_unreachable("Unknown Intrinsic Type");
17385 // Emit the node with the right value type.
17386 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17387 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17389 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17390 // Otherwise return the value from Rand, which is always 0, casted to i32.
17391 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17392 DAG.getConstant(1, dl, Op->getValueType(1)),
17393 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17394 SDValue(Result.getNode(), 1) };
17395 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17396 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17399 // Return { result, isValid, chain }.
17400 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17401 SDValue(Result.getNode(), 2));
17404 //gather(v1, mask, index, base, scale);
17405 SDValue Chain = Op.getOperand(0);
17406 SDValue Src = Op.getOperand(2);
17407 SDValue Base = Op.getOperand(3);
17408 SDValue Index = Op.getOperand(4);
17409 SDValue Mask = Op.getOperand(5);
17410 SDValue Scale = Op.getOperand(6);
17411 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17415 //scatter(base, mask, index, v1, scale);
17416 SDValue Chain = Op.getOperand(0);
17417 SDValue Base = Op.getOperand(2);
17418 SDValue Mask = Op.getOperand(3);
17419 SDValue Index = Op.getOperand(4);
17420 SDValue Src = Op.getOperand(5);
17421 SDValue Scale = Op.getOperand(6);
17422 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17426 SDValue Hint = Op.getOperand(6);
17427 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17428 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17429 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17430 SDValue Chain = Op.getOperand(0);
17431 SDValue Mask = Op.getOperand(2);
17432 SDValue Index = Op.getOperand(3);
17433 SDValue Base = Op.getOperand(4);
17434 SDValue Scale = Op.getOperand(5);
17435 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17437 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17439 SmallVector<SDValue, 2> Results;
17440 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17442 return DAG.getMergeValues(Results, dl);
17444 // Read Performance Monitoring Counters.
17446 SmallVector<SDValue, 2> Results;
17447 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17448 return DAG.getMergeValues(Results, dl);
17450 // XTEST intrinsics.
17452 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17453 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17454 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17455 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17457 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17458 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17459 Ret, SDValue(InTrans.getNode(), 1));
17463 SmallVector<SDValue, 2> Results;
17464 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17465 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17466 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17467 DAG.getConstant(-1, dl, MVT::i8));
17468 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17469 Op.getOperand(4), GenCF.getValue(1));
17470 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17471 Op.getOperand(5), MachinePointerInfo(),
17473 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17474 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17476 Results.push_back(SetCC);
17477 Results.push_back(Store);
17478 return DAG.getMergeValues(Results, dl);
17480 case COMPRESS_TO_MEM: {
17482 SDValue Mask = Op.getOperand(4);
17483 SDValue DataToCompress = Op.getOperand(3);
17484 SDValue Addr = Op.getOperand(2);
17485 SDValue Chain = Op.getOperand(0);
17487 MVT VT = DataToCompress.getSimpleValueType();
17488 if (isAllOnesConstant(Mask)) // return just a store
17489 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17490 MachinePointerInfo(), false, false,
17491 VT.getScalarSizeInBits()/8);
17493 SDValue Compressed =
17494 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17495 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17496 return DAG.getStore(Chain, dl, Compressed, Addr,
17497 MachinePointerInfo(), false, false,
17498 VT.getScalarSizeInBits()/8);
17500 case TRUNCATE_TO_MEM_VI8:
17501 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17502 case TRUNCATE_TO_MEM_VI16:
17503 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17504 case TRUNCATE_TO_MEM_VI32:
17505 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17506 case EXPAND_FROM_MEM: {
17508 SDValue Mask = Op.getOperand(4);
17509 SDValue PassThru = Op.getOperand(3);
17510 SDValue Addr = Op.getOperand(2);
17511 SDValue Chain = Op.getOperand(0);
17512 MVT VT = Op.getSimpleValueType();
17514 if (isAllOnesConstant(Mask)) // return just a load
17515 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17516 false, VT.getScalarSizeInBits()/8);
17518 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17519 false, false, false,
17520 VT.getScalarSizeInBits()/8);
17522 SDValue Results[] = {
17523 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17524 Mask, PassThru, Subtarget, DAG), Chain};
17525 return DAG.getMergeValues(Results, dl);
17530 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17531 SelectionDAG &DAG) const {
17532 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17533 MFI->setReturnAddressIsTaken(true);
17535 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17538 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17540 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17543 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17544 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17545 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17546 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17547 DAG.getNode(ISD::ADD, dl, PtrVT,
17548 FrameAddr, Offset),
17549 MachinePointerInfo(), false, false, false, 0);
17552 // Just load the return address.
17553 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17554 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17555 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17558 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17559 MachineFunction &MF = DAG.getMachineFunction();
17560 MachineFrameInfo *MFI = MF.getFrameInfo();
17561 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17562 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17563 EVT VT = Op.getValueType();
17565 MFI->setFrameAddressIsTaken(true);
17567 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17568 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17569 // is not possible to crawl up the stack without looking at the unwind codes
17571 int FrameAddrIndex = FuncInfo->getFAIndex();
17572 if (!FrameAddrIndex) {
17573 // Set up a frame object for the return address.
17574 unsigned SlotSize = RegInfo->getSlotSize();
17575 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17576 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17577 FuncInfo->setFAIndex(FrameAddrIndex);
17579 return DAG.getFrameIndex(FrameAddrIndex, VT);
17582 unsigned FrameReg =
17583 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17584 SDLoc dl(Op); // FIXME probably not meaningful
17585 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17586 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17587 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17588 "Invalid Frame Register!");
17589 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17591 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17592 MachinePointerInfo(),
17593 false, false, false, 0);
17597 // FIXME? Maybe this could be a TableGen attribute on some registers and
17598 // this table could be generated automatically from RegInfo.
17599 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17600 SelectionDAG &DAG) const {
17601 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17602 const MachineFunction &MF = DAG.getMachineFunction();
17604 unsigned Reg = StringSwitch<unsigned>(RegName)
17605 .Case("esp", X86::ESP)
17606 .Case("rsp", X86::RSP)
17607 .Case("ebp", X86::EBP)
17608 .Case("rbp", X86::RBP)
17611 if (Reg == X86::EBP || Reg == X86::RBP) {
17612 if (!TFI.hasFP(MF))
17613 report_fatal_error("register " + StringRef(RegName) +
17614 " is allocatable: function has no frame pointer");
17617 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17618 unsigned FrameReg =
17619 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17620 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17621 "Invalid Frame Register!");
17629 report_fatal_error("Invalid register name global variable");
17632 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17633 SelectionDAG &DAG) const {
17634 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17635 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17638 unsigned X86TargetLowering::getExceptionPointerRegister(
17639 const Constant *PersonalityFn) const {
17640 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17641 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17643 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17646 unsigned X86TargetLowering::getExceptionSelectorRegister(
17647 const Constant *PersonalityFn) const {
17648 // Funclet personalities don't use selectors (the runtime does the selection).
17649 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17650 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17653 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17654 SDValue Chain = Op.getOperand(0);
17655 SDValue Offset = Op.getOperand(1);
17656 SDValue Handler = Op.getOperand(2);
17659 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17660 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17661 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17662 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17663 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17664 "Invalid Frame Register!");
17665 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17666 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17668 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17669 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17671 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17672 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17674 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17676 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17677 DAG.getRegister(StoreAddrReg, PtrVT));
17680 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17681 SelectionDAG &DAG) const {
17683 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17684 DAG.getVTList(MVT::i32, MVT::Other),
17685 Op.getOperand(0), Op.getOperand(1));
17688 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17689 SelectionDAG &DAG) const {
17691 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17692 Op.getOperand(0), Op.getOperand(1));
17695 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17696 return Op.getOperand(0);
17699 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17700 SelectionDAG &DAG) const {
17701 SDValue Root = Op.getOperand(0);
17702 SDValue Trmp = Op.getOperand(1); // trampoline
17703 SDValue FPtr = Op.getOperand(2); // nested function
17704 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17707 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17708 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17710 if (Subtarget->is64Bit()) {
17711 SDValue OutChains[6];
17713 // Large code-model.
17714 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17715 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17717 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17718 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17720 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17722 // Load the pointer to the nested function into R11.
17723 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17724 SDValue Addr = Trmp;
17725 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17726 Addr, MachinePointerInfo(TrmpAddr),
17729 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17730 DAG.getConstant(2, dl, MVT::i64));
17731 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17732 MachinePointerInfo(TrmpAddr, 2),
17735 // Load the 'nest' parameter value into R10.
17736 // R10 is specified in X86CallingConv.td
17737 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17738 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17739 DAG.getConstant(10, dl, MVT::i64));
17740 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17741 Addr, MachinePointerInfo(TrmpAddr, 10),
17744 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17745 DAG.getConstant(12, dl, MVT::i64));
17746 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17747 MachinePointerInfo(TrmpAddr, 12),
17750 // Jump to the nested function.
17751 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17752 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17753 DAG.getConstant(20, dl, MVT::i64));
17754 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17755 Addr, MachinePointerInfo(TrmpAddr, 20),
17758 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17759 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17760 DAG.getConstant(22, dl, MVT::i64));
17761 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17762 Addr, MachinePointerInfo(TrmpAddr, 22),
17765 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17767 const Function *Func =
17768 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17769 CallingConv::ID CC = Func->getCallingConv();
17774 llvm_unreachable("Unsupported calling convention");
17775 case CallingConv::C:
17776 case CallingConv::X86_StdCall: {
17777 // Pass 'nest' parameter in ECX.
17778 // Must be kept in sync with X86CallingConv.td
17779 NestReg = X86::ECX;
17781 // Check that ECX wasn't needed by an 'inreg' parameter.
17782 FunctionType *FTy = Func->getFunctionType();
17783 const AttributeSet &Attrs = Func->getAttributes();
17785 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17786 unsigned InRegCount = 0;
17789 for (FunctionType::param_iterator I = FTy->param_begin(),
17790 E = FTy->param_end(); I != E; ++I, ++Idx)
17791 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17792 auto &DL = DAG.getDataLayout();
17793 // FIXME: should only count parameters that are lowered to integers.
17794 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17797 if (InRegCount > 2) {
17798 report_fatal_error("Nest register in use - reduce number of inreg"
17804 case CallingConv::X86_FastCall:
17805 case CallingConv::X86_ThisCall:
17806 case CallingConv::Fast:
17807 // Pass 'nest' parameter in EAX.
17808 // Must be kept in sync with X86CallingConv.td
17809 NestReg = X86::EAX;
17813 SDValue OutChains[4];
17814 SDValue Addr, Disp;
17816 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17817 DAG.getConstant(10, dl, MVT::i32));
17818 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17820 // This is storing the opcode for MOV32ri.
17821 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17822 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17823 OutChains[0] = DAG.getStore(Root, dl,
17824 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17825 Trmp, MachinePointerInfo(TrmpAddr),
17828 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17829 DAG.getConstant(1, dl, MVT::i32));
17830 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17831 MachinePointerInfo(TrmpAddr, 1),
17834 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17835 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17836 DAG.getConstant(5, dl, MVT::i32));
17837 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17838 Addr, MachinePointerInfo(TrmpAddr, 5),
17841 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17842 DAG.getConstant(6, dl, MVT::i32));
17843 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17844 MachinePointerInfo(TrmpAddr, 6),
17847 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17851 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17852 SelectionDAG &DAG) const {
17854 The rounding mode is in bits 11:10 of FPSR, and has the following
17856 00 Round to nearest
17861 FLT_ROUNDS, on the other hand, expects the following:
17868 To perform the conversion, we do:
17869 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17872 MachineFunction &MF = DAG.getMachineFunction();
17873 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17874 unsigned StackAlignment = TFI.getStackAlignment();
17875 MVT VT = Op.getSimpleValueType();
17878 // Save FP Control Word to stack slot
17879 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17880 SDValue StackSlot =
17881 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17883 MachineMemOperand *MMO =
17884 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17885 MachineMemOperand::MOStore, 2, 2);
17887 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17888 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17889 DAG.getVTList(MVT::Other),
17890 Ops, MVT::i16, MMO);
17892 // Load FP Control Word from stack slot
17893 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17894 MachinePointerInfo(), false, false, false, 0);
17896 // Transform as necessary
17898 DAG.getNode(ISD::SRL, DL, MVT::i16,
17899 DAG.getNode(ISD::AND, DL, MVT::i16,
17900 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17901 DAG.getConstant(11, DL, MVT::i8));
17903 DAG.getNode(ISD::SRL, DL, MVT::i16,
17904 DAG.getNode(ISD::AND, DL, MVT::i16,
17905 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17906 DAG.getConstant(9, DL, MVT::i8));
17909 DAG.getNode(ISD::AND, DL, MVT::i16,
17910 DAG.getNode(ISD::ADD, DL, MVT::i16,
17911 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17912 DAG.getConstant(1, DL, MVT::i16)),
17913 DAG.getConstant(3, DL, MVT::i16));
17915 return DAG.getNode((VT.getSizeInBits() < 16 ?
17916 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17919 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17921 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17922 // to 512-bit vector.
17923 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17924 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17925 // split the vector, perform operation on it's Lo a Hi part and
17926 // concatenate the results.
17927 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17929 MVT VT = Op.getSimpleValueType();
17930 MVT EltVT = VT.getVectorElementType();
17931 unsigned NumElems = VT.getVectorNumElements();
17933 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17934 // Extend to 512 bit vector.
17935 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17936 "Unsupported value type for operation");
17938 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17939 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17940 DAG.getUNDEF(NewVT),
17942 DAG.getIntPtrConstant(0, dl));
17943 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17945 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17946 DAG.getIntPtrConstant(0, dl));
17949 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17950 "Unsupported element type");
17952 if (16 < NumElems) {
17953 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17955 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17956 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17958 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17959 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17961 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17964 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17966 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17967 "Unsupported value type for operation");
17969 // Use native supported vector instruction vplzcntd.
17970 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17971 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17972 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17973 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17975 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17978 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17979 SelectionDAG &DAG) {
17980 MVT VT = Op.getSimpleValueType();
17982 unsigned NumBits = VT.getSizeInBits();
17985 if (VT.isVector() && Subtarget->hasAVX512())
17986 return LowerVectorCTLZ_AVX512(Op, DAG);
17988 Op = Op.getOperand(0);
17989 if (VT == MVT::i8) {
17990 // Zero extend to i32 since there is not an i8 bsr.
17992 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17995 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17996 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17997 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17999 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18002 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
18003 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18006 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18008 // Finally xor with NumBits-1.
18009 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18010 DAG.getConstant(NumBits - 1, dl, OpVT));
18013 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18017 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
18018 SelectionDAG &DAG) {
18019 MVT VT = Op.getSimpleValueType();
18021 unsigned NumBits = VT.getSizeInBits();
18024 Op = Op.getOperand(0);
18025 if (VT == MVT::i8) {
18026 // Zero extend to i32 since there is not an i8 bsr.
18028 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18031 // Issue a bsr (scan bits in reverse).
18032 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18033 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18035 // And xor with NumBits-1.
18036 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
18037 DAG.getConstant(NumBits - 1, dl, OpVT));
18040 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18044 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18045 MVT VT = Op.getSimpleValueType();
18046 unsigned NumBits = VT.getScalarSizeInBits();
18049 if (VT.isVector()) {
18050 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18052 SDValue N0 = Op.getOperand(0);
18053 SDValue Zero = DAG.getConstant(0, dl, VT);
18055 // lsb(x) = (x & -x)
18056 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
18057 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
18059 // cttz_undef(x) = (width - 1) - ctlz(lsb)
18060 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
18061 TLI.isOperationLegal(ISD::CTLZ, VT)) {
18062 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
18063 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
18064 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
18067 // cttz(x) = ctpop(lsb - 1)
18068 SDValue One = DAG.getConstant(1, dl, VT);
18069 return DAG.getNode(ISD::CTPOP, dl, VT,
18070 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
18073 assert(Op.getOpcode() == ISD::CTTZ &&
18074 "Only scalar CTTZ requires custom lowering");
18076 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18077 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18078 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
18080 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18083 DAG.getConstant(NumBits, dl, VT),
18084 DAG.getConstant(X86::COND_E, dl, MVT::i8),
18087 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18090 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18091 // ones, and then concatenate the result back.
18092 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18093 MVT VT = Op.getSimpleValueType();
18095 assert(VT.is256BitVector() && VT.isInteger() &&
18096 "Unsupported value type for operation");
18098 unsigned NumElems = VT.getVectorNumElements();
18101 // Extract the LHS vectors
18102 SDValue LHS = Op.getOperand(0);
18103 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18104 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18106 // Extract the RHS vectors
18107 SDValue RHS = Op.getOperand(1);
18108 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18109 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18111 MVT EltVT = VT.getVectorElementType();
18112 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18114 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18115 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18116 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18119 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18120 if (Op.getValueType() == MVT::i1)
18121 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18122 Op.getOperand(0), Op.getOperand(1));
18123 assert(Op.getSimpleValueType().is256BitVector() &&
18124 Op.getSimpleValueType().isInteger() &&
18125 "Only handle AVX 256-bit vector integer operation");
18126 return Lower256IntArith(Op, DAG);
18129 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18130 if (Op.getValueType() == MVT::i1)
18131 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
18132 Op.getOperand(0), Op.getOperand(1));
18133 assert(Op.getSimpleValueType().is256BitVector() &&
18134 Op.getSimpleValueType().isInteger() &&
18135 "Only handle AVX 256-bit vector integer operation");
18136 return Lower256IntArith(Op, DAG);
18139 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
18140 assert(Op.getSimpleValueType().is256BitVector() &&
18141 Op.getSimpleValueType().isInteger() &&
18142 "Only handle AVX 256-bit vector integer operation");
18143 return Lower256IntArith(Op, DAG);
18146 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18147 SelectionDAG &DAG) {
18149 MVT VT = Op.getSimpleValueType();
18152 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
18154 // Decompose 256-bit ops into smaller 128-bit ops.
18155 if (VT.is256BitVector() && !Subtarget->hasInt256())
18156 return Lower256IntArith(Op, DAG);
18158 SDValue A = Op.getOperand(0);
18159 SDValue B = Op.getOperand(1);
18161 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
18162 // pairs, multiply and truncate.
18163 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
18164 if (Subtarget->hasInt256()) {
18165 if (VT == MVT::v32i8) {
18166 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
18167 SDValue Lo = DAG.getIntPtrConstant(0, dl);
18168 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
18169 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
18170 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
18171 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
18172 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
18173 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18174 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
18175 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
18178 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
18179 return DAG.getNode(
18180 ISD::TRUNCATE, dl, VT,
18181 DAG.getNode(ISD::MUL, dl, ExVT,
18182 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
18183 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
18186 assert(VT == MVT::v16i8 &&
18187 "Pre-AVX2 support only supports v16i8 multiplication");
18188 MVT ExVT = MVT::v8i16;
18190 // Extract the lo parts and sign extend to i16
18192 if (Subtarget->hasSSE41()) {
18193 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
18194 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
18196 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
18197 -1, 4, -1, 5, -1, 6, -1, 7};
18198 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18199 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18200 ALo = DAG.getBitcast(ExVT, ALo);
18201 BLo = DAG.getBitcast(ExVT, BLo);
18202 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
18203 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
18206 // Extract the hi parts and sign extend to i16
18208 if (Subtarget->hasSSE41()) {
18209 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
18210 -1, -1, -1, -1, -1, -1, -1, -1};
18211 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18212 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18213 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
18214 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
18216 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
18217 -1, 12, -1, 13, -1, 14, -1, 15};
18218 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
18219 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
18220 AHi = DAG.getBitcast(ExVT, AHi);
18221 BHi = DAG.getBitcast(ExVT, BHi);
18222 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
18223 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
18226 // Multiply, mask the lower 8bits of the lo/hi results and pack
18227 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
18228 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
18229 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
18230 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
18231 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18234 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18235 if (VT == MVT::v4i32) {
18236 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18237 "Should not custom lower when pmuldq is available!");
18239 // Extract the odd parts.
18240 static const int UnpackMask[] = { 1, -1, 3, -1 };
18241 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18242 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18244 // Multiply the even parts.
18245 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18246 // Now multiply odd parts.
18247 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18249 Evens = DAG.getBitcast(VT, Evens);
18250 Odds = DAG.getBitcast(VT, Odds);
18252 // Merge the two vectors back together with a shuffle. This expands into 2
18254 static const int ShufMask[] = { 0, 4, 2, 6 };
18255 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18258 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18259 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18261 // Ahi = psrlqi(a, 32);
18262 // Bhi = psrlqi(b, 32);
18264 // AloBlo = pmuludq(a, b);
18265 // AloBhi = pmuludq(a, Bhi);
18266 // AhiBlo = pmuludq(Ahi, b);
18268 // AloBhi = psllqi(AloBhi, 32);
18269 // AhiBlo = psllqi(AhiBlo, 32);
18270 // return AloBlo + AloBhi + AhiBlo;
18272 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18273 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18275 SDValue AhiBlo = Ahi;
18276 SDValue AloBhi = Bhi;
18277 // Bit cast to 32-bit vectors for MULUDQ
18278 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18279 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18280 A = DAG.getBitcast(MulVT, A);
18281 B = DAG.getBitcast(MulVT, B);
18282 Ahi = DAG.getBitcast(MulVT, Ahi);
18283 Bhi = DAG.getBitcast(MulVT, Bhi);
18285 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18286 // After shifting right const values the result may be all-zero.
18287 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
18288 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18289 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18291 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
18292 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18293 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18296 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18297 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18300 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18301 assert(Subtarget->isTargetWin64() && "Unexpected target");
18302 EVT VT = Op.getValueType();
18303 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18304 "Unexpected return type for lowering");
18308 switch (Op->getOpcode()) {
18309 default: llvm_unreachable("Unexpected request for libcall!");
18310 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18311 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18312 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18313 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18314 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18315 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18319 SDValue InChain = DAG.getEntryNode();
18321 TargetLowering::ArgListTy Args;
18322 TargetLowering::ArgListEntry Entry;
18323 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18324 EVT ArgVT = Op->getOperand(i).getValueType();
18325 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18326 "Unexpected argument type for lowering");
18327 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18328 Entry.Node = StackPtr;
18329 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18331 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18332 Entry.Ty = PointerType::get(ArgTy,0);
18333 Entry.isSExt = false;
18334 Entry.isZExt = false;
18335 Args.push_back(Entry);
18338 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18339 getPointerTy(DAG.getDataLayout()));
18341 TargetLowering::CallLoweringInfo CLI(DAG);
18342 CLI.setDebugLoc(dl).setChain(InChain)
18343 .setCallee(getLibcallCallingConv(LC),
18344 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18345 Callee, std::move(Args), 0)
18346 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18348 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18349 return DAG.getBitcast(VT, CallInfo.first);
18352 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18353 SelectionDAG &DAG) {
18354 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18355 MVT VT = Op0.getSimpleValueType();
18358 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18359 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18361 // PMULxD operations multiply each even value (starting at 0) of LHS with
18362 // the related value of RHS and produce a widen result.
18363 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18364 // => <2 x i64> <ae|cg>
18366 // In other word, to have all the results, we need to perform two PMULxD:
18367 // 1. one with the even values.
18368 // 2. one with the odd values.
18369 // To achieve #2, with need to place the odd values at an even position.
18371 // Place the odd value at an even position (basically, shift all values 1
18372 // step to the left):
18373 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18374 // <a|b|c|d> => <b|undef|d|undef>
18375 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18376 // <e|f|g|h> => <f|undef|h|undef>
18377 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18379 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18381 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18382 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18384 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18385 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18386 // => <2 x i64> <ae|cg>
18387 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18388 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18389 // => <2 x i64> <bf|dh>
18390 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18392 // Shuffle it back into the right order.
18393 SDValue Highs, Lows;
18394 if (VT == MVT::v8i32) {
18395 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18396 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18397 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18398 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18400 const int HighMask[] = {1, 5, 3, 7};
18401 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18402 const int LowMask[] = {0, 4, 2, 6};
18403 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18406 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18407 // unsigned multiply.
18408 if (IsSigned && !Subtarget->hasSSE41()) {
18409 SDValue ShAmt = DAG.getConstant(
18411 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18412 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18413 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18414 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18415 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18417 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18418 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18421 // The first result of MUL_LOHI is actually the low value, followed by the
18423 SDValue Ops[] = {Lows, Highs};
18424 return DAG.getMergeValues(Ops, dl);
18427 // Return true if the required (according to Opcode) shift-imm form is natively
18428 // supported by the Subtarget
18429 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18431 if (VT.getScalarSizeInBits() < 16)
18434 if (VT.is512BitVector() &&
18435 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18438 bool LShift = VT.is128BitVector() ||
18439 (VT.is256BitVector() && Subtarget->hasInt256());
18441 bool AShift = LShift && (Subtarget->hasVLX() ||
18442 (VT != MVT::v2i64 && VT != MVT::v4i64));
18443 return (Opcode == ISD::SRA) ? AShift : LShift;
18446 // The shift amount is a variable, but it is the same for all vector lanes.
18447 // These instructions are defined together with shift-immediate.
18449 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18451 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18454 // Return true if the required (according to Opcode) variable-shift form is
18455 // natively supported by the Subtarget
18456 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18459 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18462 // vXi16 supported only on AVX-512, BWI
18463 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18466 if (VT.is512BitVector() || Subtarget->hasVLX())
18469 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18470 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18471 return (Opcode == ISD::SRA) ? AShift : LShift;
18474 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18475 const X86Subtarget *Subtarget) {
18476 MVT VT = Op.getSimpleValueType();
18478 SDValue R = Op.getOperand(0);
18479 SDValue Amt = Op.getOperand(1);
18481 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18482 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18484 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18485 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18486 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18487 SDValue Ex = DAG.getBitcast(ExVT, R);
18489 if (ShiftAmt >= 32) {
18490 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18492 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18493 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18494 ShiftAmt - 32, DAG);
18495 if (VT == MVT::v2i64)
18496 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18497 if (VT == MVT::v4i64)
18498 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18499 {9, 1, 11, 3, 13, 5, 15, 7});
18501 // SRA upper i32, SHL whole i64 and select lower i32.
18502 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18505 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18506 Lower = DAG.getBitcast(ExVT, Lower);
18507 if (VT == MVT::v2i64)
18508 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18509 if (VT == MVT::v4i64)
18510 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18511 {8, 1, 10, 3, 12, 5, 14, 7});
18513 return DAG.getBitcast(VT, Ex);
18516 // Optimize shl/srl/sra with constant shift amount.
18517 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18518 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18519 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18521 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18522 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18524 // i64 SRA needs to be performed as partial shifts.
18525 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18526 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18527 return ArithmeticShiftRight64(ShiftAmt);
18529 if (VT == MVT::v16i8 ||
18530 (Subtarget->hasInt256() && VT == MVT::v32i8) ||
18531 VT == MVT::v64i8) {
18532 unsigned NumElts = VT.getVectorNumElements();
18533 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18535 // Simple i8 add case
18536 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18537 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18539 // ashr(R, 7) === cmp_slt(R, 0)
18540 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18541 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18542 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18545 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18546 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18549 if (Op.getOpcode() == ISD::SHL) {
18550 // Make a large shift.
18551 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18553 SHL = DAG.getBitcast(VT, SHL);
18554 // Zero out the rightmost bits.
18555 return DAG.getNode(ISD::AND, dl, VT, SHL,
18556 DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, VT));
18558 if (Op.getOpcode() == ISD::SRL) {
18559 // Make a large shift.
18560 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18562 SRL = DAG.getBitcast(VT, SRL);
18563 // Zero out the leftmost bits.
18564 return DAG.getNode(ISD::AND, dl, VT, SRL,
18565 DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, VT));
18567 if (Op.getOpcode() == ISD::SRA) {
18568 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18569 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18571 SDValue Mask = DAG.getConstant(128 >> ShiftAmt, dl, VT);
18572 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18573 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18576 llvm_unreachable("Unknown shift opcode.");
18581 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18582 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18583 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18585 // Peek through any splat that was introduced for i64 shift vectorization.
18586 int SplatIndex = -1;
18587 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18588 if (SVN->isSplat()) {
18589 SplatIndex = SVN->getSplatIndex();
18590 Amt = Amt.getOperand(0);
18591 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18592 "Splat shuffle referencing second operand");
18595 if (Amt.getOpcode() != ISD::BITCAST ||
18596 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18599 Amt = Amt.getOperand(0);
18600 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18601 VT.getVectorNumElements();
18602 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18603 uint64_t ShiftAmt = 0;
18604 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18605 for (unsigned i = 0; i != Ratio; ++i) {
18606 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18610 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18613 // Check remaining shift amounts (if not a splat).
18614 if (SplatIndex < 0) {
18615 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18616 uint64_t ShAmt = 0;
18617 for (unsigned j = 0; j != Ratio; ++j) {
18618 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18622 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18624 if (ShAmt != ShiftAmt)
18629 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18630 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18632 if (Op.getOpcode() == ISD::SRA)
18633 return ArithmeticShiftRight64(ShiftAmt);
18639 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18640 const X86Subtarget* Subtarget) {
18641 MVT VT = Op.getSimpleValueType();
18643 SDValue R = Op.getOperand(0);
18644 SDValue Amt = Op.getOperand(1);
18646 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18647 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18649 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18650 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18652 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18654 MVT EltVT = VT.getVectorElementType();
18656 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18657 // Check if this build_vector node is doing a splat.
18658 // If so, then set BaseShAmt equal to the splat value.
18659 BaseShAmt = BV->getSplatValue();
18660 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18661 BaseShAmt = SDValue();
18663 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18664 Amt = Amt.getOperand(0);
18666 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18667 if (SVN && SVN->isSplat()) {
18668 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18669 SDValue InVec = Amt.getOperand(0);
18670 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18671 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18672 "Unexpected shuffle index found!");
18673 BaseShAmt = InVec.getOperand(SplatIdx);
18674 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18675 if (ConstantSDNode *C =
18676 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18677 if (C->getZExtValue() == SplatIdx)
18678 BaseShAmt = InVec.getOperand(1);
18683 // Avoid introducing an extract element from a shuffle.
18684 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18685 DAG.getIntPtrConstant(SplatIdx, dl));
18689 if (BaseShAmt.getNode()) {
18690 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18691 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18692 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18693 else if (EltVT.bitsLT(MVT::i32))
18694 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18696 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18700 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18701 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18702 Amt.getOpcode() == ISD::BITCAST &&
18703 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18704 Amt = Amt.getOperand(0);
18705 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18706 VT.getVectorNumElements();
18707 std::vector<SDValue> Vals(Ratio);
18708 for (unsigned i = 0; i != Ratio; ++i)
18709 Vals[i] = Amt.getOperand(i);
18710 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18711 for (unsigned j = 0; j != Ratio; ++j)
18712 if (Vals[j] != Amt.getOperand(i + j))
18716 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18717 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18722 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18723 SelectionDAG &DAG) {
18724 MVT VT = Op.getSimpleValueType();
18726 SDValue R = Op.getOperand(0);
18727 SDValue Amt = Op.getOperand(1);
18729 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18730 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18732 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18735 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18738 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18741 // XOP has 128-bit variable logical/arithmetic shifts.
18742 // +ve/-ve Amt = shift left/right.
18743 if (Subtarget->hasXOP() &&
18744 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18745 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18746 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18747 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18748 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18750 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18751 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18752 if (Op.getOpcode() == ISD::SRA)
18753 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18756 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18757 // shifts per-lane and then shuffle the partial results back together.
18758 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18759 // Splat the shift amounts so the scalar shifts above will catch it.
18760 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18761 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18762 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18763 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18764 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18767 // i64 vector arithmetic shift can be emulated with the transform:
18768 // M = lshr(SIGN_BIT, Amt)
18769 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18770 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18771 Op.getOpcode() == ISD::SRA) {
18772 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18773 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18774 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18775 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18776 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18780 // If possible, lower this packed shift into a vector multiply instead of
18781 // expanding it into a sequence of scalar shifts.
18782 // Do this only if the vector shift count is a constant build_vector.
18783 if (Op.getOpcode() == ISD::SHL &&
18784 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18785 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18786 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18787 SmallVector<SDValue, 8> Elts;
18788 MVT SVT = VT.getVectorElementType();
18789 unsigned SVTBits = SVT.getSizeInBits();
18790 APInt One(SVTBits, 1);
18791 unsigned NumElems = VT.getVectorNumElements();
18793 for (unsigned i=0; i !=NumElems; ++i) {
18794 SDValue Op = Amt->getOperand(i);
18795 if (Op->getOpcode() == ISD::UNDEF) {
18796 Elts.push_back(Op);
18800 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18801 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18802 uint64_t ShAmt = C.getZExtValue();
18803 if (ShAmt >= SVTBits) {
18804 Elts.push_back(DAG.getUNDEF(SVT));
18807 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18809 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18810 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18813 // Lower SHL with variable shift amount.
18814 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18815 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18817 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18818 DAG.getConstant(0x3f800000U, dl, VT));
18819 Op = DAG.getBitcast(MVT::v4f32, Op);
18820 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18821 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18824 // If possible, lower this shift as a sequence of two shifts by
18825 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18827 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18829 // Could be rewritten as:
18830 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18832 // The advantage is that the two shifts from the example would be
18833 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18834 // the vector shift into four scalar shifts plus four pairs of vector
18836 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18837 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18838 unsigned TargetOpcode = X86ISD::MOVSS;
18839 bool CanBeSimplified;
18840 // The splat value for the first packed shift (the 'X' from the example).
18841 SDValue Amt1 = Amt->getOperand(0);
18842 // The splat value for the second packed shift (the 'Y' from the example).
18843 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18844 Amt->getOperand(2);
18846 // See if it is possible to replace this node with a sequence of
18847 // two shifts followed by a MOVSS/MOVSD
18848 if (VT == MVT::v4i32) {
18849 // Check if it is legal to use a MOVSS.
18850 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18851 Amt2 == Amt->getOperand(3);
18852 if (!CanBeSimplified) {
18853 // Otherwise, check if we can still simplify this node using a MOVSD.
18854 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18855 Amt->getOperand(2) == Amt->getOperand(3);
18856 TargetOpcode = X86ISD::MOVSD;
18857 Amt2 = Amt->getOperand(2);
18860 // Do similar checks for the case where the machine value type
18862 CanBeSimplified = Amt1 == Amt->getOperand(1);
18863 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18864 CanBeSimplified = Amt2 == Amt->getOperand(i);
18866 if (!CanBeSimplified) {
18867 TargetOpcode = X86ISD::MOVSD;
18868 CanBeSimplified = true;
18869 Amt2 = Amt->getOperand(4);
18870 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18871 CanBeSimplified = Amt1 == Amt->getOperand(i);
18872 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18873 CanBeSimplified = Amt2 == Amt->getOperand(j);
18877 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18878 isa<ConstantSDNode>(Amt2)) {
18879 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18880 MVT CastVT = MVT::v4i32;
18882 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18883 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18885 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18886 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18887 if (TargetOpcode == X86ISD::MOVSD)
18888 CastVT = MVT::v2i64;
18889 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18890 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18891 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18893 return DAG.getBitcast(VT, Result);
18897 // v4i32 Non Uniform Shifts.
18898 // If the shift amount is constant we can shift each lane using the SSE2
18899 // immediate shifts, else we need to zero-extend each lane to the lower i64
18900 // and shift using the SSE2 variable shifts.
18901 // The separate results can then be blended together.
18902 if (VT == MVT::v4i32) {
18903 unsigned Opc = Op.getOpcode();
18904 SDValue Amt0, Amt1, Amt2, Amt3;
18905 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18906 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18907 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18908 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18909 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18911 // ISD::SHL is handled above but we include it here for completeness.
18914 llvm_unreachable("Unknown target vector shift node");
18916 Opc = X86ISD::VSHL;
18919 Opc = X86ISD::VSRL;
18922 Opc = X86ISD::VSRA;
18925 // The SSE2 shifts use the lower i64 as the same shift amount for
18926 // all lanes and the upper i64 is ignored. These shuffle masks
18927 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18928 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18929 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18930 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18931 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18932 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18935 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18936 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18937 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18938 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18939 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18940 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18941 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18944 if (VT == MVT::v16i8 ||
18945 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18946 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18947 unsigned ShiftOpcode = Op->getOpcode();
18949 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18950 // On SSE41 targets we make use of the fact that VSELECT lowers
18951 // to PBLENDVB which selects bytes based just on the sign bit.
18952 if (Subtarget->hasSSE41()) {
18953 V0 = DAG.getBitcast(VT, V0);
18954 V1 = DAG.getBitcast(VT, V1);
18955 Sel = DAG.getBitcast(VT, Sel);
18956 return DAG.getBitcast(SelVT,
18957 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18959 // On pre-SSE41 targets we test for the sign bit by comparing to
18960 // zero - a negative value will set all bits of the lanes to true
18961 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18962 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18963 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18964 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18967 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18968 // We can safely do this using i16 shifts as we're only interested in
18969 // the 3 lower bits of each byte.
18970 Amt = DAG.getBitcast(ExtVT, Amt);
18971 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18972 Amt = DAG.getBitcast(VT, Amt);
18974 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18975 // r = VSELECT(r, shift(r, 4), a);
18977 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18978 R = SignBitSelect(VT, Amt, M, R);
18981 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18983 // r = VSELECT(r, shift(r, 2), a);
18984 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18985 R = SignBitSelect(VT, Amt, M, R);
18988 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18990 // return VSELECT(r, shift(r, 1), a);
18991 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18992 R = SignBitSelect(VT, Amt, M, R);
18996 if (Op->getOpcode() == ISD::SRA) {
18997 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18998 // so we can correctly sign extend. We don't care what happens to the
19000 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
19001 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
19002 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
19003 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
19004 ALo = DAG.getBitcast(ExtVT, ALo);
19005 AHi = DAG.getBitcast(ExtVT, AHi);
19006 RLo = DAG.getBitcast(ExtVT, RLo);
19007 RHi = DAG.getBitcast(ExtVT, RHi);
19009 // r = VSELECT(r, shift(r, 4), a);
19010 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19011 DAG.getConstant(4, dl, ExtVT));
19012 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19013 DAG.getConstant(4, dl, ExtVT));
19014 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19015 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19018 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19019 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19021 // r = VSELECT(r, shift(r, 2), a);
19022 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19023 DAG.getConstant(2, dl, ExtVT));
19024 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19025 DAG.getConstant(2, dl, ExtVT));
19026 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19027 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19030 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
19031 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
19033 // r = VSELECT(r, shift(r, 1), a);
19034 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
19035 DAG.getConstant(1, dl, ExtVT));
19036 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
19037 DAG.getConstant(1, dl, ExtVT));
19038 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
19039 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
19041 // Logical shift the result back to the lower byte, leaving a zero upper
19043 // meaning that we can safely pack with PACKUSWB.
19045 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
19047 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
19048 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
19052 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
19053 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
19054 // solution better.
19055 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
19056 MVT ExtVT = MVT::v8i32;
19058 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
19059 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
19060 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
19061 return DAG.getNode(ISD::TRUNCATE, dl, VT,
19062 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
19065 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
19066 MVT ExtVT = MVT::v8i32;
19067 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
19068 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
19069 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
19070 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
19071 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
19072 ALo = DAG.getBitcast(ExtVT, ALo);
19073 AHi = DAG.getBitcast(ExtVT, AHi);
19074 RLo = DAG.getBitcast(ExtVT, RLo);
19075 RHi = DAG.getBitcast(ExtVT, RHi);
19076 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
19077 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
19078 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
19079 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
19080 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
19083 if (VT == MVT::v8i16) {
19084 unsigned ShiftOpcode = Op->getOpcode();
19086 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
19087 // On SSE41 targets we make use of the fact that VSELECT lowers
19088 // to PBLENDVB which selects bytes based just on the sign bit.
19089 if (Subtarget->hasSSE41()) {
19090 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
19091 V0 = DAG.getBitcast(ExtVT, V0);
19092 V1 = DAG.getBitcast(ExtVT, V1);
19093 Sel = DAG.getBitcast(ExtVT, Sel);
19094 return DAG.getBitcast(
19095 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
19097 // On pre-SSE41 targets we splat the sign bit - a negative value will
19098 // set all bits of the lanes to true and VSELECT uses that in
19099 // its OR(AND(V0,C),AND(V1,~C)) lowering.
19101 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
19102 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
19105 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
19106 if (Subtarget->hasSSE41()) {
19107 // On SSE41 targets we need to replicate the shift mask in both
19108 // bytes for PBLENDVB.
19111 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
19112 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
19114 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
19117 // r = VSELECT(r, shift(r, 8), a);
19118 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
19119 R = SignBitSelect(Amt, M, R);
19122 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19124 // r = VSELECT(r, shift(r, 4), a);
19125 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
19126 R = SignBitSelect(Amt, M, R);
19129 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19131 // r = VSELECT(r, shift(r, 2), a);
19132 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
19133 R = SignBitSelect(Amt, M, R);
19136 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
19138 // return VSELECT(r, shift(r, 1), a);
19139 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
19140 R = SignBitSelect(Amt, M, R);
19144 // Decompose 256-bit shifts into smaller 128-bit shifts.
19145 if (VT.is256BitVector()) {
19146 unsigned NumElems = VT.getVectorNumElements();
19147 MVT EltVT = VT.getVectorElementType();
19148 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19150 // Extract the two vectors
19151 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
19152 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
19154 // Recreate the shift amount vectors
19155 SDValue Amt1, Amt2;
19156 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
19157 // Constant shift amount
19158 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
19159 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
19160 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
19162 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
19163 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19165 // Variable shift amount
19166 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19167 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19170 // Issue new vector shifts for the smaller types
19171 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19172 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19174 // Concatenate the result back
19175 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19181 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
19182 SelectionDAG &DAG) {
19183 MVT VT = Op.getSimpleValueType();
19185 SDValue R = Op.getOperand(0);
19186 SDValue Amt = Op.getOperand(1);
19188 assert(VT.isVector() && "Custom lowering only for vector rotates!");
19189 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
19190 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
19192 // XOP has 128-bit vector variable + immediate rotates.
19193 // +ve/-ve Amt = rotate left/right.
19195 // Split 256-bit integers.
19196 if (VT.is256BitVector())
19197 return Lower256IntArith(Op, DAG);
19199 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
19201 // Attempt to rotate by immediate.
19202 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
19203 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
19204 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
19205 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
19206 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
19207 DAG.getConstant(RotateAmt, DL, MVT::i8));
19211 // Use general rotate by variable (per-element).
19212 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
19215 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19216 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19217 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19218 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19219 // has only one use.
19220 SDNode *N = Op.getNode();
19221 SDValue LHS = N->getOperand(0);
19222 SDValue RHS = N->getOperand(1);
19223 unsigned BaseOp = 0;
19226 switch (Op.getOpcode()) {
19227 default: llvm_unreachable("Unknown ovf instruction!");
19229 // A subtract of one will be selected as a INC. Note that INC doesn't
19230 // set CF, so we can't do this for UADDO.
19231 if (isOneConstant(RHS)) {
19232 BaseOp = X86ISD::INC;
19233 Cond = X86::COND_O;
19236 BaseOp = X86ISD::ADD;
19237 Cond = X86::COND_O;
19240 BaseOp = X86ISD::ADD;
19241 Cond = X86::COND_B;
19244 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19245 // set CF, so we can't do this for USUBO.
19246 if (isOneConstant(RHS)) {
19247 BaseOp = X86ISD::DEC;
19248 Cond = X86::COND_O;
19251 BaseOp = X86ISD::SUB;
19252 Cond = X86::COND_O;
19255 BaseOp = X86ISD::SUB;
19256 Cond = X86::COND_B;
19259 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19260 Cond = X86::COND_O;
19262 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19263 if (N->getValueType(0) == MVT::i8) {
19264 BaseOp = X86ISD::UMUL8;
19265 Cond = X86::COND_O;
19268 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19270 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19273 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19274 DAG.getConstant(X86::COND_O, DL, MVT::i32),
19275 SDValue(Sum.getNode(), 2));
19277 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19281 // Also sets EFLAGS.
19282 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19283 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19286 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19287 DAG.getConstant(Cond, DL, MVT::i32),
19288 SDValue(Sum.getNode(), 1));
19290 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19293 /// Returns true if the operand type is exactly twice the native width, and
19294 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19295 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19296 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19297 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
19298 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19301 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19302 else if (OpWidth == 128)
19303 return Subtarget->hasCmpxchg16b();
19308 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19309 return needsCmpXchgNb(SI->getValueOperand()->getType());
19312 // Note: this turns large loads into lock cmpxchg8b/16b.
19313 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19314 TargetLowering::AtomicExpansionKind
19315 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19316 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19317 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
19318 : AtomicExpansionKind::None;
19321 TargetLowering::AtomicExpansionKind
19322 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19323 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19324 Type *MemType = AI->getType();
19326 // If the operand is too big, we must see if cmpxchg8/16b is available
19327 // and default to library calls otherwise.
19328 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
19329 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
19330 : AtomicExpansionKind::None;
19333 AtomicRMWInst::BinOp Op = AI->getOperation();
19336 llvm_unreachable("Unknown atomic operation");
19337 case AtomicRMWInst::Xchg:
19338 case AtomicRMWInst::Add:
19339 case AtomicRMWInst::Sub:
19340 // It's better to use xadd, xsub or xchg for these in all cases.
19341 return AtomicExpansionKind::None;
19342 case AtomicRMWInst::Or:
19343 case AtomicRMWInst::And:
19344 case AtomicRMWInst::Xor:
19345 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19346 // prefix to a normal instruction for these operations.
19347 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19348 : AtomicExpansionKind::None;
19349 case AtomicRMWInst::Nand:
19350 case AtomicRMWInst::Max:
19351 case AtomicRMWInst::Min:
19352 case AtomicRMWInst::UMax:
19353 case AtomicRMWInst::UMin:
19354 // These always require a non-trivial set of data operations on x86. We must
19355 // use a cmpxchg loop.
19356 return AtomicExpansionKind::CmpXChg;
19360 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19361 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19362 // no-sse2). There isn't any reason to disable it if the target processor
19364 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19368 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19369 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19370 Type *MemType = AI->getType();
19371 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19372 // there is no benefit in turning such RMWs into loads, and it is actually
19373 // harmful as it introduces a mfence.
19374 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19377 auto Builder = IRBuilder<>(AI);
19378 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19379 auto SynchScope = AI->getSynchScope();
19380 // We must restrict the ordering to avoid generating loads with Release or
19381 // ReleaseAcquire orderings.
19382 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19383 auto Ptr = AI->getPointerOperand();
19385 // Before the load we need a fence. Here is an example lifted from
19386 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19389 // x.store(1, relaxed);
19390 // r1 = y.fetch_add(0, release);
19392 // y.fetch_add(42, acquire);
19393 // r2 = x.load(relaxed);
19394 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19395 // lowered to just a load without a fence. A mfence flushes the store buffer,
19396 // making the optimization clearly correct.
19397 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19398 // otherwise, we might be able to be more aggressive on relaxed idempotent
19399 // rmw. In practice, they do not look useful, so we don't try to be
19400 // especially clever.
19401 if (SynchScope == SingleThread)
19402 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19403 // the IR level, so we must wrap it in an intrinsic.
19406 if (!hasMFENCE(*Subtarget))
19407 // FIXME: it might make sense to use a locked operation here but on a
19408 // different cache-line to prevent cache-line bouncing. In practice it
19409 // is probably a small win, and x86 processors without mfence are rare
19410 // enough that we do not bother.
19414 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19415 Builder.CreateCall(MFence, {});
19417 // Finally we can emit the atomic load.
19418 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19419 AI->getType()->getPrimitiveSizeInBits());
19420 Loaded->setAtomic(Order, SynchScope);
19421 AI->replaceAllUsesWith(Loaded);
19422 AI->eraseFromParent();
19426 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19427 SelectionDAG &DAG) {
19429 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19430 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19431 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19432 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19434 // The only fence that needs an instruction is a sequentially-consistent
19435 // cross-thread fence.
19436 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19437 if (hasMFENCE(*Subtarget))
19438 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19440 SDValue Chain = Op.getOperand(0);
19441 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19443 DAG.getRegister(X86::ESP, MVT::i32), // Base
19444 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19445 DAG.getRegister(0, MVT::i32), // Index
19446 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19447 DAG.getRegister(0, MVT::i32), // Segment.
19451 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19452 return SDValue(Res, 0);
19455 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19456 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19459 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19460 SelectionDAG &DAG) {
19461 MVT T = Op.getSimpleValueType();
19465 switch(T.SimpleTy) {
19466 default: llvm_unreachable("Invalid value type!");
19467 case MVT::i8: Reg = X86::AL; size = 1; break;
19468 case MVT::i16: Reg = X86::AX; size = 2; break;
19469 case MVT::i32: Reg = X86::EAX; size = 4; break;
19471 assert(Subtarget->is64Bit() && "Node not type legal!");
19472 Reg = X86::RAX; size = 8;
19475 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19476 Op.getOperand(2), SDValue());
19477 SDValue Ops[] = { cpIn.getValue(0),
19480 DAG.getTargetConstant(size, DL, MVT::i8),
19481 cpIn.getValue(1) };
19482 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19483 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19484 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19488 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19489 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19490 MVT::i32, cpOut.getValue(2));
19491 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19492 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19495 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19496 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19497 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19501 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19502 SelectionDAG &DAG) {
19503 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19504 MVT DstVT = Op.getSimpleValueType();
19506 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19507 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19508 if (DstVT != MVT::f64)
19509 // This conversion needs to be expanded.
19512 SDValue InVec = Op->getOperand(0);
19514 unsigned NumElts = SrcVT.getVectorNumElements();
19515 MVT SVT = SrcVT.getVectorElementType();
19517 // Widen the vector in input in the case of MVT::v2i32.
19518 // Example: from MVT::v2i32 to MVT::v4i32.
19519 SmallVector<SDValue, 16> Elts;
19520 for (unsigned i = 0, e = NumElts; i != e; ++i)
19521 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19522 DAG.getIntPtrConstant(i, dl)));
19524 // Explicitly mark the extra elements as Undef.
19525 Elts.append(NumElts, DAG.getUNDEF(SVT));
19527 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19528 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19529 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19530 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19531 DAG.getIntPtrConstant(0, dl));
19534 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19535 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19536 assert((DstVT == MVT::i64 ||
19537 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19538 "Unexpected custom BITCAST");
19539 // i64 <=> MMX conversions are Legal.
19540 if (SrcVT==MVT::i64 && DstVT.isVector())
19542 if (DstVT==MVT::i64 && SrcVT.isVector())
19544 // MMX <=> MMX conversions are Legal.
19545 if (SrcVT.isVector() && DstVT.isVector())
19547 // All other conversions need to be expanded.
19551 /// Compute the horizontal sum of bytes in V for the elements of VT.
19553 /// Requires V to be a byte vector and VT to be an integer vector type with
19554 /// wider elements than V's type. The width of the elements of VT determines
19555 /// how many bytes of V are summed horizontally to produce each element of the
19557 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19558 const X86Subtarget *Subtarget,
19559 SelectionDAG &DAG) {
19561 MVT ByteVecVT = V.getSimpleValueType();
19562 MVT EltVT = VT.getVectorElementType();
19563 int NumElts = VT.getVectorNumElements();
19564 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19565 "Expected value to have byte element type.");
19566 assert(EltVT != MVT::i8 &&
19567 "Horizontal byte sum only makes sense for wider elements!");
19568 unsigned VecSize = VT.getSizeInBits();
19569 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19571 // PSADBW instruction horizontally add all bytes and leave the result in i64
19572 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19573 if (EltVT == MVT::i64) {
19574 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19575 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19576 V = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT, V, Zeros);
19577 return DAG.getBitcast(VT, V);
19580 if (EltVT == MVT::i32) {
19581 // We unpack the low half and high half into i32s interleaved with zeros so
19582 // that we can use PSADBW to horizontally sum them. The most useful part of
19583 // this is that it lines up the results of two PSADBW instructions to be
19584 // two v2i64 vectors which concatenated are the 4 population counts. We can
19585 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19586 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19587 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19588 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19590 // Do the horizontal sums into two v2i64s.
19591 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19592 MVT SadVecVT = MVT::getVectorVT(MVT::i64, VecSize / 64);
19593 Low = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19594 DAG.getBitcast(ByteVecVT, Low), Zeros);
19595 High = DAG.getNode(X86ISD::PSADBW, DL, SadVecVT,
19596 DAG.getBitcast(ByteVecVT, High), Zeros);
19598 // Merge them together.
19599 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19600 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19601 DAG.getBitcast(ShortVecVT, Low),
19602 DAG.getBitcast(ShortVecVT, High));
19604 return DAG.getBitcast(VT, V);
19607 // The only element type left is i16.
19608 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19610 // To obtain pop count for each i16 element starting from the pop count for
19611 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19612 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19613 // directly supported.
19614 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19615 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19616 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19617 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19618 DAG.getBitcast(ByteVecVT, V));
19619 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19622 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19623 const X86Subtarget *Subtarget,
19624 SelectionDAG &DAG) {
19625 MVT VT = Op.getSimpleValueType();
19626 MVT EltVT = VT.getVectorElementType();
19627 unsigned VecSize = VT.getSizeInBits();
19629 // Implement a lookup table in register by using an algorithm based on:
19630 // http://wm.ite.pl/articles/sse-popcount.html
19632 // The general idea is that every lower byte nibble in the input vector is an
19633 // index into a in-register pre-computed pop count table. We then split up the
19634 // input vector in two new ones: (1) a vector with only the shifted-right
19635 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19636 // masked out higher ones) for each byte. PSHUB is used separately with both
19637 // to index the in-register table. Next, both are added and the result is a
19638 // i8 vector where each element contains the pop count for input byte.
19640 // To obtain the pop count for elements != i8, we follow up with the same
19641 // approach and use additional tricks as described below.
19643 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19644 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19645 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19646 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19648 int NumByteElts = VecSize / 8;
19649 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19650 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19651 SmallVector<SDValue, 16> LUTVec;
19652 for (int i = 0; i < NumByteElts; ++i)
19653 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19654 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19655 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19656 DAG.getConstant(0x0F, DL, MVT::i8));
19657 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19660 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19661 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19662 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19665 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19667 // The input vector is used as the shuffle mask that index elements into the
19668 // LUT. After counting low and high nibbles, add the vector to obtain the
19669 // final pop count per i8 element.
19670 SDValue HighPopCnt =
19671 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19672 SDValue LowPopCnt =
19673 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19674 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19676 if (EltVT == MVT::i8)
19679 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19682 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19683 const X86Subtarget *Subtarget,
19684 SelectionDAG &DAG) {
19685 MVT VT = Op.getSimpleValueType();
19686 assert(VT.is128BitVector() &&
19687 "Only 128-bit vector bitmath lowering supported.");
19689 int VecSize = VT.getSizeInBits();
19690 MVT EltVT = VT.getVectorElementType();
19691 int Len = EltVT.getSizeInBits();
19693 // This is the vectorized version of the "best" algorithm from
19694 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19695 // with a minor tweak to use a series of adds + shifts instead of vector
19696 // multiplications. Implemented for all integer vector types. We only use
19697 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19698 // much faster, even faster than using native popcnt instructions.
19700 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19701 MVT VT = V.getSimpleValueType();
19702 SmallVector<SDValue, 32> Shifters(
19703 VT.getVectorNumElements(),
19704 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19705 return DAG.getNode(OpCode, DL, VT, V,
19706 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19708 auto GetMask = [&](SDValue V, APInt Mask) {
19709 MVT VT = V.getSimpleValueType();
19710 SmallVector<SDValue, 32> Masks(
19711 VT.getVectorNumElements(),
19712 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19713 return DAG.getNode(ISD::AND, DL, VT, V,
19714 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19717 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19718 // x86, so set the SRL type to have elements at least i16 wide. This is
19719 // correct because all of our SRLs are followed immediately by a mask anyways
19720 // that handles any bits that sneak into the high bits of the byte elements.
19721 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19725 // v = v - ((v >> 1) & 0x55555555...)
19727 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19728 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19729 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19731 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19732 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19733 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19734 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19735 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19737 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19738 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19739 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19740 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19742 // At this point, V contains the byte-wise population count, and we are
19743 // merely doing a horizontal sum if necessary to get the wider element
19745 if (EltVT == MVT::i8)
19748 return LowerHorizontalByteSum(
19749 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19753 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19754 SelectionDAG &DAG) {
19755 MVT VT = Op.getSimpleValueType();
19756 // FIXME: Need to add AVX-512 support here!
19757 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19758 "Unknown CTPOP type to handle");
19759 SDLoc DL(Op.getNode());
19760 SDValue Op0 = Op.getOperand(0);
19762 if (!Subtarget->hasSSSE3()) {
19763 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19764 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19765 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19768 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19769 unsigned NumElems = VT.getVectorNumElements();
19771 // Extract each 128-bit vector, compute pop count and concat the result.
19772 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19773 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19775 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19776 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19777 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19780 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19783 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19784 SelectionDAG &DAG) {
19785 assert(Op.getSimpleValueType().isVector() &&
19786 "We only do custom lowering for vector population count.");
19787 return LowerVectorCTPOP(Op, Subtarget, DAG);
19790 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19791 SDNode *Node = Op.getNode();
19793 EVT T = Node->getValueType(0);
19794 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19795 DAG.getConstant(0, dl, T), Node->getOperand(2));
19796 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19797 cast<AtomicSDNode>(Node)->getMemoryVT(),
19798 Node->getOperand(0),
19799 Node->getOperand(1), negOp,
19800 cast<AtomicSDNode>(Node)->getMemOperand(),
19801 cast<AtomicSDNode>(Node)->getOrdering(),
19802 cast<AtomicSDNode>(Node)->getSynchScope());
19805 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19806 SDNode *Node = Op.getNode();
19808 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19810 // Convert seq_cst store -> xchg
19811 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19812 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19813 // (The only way to get a 16-byte store is cmpxchg16b)
19814 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19815 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19816 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19817 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19818 cast<AtomicSDNode>(Node)->getMemoryVT(),
19819 Node->getOperand(0),
19820 Node->getOperand(1), Node->getOperand(2),
19821 cast<AtomicSDNode>(Node)->getMemOperand(),
19822 cast<AtomicSDNode>(Node)->getOrdering(),
19823 cast<AtomicSDNode>(Node)->getSynchScope());
19824 return Swap.getValue(1);
19826 // Other atomic stores have a simple pattern.
19830 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19831 MVT VT = Op.getNode()->getSimpleValueType(0);
19833 // Let legalize expand this if it isn't a legal type yet.
19834 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19837 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19840 bool ExtraOp = false;
19841 switch (Op.getOpcode()) {
19842 default: llvm_unreachable("Invalid code");
19843 case ISD::ADDC: Opc = X86ISD::ADD; break;
19844 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19845 case ISD::SUBC: Opc = X86ISD::SUB; break;
19846 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19850 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19852 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19853 Op.getOperand(1), Op.getOperand(2));
19856 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19857 SelectionDAG &DAG) {
19858 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19860 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19861 // which returns the values as { float, float } (in XMM0) or
19862 // { double, double } (which is returned in XMM0, XMM1).
19864 SDValue Arg = Op.getOperand(0);
19865 EVT ArgVT = Arg.getValueType();
19866 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19868 TargetLowering::ArgListTy Args;
19869 TargetLowering::ArgListEntry Entry;
19873 Entry.isSExt = false;
19874 Entry.isZExt = false;
19875 Args.push_back(Entry);
19877 bool isF64 = ArgVT == MVT::f64;
19878 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19879 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19880 // the results are returned via SRet in memory.
19881 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19882 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19884 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19886 Type *RetTy = isF64
19887 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19888 : (Type*)VectorType::get(ArgTy, 4);
19890 TargetLowering::CallLoweringInfo CLI(DAG);
19891 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19892 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19894 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19897 // Returned in xmm0 and xmm1.
19898 return CallResult.first;
19900 // Returned in bits 0:31 and 32:64 xmm0.
19901 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19902 CallResult.first, DAG.getIntPtrConstant(0, dl));
19903 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19904 CallResult.first, DAG.getIntPtrConstant(1, dl));
19905 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19906 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19909 /// Widen a vector input to a vector of NVT. The
19910 /// input vector must have the same element type as NVT.
19911 static SDValue ExtendToType(SDValue InOp, MVT NVT, SelectionDAG &DAG,
19912 bool FillWithZeroes = false) {
19913 // Check if InOp already has the right width.
19914 MVT InVT = InOp.getSimpleValueType();
19918 if (InOp.isUndef())
19919 return DAG.getUNDEF(NVT);
19921 assert(InVT.getVectorElementType() == NVT.getVectorElementType() &&
19922 "input and widen element type must match");
19924 unsigned InNumElts = InVT.getVectorNumElements();
19925 unsigned WidenNumElts = NVT.getVectorNumElements();
19926 assert(WidenNumElts > InNumElts && WidenNumElts % InNumElts == 0 &&
19927 "Unexpected request for vector widening");
19929 EVT EltVT = NVT.getVectorElementType();
19932 if (InOp.getOpcode() == ISD::CONCAT_VECTORS &&
19933 InOp.getNumOperands() == 2) {
19934 SDValue N1 = InOp.getOperand(1);
19935 if ((ISD::isBuildVectorAllZeros(N1.getNode()) && FillWithZeroes) ||
19937 InOp = InOp.getOperand(0);
19938 InVT = InOp.getSimpleValueType();
19939 InNumElts = InVT.getVectorNumElements();
19942 if (ISD::isBuildVectorOfConstantSDNodes(InOp.getNode()) ||
19943 ISD::isBuildVectorOfConstantFPSDNodes(InOp.getNode())) {
19944 SmallVector<SDValue, 16> Ops;
19945 for (unsigned i = 0; i < InNumElts; ++i)
19946 Ops.push_back(InOp.getOperand(i));
19948 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, EltVT) :
19949 DAG.getUNDEF(EltVT);
19950 for (unsigned i = 0; i < WidenNumElts - InNumElts; ++i)
19951 Ops.push_back(FillVal);
19952 return DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Ops);
19954 SDValue FillVal = FillWithZeroes ? DAG.getConstant(0, dl, NVT) :
19956 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NVT, FillVal,
19957 InOp, DAG.getIntPtrConstant(0, dl));
19960 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19961 SelectionDAG &DAG) {
19962 assert(Subtarget->hasAVX512() &&
19963 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19965 // X86 scatter kills mask register, so its type should be added to
19966 // the list of return values.
19967 // If the "scatter" has 2 return values, it is already handled.
19968 if (Op.getNode()->getNumValues() == 2)
19971 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19972 SDValue Src = N->getValue();
19973 MVT VT = Src.getSimpleValueType();
19974 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19977 SDValue NewScatter;
19978 SDValue Index = N->getIndex();
19979 SDValue Mask = N->getMask();
19980 SDValue Chain = N->getChain();
19981 SDValue BasePtr = N->getBasePtr();
19982 MVT MemVT = N->getMemoryVT().getSimpleVT();
19983 MVT IndexVT = Index.getSimpleValueType();
19984 MVT MaskVT = Mask.getSimpleValueType();
19986 if (MemVT.getScalarSizeInBits() < VT.getScalarSizeInBits()) {
19987 // The v2i32 value was promoted to v2i64.
19988 // Now we "redo" the type legalizer's work and widen the original
19989 // v2i32 value to v4i32. The original v2i32 is retrieved from v2i64
19991 assert((MemVT == MVT::v2i32 && VT == MVT::v2i64) &&
19992 "Unexpected memory type");
19993 int ShuffleMask[] = {0, 2, -1, -1};
19994 Src = DAG.getVectorShuffle(MVT::v4i32, dl, DAG.getBitcast(MVT::v4i32, Src),
19995 DAG.getUNDEF(MVT::v4i32), ShuffleMask);
19996 // Now we have 4 elements instead of 2.
19997 // Expand the index.
19998 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), 4);
19999 Index = ExtendToType(Index, NewIndexVT, DAG);
20001 // Expand the mask with zeroes
20002 // Mask may be <2 x i64> or <2 x i1> at this moment
20003 assert((MaskVT == MVT::v2i1 || MaskVT == MVT::v2i64) &&
20004 "Unexpected mask type");
20005 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), 4);
20006 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20010 unsigned NumElts = VT.getVectorNumElements();
20011 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20012 !Index.getSimpleValueType().is512BitVector()) {
20013 // AVX512F supports only 512-bit vectors. Or data or index should
20014 // be 512 bit wide. If now the both index and data are 256-bit, but
20015 // the vector contains 8 elements, we just sign-extend the index
20016 if (IndexVT == MVT::v8i32)
20017 // Just extend index
20018 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20020 // The minimal number of elts in scatter is 8
20023 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20024 // Use original index here, do not modify the index twice
20025 Index = ExtendToType(N->getIndex(), NewIndexVT, DAG);
20026 if (IndexVT.getScalarType() == MVT::i32)
20027 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20030 // At this point we have promoted mask operand
20031 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20032 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20033 // Use the original mask here, do not modify the mask twice
20034 Mask = ExtendToType(N->getMask(), ExtMaskVT, DAG, true);
20036 // The value that should be stored
20037 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20038 Src = ExtendToType(Src, NewVT, DAG);
20041 // If the mask is "wide" at this point - truncate it to i1 vector
20042 MVT BitMaskVT = MVT::getVectorVT(MVT::i1, NumElts);
20043 Mask = DAG.getNode(ISD::TRUNCATE, dl, BitMaskVT, Mask);
20045 // The mask is killed by scatter, add it to the values
20046 SDVTList VTs = DAG.getVTList(BitMaskVT, MVT::Other);
20047 SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
20048 NewScatter = DAG.getMaskedScatter(VTs, N->getMemoryVT(), dl, Ops,
20049 N->getMemOperand());
20050 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
20051 return SDValue(NewScatter.getNode(), 0);
20054 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget *Subtarget,
20055 SelectionDAG &DAG) {
20057 MaskedLoadSDNode *N = cast<MaskedLoadSDNode>(Op.getNode());
20058 MVT VT = Op.getSimpleValueType();
20059 SDValue Mask = N->getMask();
20062 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20063 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20064 // This operation is legal for targets with VLX, but without
20065 // VLX the vector should be widened to 512 bit
20066 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20067 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20068 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20069 SDValue Src0 = N->getSrc0();
20070 Src0 = ExtendToType(Src0, WideDataVT, DAG);
20071 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20072 SDValue NewLoad = DAG.getMaskedLoad(WideDataVT, dl, N->getChain(),
20073 N->getBasePtr(), Mask, Src0,
20074 N->getMemoryVT(), N->getMemOperand(),
20075 N->getExtensionType());
20077 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20078 NewLoad.getValue(0),
20079 DAG.getIntPtrConstant(0, dl));
20080 SDValue RetOps[] = {Exract, NewLoad.getValue(1)};
20081 return DAG.getMergeValues(RetOps, dl);
20086 static SDValue LowerMSTORE(SDValue Op, const X86Subtarget *Subtarget,
20087 SelectionDAG &DAG) {
20088 MaskedStoreSDNode *N = cast<MaskedStoreSDNode>(Op.getNode());
20089 SDValue DataToStore = N->getValue();
20090 MVT VT = DataToStore.getSimpleValueType();
20091 SDValue Mask = N->getMask();
20094 if (Subtarget->hasAVX512() && !Subtarget->hasVLX() &&
20095 !VT.is512BitVector() && Mask.getValueType() == MVT::v8i1) {
20096 // This operation is legal for targets with VLX, but without
20097 // VLX the vector should be widened to 512 bit
20098 unsigned NumEltsInWideVec = 512/VT.getScalarSizeInBits();
20099 MVT WideDataVT = MVT::getVectorVT(VT.getScalarType(), NumEltsInWideVec);
20100 MVT WideMaskVT = MVT::getVectorVT(MVT::i1, NumEltsInWideVec);
20101 DataToStore = ExtendToType(DataToStore, WideDataVT, DAG);
20102 Mask = ExtendToType(Mask, WideMaskVT, DAG, true);
20103 return DAG.getMaskedStore(N->getChain(), dl, DataToStore, N->getBasePtr(),
20104 Mask, N->getMemoryVT(), N->getMemOperand(),
20105 N->isTruncatingStore());
20110 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
20111 SelectionDAG &DAG) {
20112 assert(Subtarget->hasAVX512() &&
20113 "MGATHER/MSCATTER are supported on AVX-512 arch only");
20115 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
20117 MVT VT = Op.getSimpleValueType();
20118 SDValue Index = N->getIndex();
20119 SDValue Mask = N->getMask();
20120 SDValue Src0 = N->getValue();
20121 MVT IndexVT = Index.getSimpleValueType();
20122 MVT MaskVT = Mask.getSimpleValueType();
20124 unsigned NumElts = VT.getVectorNumElements();
20125 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
20127 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
20128 !Index.getSimpleValueType().is512BitVector()) {
20129 // AVX512F supports only 512-bit vectors. Or data or index should
20130 // be 512 bit wide. If now the both index and data are 256-bit, but
20131 // the vector contains 8 elements, we just sign-extend the index
20132 if (NumElts == 8) {
20133 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20134 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
20135 N->getOperand(3), Index };
20136 DAG.UpdateNodeOperands(N, Ops);
20140 // Minimal number of elements in Gather
20143 MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
20144 Index = ExtendToType(Index, NewIndexVT, DAG);
20145 if (IndexVT.getScalarType() == MVT::i32)
20146 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
20149 MVT MaskBitVT = MVT::getVectorVT(MVT::i1, NumElts);
20150 // At this point we have promoted mask operand
20151 assert(MaskVT.getScalarSizeInBits() >= 32 && "unexpected mask type");
20152 MVT ExtMaskVT = MVT::getVectorVT(MaskVT.getScalarType(), NumElts);
20153 Mask = ExtendToType(Mask, ExtMaskVT, DAG, true);
20154 Mask = DAG.getNode(ISD::TRUNCATE, dl, MaskBitVT, Mask);
20156 // The pass-thru value
20157 MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
20158 Src0 = ExtendToType(Src0, NewVT, DAG);
20160 SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
20161 SDValue NewGather = DAG.getMaskedGather(DAG.getVTList(NewVT, MVT::Other),
20162 N->getMemoryVT(), dl, Ops,
20163 N->getMemOperand());
20164 SDValue Exract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
20165 NewGather.getValue(0),
20166 DAG.getIntPtrConstant(0, dl));
20167 SDValue RetOps[] = {Exract, NewGather.getValue(1)};
20168 return DAG.getMergeValues(RetOps, dl);
20173 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
20174 SelectionDAG &DAG) const {
20175 // TODO: Eventually, the lowering of these nodes should be informed by or
20176 // deferred to the GC strategy for the function in which they appear. For
20177 // now, however, they must be lowered to something. Since they are logically
20178 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20179 // require special handling for these nodes), lower them as literal NOOPs for
20181 SmallVector<SDValue, 2> Ops;
20183 Ops.push_back(Op.getOperand(0));
20184 if (Op->getGluedNode())
20185 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20188 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20189 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20194 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
20195 SelectionDAG &DAG) const {
20196 // TODO: Eventually, the lowering of these nodes should be informed by or
20197 // deferred to the GC strategy for the function in which they appear. For
20198 // now, however, they must be lowered to something. Since they are logically
20199 // no-ops in the case of a null GC strategy (or a GC strategy which does not
20200 // require special handling for these nodes), lower them as literal NOOPs for
20202 SmallVector<SDValue, 2> Ops;
20204 Ops.push_back(Op.getOperand(0));
20205 if (Op->getGluedNode())
20206 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
20209 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
20210 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
20215 /// LowerOperation - Provide custom lowering hooks for some operations.
20217 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
20218 switch (Op.getOpcode()) {
20219 default: llvm_unreachable("Should not custom lower this!");
20220 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
20221 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
20222 return LowerCMP_SWAP(Op, Subtarget, DAG);
20223 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
20224 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
20225 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
20226 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
20227 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
20228 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
20229 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
20230 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
20231 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
20232 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
20233 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
20234 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
20235 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
20236 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
20237 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
20238 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
20239 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
20240 case ISD::SHL_PARTS:
20241 case ISD::SRA_PARTS:
20242 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
20243 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
20244 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
20245 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
20246 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
20247 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
20248 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
20249 case ISD::SIGN_EXTEND_VECTOR_INREG:
20250 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
20251 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
20252 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
20253 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
20254 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
20256 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
20257 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
20258 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
20259 case ISD::SETCC: return LowerSETCC(Op, DAG);
20260 case ISD::SETCCE: return LowerSETCCE(Op, DAG);
20261 case ISD::SELECT: return LowerSELECT(Op, DAG);
20262 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
20263 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
20264 case ISD::VASTART: return LowerVASTART(Op, DAG);
20265 case ISD::VAARG: return LowerVAARG(Op, DAG);
20266 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
20267 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
20268 case ISD::INTRINSIC_VOID:
20269 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
20270 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
20271 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
20272 case ISD::FRAME_TO_ARGS_OFFSET:
20273 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
20274 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
20275 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
20276 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
20277 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
20278 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
20279 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
20280 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
20281 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
20282 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
20284 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
20285 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
20286 case ISD::UMUL_LOHI:
20287 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
20288 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
20291 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
20297 case ISD::UMULO: return LowerXALUO(Op, DAG);
20298 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
20299 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
20303 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
20304 case ISD::ADD: return LowerADD(Op, DAG);
20305 case ISD::SUB: return LowerSUB(Op, DAG);
20309 case ISD::UMIN: return LowerMINMAX(Op, DAG);
20310 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
20311 case ISD::MLOAD: return LowerMLOAD(Op, Subtarget, DAG);
20312 case ISD::MSTORE: return LowerMSTORE(Op, Subtarget, DAG);
20313 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
20314 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
20315 case ISD::GC_TRANSITION_START:
20316 return LowerGC_TRANSITION_START(Op, DAG);
20317 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
20321 /// ReplaceNodeResults - Replace a node with an illegal result type
20322 /// with a new node built out of custom code.
20323 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
20324 SmallVectorImpl<SDValue>&Results,
20325 SelectionDAG &DAG) const {
20327 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20328 switch (N->getOpcode()) {
20330 llvm_unreachable("Do not know how to custom type legalize this operation!");
20331 case X86ISD::AVG: {
20332 // Legalize types for X86ISD::AVG by expanding vectors.
20333 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20335 auto InVT = N->getValueType(0);
20336 auto InVTSize = InVT.getSizeInBits();
20337 const unsigned RegSize =
20338 (InVTSize > 128) ? ((InVTSize > 256) ? 512 : 256) : 128;
20339 assert((!Subtarget->hasAVX512() || RegSize < 512) &&
20340 "512-bit vector requires AVX512");
20341 assert((!Subtarget->hasAVX2() || RegSize < 256) &&
20342 "256-bit vector requires AVX2");
20344 auto ElemVT = InVT.getVectorElementType();
20345 auto RegVT = EVT::getVectorVT(*DAG.getContext(), ElemVT,
20346 RegSize / ElemVT.getSizeInBits());
20347 assert(RegSize % InVT.getSizeInBits() == 0);
20348 unsigned NumConcat = RegSize / InVT.getSizeInBits();
20350 SmallVector<SDValue, 16> Ops(NumConcat, DAG.getUNDEF(InVT));
20351 Ops[0] = N->getOperand(0);
20352 SDValue InVec0 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20353 Ops[0] = N->getOperand(1);
20354 SDValue InVec1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, RegVT, Ops);
20356 SDValue Res = DAG.getNode(X86ISD::AVG, dl, RegVT, InVec0, InVec1);
20357 Results.push_back(DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, InVT, Res,
20358 DAG.getIntPtrConstant(0, dl)));
20361 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
20362 case X86ISD::FMINC:
20364 case X86ISD::FMAXC:
20365 case X86ISD::FMAX: {
20366 EVT VT = N->getValueType(0);
20367 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
20368 SDValue UNDEF = DAG.getUNDEF(VT);
20369 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20370 N->getOperand(0), UNDEF);
20371 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
20372 N->getOperand(1), UNDEF);
20373 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
20376 case ISD::SIGN_EXTEND_INREG:
20381 // We don't want to expand or promote these.
20388 case ISD::UDIVREM: {
20389 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
20390 Results.push_back(V);
20393 case ISD::FP_TO_SINT:
20394 case ISD::FP_TO_UINT: {
20395 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
20397 std::pair<SDValue,SDValue> Vals =
20398 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
20399 SDValue FIST = Vals.first, StackSlot = Vals.second;
20400 if (FIST.getNode()) {
20401 EVT VT = N->getValueType(0);
20402 // Return a load from the stack slot.
20403 if (StackSlot.getNode())
20404 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
20405 MachinePointerInfo(),
20406 false, false, false, 0));
20408 Results.push_back(FIST);
20412 case ISD::UINT_TO_FP: {
20413 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20414 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
20415 N->getValueType(0) != MVT::v2f32)
20417 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
20419 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
20421 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
20422 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
20423 DAG.getBitcast(MVT::v2i64, VBias));
20424 Or = DAG.getBitcast(MVT::v2f64, Or);
20425 // TODO: Are there any fast-math-flags to propagate here?
20426 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
20427 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
20430 case ISD::FP_ROUND: {
20431 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
20433 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
20434 Results.push_back(V);
20437 case ISD::FP_EXTEND: {
20438 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
20439 // No other ValueType for FP_EXTEND should reach this point.
20440 assert(N->getValueType(0) == MVT::v2f32 &&
20441 "Do not know how to legalize this Node");
20444 case ISD::INTRINSIC_W_CHAIN: {
20445 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
20447 default : llvm_unreachable("Do not know how to custom type "
20448 "legalize this intrinsic operation!");
20449 case Intrinsic::x86_rdtsc:
20450 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20452 case Intrinsic::x86_rdtscp:
20453 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
20455 case Intrinsic::x86_rdpmc:
20456 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
20459 case ISD::INTRINSIC_WO_CHAIN: {
20460 if (SDValue V = LowerINTRINSIC_WO_CHAIN(SDValue(N, 0), Subtarget, DAG))
20461 Results.push_back(V);
20464 case ISD::READCYCLECOUNTER: {
20465 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
20468 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
20469 EVT T = N->getValueType(0);
20470 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
20471 bool Regs64bit = T == MVT::i128;
20472 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
20473 SDValue cpInL, cpInH;
20474 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20475 DAG.getConstant(0, dl, HalfT));
20476 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
20477 DAG.getConstant(1, dl, HalfT));
20478 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
20479 Regs64bit ? X86::RAX : X86::EAX,
20481 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
20482 Regs64bit ? X86::RDX : X86::EDX,
20483 cpInH, cpInL.getValue(1));
20484 SDValue swapInL, swapInH;
20485 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20486 DAG.getConstant(0, dl, HalfT));
20487 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
20488 DAG.getConstant(1, dl, HalfT));
20489 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
20490 Regs64bit ? X86::RBX : X86::EBX,
20491 swapInL, cpInH.getValue(1));
20492 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
20493 Regs64bit ? X86::RCX : X86::ECX,
20494 swapInH, swapInL.getValue(1));
20495 SDValue Ops[] = { swapInH.getValue(0),
20497 swapInH.getValue(1) };
20498 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20499 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20500 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20501 X86ISD::LCMPXCHG8_DAG;
20502 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20503 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20504 Regs64bit ? X86::RAX : X86::EAX,
20505 HalfT, Result.getValue(1));
20506 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20507 Regs64bit ? X86::RDX : X86::EDX,
20508 HalfT, cpOutL.getValue(2));
20509 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20511 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20512 MVT::i32, cpOutH.getValue(2));
20514 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20515 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
20516 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20518 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20519 Results.push_back(Success);
20520 Results.push_back(EFLAGS.getValue(1));
20523 case ISD::ATOMIC_SWAP:
20524 case ISD::ATOMIC_LOAD_ADD:
20525 case ISD::ATOMIC_LOAD_SUB:
20526 case ISD::ATOMIC_LOAD_AND:
20527 case ISD::ATOMIC_LOAD_OR:
20528 case ISD::ATOMIC_LOAD_XOR:
20529 case ISD::ATOMIC_LOAD_NAND:
20530 case ISD::ATOMIC_LOAD_MIN:
20531 case ISD::ATOMIC_LOAD_MAX:
20532 case ISD::ATOMIC_LOAD_UMIN:
20533 case ISD::ATOMIC_LOAD_UMAX:
20534 case ISD::ATOMIC_LOAD: {
20535 // Delegate to generic TypeLegalization. Situations we can really handle
20536 // should have already been dealt with by AtomicExpandPass.cpp.
20539 case ISD::BITCAST: {
20540 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20541 EVT DstVT = N->getValueType(0);
20542 EVT SrcVT = N->getOperand(0)->getValueType(0);
20544 if (SrcVT != MVT::f64 ||
20545 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20548 unsigned NumElts = DstVT.getVectorNumElements();
20549 EVT SVT = DstVT.getVectorElementType();
20550 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20551 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20552 MVT::v2f64, N->getOperand(0));
20553 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
20555 if (ExperimentalVectorWideningLegalization) {
20556 // If we are legalizing vectors by widening, we already have the desired
20557 // legal vector type, just return it.
20558 Results.push_back(ToVecInt);
20562 SmallVector<SDValue, 8> Elts;
20563 for (unsigned i = 0, e = NumElts; i != e; ++i)
20564 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20565 ToVecInt, DAG.getIntPtrConstant(i, dl)));
20567 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20572 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20573 switch ((X86ISD::NodeType)Opcode) {
20574 case X86ISD::FIRST_NUMBER: break;
20575 case X86ISD::BSF: return "X86ISD::BSF";
20576 case X86ISD::BSR: return "X86ISD::BSR";
20577 case X86ISD::SHLD: return "X86ISD::SHLD";
20578 case X86ISD::SHRD: return "X86ISD::SHRD";
20579 case X86ISD::FAND: return "X86ISD::FAND";
20580 case X86ISD::FANDN: return "X86ISD::FANDN";
20581 case X86ISD::FOR: return "X86ISD::FOR";
20582 case X86ISD::FXOR: return "X86ISD::FXOR";
20583 case X86ISD::FILD: return "X86ISD::FILD";
20584 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20585 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20586 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20587 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20588 case X86ISD::FLD: return "X86ISD::FLD";
20589 case X86ISD::FST: return "X86ISD::FST";
20590 case X86ISD::CALL: return "X86ISD::CALL";
20591 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20592 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20593 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20594 case X86ISD::BT: return "X86ISD::BT";
20595 case X86ISD::CMP: return "X86ISD::CMP";
20596 case X86ISD::COMI: return "X86ISD::COMI";
20597 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20598 case X86ISD::CMPM: return "X86ISD::CMPM";
20599 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20600 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20601 case X86ISD::SETCC: return "X86ISD::SETCC";
20602 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20603 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20604 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20605 case X86ISD::CMOV: return "X86ISD::CMOV";
20606 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20607 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20608 case X86ISD::IRET: return "X86ISD::IRET";
20609 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20610 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20611 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20612 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20613 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20614 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20615 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20616 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20617 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20618 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20619 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20620 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20621 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20622 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20623 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20624 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20625 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20626 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20627 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20628 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20629 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20630 case X86ISD::HADD: return "X86ISD::HADD";
20631 case X86ISD::HSUB: return "X86ISD::HSUB";
20632 case X86ISD::FHADD: return "X86ISD::FHADD";
20633 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20634 case X86ISD::ABS: return "X86ISD::ABS";
20635 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20636 case X86ISD::FMAX: return "X86ISD::FMAX";
20637 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20638 case X86ISD::FMIN: return "X86ISD::FMIN";
20639 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20640 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20641 case X86ISD::FMINC: return "X86ISD::FMINC";
20642 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20643 case X86ISD::FRCP: return "X86ISD::FRCP";
20644 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20645 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20646 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20647 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20648 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20649 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20650 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20651 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20652 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20653 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20654 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20655 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20656 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20657 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20658 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20659 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20660 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20661 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20662 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20663 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20664 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20665 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20666 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20667 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20668 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20669 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20670 case X86ISD::CVT2MASK: return "X86ISD::CVT2MASK";
20671 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20672 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20673 case X86ISD::VSHL: return "X86ISD::VSHL";
20674 case X86ISD::VSRL: return "X86ISD::VSRL";
20675 case X86ISD::VSRA: return "X86ISD::VSRA";
20676 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20677 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20678 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20679 case X86ISD::CMPP: return "X86ISD::CMPP";
20680 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20681 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20682 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20683 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20684 case X86ISD::ADD: return "X86ISD::ADD";
20685 case X86ISD::SUB: return "X86ISD::SUB";
20686 case X86ISD::ADC: return "X86ISD::ADC";
20687 case X86ISD::SBB: return "X86ISD::SBB";
20688 case X86ISD::SMUL: return "X86ISD::SMUL";
20689 case X86ISD::UMUL: return "X86ISD::UMUL";
20690 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20691 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20692 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20693 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20694 case X86ISD::INC: return "X86ISD::INC";
20695 case X86ISD::DEC: return "X86ISD::DEC";
20696 case X86ISD::OR: return "X86ISD::OR";
20697 case X86ISD::XOR: return "X86ISD::XOR";
20698 case X86ISD::AND: return "X86ISD::AND";
20699 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20700 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20701 case X86ISD::PTEST: return "X86ISD::PTEST";
20702 case X86ISD::TESTP: return "X86ISD::TESTP";
20703 case X86ISD::TESTM: return "X86ISD::TESTM";
20704 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20705 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20706 case X86ISD::KTEST: return "X86ISD::KTEST";
20707 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20708 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20709 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20710 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20711 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20712 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20713 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20714 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20715 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20716 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20717 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20718 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20719 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20720 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20721 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20722 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20723 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20724 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20725 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20726 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20727 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20728 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20729 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20730 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20731 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20732 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20733 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20734 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20735 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20736 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20737 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20738 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20739 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20740 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20741 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20742 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20743 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20744 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20745 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20746 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20747 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20748 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20749 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20750 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20751 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20752 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20753 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20754 case X86ISD::SAHF: return "X86ISD::SAHF";
20755 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20756 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20757 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20758 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20759 case X86ISD::VPROT: return "X86ISD::VPROT";
20760 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20761 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20762 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20763 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20764 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20765 case X86ISD::FMADD: return "X86ISD::FMADD";
20766 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20767 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20768 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20769 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20770 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20771 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20772 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20773 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20774 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20775 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20776 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20777 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20778 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20779 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20780 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20781 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20782 case X86ISD::XTEST: return "X86ISD::XTEST";
20783 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20784 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20785 case X86ISD::SELECT: return "X86ISD::SELECT";
20786 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20787 case X86ISD::RCP28: return "X86ISD::RCP28";
20788 case X86ISD::EXP2: return "X86ISD::EXP2";
20789 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20790 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20791 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20792 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20793 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20794 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20795 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20796 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20797 case X86ISD::ADDS: return "X86ISD::ADDS";
20798 case X86ISD::SUBS: return "X86ISD::SUBS";
20799 case X86ISD::AVG: return "X86ISD::AVG";
20800 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20801 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20802 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20803 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20804 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20805 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20806 case X86ISD::VFPCLASSS: return "X86ISD::VFPCLASSS";
20811 // isLegalAddressingMode - Return true if the addressing mode represented
20812 // by AM is legal for this target, for a load/store of the specified type.
20813 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20814 const AddrMode &AM, Type *Ty,
20815 unsigned AS) const {
20816 // X86 supports extremely general addressing modes.
20817 CodeModel::Model M = getTargetMachine().getCodeModel();
20818 Reloc::Model R = getTargetMachine().getRelocationModel();
20820 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20821 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20826 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20828 // If a reference to this global requires an extra load, we can't fold it.
20829 if (isGlobalStubReference(GVFlags))
20832 // If BaseGV requires a register for the PIC base, we cannot also have a
20833 // BaseReg specified.
20834 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20837 // If lower 4G is not available, then we must use rip-relative addressing.
20838 if ((M != CodeModel::Small || R != Reloc::Static) &&
20839 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20843 switch (AM.Scale) {
20849 // These scales always work.
20854 // These scales are formed with basereg+scalereg. Only accept if there is
20859 default: // Other stuff never works.
20866 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20867 unsigned Bits = Ty->getScalarSizeInBits();
20869 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20870 // particularly cheaper than those without.
20874 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20875 // variable shifts just as cheap as scalar ones.
20876 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20879 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20880 // fully general vector.
20884 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20885 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20887 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20888 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20889 return NumBits1 > NumBits2;
20892 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20893 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20896 if (!isTypeLegal(EVT::getEVT(Ty1)))
20899 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20901 // Assuming the caller doesn't have a zeroext or signext return parameter,
20902 // truncation all the way down to i1 is valid.
20906 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20907 return isInt<32>(Imm);
20910 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20911 // Can also use sub to handle negated immediates.
20912 return isInt<32>(Imm);
20915 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20916 if (!VT1.isInteger() || !VT2.isInteger())
20918 unsigned NumBits1 = VT1.getSizeInBits();
20919 unsigned NumBits2 = VT2.getSizeInBits();
20920 return NumBits1 > NumBits2;
20923 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20924 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20925 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20928 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20929 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20930 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20933 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20934 EVT VT1 = Val.getValueType();
20935 if (isZExtFree(VT1, VT2))
20938 if (Val.getOpcode() != ISD::LOAD)
20941 if (!VT1.isSimple() || !VT1.isInteger() ||
20942 !VT2.isSimple() || !VT2.isInteger())
20945 switch (VT1.getSimpleVT().SimpleTy) {
20950 // X86 has 8, 16, and 32-bit zero-extending loads.
20957 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20960 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20961 if (!Subtarget->hasAnyFMA())
20964 VT = VT.getScalarType();
20966 if (!VT.isSimple())
20969 switch (VT.getSimpleVT().SimpleTy) {
20980 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20981 // i16 instructions are longer (0x66 prefix) and potentially slower.
20982 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20985 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20986 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20987 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20988 /// are assumed to be legal.
20990 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20992 if (!VT.isSimple())
20995 // Not for i1 vectors
20996 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20999 // Very little shuffling can be done for 64-bit vectors right now.
21000 if (VT.getSimpleVT().getSizeInBits() == 64)
21003 // We only care that the types being shuffled are legal. The lowering can
21004 // handle any possible shuffle mask that results.
21005 return isTypeLegal(VT.getSimpleVT());
21009 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
21011 // Just delegate to the generic legality, clear masks aren't special.
21012 return isShuffleMaskLegal(Mask, VT);
21015 //===----------------------------------------------------------------------===//
21016 // X86 Scheduler Hooks
21017 //===----------------------------------------------------------------------===//
21019 /// Utility function to emit xbegin specifying the start of an RTM region.
21020 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
21021 const TargetInstrInfo *TII) {
21022 DebugLoc DL = MI->getDebugLoc();
21024 const BasicBlock *BB = MBB->getBasicBlock();
21025 MachineFunction::iterator I = ++MBB->getIterator();
21027 // For the v = xbegin(), we generate
21038 MachineBasicBlock *thisMBB = MBB;
21039 MachineFunction *MF = MBB->getParent();
21040 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21041 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21042 MF->insert(I, mainMBB);
21043 MF->insert(I, sinkMBB);
21045 // Transfer the remainder of BB and its successor edges to sinkMBB.
21046 sinkMBB->splice(sinkMBB->begin(), MBB,
21047 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21048 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21052 // # fallthrough to mainMBB
21053 // # abortion to sinkMBB
21054 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
21055 thisMBB->addSuccessor(mainMBB);
21056 thisMBB->addSuccessor(sinkMBB);
21060 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
21061 mainMBB->addSuccessor(sinkMBB);
21064 // EAX is live into the sinkMBB
21065 sinkMBB->addLiveIn(X86::EAX);
21066 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21067 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21070 MI->eraseFromParent();
21074 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
21075 // or XMM0_V32I8 in AVX all of this code can be replaced with that
21076 // in the .td file.
21077 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
21078 const TargetInstrInfo *TII) {
21080 switch (MI->getOpcode()) {
21081 default: llvm_unreachable("illegal opcode!");
21082 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
21083 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
21084 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
21085 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
21086 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
21087 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
21088 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
21089 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
21092 DebugLoc dl = MI->getDebugLoc();
21093 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21095 unsigned NumArgs = MI->getNumOperands();
21096 for (unsigned i = 1; i < NumArgs; ++i) {
21097 MachineOperand &Op = MI->getOperand(i);
21098 if (!(Op.isReg() && Op.isImplicit()))
21099 MIB.addOperand(Op);
21101 if (MI->hasOneMemOperand())
21102 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21104 BuildMI(*BB, MI, dl,
21105 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21106 .addReg(X86::XMM0);
21108 MI->eraseFromParent();
21112 // FIXME: Custom handling because TableGen doesn't support multiple implicit
21113 // defs in an instruction pattern
21114 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
21115 const TargetInstrInfo *TII) {
21117 switch (MI->getOpcode()) {
21118 default: llvm_unreachable("illegal opcode!");
21119 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
21120 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
21121 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
21122 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
21123 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
21124 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
21125 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
21126 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
21129 DebugLoc dl = MI->getDebugLoc();
21130 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
21132 unsigned NumArgs = MI->getNumOperands(); // remove the results
21133 for (unsigned i = 1; i < NumArgs; ++i) {
21134 MachineOperand &Op = MI->getOperand(i);
21135 if (!(Op.isReg() && Op.isImplicit()))
21136 MIB.addOperand(Op);
21138 if (MI->hasOneMemOperand())
21139 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
21141 BuildMI(*BB, MI, dl,
21142 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21145 MI->eraseFromParent();
21149 static MachineBasicBlock *EmitWRPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21150 const X86Subtarget *Subtarget) {
21151 DebugLoc dl = MI->getDebugLoc();
21152 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21154 // insert input VAL into EAX
21155 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
21156 .addReg(MI->getOperand(0).getReg());
21157 // insert zero to ECX
21158 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21161 // insert zero to EDX
21162 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::EDX)
21165 // insert WRPKRU instruction
21166 BuildMI(*BB, MI, dl, TII->get(X86::WRPKRUr));
21168 MI->eraseFromParent(); // The pseudo is gone now.
21172 static MachineBasicBlock *EmitRDPKRU(MachineInstr *MI, MachineBasicBlock *BB,
21173 const X86Subtarget *Subtarget) {
21174 DebugLoc dl = MI->getDebugLoc();
21175 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21177 // insert zero to ECX
21178 BuildMI(*BB, MI, dl, TII->get(X86::XOR32rr), X86::ECX)
21181 // insert RDPKRU instruction
21182 BuildMI(*BB, MI, dl, TII->get(X86::RDPKRUr));
21183 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
21186 MI->eraseFromParent(); // The pseudo is gone now.
21190 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
21191 const X86Subtarget *Subtarget) {
21192 DebugLoc dl = MI->getDebugLoc();
21193 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21194 // Address into RAX/EAX, other two args into ECX, EDX.
21195 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
21196 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
21197 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
21198 for (int i = 0; i < X86::AddrNumOperands; ++i)
21199 MIB.addOperand(MI->getOperand(i));
21201 unsigned ValOps = X86::AddrNumOperands;
21202 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
21203 .addReg(MI->getOperand(ValOps).getReg());
21204 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
21205 .addReg(MI->getOperand(ValOps+1).getReg());
21207 // The instruction doesn't actually take any operands though.
21208 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
21210 MI->eraseFromParent(); // The pseudo is gone now.
21214 MachineBasicBlock *
21215 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
21216 MachineBasicBlock *MBB) const {
21217 // Emit va_arg instruction on X86-64.
21219 // Operands to this pseudo-instruction:
21220 // 0 ) Output : destination address (reg)
21221 // 1-5) Input : va_list address (addr, i64mem)
21222 // 6 ) ArgSize : Size (in bytes) of vararg type
21223 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
21224 // 8 ) Align : Alignment of type
21225 // 9 ) EFLAGS (implicit-def)
21227 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
21228 static_assert(X86::AddrNumOperands == 5,
21229 "VAARG_64 assumes 5 address operands");
21231 unsigned DestReg = MI->getOperand(0).getReg();
21232 MachineOperand &Base = MI->getOperand(1);
21233 MachineOperand &Scale = MI->getOperand(2);
21234 MachineOperand &Index = MI->getOperand(3);
21235 MachineOperand &Disp = MI->getOperand(4);
21236 MachineOperand &Segment = MI->getOperand(5);
21237 unsigned ArgSize = MI->getOperand(6).getImm();
21238 unsigned ArgMode = MI->getOperand(7).getImm();
21239 unsigned Align = MI->getOperand(8).getImm();
21241 // Memory Reference
21242 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
21243 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21244 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21246 // Machine Information
21247 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21248 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
21249 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
21250 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
21251 DebugLoc DL = MI->getDebugLoc();
21253 // struct va_list {
21256 // i64 overflow_area (address)
21257 // i64 reg_save_area (address)
21259 // sizeof(va_list) = 24
21260 // alignment(va_list) = 8
21262 unsigned TotalNumIntRegs = 6;
21263 unsigned TotalNumXMMRegs = 8;
21264 bool UseGPOffset = (ArgMode == 1);
21265 bool UseFPOffset = (ArgMode == 2);
21266 unsigned MaxOffset = TotalNumIntRegs * 8 +
21267 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
21269 /* Align ArgSize to a multiple of 8 */
21270 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
21271 bool NeedsAlign = (Align > 8);
21273 MachineBasicBlock *thisMBB = MBB;
21274 MachineBasicBlock *overflowMBB;
21275 MachineBasicBlock *offsetMBB;
21276 MachineBasicBlock *endMBB;
21278 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
21279 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
21280 unsigned OffsetReg = 0;
21282 if (!UseGPOffset && !UseFPOffset) {
21283 // If we only pull from the overflow region, we don't create a branch.
21284 // We don't need to alter control flow.
21285 OffsetDestReg = 0; // unused
21286 OverflowDestReg = DestReg;
21288 offsetMBB = nullptr;
21289 overflowMBB = thisMBB;
21292 // First emit code to check if gp_offset (or fp_offset) is below the bound.
21293 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
21294 // If not, pull from overflow_area. (branch to overflowMBB)
21299 // offsetMBB overflowMBB
21304 // Registers for the PHI in endMBB
21305 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
21306 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
21308 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21309 MachineFunction *MF = MBB->getParent();
21310 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21311 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21312 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21314 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21316 // Insert the new basic blocks
21317 MF->insert(MBBIter, offsetMBB);
21318 MF->insert(MBBIter, overflowMBB);
21319 MF->insert(MBBIter, endMBB);
21321 // Transfer the remainder of MBB and its successor edges to endMBB.
21322 endMBB->splice(endMBB->begin(), thisMBB,
21323 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
21324 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
21326 // Make offsetMBB and overflowMBB successors of thisMBB
21327 thisMBB->addSuccessor(offsetMBB);
21328 thisMBB->addSuccessor(overflowMBB);
21330 // endMBB is a successor of both offsetMBB and overflowMBB
21331 offsetMBB->addSuccessor(endMBB);
21332 overflowMBB->addSuccessor(endMBB);
21334 // Load the offset value into a register
21335 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21336 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
21340 .addDisp(Disp, UseFPOffset ? 4 : 0)
21341 .addOperand(Segment)
21342 .setMemRefs(MMOBegin, MMOEnd);
21344 // Check if there is enough room left to pull this argument.
21345 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
21347 .addImm(MaxOffset + 8 - ArgSizeA8);
21349 // Branch to "overflowMBB" if offset >= max
21350 // Fall through to "offsetMBB" otherwise
21351 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
21352 .addMBB(overflowMBB);
21355 // In offsetMBB, emit code to use the reg_save_area.
21357 assert(OffsetReg != 0);
21359 // Read the reg_save_area address.
21360 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
21361 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
21366 .addOperand(Segment)
21367 .setMemRefs(MMOBegin, MMOEnd);
21369 // Zero-extend the offset
21370 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
21371 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
21374 .addImm(X86::sub_32bit);
21376 // Add the offset to the reg_save_area to get the final address.
21377 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
21378 .addReg(OffsetReg64)
21379 .addReg(RegSaveReg);
21381 // Compute the offset for the next argument
21382 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
21383 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
21385 .addImm(UseFPOffset ? 16 : 8);
21387 // Store it back into the va_list.
21388 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
21392 .addDisp(Disp, UseFPOffset ? 4 : 0)
21393 .addOperand(Segment)
21394 .addReg(NextOffsetReg)
21395 .setMemRefs(MMOBegin, MMOEnd);
21398 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
21403 // Emit code to use overflow area
21406 // Load the overflow_area address into a register.
21407 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
21408 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
21413 .addOperand(Segment)
21414 .setMemRefs(MMOBegin, MMOEnd);
21416 // If we need to align it, do so. Otherwise, just copy the address
21417 // to OverflowDestReg.
21419 // Align the overflow address
21420 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
21421 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
21423 // aligned_addr = (addr + (align-1)) & ~(align-1)
21424 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
21425 .addReg(OverflowAddrReg)
21428 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
21430 .addImm(~(uint64_t)(Align-1));
21432 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
21433 .addReg(OverflowAddrReg);
21436 // Compute the next overflow address after this argument.
21437 // (the overflow address should be kept 8-byte aligned)
21438 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
21439 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
21440 .addReg(OverflowDestReg)
21441 .addImm(ArgSizeA8);
21443 // Store the new overflow address.
21444 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
21449 .addOperand(Segment)
21450 .addReg(NextAddrReg)
21451 .setMemRefs(MMOBegin, MMOEnd);
21453 // If we branched, emit the PHI to the front of endMBB.
21455 BuildMI(*endMBB, endMBB->begin(), DL,
21456 TII->get(X86::PHI), DestReg)
21457 .addReg(OffsetDestReg).addMBB(offsetMBB)
21458 .addReg(OverflowDestReg).addMBB(overflowMBB);
21461 // Erase the pseudo instruction
21462 MI->eraseFromParent();
21467 MachineBasicBlock *
21468 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
21470 MachineBasicBlock *MBB) const {
21471 // Emit code to save XMM registers to the stack. The ABI says that the
21472 // number of registers to save is given in %al, so it's theoretically
21473 // possible to do an indirect jump trick to avoid saving all of them,
21474 // however this code takes a simpler approach and just executes all
21475 // of the stores if %al is non-zero. It's less code, and it's probably
21476 // easier on the hardware branch predictor, and stores aren't all that
21477 // expensive anyway.
21479 // Create the new basic blocks. One block contains all the XMM stores,
21480 // and one block is the final destination regardless of whether any
21481 // stores were performed.
21482 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
21483 MachineFunction *F = MBB->getParent();
21484 MachineFunction::iterator MBBIter = ++MBB->getIterator();
21485 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
21486 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
21487 F->insert(MBBIter, XMMSaveMBB);
21488 F->insert(MBBIter, EndMBB);
21490 // Transfer the remainder of MBB and its successor edges to EndMBB.
21491 EndMBB->splice(EndMBB->begin(), MBB,
21492 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21493 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
21495 // The original block will now fall through to the XMM save block.
21496 MBB->addSuccessor(XMMSaveMBB);
21497 // The XMMSaveMBB will fall through to the end block.
21498 XMMSaveMBB->addSuccessor(EndMBB);
21500 // Now add the instructions.
21501 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21502 DebugLoc DL = MI->getDebugLoc();
21504 unsigned CountReg = MI->getOperand(0).getReg();
21505 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
21506 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
21508 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
21509 // If %al is 0, branch around the XMM save block.
21510 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
21511 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
21512 MBB->addSuccessor(EndMBB);
21515 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
21516 // that was just emitted, but clearly shouldn't be "saved".
21517 assert((MI->getNumOperands() <= 3 ||
21518 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
21519 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
21520 && "Expected last argument to be EFLAGS");
21521 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
21522 // In the XMM save block, save all the XMM argument registers.
21523 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
21524 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
21525 MachineMemOperand *MMO = F->getMachineMemOperand(
21526 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
21527 MachineMemOperand::MOStore,
21528 /*Size=*/16, /*Align=*/16);
21529 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
21530 .addFrameIndex(RegSaveFrameIndex)
21531 .addImm(/*Scale=*/1)
21532 .addReg(/*IndexReg=*/0)
21533 .addImm(/*Disp=*/Offset)
21534 .addReg(/*Segment=*/0)
21535 .addReg(MI->getOperand(i).getReg())
21536 .addMemOperand(MMO);
21539 MI->eraseFromParent(); // The pseudo instruction is gone now.
21544 // The EFLAGS operand of SelectItr might be missing a kill marker
21545 // because there were multiple uses of EFLAGS, and ISel didn't know
21546 // which to mark. Figure out whether SelectItr should have had a
21547 // kill marker, and set it if it should. Returns the correct kill
21549 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21550 MachineBasicBlock* BB,
21551 const TargetRegisterInfo* TRI) {
21552 // Scan forward through BB for a use/def of EFLAGS.
21553 MachineBasicBlock::iterator miI(std::next(SelectItr));
21554 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21555 const MachineInstr& mi = *miI;
21556 if (mi.readsRegister(X86::EFLAGS))
21558 if (mi.definesRegister(X86::EFLAGS))
21559 break; // Should have kill-flag - update below.
21562 // If we hit the end of the block, check whether EFLAGS is live into a
21564 if (miI == BB->end()) {
21565 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21566 sEnd = BB->succ_end();
21567 sItr != sEnd; ++sItr) {
21568 MachineBasicBlock* succ = *sItr;
21569 if (succ->isLiveIn(X86::EFLAGS))
21574 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21575 // out. SelectMI should have a kill flag on EFLAGS.
21576 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21580 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
21581 // together with other CMOV pseudo-opcodes into a single basic-block with
21582 // conditional jump around it.
21583 static bool isCMOVPseudo(MachineInstr *MI) {
21584 switch (MI->getOpcode()) {
21585 case X86::CMOV_FR32:
21586 case X86::CMOV_FR64:
21587 case X86::CMOV_GR8:
21588 case X86::CMOV_GR16:
21589 case X86::CMOV_GR32:
21590 case X86::CMOV_RFP32:
21591 case X86::CMOV_RFP64:
21592 case X86::CMOV_RFP80:
21593 case X86::CMOV_V2F64:
21594 case X86::CMOV_V2I64:
21595 case X86::CMOV_V4F32:
21596 case X86::CMOV_V4F64:
21597 case X86::CMOV_V4I64:
21598 case X86::CMOV_V16F32:
21599 case X86::CMOV_V8F32:
21600 case X86::CMOV_V8F64:
21601 case X86::CMOV_V8I64:
21602 case X86::CMOV_V8I1:
21603 case X86::CMOV_V16I1:
21604 case X86::CMOV_V32I1:
21605 case X86::CMOV_V64I1:
21613 MachineBasicBlock *
21614 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21615 MachineBasicBlock *BB) const {
21616 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21617 DebugLoc DL = MI->getDebugLoc();
21619 // To "insert" a SELECT_CC instruction, we actually have to insert the
21620 // diamond control-flow pattern. The incoming instruction knows the
21621 // destination vreg to set, the condition code register to branch on, the
21622 // true/false values to select between, and a branch opcode to use.
21623 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21624 MachineFunction::iterator It = ++BB->getIterator();
21629 // cmpTY ccX, r1, r2
21631 // fallthrough --> copy0MBB
21632 MachineBasicBlock *thisMBB = BB;
21633 MachineFunction *F = BB->getParent();
21635 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21636 // as described above, by inserting a BB, and then making a PHI at the join
21637 // point to select the true and false operands of the CMOV in the PHI.
21639 // The code also handles two different cases of multiple CMOV opcodes
21643 // In this case, there are multiple CMOVs in a row, all which are based on
21644 // the same condition setting (or the exact opposite condition setting).
21645 // In this case we can lower all the CMOVs using a single inserted BB, and
21646 // then make a number of PHIs at the join point to model the CMOVs. The only
21647 // trickiness here, is that in a case like:
21649 // t2 = CMOV cond1 t1, f1
21650 // t3 = CMOV cond1 t2, f2
21652 // when rewriting this into PHIs, we have to perform some renaming on the
21653 // temps since you cannot have a PHI operand refer to a PHI result earlier
21654 // in the same block. The "simple" but wrong lowering would be:
21656 // t2 = PHI t1(BB1), f1(BB2)
21657 // t3 = PHI t2(BB1), f2(BB2)
21659 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21660 // renaming is to note that on the path through BB1, t2 is really just a
21661 // copy of t1, and do that renaming, properly generating:
21663 // t2 = PHI t1(BB1), f1(BB2)
21664 // t3 = PHI t1(BB1), f2(BB2)
21666 // Case 2, we lower cascaded CMOVs such as
21668 // (CMOV (CMOV F, T, cc1), T, cc2)
21670 // to two successives branches. For that, we look for another CMOV as the
21671 // following instruction.
21673 // Without this, we would add a PHI between the two jumps, which ends up
21674 // creating a few copies all around. For instance, for
21676 // (sitofp (zext (fcmp une)))
21678 // we would generate:
21680 // ucomiss %xmm1, %xmm0
21681 // movss <1.0f>, %xmm0
21682 // movaps %xmm0, %xmm1
21684 // xorps %xmm1, %xmm1
21687 // movaps %xmm1, %xmm0
21691 // because this custom-inserter would have generated:
21703 // A: X = ...; Y = ...
21705 // C: Z = PHI [X, A], [Y, B]
21707 // E: PHI [X, C], [Z, D]
21709 // If we lower both CMOVs in a single step, we can instead generate:
21721 // A: X = ...; Y = ...
21723 // E: PHI [X, A], [X, C], [Y, D]
21725 // Which, in our sitofp/fcmp example, gives us something like:
21727 // ucomiss %xmm1, %xmm0
21728 // movss <1.0f>, %xmm0
21731 // xorps %xmm0, %xmm0
21735 MachineInstr *CascadedCMOV = nullptr;
21736 MachineInstr *LastCMOV = MI;
21737 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21738 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21739 MachineBasicBlock::iterator NextMIIt =
21740 std::next(MachineBasicBlock::iterator(MI));
21742 // Check for case 1, where there are multiple CMOVs with the same condition
21743 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21744 // number of jumps the most.
21746 if (isCMOVPseudo(MI)) {
21747 // See if we have a string of CMOVS with the same condition.
21748 while (NextMIIt != BB->end() &&
21749 isCMOVPseudo(NextMIIt) &&
21750 (NextMIIt->getOperand(3).getImm() == CC ||
21751 NextMIIt->getOperand(3).getImm() == OppCC)) {
21752 LastCMOV = &*NextMIIt;
21757 // This checks for case 2, but only do this if we didn't already find
21758 // case 1, as indicated by LastCMOV == MI.
21759 if (LastCMOV == MI &&
21760 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21761 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21762 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21763 CascadedCMOV = &*NextMIIt;
21766 MachineBasicBlock *jcc1MBB = nullptr;
21768 // If we have a cascaded CMOV, we lower it to two successive branches to
21769 // the same block. EFLAGS is used by both, so mark it as live in the second.
21770 if (CascadedCMOV) {
21771 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21772 F->insert(It, jcc1MBB);
21773 jcc1MBB->addLiveIn(X86::EFLAGS);
21776 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21777 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21778 F->insert(It, copy0MBB);
21779 F->insert(It, sinkMBB);
21781 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21782 // live into the sink and copy blocks.
21783 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21785 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21786 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21787 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21788 copy0MBB->addLiveIn(X86::EFLAGS);
21789 sinkMBB->addLiveIn(X86::EFLAGS);
21792 // Transfer the remainder of BB and its successor edges to sinkMBB.
21793 sinkMBB->splice(sinkMBB->begin(), BB,
21794 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21795 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21797 // Add the true and fallthrough blocks as its successors.
21798 if (CascadedCMOV) {
21799 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21800 BB->addSuccessor(jcc1MBB);
21802 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21803 // jump to the sinkMBB.
21804 jcc1MBB->addSuccessor(copy0MBB);
21805 jcc1MBB->addSuccessor(sinkMBB);
21807 BB->addSuccessor(copy0MBB);
21810 // The true block target of the first (or only) branch is always sinkMBB.
21811 BB->addSuccessor(sinkMBB);
21813 // Create the conditional branch instruction.
21814 unsigned Opc = X86::GetCondBranchFromCond(CC);
21815 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21817 if (CascadedCMOV) {
21818 unsigned Opc2 = X86::GetCondBranchFromCond(
21819 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21820 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21824 // %FalseValue = ...
21825 // # fallthrough to sinkMBB
21826 copy0MBB->addSuccessor(sinkMBB);
21829 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21831 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21832 MachineBasicBlock::iterator MIItEnd =
21833 std::next(MachineBasicBlock::iterator(LastCMOV));
21834 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21835 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21836 MachineInstrBuilder MIB;
21838 // As we are creating the PHIs, we have to be careful if there is more than
21839 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21840 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21841 // That also means that PHI construction must work forward from earlier to
21842 // later, and that the code must maintain a mapping from earlier PHI's
21843 // destination registers, and the registers that went into the PHI.
21845 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21846 unsigned DestReg = MIIt->getOperand(0).getReg();
21847 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21848 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21850 // If this CMOV we are generating is the opposite condition from
21851 // the jump we generated, then we have to swap the operands for the
21852 // PHI that is going to be generated.
21853 if (MIIt->getOperand(3).getImm() == OppCC)
21854 std::swap(Op1Reg, Op2Reg);
21856 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21857 Op1Reg = RegRewriteTable[Op1Reg].first;
21859 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21860 Op2Reg = RegRewriteTable[Op2Reg].second;
21862 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21863 TII->get(X86::PHI), DestReg)
21864 .addReg(Op1Reg).addMBB(copy0MBB)
21865 .addReg(Op2Reg).addMBB(thisMBB);
21867 // Add this PHI to the rewrite table.
21868 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21871 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21872 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21873 if (CascadedCMOV) {
21874 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21875 // Copy the PHI result to the register defined by the second CMOV.
21876 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21877 DL, TII->get(TargetOpcode::COPY),
21878 CascadedCMOV->getOperand(0).getReg())
21879 .addReg(MI->getOperand(0).getReg());
21880 CascadedCMOV->eraseFromParent();
21883 // Now remove the CMOV(s).
21884 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21885 (MIIt++)->eraseFromParent();
21890 MachineBasicBlock *
21891 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21892 MachineBasicBlock *BB) const {
21893 // Combine the following atomic floating-point modification pattern:
21894 // a.store(reg OP a.load(acquire), release)
21895 // Transform them into:
21896 // OPss (%gpr), %xmm
21897 // movss %xmm, (%gpr)
21898 // Or sd equivalent for 64-bit operations.
21900 switch (MI->getOpcode()) {
21901 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21902 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21903 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21905 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21906 DebugLoc DL = MI->getDebugLoc();
21907 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21908 MachineOperand MSrc = MI->getOperand(0);
21909 unsigned VSrc = MI->getOperand(5).getReg();
21910 const MachineOperand &Disp = MI->getOperand(3);
21911 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21912 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21913 if (hasDisp && MSrc.isReg())
21914 MSrc.setIsKill(false);
21915 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21916 .addOperand(/*Base=*/MSrc)
21917 .addImm(/*Scale=*/1)
21918 .addReg(/*Index=*/0)
21919 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21921 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21922 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21924 .addOperand(/*Base=*/MSrc)
21925 .addImm(/*Scale=*/1)
21926 .addReg(/*Index=*/0)
21927 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21928 .addReg(/*Segment=*/0);
21929 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21930 MI->eraseFromParent(); // The pseudo instruction is gone now.
21934 MachineBasicBlock *
21935 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21936 MachineBasicBlock *BB) const {
21937 MachineFunction *MF = BB->getParent();
21938 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21939 DebugLoc DL = MI->getDebugLoc();
21940 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21942 assert(MF->shouldSplitStack());
21944 const bool Is64Bit = Subtarget->is64Bit();
21945 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21947 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21948 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21951 // ... [Till the alloca]
21952 // If stacklet is not large enough, jump to mallocMBB
21955 // Allocate by subtracting from RSP
21956 // Jump to continueMBB
21959 // Allocate by call to runtime
21963 // [rest of original BB]
21966 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21967 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21968 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21970 MachineRegisterInfo &MRI = MF->getRegInfo();
21971 const TargetRegisterClass *AddrRegClass =
21972 getRegClassFor(getPointerTy(MF->getDataLayout()));
21974 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21975 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21976 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21977 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21978 sizeVReg = MI->getOperand(1).getReg(),
21979 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21981 MachineFunction::iterator MBBIter = ++BB->getIterator();
21983 MF->insert(MBBIter, bumpMBB);
21984 MF->insert(MBBIter, mallocMBB);
21985 MF->insert(MBBIter, continueMBB);
21987 continueMBB->splice(continueMBB->begin(), BB,
21988 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21989 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21991 // Add code to the main basic block to check if the stack limit has been hit,
21992 // and if so, jump to mallocMBB otherwise to bumpMBB.
21993 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21994 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21995 .addReg(tmpSPVReg).addReg(sizeVReg);
21996 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21997 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21998 .addReg(SPLimitVReg);
21999 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
22001 // bumpMBB simply decreases the stack pointer, since we know the current
22002 // stacklet has enough space.
22003 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
22004 .addReg(SPLimitVReg);
22005 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
22006 .addReg(SPLimitVReg);
22007 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22009 // Calls into a routine in libgcc to allocate more space from the heap.
22010 const uint32_t *RegMask =
22011 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
22013 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
22015 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22016 .addExternalSymbol("__morestack_allocate_stack_space")
22017 .addRegMask(RegMask)
22018 .addReg(X86::RDI, RegState::Implicit)
22019 .addReg(X86::RAX, RegState::ImplicitDefine);
22020 } else if (Is64Bit) {
22021 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
22023 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
22024 .addExternalSymbol("__morestack_allocate_stack_space")
22025 .addRegMask(RegMask)
22026 .addReg(X86::EDI, RegState::Implicit)
22027 .addReg(X86::EAX, RegState::ImplicitDefine);
22029 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
22031 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
22032 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
22033 .addExternalSymbol("__morestack_allocate_stack_space")
22034 .addRegMask(RegMask)
22035 .addReg(X86::EAX, RegState::ImplicitDefine);
22039 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
22042 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
22043 .addReg(IsLP64 ? X86::RAX : X86::EAX);
22044 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
22046 // Set up the CFG correctly.
22047 BB->addSuccessor(bumpMBB);
22048 BB->addSuccessor(mallocMBB);
22049 mallocMBB->addSuccessor(continueMBB);
22050 bumpMBB->addSuccessor(continueMBB);
22052 // Take care of the PHI nodes.
22053 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
22054 MI->getOperand(0).getReg())
22055 .addReg(mallocPtrVReg).addMBB(mallocMBB)
22056 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
22058 // Delete the original pseudo instruction.
22059 MI->eraseFromParent();
22062 return continueMBB;
22065 MachineBasicBlock *
22066 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
22067 MachineBasicBlock *BB) const {
22068 assert(!Subtarget->isTargetMachO());
22069 DebugLoc DL = MI->getDebugLoc();
22070 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
22071 *BB->getParent(), *BB, MI, DL, false);
22072 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
22073 MI->eraseFromParent(); // The pseudo instruction is gone now.
22077 MachineBasicBlock *
22078 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
22079 MachineBasicBlock *BB) const {
22080 MachineFunction *MF = BB->getParent();
22081 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22082 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
22083 DebugLoc DL = MI->getDebugLoc();
22085 assert(!isAsynchronousEHPersonality(
22086 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
22087 "SEH does not use catchret!");
22089 // Only 32-bit EH needs to worry about manually restoring stack pointers.
22090 if (!Subtarget->is32Bit())
22093 // C++ EH creates a new target block to hold the restore code, and wires up
22094 // the new block to the return destination with a normal JMP_4.
22095 MachineBasicBlock *RestoreMBB =
22096 MF->CreateMachineBasicBlock(BB->getBasicBlock());
22097 assert(BB->succ_size() == 1);
22098 MF->insert(std::next(BB->getIterator()), RestoreMBB);
22099 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
22100 BB->addSuccessor(RestoreMBB);
22101 MI->getOperand(0).setMBB(RestoreMBB);
22103 auto RestoreMBBI = RestoreMBB->begin();
22104 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
22105 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
22109 MachineBasicBlock *
22110 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
22111 MachineBasicBlock *BB) const {
22112 MachineFunction *MF = BB->getParent();
22113 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
22114 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
22115 // Only 32-bit SEH requires special handling for catchpad.
22116 if (IsSEH && Subtarget->is32Bit()) {
22117 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22118 DebugLoc DL = MI->getDebugLoc();
22119 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
22121 MI->eraseFromParent();
22125 MachineBasicBlock *
22126 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
22127 MachineBasicBlock *BB) const {
22128 // This is pretty easy. We're taking the value that we received from
22129 // our load from the relocation, sticking it in either RDI (x86-64)
22130 // or EAX and doing an indirect call. The return value will then
22131 // be in the normal return register.
22132 MachineFunction *F = BB->getParent();
22133 const X86InstrInfo *TII = Subtarget->getInstrInfo();
22134 DebugLoc DL = MI->getDebugLoc();
22136 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
22137 assert(MI->getOperand(3).isGlobal() && "This should be a global");
22139 // Get a register mask for the lowered call.
22140 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
22141 // proper register mask.
22142 const uint32_t *RegMask =
22143 Subtarget->is64Bit() ?
22144 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
22145 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
22146 if (Subtarget->is64Bit()) {
22147 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22148 TII->get(X86::MOV64rm), X86::RDI)
22150 .addImm(0).addReg(0)
22151 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22152 MI->getOperand(3).getTargetFlags())
22154 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
22155 addDirectMem(MIB, X86::RDI);
22156 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
22157 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
22158 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22159 TII->get(X86::MOV32rm), X86::EAX)
22161 .addImm(0).addReg(0)
22162 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22163 MI->getOperand(3).getTargetFlags())
22165 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22166 addDirectMem(MIB, X86::EAX);
22167 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22169 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
22170 TII->get(X86::MOV32rm), X86::EAX)
22171 .addReg(TII->getGlobalBaseReg(F))
22172 .addImm(0).addReg(0)
22173 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
22174 MI->getOperand(3).getTargetFlags())
22176 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
22177 addDirectMem(MIB, X86::EAX);
22178 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
22181 MI->eraseFromParent(); // The pseudo instruction is gone now.
22185 MachineBasicBlock *
22186 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
22187 MachineBasicBlock *MBB) const {
22188 DebugLoc DL = MI->getDebugLoc();
22189 MachineFunction *MF = MBB->getParent();
22190 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22191 MachineRegisterInfo &MRI = MF->getRegInfo();
22193 const BasicBlock *BB = MBB->getBasicBlock();
22194 MachineFunction::iterator I = ++MBB->getIterator();
22196 // Memory Reference
22197 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22198 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22201 unsigned MemOpndSlot = 0;
22203 unsigned CurOp = 0;
22205 DstReg = MI->getOperand(CurOp++).getReg();
22206 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
22207 assert(RC->hasType(MVT::i32) && "Invalid destination!");
22208 unsigned mainDstReg = MRI.createVirtualRegister(RC);
22209 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
22211 MemOpndSlot = CurOp;
22213 MVT PVT = getPointerTy(MF->getDataLayout());
22214 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22215 "Invalid Pointer Size!");
22217 // For v = setjmp(buf), we generate
22220 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
22221 // SjLjSetup restoreMBB
22227 // v = phi(main, restore)
22230 // if base pointer being used, load it from frame
22233 MachineBasicBlock *thisMBB = MBB;
22234 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
22235 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
22236 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
22237 MF->insert(I, mainMBB);
22238 MF->insert(I, sinkMBB);
22239 MF->push_back(restoreMBB);
22240 restoreMBB->setHasAddressTaken();
22242 MachineInstrBuilder MIB;
22244 // Transfer the remainder of BB and its successor edges to sinkMBB.
22245 sinkMBB->splice(sinkMBB->begin(), MBB,
22246 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
22247 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
22250 unsigned PtrStoreOpc = 0;
22251 unsigned LabelReg = 0;
22252 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22253 Reloc::Model RM = MF->getTarget().getRelocationModel();
22254 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
22255 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
22257 // Prepare IP either in reg or imm.
22258 if (!UseImmLabel) {
22259 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
22260 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
22261 LabelReg = MRI.createVirtualRegister(PtrRC);
22262 if (Subtarget->is64Bit()) {
22263 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
22267 .addMBB(restoreMBB)
22270 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
22271 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
22272 .addReg(XII->getGlobalBaseReg(MF))
22275 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
22279 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
22281 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
22282 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22283 if (i == X86::AddrDisp)
22284 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
22286 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
22289 MIB.addReg(LabelReg);
22291 MIB.addMBB(restoreMBB);
22292 MIB.setMemRefs(MMOBegin, MMOEnd);
22294 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
22295 .addMBB(restoreMBB);
22297 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22298 MIB.addRegMask(RegInfo->getNoPreservedMask());
22299 thisMBB->addSuccessor(mainMBB);
22300 thisMBB->addSuccessor(restoreMBB);
22304 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
22305 mainMBB->addSuccessor(sinkMBB);
22308 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
22309 TII->get(X86::PHI), DstReg)
22310 .addReg(mainDstReg).addMBB(mainMBB)
22311 .addReg(restoreDstReg).addMBB(restoreMBB);
22314 if (RegInfo->hasBasePointer(*MF)) {
22315 const bool Uses64BitFramePtr =
22316 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
22317 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
22318 X86FI->setRestoreBasePointer(MF);
22319 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
22320 unsigned BasePtr = RegInfo->getBaseRegister();
22321 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
22322 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
22323 FramePtr, true, X86FI->getRestoreBasePointerOffset())
22324 .setMIFlag(MachineInstr::FrameSetup);
22326 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
22327 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
22328 restoreMBB->addSuccessor(sinkMBB);
22330 MI->eraseFromParent();
22334 MachineBasicBlock *
22335 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
22336 MachineBasicBlock *MBB) const {
22337 DebugLoc DL = MI->getDebugLoc();
22338 MachineFunction *MF = MBB->getParent();
22339 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22340 MachineRegisterInfo &MRI = MF->getRegInfo();
22342 // Memory Reference
22343 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
22344 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
22346 MVT PVT = getPointerTy(MF->getDataLayout());
22347 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
22348 "Invalid Pointer Size!");
22350 const TargetRegisterClass *RC =
22351 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
22352 unsigned Tmp = MRI.createVirtualRegister(RC);
22353 // Since FP is only updated here but NOT referenced, it's treated as GPR.
22354 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
22355 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
22356 unsigned SP = RegInfo->getStackRegister();
22358 MachineInstrBuilder MIB;
22360 const int64_t LabelOffset = 1 * PVT.getStoreSize();
22361 const int64_t SPOffset = 2 * PVT.getStoreSize();
22363 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
22364 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
22367 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
22368 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
22369 MIB.addOperand(MI->getOperand(i));
22370 MIB.setMemRefs(MMOBegin, MMOEnd);
22372 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
22373 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22374 if (i == X86::AddrDisp)
22375 MIB.addDisp(MI->getOperand(i), LabelOffset);
22377 MIB.addOperand(MI->getOperand(i));
22379 MIB.setMemRefs(MMOBegin, MMOEnd);
22381 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
22382 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
22383 if (i == X86::AddrDisp)
22384 MIB.addDisp(MI->getOperand(i), SPOffset);
22386 MIB.addOperand(MI->getOperand(i));
22388 MIB.setMemRefs(MMOBegin, MMOEnd);
22390 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
22392 MI->eraseFromParent();
22396 // Replace 213-type (isel default) FMA3 instructions with 231-type for
22397 // accumulator loops. Writing back to the accumulator allows the coalescer
22398 // to remove extra copies in the loop.
22399 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
22400 MachineBasicBlock *
22401 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
22402 MachineBasicBlock *MBB) const {
22403 MachineOperand &AddendOp = MI->getOperand(3);
22405 // Bail out early if the addend isn't a register - we can't switch these.
22406 if (!AddendOp.isReg())
22409 MachineFunction &MF = *MBB->getParent();
22410 MachineRegisterInfo &MRI = MF.getRegInfo();
22412 // Check whether the addend is defined by a PHI:
22413 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
22414 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
22415 if (!AddendDef.isPHI())
22418 // Look for the following pattern:
22420 // %addend = phi [%entry, 0], [%loop, %result]
22422 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
22426 // %addend = phi [%entry, 0], [%loop, %result]
22428 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
22430 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
22431 assert(AddendDef.getOperand(i).isReg());
22432 MachineOperand PHISrcOp = AddendDef.getOperand(i);
22433 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
22434 if (&PHISrcInst == MI) {
22435 // Found a matching instruction.
22436 unsigned NewFMAOpc = 0;
22437 switch (MI->getOpcode()) {
22438 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
22439 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
22440 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
22441 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
22442 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
22443 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
22444 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
22445 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
22446 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
22447 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
22448 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
22449 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
22450 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
22451 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
22452 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
22453 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
22454 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
22455 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
22456 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
22457 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
22459 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
22460 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
22461 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
22462 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
22463 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
22464 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
22465 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
22466 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
22467 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
22468 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
22469 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
22470 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
22471 default: llvm_unreachable("Unrecognized FMA variant.");
22474 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
22475 MachineInstrBuilder MIB =
22476 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
22477 .addOperand(MI->getOperand(0))
22478 .addOperand(MI->getOperand(3))
22479 .addOperand(MI->getOperand(2))
22480 .addOperand(MI->getOperand(1));
22481 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
22482 MI->eraseFromParent();
22489 MachineBasicBlock *
22490 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
22491 MachineBasicBlock *BB) const {
22492 switch (MI->getOpcode()) {
22493 default: llvm_unreachable("Unexpected instr type to insert");
22494 case X86::TAILJMPd64:
22495 case X86::TAILJMPr64:
22496 case X86::TAILJMPm64:
22497 case X86::TAILJMPd64_REX:
22498 case X86::TAILJMPr64_REX:
22499 case X86::TAILJMPm64_REX:
22500 llvm_unreachable("TAILJMP64 would not be touched here.");
22501 case X86::TCRETURNdi64:
22502 case X86::TCRETURNri64:
22503 case X86::TCRETURNmi64:
22505 case X86::WIN_ALLOCA:
22506 return EmitLoweredWinAlloca(MI, BB);
22507 case X86::CATCHRET:
22508 return EmitLoweredCatchRet(MI, BB);
22509 case X86::CATCHPAD:
22510 return EmitLoweredCatchPad(MI, BB);
22511 case X86::SEG_ALLOCA_32:
22512 case X86::SEG_ALLOCA_64:
22513 return EmitLoweredSegAlloca(MI, BB);
22514 case X86::TLSCall_32:
22515 case X86::TLSCall_64:
22516 return EmitLoweredTLSCall(MI, BB);
22517 case X86::CMOV_FR32:
22518 case X86::CMOV_FR64:
22519 case X86::CMOV_FR128:
22520 case X86::CMOV_GR8:
22521 case X86::CMOV_GR16:
22522 case X86::CMOV_GR32:
22523 case X86::CMOV_RFP32:
22524 case X86::CMOV_RFP64:
22525 case X86::CMOV_RFP80:
22526 case X86::CMOV_V2F64:
22527 case X86::CMOV_V2I64:
22528 case X86::CMOV_V4F32:
22529 case X86::CMOV_V4F64:
22530 case X86::CMOV_V4I64:
22531 case X86::CMOV_V16F32:
22532 case X86::CMOV_V8F32:
22533 case X86::CMOV_V8F64:
22534 case X86::CMOV_V8I64:
22535 case X86::CMOV_V8I1:
22536 case X86::CMOV_V16I1:
22537 case X86::CMOV_V32I1:
22538 case X86::CMOV_V64I1:
22539 return EmitLoweredSelect(MI, BB);
22541 case X86::RDFLAGS32:
22542 case X86::RDFLAGS64: {
22543 DebugLoc DL = MI->getDebugLoc();
22544 MachineFunction *MF = BB->getParent();
22545 MF->getFrameInfo()->setHasOpaqueSPAdjustment(true);
22546 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22548 MI->getOpcode() == X86::RDFLAGS32 ? X86::PUSHF32 : X86::PUSHF64;
22550 MI->getOpcode() == X86::RDFLAGS32 ? X86::POP32r : X86::POP64r;
22551 BuildMI(*BB, MI, DL, TII->get(PushF));
22552 BuildMI(*BB, MI, DL, TII->get(Pop), MI->getOperand(0).getReg());
22554 MI->eraseFromParent(); // The pseudo is gone now.
22558 case X86::WRFLAGS32:
22559 case X86::WRFLAGS64: {
22560 DebugLoc DL = MI->getDebugLoc();
22561 MachineFunction *MF = BB->getParent();
22562 MF->getFrameInfo()->setHasOpaqueSPAdjustment(true);
22563 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22565 MI->getOpcode() == X86::WRFLAGS32 ? X86::PUSH32r : X86::PUSH64r;
22567 MI->getOpcode() == X86::WRFLAGS32 ? X86::POPF32 : X86::POPF64;
22568 BuildMI(*BB, MI, DL, TII->get(Push)).addReg(MI->getOperand(0).getReg());
22569 BuildMI(*BB, MI, DL, TII->get(PopF));
22571 MI->eraseFromParent(); // The pseudo is gone now.
22575 case X86::RELEASE_FADD32mr:
22576 case X86::RELEASE_FADD64mr:
22577 return EmitLoweredAtomicFP(MI, BB);
22579 case X86::FP32_TO_INT16_IN_MEM:
22580 case X86::FP32_TO_INT32_IN_MEM:
22581 case X86::FP32_TO_INT64_IN_MEM:
22582 case X86::FP64_TO_INT16_IN_MEM:
22583 case X86::FP64_TO_INT32_IN_MEM:
22584 case X86::FP64_TO_INT64_IN_MEM:
22585 case X86::FP80_TO_INT16_IN_MEM:
22586 case X86::FP80_TO_INT32_IN_MEM:
22587 case X86::FP80_TO_INT64_IN_MEM: {
22588 MachineFunction *F = BB->getParent();
22589 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
22590 DebugLoc DL = MI->getDebugLoc();
22592 // Change the floating point control register to use "round towards zero"
22593 // mode when truncating to an integer value.
22594 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
22595 addFrameReference(BuildMI(*BB, MI, DL,
22596 TII->get(X86::FNSTCW16m)), CWFrameIdx);
22598 // Load the old value of the high byte of the control word...
22600 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
22601 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
22604 // Set the high part to be round to zero...
22605 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
22608 // Reload the modified control word now...
22609 addFrameReference(BuildMI(*BB, MI, DL,
22610 TII->get(X86::FLDCW16m)), CWFrameIdx);
22612 // Restore the memory image of control word to original value
22613 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
22616 // Get the X86 opcode to use.
22618 switch (MI->getOpcode()) {
22619 default: llvm_unreachable("illegal opcode!");
22620 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
22621 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
22622 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
22623 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
22624 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
22625 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
22626 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
22627 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
22628 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
22632 MachineOperand &Op = MI->getOperand(0);
22634 AM.BaseType = X86AddressMode::RegBase;
22635 AM.Base.Reg = Op.getReg();
22637 AM.BaseType = X86AddressMode::FrameIndexBase;
22638 AM.Base.FrameIndex = Op.getIndex();
22640 Op = MI->getOperand(1);
22642 AM.Scale = Op.getImm();
22643 Op = MI->getOperand(2);
22645 AM.IndexReg = Op.getImm();
22646 Op = MI->getOperand(3);
22647 if (Op.isGlobal()) {
22648 AM.GV = Op.getGlobal();
22650 AM.Disp = Op.getImm();
22652 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
22653 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
22655 // Reload the original control word now.
22656 addFrameReference(BuildMI(*BB, MI, DL,
22657 TII->get(X86::FLDCW16m)), CWFrameIdx);
22659 MI->eraseFromParent(); // The pseudo instruction is gone now.
22662 // String/text processing lowering.
22663 case X86::PCMPISTRM128REG:
22664 case X86::VPCMPISTRM128REG:
22665 case X86::PCMPISTRM128MEM:
22666 case X86::VPCMPISTRM128MEM:
22667 case X86::PCMPESTRM128REG:
22668 case X86::VPCMPESTRM128REG:
22669 case X86::PCMPESTRM128MEM:
22670 case X86::VPCMPESTRM128MEM:
22671 assert(Subtarget->hasSSE42() &&
22672 "Target must have SSE4.2 or AVX features enabled");
22673 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
22675 // String/text processing lowering.
22676 case X86::PCMPISTRIREG:
22677 case X86::VPCMPISTRIREG:
22678 case X86::PCMPISTRIMEM:
22679 case X86::VPCMPISTRIMEM:
22680 case X86::PCMPESTRIREG:
22681 case X86::VPCMPESTRIREG:
22682 case X86::PCMPESTRIMEM:
22683 case X86::VPCMPESTRIMEM:
22684 assert(Subtarget->hasSSE42() &&
22685 "Target must have SSE4.2 or AVX features enabled");
22686 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
22688 // Thread synchronization.
22690 return EmitMonitor(MI, BB, Subtarget);
22693 return EmitWRPKRU(MI, BB, Subtarget);
22695 return EmitRDPKRU(MI, BB, Subtarget);
22698 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
22700 case X86::VASTART_SAVE_XMM_REGS:
22701 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22703 case X86::VAARG_64:
22704 return EmitVAARG64WithCustomInserter(MI, BB);
22706 case X86::EH_SjLj_SetJmp32:
22707 case X86::EH_SjLj_SetJmp64:
22708 return emitEHSjLjSetJmp(MI, BB);
22710 case X86::EH_SjLj_LongJmp32:
22711 case X86::EH_SjLj_LongJmp64:
22712 return emitEHSjLjLongJmp(MI, BB);
22714 case TargetOpcode::STATEPOINT:
22715 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22716 // this point in the process. We diverge later.
22717 return emitPatchPoint(MI, BB);
22719 case TargetOpcode::STACKMAP:
22720 case TargetOpcode::PATCHPOINT:
22721 return emitPatchPoint(MI, BB);
22723 case X86::VFMADDPDr213r:
22724 case X86::VFMADDPSr213r:
22725 case X86::VFMADDSDr213r:
22726 case X86::VFMADDSSr213r:
22727 case X86::VFMSUBPDr213r:
22728 case X86::VFMSUBPSr213r:
22729 case X86::VFMSUBSDr213r:
22730 case X86::VFMSUBSSr213r:
22731 case X86::VFNMADDPDr213r:
22732 case X86::VFNMADDPSr213r:
22733 case X86::VFNMADDSDr213r:
22734 case X86::VFNMADDSSr213r:
22735 case X86::VFNMSUBPDr213r:
22736 case X86::VFNMSUBPSr213r:
22737 case X86::VFNMSUBSDr213r:
22738 case X86::VFNMSUBSSr213r:
22739 case X86::VFMADDSUBPDr213r:
22740 case X86::VFMADDSUBPSr213r:
22741 case X86::VFMSUBADDPDr213r:
22742 case X86::VFMSUBADDPSr213r:
22743 case X86::VFMADDPDr213rY:
22744 case X86::VFMADDPSr213rY:
22745 case X86::VFMSUBPDr213rY:
22746 case X86::VFMSUBPSr213rY:
22747 case X86::VFNMADDPDr213rY:
22748 case X86::VFNMADDPSr213rY:
22749 case X86::VFNMSUBPDr213rY:
22750 case X86::VFNMSUBPSr213rY:
22751 case X86::VFMADDSUBPDr213rY:
22752 case X86::VFMADDSUBPSr213rY:
22753 case X86::VFMSUBADDPDr213rY:
22754 case X86::VFMSUBADDPSr213rY:
22755 return emitFMA3Instr(MI, BB);
22759 //===----------------------------------------------------------------------===//
22760 // X86 Optimization Hooks
22761 //===----------------------------------------------------------------------===//
22763 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22766 const SelectionDAG &DAG,
22767 unsigned Depth) const {
22768 unsigned BitWidth = KnownZero.getBitWidth();
22769 unsigned Opc = Op.getOpcode();
22770 assert((Opc >= ISD::BUILTIN_OP_END ||
22771 Opc == ISD::INTRINSIC_WO_CHAIN ||
22772 Opc == ISD::INTRINSIC_W_CHAIN ||
22773 Opc == ISD::INTRINSIC_VOID) &&
22774 "Should use MaskedValueIsZero if you don't know whether Op"
22775 " is a target node!");
22777 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22791 // These nodes' second result is a boolean.
22792 if (Op.getResNo() == 0)
22795 case X86ISD::SETCC:
22796 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22798 case ISD::INTRINSIC_WO_CHAIN: {
22799 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22800 unsigned NumLoBits = 0;
22803 case Intrinsic::x86_sse_movmsk_ps:
22804 case Intrinsic::x86_avx_movmsk_ps_256:
22805 case Intrinsic::x86_sse2_movmsk_pd:
22806 case Intrinsic::x86_avx_movmsk_pd_256:
22807 case Intrinsic::x86_mmx_pmovmskb:
22808 case Intrinsic::x86_sse2_pmovmskb_128:
22809 case Intrinsic::x86_avx2_pmovmskb: {
22810 // High bits of movmskp{s|d}, pmovmskb are known zero.
22812 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22813 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22814 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22815 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22816 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22817 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22818 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22819 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22821 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22830 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22832 const SelectionDAG &,
22833 unsigned Depth) const {
22834 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22835 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22836 return Op.getValueType().getScalarSizeInBits();
22842 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22843 /// node is a GlobalAddress + offset.
22844 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22845 const GlobalValue* &GA,
22846 int64_t &Offset) const {
22847 if (N->getOpcode() == X86ISD::Wrapper) {
22848 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22849 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22850 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22854 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22857 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22858 /// FIXME: This could be expanded to support 512 bit vectors as well.
22859 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22860 TargetLowering::DAGCombinerInfo &DCI,
22861 const X86Subtarget* Subtarget) {
22863 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22864 SDValue V1 = SVOp->getOperand(0);
22865 SDValue V2 = SVOp->getOperand(1);
22866 MVT VT = SVOp->getSimpleValueType(0);
22867 unsigned NumElems = VT.getVectorNumElements();
22869 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22870 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22874 // V UNDEF BUILD_VECTOR UNDEF
22876 // CONCAT_VECTOR CONCAT_VECTOR
22879 // RESULT: V + zero extended
22881 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22882 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22883 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22886 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22889 // To match the shuffle mask, the first half of the mask should
22890 // be exactly the first vector, and all the rest a splat with the
22891 // first element of the second one.
22892 for (unsigned i = 0; i != NumElems/2; ++i)
22893 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22894 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22897 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22898 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22899 if (Ld->hasNUsesOfValue(1, 0)) {
22900 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22901 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22903 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22905 Ld->getPointerInfo(),
22906 Ld->getAlignment(),
22907 false/*isVolatile*/, true/*ReadMem*/,
22908 false/*WriteMem*/);
22910 // Make sure the newly-created LOAD is in the same position as Ld in
22911 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22912 // and update uses of Ld's output chain to use the TokenFactor.
22913 if (Ld->hasAnyUseOfValue(1)) {
22914 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22915 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22916 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22917 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22918 SDValue(ResNode.getNode(), 1));
22921 return DAG.getBitcast(VT, ResNode);
22925 // Emit a zeroed vector and insert the desired subvector on its
22927 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22928 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22929 return DCI.CombineTo(N, InsV);
22935 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22938 /// This is the leaf of the recursive combinine below. When we have found some
22939 /// chain of single-use x86 shuffle instructions and accumulated the combined
22940 /// shuffle mask represented by them, this will try to pattern match that mask
22941 /// into either a single instruction if there is a special purpose instruction
22942 /// for this operation, or into a PSHUFB instruction which is a fully general
22943 /// instruction but should only be used to replace chains over a certain depth.
22944 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22945 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22946 TargetLowering::DAGCombinerInfo &DCI,
22947 const X86Subtarget *Subtarget) {
22948 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22950 // Find the operand that enters the chain. Note that multiple uses are OK
22951 // here, we're not going to remove the operand we find.
22952 SDValue Input = Op.getOperand(0);
22953 while (Input.getOpcode() == ISD::BITCAST)
22954 Input = Input.getOperand(0);
22956 MVT VT = Input.getSimpleValueType();
22957 MVT RootVT = Root.getSimpleValueType();
22960 if (Mask.size() == 1) {
22961 int Index = Mask[0];
22962 assert((Index >= 0 || Index == SM_SentinelUndef ||
22963 Index == SM_SentinelZero) &&
22964 "Invalid shuffle index found!");
22966 // We may end up with an accumulated mask of size 1 as a result of
22967 // widening of shuffle operands (see function canWidenShuffleElements).
22968 // If the only shuffle index is equal to SM_SentinelZero then propagate
22969 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22970 // mask, and therefore the entire chain of shuffles can be folded away.
22971 if (Index == SM_SentinelZero)
22972 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22974 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22979 // Use the float domain if the operand type is a floating point type.
22980 bool FloatDomain = VT.isFloatingPoint();
22982 // For floating point shuffles, we don't have free copies in the shuffle
22983 // instructions or the ability to load as part of the instruction, so
22984 // canonicalize their shuffles to UNPCK or MOV variants.
22986 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22987 // vectors because it can have a load folded into it that UNPCK cannot. This
22988 // doesn't preclude something switching to the shorter encoding post-RA.
22990 // FIXME: Should teach these routines about AVX vector widths.
22991 if (FloatDomain && VT.is128BitVector()) {
22992 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22993 bool Lo = Mask.equals({0, 0});
22996 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22997 // is no slower than UNPCKLPD but has the option to fold the input operand
22998 // into even an unaligned memory load.
22999 if (Lo && Subtarget->hasSSE3()) {
23000 Shuffle = X86ISD::MOVDDUP;
23001 ShuffleVT = MVT::v2f64;
23003 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
23004 // than the UNPCK variants.
23005 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
23006 ShuffleVT = MVT::v4f32;
23008 if (Depth == 1 && Root->getOpcode() == Shuffle)
23009 return false; // Nothing to do!
23010 Op = DAG.getBitcast(ShuffleVT, Input);
23011 DCI.AddToWorklist(Op.getNode());
23012 if (Shuffle == X86ISD::MOVDDUP)
23013 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23015 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23016 DCI.AddToWorklist(Op.getNode());
23017 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23021 if (Subtarget->hasSSE3() &&
23022 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
23023 bool Lo = Mask.equals({0, 0, 2, 2});
23024 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
23025 MVT ShuffleVT = MVT::v4f32;
23026 if (Depth == 1 && Root->getOpcode() == Shuffle)
23027 return false; // Nothing to do!
23028 Op = DAG.getBitcast(ShuffleVT, Input);
23029 DCI.AddToWorklist(Op.getNode());
23030 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
23031 DCI.AddToWorklist(Op.getNode());
23032 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23036 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
23037 bool Lo = Mask.equals({0, 0, 1, 1});
23038 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23039 MVT ShuffleVT = MVT::v4f32;
23040 if (Depth == 1 && Root->getOpcode() == Shuffle)
23041 return false; // Nothing to do!
23042 Op = DAG.getBitcast(ShuffleVT, Input);
23043 DCI.AddToWorklist(Op.getNode());
23044 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23045 DCI.AddToWorklist(Op.getNode());
23046 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23052 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
23053 // variants as none of these have single-instruction variants that are
23054 // superior to the UNPCK formulation.
23055 if (!FloatDomain && VT.is128BitVector() &&
23056 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23057 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
23058 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
23060 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
23061 bool Lo = Mask[0] == 0;
23062 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
23063 if (Depth == 1 && Root->getOpcode() == Shuffle)
23064 return false; // Nothing to do!
23066 switch (Mask.size()) {
23068 ShuffleVT = MVT::v8i16;
23071 ShuffleVT = MVT::v16i8;
23074 llvm_unreachable("Impossible mask size!");
23076 Op = DAG.getBitcast(ShuffleVT, Input);
23077 DCI.AddToWorklist(Op.getNode());
23078 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
23079 DCI.AddToWorklist(Op.getNode());
23080 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23085 // Don't try to re-form single instruction chains under any circumstances now
23086 // that we've done encoding canonicalization for them.
23090 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
23091 // can replace them with a single PSHUFB instruction profitably. Intel's
23092 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
23093 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
23094 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
23095 SmallVector<SDValue, 16> PSHUFBMask;
23096 int NumBytes = VT.getSizeInBits() / 8;
23097 int Ratio = NumBytes / Mask.size();
23098 for (int i = 0; i < NumBytes; ++i) {
23099 if (Mask[i / Ratio] == SM_SentinelUndef) {
23100 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
23103 int M = Mask[i / Ratio] != SM_SentinelZero
23104 ? Ratio * Mask[i / Ratio] + i % Ratio
23106 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
23108 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
23109 Op = DAG.getBitcast(ByteVT, Input);
23110 DCI.AddToWorklist(Op.getNode());
23111 SDValue PSHUFBMaskOp =
23112 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
23113 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
23114 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
23115 DCI.AddToWorklist(Op.getNode());
23116 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
23121 // Failed to find any combines.
23125 /// \brief Fully generic combining of x86 shuffle instructions.
23127 /// This should be the last combine run over the x86 shuffle instructions. Once
23128 /// they have been fully optimized, this will recursively consider all chains
23129 /// of single-use shuffle instructions, build a generic model of the cumulative
23130 /// shuffle operation, and check for simpler instructions which implement this
23131 /// operation. We use this primarily for two purposes:
23133 /// 1) Collapse generic shuffles to specialized single instructions when
23134 /// equivalent. In most cases, this is just an encoding size win, but
23135 /// sometimes we will collapse multiple generic shuffles into a single
23136 /// special-purpose shuffle.
23137 /// 2) Look for sequences of shuffle instructions with 3 or more total
23138 /// instructions, and replace them with the slightly more expensive SSSE3
23139 /// PSHUFB instruction if available. We do this as the last combining step
23140 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
23141 /// a suitable short sequence of other instructions. The PHUFB will either
23142 /// use a register or have to read from memory and so is slightly (but only
23143 /// slightly) more expensive than the other shuffle instructions.
23145 /// Because this is inherently a quadratic operation (for each shuffle in
23146 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
23147 /// This should never be an issue in practice as the shuffle lowering doesn't
23148 /// produce sequences of more than 8 instructions.
23150 /// FIXME: We will currently miss some cases where the redundant shuffling
23151 /// would simplify under the threshold for PSHUFB formation because of
23152 /// combine-ordering. To fix this, we should do the redundant instruction
23153 /// combining in this recursive walk.
23154 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
23155 ArrayRef<int> RootMask,
23156 int Depth, bool HasPSHUFB,
23158 TargetLowering::DAGCombinerInfo &DCI,
23159 const X86Subtarget *Subtarget) {
23160 // Bound the depth of our recursive combine because this is ultimately
23161 // quadratic in nature.
23165 // Directly rip through bitcasts to find the underlying operand.
23166 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
23167 Op = Op.getOperand(0);
23169 MVT VT = Op.getSimpleValueType();
23170 if (!VT.isVector())
23171 return false; // Bail if we hit a non-vector.
23173 assert(Root.getSimpleValueType().isVector() &&
23174 "Shuffles operate on vector types!");
23175 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
23176 "Can only combine shuffles of the same vector register size.");
23178 if (!isTargetShuffle(Op.getOpcode()))
23180 SmallVector<int, 16> OpMask;
23182 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
23183 // We only can combine unary shuffles which we can decode the mask for.
23184 if (!HaveMask || !IsUnary)
23187 assert(VT.getVectorNumElements() == OpMask.size() &&
23188 "Different mask size from vector size!");
23189 assert(((RootMask.size() > OpMask.size() &&
23190 RootMask.size() % OpMask.size() == 0) ||
23191 (OpMask.size() > RootMask.size() &&
23192 OpMask.size() % RootMask.size() == 0) ||
23193 OpMask.size() == RootMask.size()) &&
23194 "The smaller number of elements must divide the larger.");
23195 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
23196 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
23197 assert(((RootRatio == 1 && OpRatio == 1) ||
23198 (RootRatio == 1) != (OpRatio == 1)) &&
23199 "Must not have a ratio for both incoming and op masks!");
23201 SmallVector<int, 16> Mask;
23202 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
23204 // Merge this shuffle operation's mask into our accumulated mask. Note that
23205 // this shuffle's mask will be the first applied to the input, followed by the
23206 // root mask to get us all the way to the root value arrangement. The reason
23207 // for this order is that we are recursing up the operation chain.
23208 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
23209 int RootIdx = i / RootRatio;
23210 if (RootMask[RootIdx] < 0) {
23211 // This is a zero or undef lane, we're done.
23212 Mask.push_back(RootMask[RootIdx]);
23216 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
23217 int OpIdx = RootMaskedIdx / OpRatio;
23218 if (OpMask[OpIdx] < 0) {
23219 // The incoming lanes are zero or undef, it doesn't matter which ones we
23221 Mask.push_back(OpMask[OpIdx]);
23225 // Ok, we have non-zero lanes, map them through.
23226 Mask.push_back(OpMask[OpIdx] * OpRatio +
23227 RootMaskedIdx % OpRatio);
23230 // See if we can recurse into the operand to combine more things.
23231 switch (Op.getOpcode()) {
23232 case X86ISD::PSHUFB:
23234 case X86ISD::PSHUFD:
23235 case X86ISD::PSHUFHW:
23236 case X86ISD::PSHUFLW:
23237 if (Op.getOperand(0).hasOneUse() &&
23238 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23239 HasPSHUFB, DAG, DCI, Subtarget))
23243 case X86ISD::UNPCKL:
23244 case X86ISD::UNPCKH:
23245 assert(Op.getOperand(0) == Op.getOperand(1) &&
23246 "We only combine unary shuffles!");
23247 // We can't check for single use, we have to check that this shuffle is the
23249 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
23250 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
23251 HasPSHUFB, DAG, DCI, Subtarget))
23256 // Minor canonicalization of the accumulated shuffle mask to make it easier
23257 // to match below. All this does is detect masks with squential pairs of
23258 // elements, and shrink them to the half-width mask. It does this in a loop
23259 // so it will reduce the size of the mask to the minimal width mask which
23260 // performs an equivalent shuffle.
23261 SmallVector<int, 16> WidenedMask;
23262 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
23263 Mask = std::move(WidenedMask);
23264 WidenedMask.clear();
23267 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
23271 /// \brief Get the PSHUF-style mask from PSHUF node.
23273 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
23274 /// PSHUF-style masks that can be reused with such instructions.
23275 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
23276 MVT VT = N.getSimpleValueType();
23277 SmallVector<int, 4> Mask;
23279 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
23283 // If we have more than 128-bits, only the low 128-bits of shuffle mask
23284 // matter. Check that the upper masks are repeats and remove them.
23285 if (VT.getSizeInBits() > 128) {
23286 int LaneElts = 128 / VT.getScalarSizeInBits();
23288 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
23289 for (int j = 0; j < LaneElts; ++j)
23290 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
23291 "Mask doesn't repeat in high 128-bit lanes!");
23293 Mask.resize(LaneElts);
23296 switch (N.getOpcode()) {
23297 case X86ISD::PSHUFD:
23299 case X86ISD::PSHUFLW:
23302 case X86ISD::PSHUFHW:
23303 Mask.erase(Mask.begin(), Mask.begin() + 4);
23304 for (int &M : Mask)
23308 llvm_unreachable("No valid shuffle instruction found!");
23312 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
23314 /// We walk up the chain and look for a combinable shuffle, skipping over
23315 /// shuffles that we could hoist this shuffle's transformation past without
23316 /// altering anything.
23318 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
23320 TargetLowering::DAGCombinerInfo &DCI) {
23321 assert(N.getOpcode() == X86ISD::PSHUFD &&
23322 "Called with something other than an x86 128-bit half shuffle!");
23325 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
23326 // of the shuffles in the chain so that we can form a fresh chain to replace
23328 SmallVector<SDValue, 8> Chain;
23329 SDValue V = N.getOperand(0);
23330 for (; V.hasOneUse(); V = V.getOperand(0)) {
23331 switch (V.getOpcode()) {
23333 return SDValue(); // Nothing combined!
23336 // Skip bitcasts as we always know the type for the target specific
23340 case X86ISD::PSHUFD:
23341 // Found another dword shuffle.
23344 case X86ISD::PSHUFLW:
23345 // Check that the low words (being shuffled) are the identity in the
23346 // dword shuffle, and the high words are self-contained.
23347 if (Mask[0] != 0 || Mask[1] != 1 ||
23348 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
23351 Chain.push_back(V);
23354 case X86ISD::PSHUFHW:
23355 // Check that the high words (being shuffled) are the identity in the
23356 // dword shuffle, and the low words are self-contained.
23357 if (Mask[2] != 2 || Mask[3] != 3 ||
23358 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
23361 Chain.push_back(V);
23364 case X86ISD::UNPCKL:
23365 case X86ISD::UNPCKH:
23366 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
23367 // shuffle into a preceding word shuffle.
23368 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
23369 V.getSimpleValueType().getVectorElementType() != MVT::i16)
23372 // Search for a half-shuffle which we can combine with.
23373 unsigned CombineOp =
23374 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
23375 if (V.getOperand(0) != V.getOperand(1) ||
23376 !V->isOnlyUserOf(V.getOperand(0).getNode()))
23378 Chain.push_back(V);
23379 V = V.getOperand(0);
23381 switch (V.getOpcode()) {
23383 return SDValue(); // Nothing to combine.
23385 case X86ISD::PSHUFLW:
23386 case X86ISD::PSHUFHW:
23387 if (V.getOpcode() == CombineOp)
23390 Chain.push_back(V);
23394 V = V.getOperand(0);
23398 } while (V.hasOneUse());
23401 // Break out of the loop if we break out of the switch.
23405 if (!V.hasOneUse())
23406 // We fell out of the loop without finding a viable combining instruction.
23409 // Merge this node's mask and our incoming mask.
23410 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23411 for (int &M : Mask)
23413 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
23414 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23416 // Rebuild the chain around this new shuffle.
23417 while (!Chain.empty()) {
23418 SDValue W = Chain.pop_back_val();
23420 if (V.getValueType() != W.getOperand(0).getValueType())
23421 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
23423 switch (W.getOpcode()) {
23425 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
23427 case X86ISD::UNPCKL:
23428 case X86ISD::UNPCKH:
23429 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
23432 case X86ISD::PSHUFD:
23433 case X86ISD::PSHUFLW:
23434 case X86ISD::PSHUFHW:
23435 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
23439 if (V.getValueType() != N.getValueType())
23440 V = DAG.getBitcast(N.getValueType(), V);
23442 // Return the new chain to replace N.
23446 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
23449 /// We walk up the chain, skipping shuffles of the other half and looking
23450 /// through shuffles which switch halves trying to find a shuffle of the same
23451 /// pair of dwords.
23452 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
23454 TargetLowering::DAGCombinerInfo &DCI) {
23456 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
23457 "Called with something other than an x86 128-bit half shuffle!");
23459 unsigned CombineOpcode = N.getOpcode();
23461 // Walk up a single-use chain looking for a combinable shuffle.
23462 SDValue V = N.getOperand(0);
23463 for (; V.hasOneUse(); V = V.getOperand(0)) {
23464 switch (V.getOpcode()) {
23466 return false; // Nothing combined!
23469 // Skip bitcasts as we always know the type for the target specific
23473 case X86ISD::PSHUFLW:
23474 case X86ISD::PSHUFHW:
23475 if (V.getOpcode() == CombineOpcode)
23478 // Other-half shuffles are no-ops.
23481 // Break out of the loop if we break out of the switch.
23485 if (!V.hasOneUse())
23486 // We fell out of the loop without finding a viable combining instruction.
23489 // Combine away the bottom node as its shuffle will be accumulated into
23490 // a preceding shuffle.
23491 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23493 // Record the old value.
23496 // Merge this node's mask and our incoming mask (adjusted to account for all
23497 // the pshufd instructions encountered).
23498 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23499 for (int &M : Mask)
23501 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
23502 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
23504 // Check that the shuffles didn't cancel each other out. If not, we need to
23505 // combine to the new one.
23507 // Replace the combinable shuffle with the combined one, updating all users
23508 // so that we re-evaluate the chain here.
23509 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
23514 /// \brief Try to combine x86 target specific shuffles.
23515 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
23516 TargetLowering::DAGCombinerInfo &DCI,
23517 const X86Subtarget *Subtarget) {
23519 MVT VT = N.getSimpleValueType();
23520 SmallVector<int, 4> Mask;
23522 switch (N.getOpcode()) {
23523 case X86ISD::PSHUFD:
23524 case X86ISD::PSHUFLW:
23525 case X86ISD::PSHUFHW:
23526 Mask = getPSHUFShuffleMask(N);
23527 assert(Mask.size() == 4);
23529 case X86ISD::UNPCKL: {
23530 // Combine X86ISD::UNPCKL and ISD::VECTOR_SHUFFLE into X86ISD::UNPCKH, in
23531 // which X86ISD::UNPCKL has a ISD::UNDEF operand, and ISD::VECTOR_SHUFFLE
23532 // moves upper half elements into the lower half part. For example:
23534 // t2: v16i8 = vector_shuffle<8,9,10,11,12,13,14,15,u,u,u,u,u,u,u,u> t1,
23536 // t3: v16i8 = X86ISD::UNPCKL undef:v16i8, t2
23538 // will be combined to:
23540 // t3: v16i8 = X86ISD::UNPCKH undef:v16i8, t1
23542 // This is only for 128-bit vectors. From SSE4.1 onward this combine may not
23543 // happen due to advanced instructions.
23544 if (!VT.is128BitVector())
23547 auto Op0 = N.getOperand(0);
23548 auto Op1 = N.getOperand(1);
23549 if (Op0.getOpcode() == ISD::UNDEF &&
23550 Op1.getNode()->getOpcode() == ISD::VECTOR_SHUFFLE) {
23551 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op1.getNode())->getMask();
23553 unsigned NumElts = VT.getVectorNumElements();
23554 SmallVector<int, 8> ExpectedMask(NumElts, -1);
23555 std::iota(ExpectedMask.begin(), ExpectedMask.begin() + NumElts / 2,
23558 auto ShufOp = Op1.getOperand(0);
23559 if (isShuffleEquivalent(Op1, ShufOp, Mask, ExpectedMask))
23560 return DAG.getNode(X86ISD::UNPCKH, DL, VT, N.getOperand(0), ShufOp);
23564 case X86ISD::BLENDI: {
23565 SDValue V0 = N->getOperand(0);
23566 SDValue V1 = N->getOperand(1);
23567 assert(VT == V0.getSimpleValueType() && VT == V1.getSimpleValueType() &&
23568 "Unexpected input vector types");
23570 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
23571 // operands and changing the mask to 1. This saves us a bunch of
23572 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
23573 // x86InstrInfo knows how to commute this back after instruction selection
23574 // if it would help register allocation.
23576 // TODO: If optimizing for size or a processor that doesn't suffer from
23577 // partial register update stalls, this should be transformed into a MOVSD
23578 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
23580 if (VT == MVT::v2f64)
23581 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
23582 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
23583 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
23584 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23593 // Nuke no-op shuffles that show up after combining.
23594 if (isNoopShuffleMask(Mask))
23595 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
23597 // Look for simplifications involving one or two shuffle instructions.
23598 SDValue V = N.getOperand(0);
23599 switch (N.getOpcode()) {
23602 case X86ISD::PSHUFLW:
23603 case X86ISD::PSHUFHW:
23604 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
23606 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
23607 return SDValue(); // We combined away this shuffle, so we're done.
23609 // See if this reduces to a PSHUFD which is no more expensive and can
23610 // combine with more operations. Note that it has to at least flip the
23611 // dwords as otherwise it would have been removed as a no-op.
23612 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
23613 int DMask[] = {0, 1, 2, 3};
23614 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
23615 DMask[DOffset + 0] = DOffset + 1;
23616 DMask[DOffset + 1] = DOffset + 0;
23617 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
23618 V = DAG.getBitcast(DVT, V);
23619 DCI.AddToWorklist(V.getNode());
23620 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
23621 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
23622 DCI.AddToWorklist(V.getNode());
23623 return DAG.getBitcast(VT, V);
23626 // Look for shuffle patterns which can be implemented as a single unpack.
23627 // FIXME: This doesn't handle the location of the PSHUFD generically, and
23628 // only works when we have a PSHUFD followed by two half-shuffles.
23629 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
23630 (V.getOpcode() == X86ISD::PSHUFLW ||
23631 V.getOpcode() == X86ISD::PSHUFHW) &&
23632 V.getOpcode() != N.getOpcode() &&
23634 SDValue D = V.getOperand(0);
23635 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
23636 D = D.getOperand(0);
23637 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
23638 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
23639 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
23640 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23641 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
23643 for (int i = 0; i < 4; ++i) {
23644 WordMask[i + NOffset] = Mask[i] + NOffset;
23645 WordMask[i + VOffset] = VMask[i] + VOffset;
23647 // Map the word mask through the DWord mask.
23649 for (int i = 0; i < 8; ++i)
23650 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
23651 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
23652 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
23653 // We can replace all three shuffles with an unpack.
23654 V = DAG.getBitcast(VT, D.getOperand(0));
23655 DCI.AddToWorklist(V.getNode());
23656 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
23665 case X86ISD::PSHUFD:
23666 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
23675 /// \brief Try to combine a shuffle into a target-specific add-sub node.
23677 /// We combine this directly on the abstract vector shuffle nodes so it is
23678 /// easier to generically match. We also insert dummy vector shuffle nodes for
23679 /// the operands which explicitly discard the lanes which are unused by this
23680 /// operation to try to flow through the rest of the combiner the fact that
23681 /// they're unused.
23682 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
23684 EVT VT = N->getValueType(0);
23686 // We only handle target-independent shuffles.
23687 // FIXME: It would be easy and harmless to use the target shuffle mask
23688 // extraction tool to support more.
23689 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
23692 auto *SVN = cast<ShuffleVectorSDNode>(N);
23693 SmallVector<int, 8> Mask;
23694 for (int M : SVN->getMask())
23697 SDValue V1 = N->getOperand(0);
23698 SDValue V2 = N->getOperand(1);
23700 // We require the first shuffle operand to be the FSUB node, and the second to
23701 // be the FADD node.
23702 if (V1.getOpcode() == ISD::FADD && V2.getOpcode() == ISD::FSUB) {
23703 ShuffleVectorSDNode::commuteMask(Mask);
23705 } else if (V1.getOpcode() != ISD::FSUB || V2.getOpcode() != ISD::FADD)
23708 // If there are other uses of these operations we can't fold them.
23709 if (!V1->hasOneUse() || !V2->hasOneUse())
23712 // Ensure that both operations have the same operands. Note that we can
23713 // commute the FADD operands.
23714 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
23715 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
23716 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23719 // We're looking for blends between FADD and FSUB nodes. We insist on these
23720 // nodes being lined up in a specific expected pattern.
23721 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23722 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23723 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23726 // Only specific types are legal at this point, assert so we notice if and
23727 // when these change.
23728 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23729 VT == MVT::v4f64) &&
23730 "Unknown vector type encountered!");
23732 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23735 /// PerformShuffleCombine - Performs several different shuffle combines.
23736 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23737 TargetLowering::DAGCombinerInfo &DCI,
23738 const X86Subtarget *Subtarget) {
23740 SDValue N0 = N->getOperand(0);
23741 SDValue N1 = N->getOperand(1);
23742 EVT VT = N->getValueType(0);
23744 // Don't create instructions with illegal types after legalize types has run.
23745 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23746 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23749 // If we have legalized the vector types, look for blends of FADD and FSUB
23750 // nodes that we can fuse into an ADDSUB node.
23751 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23752 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23755 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23756 if (TLI.isTypeLegal(VT) && Subtarget->hasFp256() && VT.is256BitVector() &&
23757 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23758 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23760 // During Type Legalization, when promoting illegal vector types,
23761 // the backend might introduce new shuffle dag nodes and bitcasts.
23763 // This code performs the following transformation:
23764 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23765 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23767 // We do this only if both the bitcast and the BINOP dag nodes have
23768 // one use. Also, perform this transformation only if the new binary
23769 // operation is legal. This is to avoid introducing dag nodes that
23770 // potentially need to be further expanded (or custom lowered) into a
23771 // less optimal sequence of dag nodes.
23772 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23773 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23774 N0.getOpcode() == ISD::BITCAST) {
23775 SDValue BC0 = N0.getOperand(0);
23776 EVT SVT = BC0.getValueType();
23777 unsigned Opcode = BC0.getOpcode();
23778 unsigned NumElts = VT.getVectorNumElements();
23780 if (BC0.hasOneUse() && SVT.isVector() &&
23781 SVT.getVectorNumElements() * 2 == NumElts &&
23782 TLI.isOperationLegal(Opcode, VT)) {
23783 bool CanFold = false;
23795 unsigned SVTNumElts = SVT.getVectorNumElements();
23796 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23797 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23798 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23799 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23800 CanFold = SVOp->getMaskElt(i) < 0;
23803 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23804 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23805 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23806 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23811 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23812 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23813 // consecutive, non-overlapping, and in the right order.
23814 SmallVector<SDValue, 16> Elts;
23815 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23816 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23818 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23821 if (isTargetShuffle(N->getOpcode())) {
23823 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23824 if (Shuffle.getNode())
23827 // Try recursively combining arbitrary sequences of x86 shuffle
23828 // instructions into higher-order shuffles. We do this after combining
23829 // specific PSHUF instruction sequences into their minimal form so that we
23830 // can evaluate how many specialized shuffle instructions are involved in
23831 // a particular chain.
23832 SmallVector<int, 1> NonceMask; // Just a placeholder.
23833 NonceMask.push_back(0);
23834 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23835 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23837 return SDValue(); // This routine will use CombineTo to replace N.
23843 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23844 /// specific shuffle of a load can be folded into a single element load.
23845 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23846 /// shuffles have been custom lowered so we need to handle those here.
23847 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23848 TargetLowering::DAGCombinerInfo &DCI) {
23849 if (DCI.isBeforeLegalizeOps())
23852 SDValue InVec = N->getOperand(0);
23853 SDValue EltNo = N->getOperand(1);
23855 if (!isa<ConstantSDNode>(EltNo))
23858 EVT OriginalVT = InVec.getValueType();
23860 if (InVec.getOpcode() == ISD::BITCAST) {
23861 // Don't duplicate a load with other uses.
23862 if (!InVec.hasOneUse())
23864 EVT BCVT = InVec.getOperand(0).getValueType();
23865 if (!BCVT.isVector() ||
23866 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23868 InVec = InVec.getOperand(0);
23871 EVT CurrentVT = InVec.getValueType();
23873 if (!isTargetShuffle(InVec.getOpcode()))
23876 // Don't duplicate a load with other uses.
23877 if (!InVec.hasOneUse())
23880 SmallVector<int, 16> ShuffleMask;
23882 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23883 ShuffleMask, UnaryShuffle))
23886 // Select the input vector, guarding against out of range extract vector.
23887 unsigned NumElems = CurrentVT.getVectorNumElements();
23888 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23889 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23890 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23891 : InVec.getOperand(1);
23893 // If inputs to shuffle are the same for both ops, then allow 2 uses
23894 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23895 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23897 if (LdNode.getOpcode() == ISD::BITCAST) {
23898 // Don't duplicate a load with other uses.
23899 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23902 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23903 LdNode = LdNode.getOperand(0);
23906 if (!ISD::isNormalLoad(LdNode.getNode()))
23909 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23911 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23914 EVT EltVT = N->getValueType(0);
23915 // If there's a bitcast before the shuffle, check if the load type and
23916 // alignment is valid.
23917 unsigned Align = LN0->getAlignment();
23918 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23919 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23920 EltVT.getTypeForEVT(*DAG.getContext()));
23922 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23925 // All checks match so transform back to vector_shuffle so that DAG combiner
23926 // can finish the job
23929 // Create shuffle node taking into account the case that its a unary shuffle
23930 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23931 : InVec.getOperand(1);
23932 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23933 InVec.getOperand(0), Shuffle,
23935 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23936 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23940 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23941 const X86Subtarget *Subtarget) {
23942 SDValue N0 = N->getOperand(0);
23943 EVT VT = N->getValueType(0);
23945 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23946 // special and don't usually play with other vector types, it's better to
23947 // handle them early to be sure we emit efficient code by avoiding
23948 // store-load conversions.
23949 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23950 N0.getValueType() == MVT::v2i32 &&
23951 isNullConstant(N0.getOperand(1))) {
23952 SDValue N00 = N0->getOperand(0);
23953 if (N00.getValueType() == MVT::i32)
23954 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23957 // Convert a bitcasted integer logic operation that has one bitcasted
23958 // floating-point operand and one constant operand into a floating-point
23959 // logic operation. This may create a load of the constant, but that is
23960 // cheaper than materializing the constant in an integer register and
23961 // transferring it to an SSE register or transferring the SSE operand to
23962 // integer register and back.
23964 switch (N0.getOpcode()) {
23965 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23966 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23967 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23968 default: return SDValue();
23970 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23971 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23972 isa<ConstantSDNode>(N0.getOperand(1)) &&
23973 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23974 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23975 SDValue N000 = N0.getOperand(0).getOperand(0);
23976 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23977 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23983 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23984 /// generation and convert it from being a bunch of shuffles and extracts
23985 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23986 /// storing the value and loading scalars back, while for x64 we should
23987 /// use 64-bit extracts and shifts.
23988 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23989 TargetLowering::DAGCombinerInfo &DCI) {
23990 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23993 SDValue InputVector = N->getOperand(0);
23994 SDLoc dl(InputVector);
23995 // Detect mmx to i32 conversion through a v2i32 elt extract.
23996 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23997 N->getValueType(0) == MVT::i32 &&
23998 InputVector.getValueType() == MVT::v2i32) {
24000 // The bitcast source is a direct mmx result.
24001 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
24002 if (MMXSrc.getValueType() == MVT::x86mmx)
24003 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24004 N->getValueType(0),
24005 InputVector.getNode()->getOperand(0));
24007 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
24008 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
24009 MMXSrc.getValueType() == MVT::i64) {
24010 SDValue MMXSrcOp = MMXSrc.getOperand(0);
24011 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
24012 MMXSrcOp.getValueType() == MVT::v1i64 &&
24013 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
24014 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
24015 N->getValueType(0), MMXSrcOp.getOperand(0));
24019 EVT VT = N->getValueType(0);
24021 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
24022 InputVector.getOpcode() == ISD::BITCAST &&
24023 isa<ConstantSDNode>(InputVector.getOperand(0))) {
24024 uint64_t ExtractedElt =
24025 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
24026 uint64_t InputValue =
24027 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
24028 uint64_t Res = (InputValue >> ExtractedElt) & 1;
24029 return DAG.getConstant(Res, dl, MVT::i1);
24031 // Only operate on vectors of 4 elements, where the alternative shuffling
24032 // gets to be more expensive.
24033 if (InputVector.getValueType() != MVT::v4i32)
24036 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
24037 // single use which is a sign-extend or zero-extend, and all elements are
24039 SmallVector<SDNode *, 4> Uses;
24040 unsigned ExtractedElements = 0;
24041 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
24042 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
24043 if (UI.getUse().getResNo() != InputVector.getResNo())
24046 SDNode *Extract = *UI;
24047 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
24050 if (Extract->getValueType(0) != MVT::i32)
24052 if (!Extract->hasOneUse())
24054 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
24055 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
24057 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
24060 // Record which element was extracted.
24061 ExtractedElements |=
24062 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
24064 Uses.push_back(Extract);
24067 // If not all the elements were used, this may not be worthwhile.
24068 if (ExtractedElements != 15)
24071 // Ok, we've now decided to do the transformation.
24072 // If 64-bit shifts are legal, use the extract-shift sequence,
24073 // otherwise bounce the vector off the cache.
24074 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24077 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
24078 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
24079 auto &DL = DAG.getDataLayout();
24080 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
24081 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24082 DAG.getConstant(0, dl, VecIdxTy));
24083 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
24084 DAG.getConstant(1, dl, VecIdxTy));
24086 SDValue ShAmt = DAG.getConstant(
24087 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
24088 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
24089 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24090 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
24091 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
24092 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
24093 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
24095 // Store the value to a temporary stack slot.
24096 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
24097 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
24098 MachinePointerInfo(), false, false, 0);
24100 EVT ElementType = InputVector.getValueType().getVectorElementType();
24101 unsigned EltSize = ElementType.getSizeInBits() / 8;
24103 // Replace each use (extract) with a load of the appropriate element.
24104 for (unsigned i = 0; i < 4; ++i) {
24105 uint64_t Offset = EltSize * i;
24106 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
24107 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
24109 SDValue ScalarAddr =
24110 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
24112 // Load the scalar.
24113 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
24114 ScalarAddr, MachinePointerInfo(),
24115 false, false, false, 0);
24120 // Replace the extracts
24121 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
24122 UE = Uses.end(); UI != UE; ++UI) {
24123 SDNode *Extract = *UI;
24125 SDValue Idx = Extract->getOperand(1);
24126 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
24127 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
24130 // The replacement was made in place; don't return anything.
24135 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
24136 const X86Subtarget *Subtarget) {
24138 SDValue Cond = N->getOperand(0);
24139 SDValue LHS = N->getOperand(1);
24140 SDValue RHS = N->getOperand(2);
24142 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
24143 SDValue CondSrc = Cond->getOperand(0);
24144 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
24145 Cond = CondSrc->getOperand(0);
24148 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
24151 // A vselect where all conditions and data are constants can be optimized into
24152 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
24153 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
24154 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
24157 unsigned MaskValue = 0;
24158 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
24161 MVT VT = N->getSimpleValueType(0);
24162 unsigned NumElems = VT.getVectorNumElements();
24163 SmallVector<int, 8> ShuffleMask(NumElems, -1);
24164 for (unsigned i = 0; i < NumElems; ++i) {
24165 // Be sure we emit undef where we can.
24166 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
24167 ShuffleMask[i] = -1;
24169 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
24172 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24173 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
24175 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
24178 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
24180 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
24181 TargetLowering::DAGCombinerInfo &DCI,
24182 const X86Subtarget *Subtarget) {
24184 SDValue Cond = N->getOperand(0);
24185 // Get the LHS/RHS of the select.
24186 SDValue LHS = N->getOperand(1);
24187 SDValue RHS = N->getOperand(2);
24188 EVT VT = LHS.getValueType();
24189 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24191 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
24192 // instructions match the semantics of the common C idiom x<y?x:y but not
24193 // x<=y?x:y, because of how they handle negative zero (which can be
24194 // ignored in unsafe-math mode).
24195 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
24196 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
24197 VT != MVT::f80 && VT != MVT::f128 &&
24198 (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
24199 (Subtarget->hasSSE2() ||
24200 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
24201 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24203 unsigned Opcode = 0;
24204 // Check for x CC y ? x : y.
24205 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24206 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24210 // Converting this to a min would handle NaNs incorrectly, and swapping
24211 // the operands would cause it to handle comparisons between positive
24212 // and negative zero incorrectly.
24213 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24214 if (!DAG.getTarget().Options.UnsafeFPMath &&
24215 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24217 std::swap(LHS, RHS);
24219 Opcode = X86ISD::FMIN;
24222 // Converting this to a min would handle comparisons between positive
24223 // and negative zero incorrectly.
24224 if (!DAG.getTarget().Options.UnsafeFPMath &&
24225 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24227 Opcode = X86ISD::FMIN;
24230 // Converting this to a min would handle both negative zeros and NaNs
24231 // incorrectly, but we can swap the operands to fix both.
24232 std::swap(LHS, RHS);
24236 Opcode = X86ISD::FMIN;
24240 // Converting this to a max would handle comparisons between positive
24241 // and negative zero incorrectly.
24242 if (!DAG.getTarget().Options.UnsafeFPMath &&
24243 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
24245 Opcode = X86ISD::FMAX;
24248 // Converting this to a max would handle NaNs incorrectly, and swapping
24249 // the operands would cause it to handle comparisons between positive
24250 // and negative zero incorrectly.
24251 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
24252 if (!DAG.getTarget().Options.UnsafeFPMath &&
24253 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
24255 std::swap(LHS, RHS);
24257 Opcode = X86ISD::FMAX;
24260 // Converting this to a max would handle both negative zeros and NaNs
24261 // incorrectly, but we can swap the operands to fix both.
24262 std::swap(LHS, RHS);
24266 Opcode = X86ISD::FMAX;
24269 // Check for x CC y ? y : x -- a min/max with reversed arms.
24270 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
24271 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
24275 // Converting this to a min would handle comparisons between positive
24276 // and negative zero incorrectly, and swapping the operands would
24277 // cause it to handle NaNs incorrectly.
24278 if (!DAG.getTarget().Options.UnsafeFPMath &&
24279 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
24280 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24282 std::swap(LHS, RHS);
24284 Opcode = X86ISD::FMIN;
24287 // Converting this to a min would handle NaNs incorrectly.
24288 if (!DAG.getTarget().Options.UnsafeFPMath &&
24289 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
24291 Opcode = X86ISD::FMIN;
24294 // Converting this to a min would handle both negative zeros and NaNs
24295 // incorrectly, but we can swap the operands to fix both.
24296 std::swap(LHS, RHS);
24300 Opcode = X86ISD::FMIN;
24304 // Converting this to a max would handle NaNs incorrectly.
24305 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24307 Opcode = X86ISD::FMAX;
24310 // Converting this to a max would handle comparisons between positive
24311 // and negative zero incorrectly, and swapping the operands would
24312 // cause it to handle NaNs incorrectly.
24313 if (!DAG.getTarget().Options.UnsafeFPMath &&
24314 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
24315 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
24317 std::swap(LHS, RHS);
24319 Opcode = X86ISD::FMAX;
24322 // Converting this to a max would handle both negative zeros and NaNs
24323 // incorrectly, but we can swap the operands to fix both.
24324 std::swap(LHS, RHS);
24328 Opcode = X86ISD::FMAX;
24334 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
24337 EVT CondVT = Cond.getValueType();
24338 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
24339 CondVT.getVectorElementType() == MVT::i1) {
24340 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
24341 // lowering on KNL. In this case we convert it to
24342 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
24343 // The same situation for all 128 and 256-bit vectors of i8 and i16.
24344 // Since SKX these selects have a proper lowering.
24345 EVT OpVT = LHS.getValueType();
24346 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
24347 (OpVT.getVectorElementType() == MVT::i8 ||
24348 OpVT.getVectorElementType() == MVT::i16) &&
24349 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
24350 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
24351 DCI.AddToWorklist(Cond.getNode());
24352 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
24355 // If this is a select between two integer constants, try to do some
24357 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
24358 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
24359 // Don't do this for crazy integer types.
24360 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
24361 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
24362 // so that TrueC (the true value) is larger than FalseC.
24363 bool NeedsCondInvert = false;
24365 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
24366 // Efficiently invertible.
24367 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
24368 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
24369 isa<ConstantSDNode>(Cond.getOperand(1))))) {
24370 NeedsCondInvert = true;
24371 std::swap(TrueC, FalseC);
24374 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
24375 if (FalseC->getAPIntValue() == 0 &&
24376 TrueC->getAPIntValue().isPowerOf2()) {
24377 if (NeedsCondInvert) // Invert the condition if needed.
24378 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24379 DAG.getConstant(1, DL, Cond.getValueType()));
24381 // Zero extend the condition if needed.
24382 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
24384 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24385 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
24386 DAG.getConstant(ShAmt, DL, MVT::i8));
24389 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
24390 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24391 if (NeedsCondInvert) // Invert the condition if needed.
24392 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24393 DAG.getConstant(1, DL, Cond.getValueType()));
24395 // Zero extend the condition if needed.
24396 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24397 FalseC->getValueType(0), Cond);
24398 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24399 SDValue(FalseC, 0));
24402 // Optimize cases that will turn into an LEA instruction. This requires
24403 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24404 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24405 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24406 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24408 bool isFastMultiplier = false;
24410 switch ((unsigned char)Diff) {
24412 case 1: // result = add base, cond
24413 case 2: // result = lea base( , cond*2)
24414 case 3: // result = lea base(cond, cond*2)
24415 case 4: // result = lea base( , cond*4)
24416 case 5: // result = lea base(cond, cond*4)
24417 case 8: // result = lea base( , cond*8)
24418 case 9: // result = lea base(cond, cond*8)
24419 isFastMultiplier = true;
24424 if (isFastMultiplier) {
24425 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24426 if (NeedsCondInvert) // Invert the condition if needed.
24427 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
24428 DAG.getConstant(1, DL, Cond.getValueType()));
24430 // Zero extend the condition if needed.
24431 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24433 // Scale the condition by the difference.
24435 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24436 DAG.getConstant(Diff, DL,
24437 Cond.getValueType()));
24439 // Add the base if non-zero.
24440 if (FalseC->getAPIntValue() != 0)
24441 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24442 SDValue(FalseC, 0));
24449 // Canonicalize max and min:
24450 // (x > y) ? x : y -> (x >= y) ? x : y
24451 // (x < y) ? x : y -> (x <= y) ? x : y
24452 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
24453 // the need for an extra compare
24454 // against zero. e.g.
24455 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
24457 // testl %edi, %edi
24459 // cmovgl %edi, %eax
24463 // cmovsl %eax, %edi
24464 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
24465 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
24466 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
24467 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24472 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
24473 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
24474 Cond.getOperand(0), Cond.getOperand(1), NewCC);
24475 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
24480 // Early exit check
24481 if (!TLI.isTypeLegal(VT))
24484 // Match VSELECTs into subs with unsigned saturation.
24485 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
24486 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
24487 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
24488 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
24489 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
24491 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
24492 // left side invert the predicate to simplify logic below.
24494 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
24496 CC = ISD::getSetCCInverse(CC, true);
24497 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
24501 if (Other.getNode() && Other->getNumOperands() == 2 &&
24502 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
24503 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
24504 SDValue CondRHS = Cond->getOperand(1);
24506 // Look for a general sub with unsigned saturation first.
24507 // x >= y ? x-y : 0 --> subus x, y
24508 // x > y ? x-y : 0 --> subus x, y
24509 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
24510 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
24511 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
24513 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
24514 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
24515 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
24516 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
24517 // If the RHS is a constant we have to reverse the const
24518 // canonicalization.
24519 // x > C-1 ? x+-C : 0 --> subus x, C
24520 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
24521 CondRHSConst->getAPIntValue() ==
24522 (-OpRHSConst->getAPIntValue() - 1))
24523 return DAG.getNode(
24524 X86ISD::SUBUS, DL, VT, OpLHS,
24525 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
24527 // Another special case: If C was a sign bit, the sub has been
24528 // canonicalized into a xor.
24529 // FIXME: Would it be better to use computeKnownBits to determine
24530 // whether it's safe to decanonicalize the xor?
24531 // x s< 0 ? x^C : 0 --> subus x, C
24532 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
24533 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
24534 OpRHSConst->getAPIntValue().isSignBit())
24535 // Note that we have to rebuild the RHS constant here to ensure we
24536 // don't rely on particular values of undef lanes.
24537 return DAG.getNode(
24538 X86ISD::SUBUS, DL, VT, OpLHS,
24539 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
24544 // Simplify vector selection if condition value type matches vselect
24546 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
24547 assert(Cond.getValueType().isVector() &&
24548 "vector select expects a vector selector!");
24550 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
24551 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
24553 // Try invert the condition if true value is not all 1s and false value
24555 if (!TValIsAllOnes && !FValIsAllZeros &&
24556 // Check if the selector will be produced by CMPP*/PCMP*
24557 Cond.getOpcode() == ISD::SETCC &&
24558 // Check if SETCC has already been promoted
24559 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
24561 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
24562 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
24564 if (TValIsAllZeros || FValIsAllOnes) {
24565 SDValue CC = Cond.getOperand(2);
24566 ISD::CondCode NewCC =
24567 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
24568 Cond.getOperand(0).getValueType().isInteger());
24569 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
24570 std::swap(LHS, RHS);
24571 TValIsAllOnes = FValIsAllOnes;
24572 FValIsAllZeros = TValIsAllZeros;
24576 if (TValIsAllOnes || FValIsAllZeros) {
24579 if (TValIsAllOnes && FValIsAllZeros)
24581 else if (TValIsAllOnes)
24583 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
24584 else if (FValIsAllZeros)
24585 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
24586 DAG.getBitcast(CondVT, LHS));
24588 return DAG.getBitcast(VT, Ret);
24592 // We should generate an X86ISD::BLENDI from a vselect if its argument
24593 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
24594 // constants. This specific pattern gets generated when we split a
24595 // selector for a 512 bit vector in a machine without AVX512 (but with
24596 // 256-bit vectors), during legalization:
24598 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
24600 // Iff we find this pattern and the build_vectors are built from
24601 // constants, we translate the vselect into a shuffle_vector that we
24602 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
24603 if ((N->getOpcode() == ISD::VSELECT ||
24604 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
24605 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
24606 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
24607 if (Shuffle.getNode())
24611 // If this is a *dynamic* select (non-constant condition) and we can match
24612 // this node with one of the variable blend instructions, restructure the
24613 // condition so that the blends can use the high bit of each element and use
24614 // SimplifyDemandedBits to simplify the condition operand.
24615 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
24616 !DCI.isBeforeLegalize() &&
24617 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
24618 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
24620 // Don't optimize vector selects that map to mask-registers.
24624 // We can only handle the cases where VSELECT is directly legal on the
24625 // subtarget. We custom lower VSELECT nodes with constant conditions and
24626 // this makes it hard to see whether a dynamic VSELECT will correctly
24627 // lower, so we both check the operation's status and explicitly handle the
24628 // cases where a *dynamic* blend will fail even though a constant-condition
24629 // blend could be custom lowered.
24630 // FIXME: We should find a better way to handle this class of problems.
24631 // Potentially, we should combine constant-condition vselect nodes
24632 // pre-legalization into shuffles and not mark as many types as custom
24634 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
24636 // FIXME: We don't support i16-element blends currently. We could and
24637 // should support them by making *all* the bits in the condition be set
24638 // rather than just the high bit and using an i8-element blend.
24639 if (VT.getVectorElementType() == MVT::i16)
24641 // Dynamic blending was only available from SSE4.1 onward.
24642 if (VT.is128BitVector() && !Subtarget->hasSSE41())
24644 // Byte blends are only available in AVX2
24645 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
24648 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
24649 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
24651 APInt KnownZero, KnownOne;
24652 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
24653 DCI.isBeforeLegalizeOps());
24654 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
24655 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
24657 // If we changed the computation somewhere in the DAG, this change
24658 // will affect all users of Cond.
24659 // Make sure it is fine and update all the nodes so that we do not
24660 // use the generic VSELECT anymore. Otherwise, we may perform
24661 // wrong optimizations as we messed up with the actual expectation
24662 // for the vector boolean values.
24663 if (Cond != TLO.Old) {
24664 // Check all uses of that condition operand to check whether it will be
24665 // consumed by non-BLEND instructions, which may depend on all bits are
24667 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24669 if (I->getOpcode() != ISD::VSELECT)
24670 // TODO: Add other opcodes eventually lowered into BLEND.
24673 // Update all the users of the condition, before committing the change,
24674 // so that the VSELECT optimizations that expect the correct vector
24675 // boolean value will not be triggered.
24676 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
24678 DAG.ReplaceAllUsesOfValueWith(
24680 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
24681 Cond, I->getOperand(1), I->getOperand(2)));
24682 DCI.CommitTargetLoweringOpt(TLO);
24685 // At this point, only Cond is changed. Change the condition
24686 // just for N to keep the opportunity to optimize all other
24687 // users their own way.
24688 DAG.ReplaceAllUsesOfValueWith(
24690 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
24691 TLO.New, N->getOperand(1), N->getOperand(2)));
24699 // Check whether a boolean test is testing a boolean value generated by
24700 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
24703 // Simplify the following patterns:
24704 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
24705 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
24706 // to (Op EFLAGS Cond)
24708 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
24709 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
24710 // to (Op EFLAGS !Cond)
24712 // where Op could be BRCOND or CMOV.
24714 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
24715 // Quit if not CMP and SUB with its value result used.
24716 if (Cmp.getOpcode() != X86ISD::CMP &&
24717 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24720 // Quit if not used as a boolean value.
24721 if (CC != X86::COND_E && CC != X86::COND_NE)
24724 // Check CMP operands. One of them should be 0 or 1 and the other should be
24725 // an SetCC or extended from it.
24726 SDValue Op1 = Cmp.getOperand(0);
24727 SDValue Op2 = Cmp.getOperand(1);
24730 const ConstantSDNode* C = nullptr;
24731 bool needOppositeCond = (CC == X86::COND_E);
24732 bool checkAgainstTrue = false; // Is it a comparison against 1?
24734 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24736 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24738 else // Quit if all operands are not constants.
24741 if (C->getZExtValue() == 1) {
24742 needOppositeCond = !needOppositeCond;
24743 checkAgainstTrue = true;
24744 } else if (C->getZExtValue() != 0)
24745 // Quit if the constant is neither 0 or 1.
24748 bool truncatedToBoolWithAnd = false;
24749 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24750 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24751 SetCC.getOpcode() == ISD::TRUNCATE ||
24752 SetCC.getOpcode() == ISD::AND) {
24753 if (SetCC.getOpcode() == ISD::AND) {
24755 if (isOneConstant(SetCC.getOperand(0)))
24757 if (isOneConstant(SetCC.getOperand(1)))
24761 SetCC = SetCC.getOperand(OpIdx);
24762 truncatedToBoolWithAnd = true;
24764 SetCC = SetCC.getOperand(0);
24767 switch (SetCC.getOpcode()) {
24768 case X86ISD::SETCC_CARRY:
24769 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24770 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24771 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24772 // truncated to i1 using 'and'.
24773 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24775 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24776 "Invalid use of SETCC_CARRY!");
24778 case X86ISD::SETCC:
24779 // Set the condition code or opposite one if necessary.
24780 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24781 if (needOppositeCond)
24782 CC = X86::GetOppositeBranchCondition(CC);
24783 return SetCC.getOperand(1);
24784 case X86ISD::CMOV: {
24785 // Check whether false/true value has canonical one, i.e. 0 or 1.
24786 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24787 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24788 // Quit if true value is not a constant.
24791 // Quit if false value is not a constant.
24793 SDValue Op = SetCC.getOperand(0);
24794 // Skip 'zext' or 'trunc' node.
24795 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24796 Op.getOpcode() == ISD::TRUNCATE)
24797 Op = Op.getOperand(0);
24798 // A special case for rdrand/rdseed, where 0 is set if false cond is
24800 if ((Op.getOpcode() != X86ISD::RDRAND &&
24801 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24804 // Quit if false value is not the constant 0 or 1.
24805 bool FValIsFalse = true;
24806 if (FVal && FVal->getZExtValue() != 0) {
24807 if (FVal->getZExtValue() != 1)
24809 // If FVal is 1, opposite cond is needed.
24810 needOppositeCond = !needOppositeCond;
24811 FValIsFalse = false;
24813 // Quit if TVal is not the constant opposite of FVal.
24814 if (FValIsFalse && TVal->getZExtValue() != 1)
24816 if (!FValIsFalse && TVal->getZExtValue() != 0)
24818 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24819 if (needOppositeCond)
24820 CC = X86::GetOppositeBranchCondition(CC);
24821 return SetCC.getOperand(3);
24828 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24830 /// (X86or (X86setcc) (X86setcc))
24831 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24832 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24833 X86::CondCode &CC1, SDValue &Flags,
24835 if (Cond->getOpcode() == X86ISD::CMP) {
24836 if (!isNullConstant(Cond->getOperand(1)))
24839 Cond = Cond->getOperand(0);
24844 SDValue SetCC0, SetCC1;
24845 switch (Cond->getOpcode()) {
24846 default: return false;
24853 SetCC0 = Cond->getOperand(0);
24854 SetCC1 = Cond->getOperand(1);
24858 // Make sure we have SETCC nodes, using the same flags value.
24859 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24860 SetCC1.getOpcode() != X86ISD::SETCC ||
24861 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24864 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24865 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24866 Flags = SetCC0->getOperand(1);
24870 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24871 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24872 TargetLowering::DAGCombinerInfo &DCI,
24873 const X86Subtarget *Subtarget) {
24876 // If the flag operand isn't dead, don't touch this CMOV.
24877 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24880 SDValue FalseOp = N->getOperand(0);
24881 SDValue TrueOp = N->getOperand(1);
24882 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24883 SDValue Cond = N->getOperand(3);
24885 if (CC == X86::COND_E || CC == X86::COND_NE) {
24886 switch (Cond.getOpcode()) {
24890 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24891 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24892 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24898 Flags = checkBoolTestSetCCCombine(Cond, CC);
24899 if (Flags.getNode() &&
24900 // Extra check as FCMOV only supports a subset of X86 cond.
24901 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24902 SDValue Ops[] = { FalseOp, TrueOp,
24903 DAG.getConstant(CC, DL, MVT::i8), Flags };
24904 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24907 // If this is a select between two integer constants, try to do some
24908 // optimizations. Note that the operands are ordered the opposite of SELECT
24910 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24911 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24912 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24913 // larger than FalseC (the false value).
24914 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24915 CC = X86::GetOppositeBranchCondition(CC);
24916 std::swap(TrueC, FalseC);
24917 std::swap(TrueOp, FalseOp);
24920 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24921 // This is efficient for any integer data type (including i8/i16) and
24923 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24924 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24925 DAG.getConstant(CC, DL, MVT::i8), Cond);
24927 // Zero extend the condition if needed.
24928 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24930 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24931 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24932 DAG.getConstant(ShAmt, DL, MVT::i8));
24933 if (N->getNumValues() == 2) // Dead flag value?
24934 return DCI.CombineTo(N, Cond, SDValue());
24938 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24939 // for any integer data type, including i8/i16.
24940 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24941 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24942 DAG.getConstant(CC, DL, MVT::i8), Cond);
24944 // Zero extend the condition if needed.
24945 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24946 FalseC->getValueType(0), Cond);
24947 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24948 SDValue(FalseC, 0));
24950 if (N->getNumValues() == 2) // Dead flag value?
24951 return DCI.CombineTo(N, Cond, SDValue());
24955 // Optimize cases that will turn into an LEA instruction. This requires
24956 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24957 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24958 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24959 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24961 bool isFastMultiplier = false;
24963 switch ((unsigned char)Diff) {
24965 case 1: // result = add base, cond
24966 case 2: // result = lea base( , cond*2)
24967 case 3: // result = lea base(cond, cond*2)
24968 case 4: // result = lea base( , cond*4)
24969 case 5: // result = lea base(cond, cond*4)
24970 case 8: // result = lea base( , cond*8)
24971 case 9: // result = lea base(cond, cond*8)
24972 isFastMultiplier = true;
24977 if (isFastMultiplier) {
24978 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24979 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24980 DAG.getConstant(CC, DL, MVT::i8), Cond);
24981 // Zero extend the condition if needed.
24982 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24984 // Scale the condition by the difference.
24986 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24987 DAG.getConstant(Diff, DL, Cond.getValueType()));
24989 // Add the base if non-zero.
24990 if (FalseC->getAPIntValue() != 0)
24991 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24992 SDValue(FalseC, 0));
24993 if (N->getNumValues() == 2) // Dead flag value?
24994 return DCI.CombineTo(N, Cond, SDValue());
25001 // Handle these cases:
25002 // (select (x != c), e, c) -> select (x != c), e, x),
25003 // (select (x == c), c, e) -> select (x == c), x, e)
25004 // where the c is an integer constant, and the "select" is the combination
25005 // of CMOV and CMP.
25007 // The rationale for this change is that the conditional-move from a constant
25008 // needs two instructions, however, conditional-move from a register needs
25009 // only one instruction.
25011 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
25012 // some instruction-combining opportunities. This opt needs to be
25013 // postponed as late as possible.
25015 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
25016 // the DCI.xxxx conditions are provided to postpone the optimization as
25017 // late as possible.
25019 ConstantSDNode *CmpAgainst = nullptr;
25020 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
25021 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
25022 !isa<ConstantSDNode>(Cond.getOperand(0))) {
25024 if (CC == X86::COND_NE &&
25025 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
25026 CC = X86::GetOppositeBranchCondition(CC);
25027 std::swap(TrueOp, FalseOp);
25030 if (CC == X86::COND_E &&
25031 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
25032 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
25033 DAG.getConstant(CC, DL, MVT::i8), Cond };
25034 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
25039 // Fold and/or of setcc's to double CMOV:
25040 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
25041 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
25043 // This combine lets us generate:
25044 // cmovcc1 (jcc1 if we don't have CMOV)
25050 // cmovne (jne if we don't have CMOV)
25051 // When we can't use the CMOV instruction, it might increase branch
25053 // When we can use CMOV, or when there is no mispredict, this improves
25054 // throughput and reduces register pressure.
25056 if (CC == X86::COND_NE) {
25058 X86::CondCode CC0, CC1;
25060 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
25062 std::swap(FalseOp, TrueOp);
25063 CC0 = X86::GetOppositeBranchCondition(CC0);
25064 CC1 = X86::GetOppositeBranchCondition(CC1);
25067 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
25069 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
25070 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
25071 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
25072 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
25080 /// PerformMulCombine - Optimize a single multiply with constant into two
25081 /// in order to implement it with two cheaper instructions, e.g.
25082 /// LEA + SHL, LEA + LEA.
25083 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
25084 TargetLowering::DAGCombinerInfo &DCI) {
25085 // An imul is usually smaller than the alternative sequence.
25086 if (DAG.getMachineFunction().getFunction()->optForMinSize())
25089 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
25092 EVT VT = N->getValueType(0);
25093 if (VT != MVT::i64 && VT != MVT::i32)
25096 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
25099 uint64_t MulAmt = C->getZExtValue();
25100 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
25103 uint64_t MulAmt1 = 0;
25104 uint64_t MulAmt2 = 0;
25105 if ((MulAmt % 9) == 0) {
25107 MulAmt2 = MulAmt / 9;
25108 } else if ((MulAmt % 5) == 0) {
25110 MulAmt2 = MulAmt / 5;
25111 } else if ((MulAmt % 3) == 0) {
25113 MulAmt2 = MulAmt / 3;
25119 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
25121 if (isPowerOf2_64(MulAmt2) &&
25122 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
25123 // If second multiplifer is pow2, issue it first. We want the multiply by
25124 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
25126 std::swap(MulAmt1, MulAmt2);
25128 if (isPowerOf2_64(MulAmt1))
25129 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25130 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
25132 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
25133 DAG.getConstant(MulAmt1, DL, VT));
25135 if (isPowerOf2_64(MulAmt2))
25136 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
25137 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
25139 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
25140 DAG.getConstant(MulAmt2, DL, VT));
25144 assert(MulAmt != 0 && MulAmt != (VT == MVT::i64 ? UINT64_MAX : UINT32_MAX)
25145 && "Both cases that could cause potential overflows should have "
25146 "already been handled.");
25147 if (isPowerOf2_64(MulAmt - 1))
25148 // (mul x, 2^N + 1) => (add (shl x, N), x)
25149 NewMul = DAG.getNode(ISD::ADD, DL, VT, N->getOperand(0),
25150 DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
25151 DAG.getConstant(Log2_64(MulAmt - 1), DL,
25154 else if (isPowerOf2_64(MulAmt + 1))
25155 // (mul x, 2^N - 1) => (sub (shl x, N), x)
25156 NewMul = DAG.getNode(ISD::SUB, DL, VT, DAG.getNode(ISD::SHL, DL, VT,
25158 DAG.getConstant(Log2_64(MulAmt + 1),
25159 DL, MVT::i8)), N->getOperand(0));
25163 // Do not add new nodes to DAG combiner worklist.
25164 DCI.CombineTo(N, NewMul, false);
25169 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
25170 SDValue N0 = N->getOperand(0);
25171 SDValue N1 = N->getOperand(1);
25172 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
25173 EVT VT = N0.getValueType();
25175 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
25176 // since the result of setcc_c is all zero's or all ones.
25177 if (VT.isInteger() && !VT.isVector() &&
25178 N1C && N0.getOpcode() == ISD::AND &&
25179 N0.getOperand(1).getOpcode() == ISD::Constant) {
25180 SDValue N00 = N0.getOperand(0);
25181 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
25182 APInt ShAmt = N1C->getAPIntValue();
25183 Mask = Mask.shl(ShAmt);
25184 bool MaskOK = false;
25185 // We can handle cases concerning bit-widening nodes containing setcc_c if
25186 // we carefully interrogate the mask to make sure we are semantics
25188 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
25189 // of the underlying setcc_c operation if the setcc_c was zero extended.
25190 // Consider the following example:
25191 // zext(setcc_c) -> i32 0x0000FFFF
25192 // c1 -> i32 0x0000FFFF
25193 // c2 -> i32 0x00000001
25194 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
25195 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
25196 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25198 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
25199 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25201 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
25202 N00.getOpcode() == ISD::ANY_EXTEND) &&
25203 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
25204 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
25206 if (MaskOK && Mask != 0) {
25208 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
25212 // Hardware support for vector shifts is sparse which makes us scalarize the
25213 // vector operations in many cases. Also, on sandybridge ADD is faster than
25215 // (shl V, 1) -> add V,V
25216 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
25217 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
25218 assert(N0.getValueType().isVector() && "Invalid vector shift type");
25219 // We shift all of the values by one. In many cases we do not have
25220 // hardware support for this operation. This is better expressed as an ADD
25222 if (N1SplatC->getAPIntValue() == 1)
25223 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
25229 static SDValue PerformSRACombine(SDNode *N, SelectionDAG &DAG) {
25230 SDValue N0 = N->getOperand(0);
25231 SDValue N1 = N->getOperand(1);
25232 EVT VT = N0.getValueType();
25233 unsigned Size = VT.getSizeInBits();
25235 // fold (ashr (shl, a, [56,48,32,24,16]), SarConst)
25236 // into (shl, (sext (a), [56,48,32,24,16] - SarConst)) or
25237 // into (lshr, (sext (a), SarConst - [56,48,32,24,16]))
25238 // depending on sign of (SarConst - [56,48,32,24,16])
25240 // sexts in X86 are MOVs. The MOVs have the same code size
25241 // as above SHIFTs (only SHIFT on 1 has lower code size).
25242 // However the MOVs have 2 advantages to a SHIFT:
25243 // 1. MOVs can write to a register that differs from source
25244 // 2. MOVs accept memory operands
25246 if (!VT.isInteger() || VT.isVector() || N1.getOpcode() != ISD::Constant ||
25247 N0.getOpcode() != ISD::SHL || !N0.hasOneUse() ||
25248 N0.getOperand(1).getOpcode() != ISD::Constant)
25251 SDValue N00 = N0.getOperand(0);
25252 SDValue N01 = N0.getOperand(1);
25253 APInt ShlConst = (cast<ConstantSDNode>(N01))->getAPIntValue();
25254 APInt SarConst = (cast<ConstantSDNode>(N1))->getAPIntValue();
25255 EVT CVT = N1.getValueType();
25257 if (SarConst.isNegative())
25260 for (MVT SVT : MVT::integer_valuetypes()) {
25261 unsigned ShiftSize = SVT.getSizeInBits();
25262 // skipping types without corresponding sext/zext and
25263 // ShlConst that is not one of [56,48,32,24,16]
25264 if (ShiftSize < 8 || ShiftSize > 64 || ShlConst != Size - ShiftSize)
25268 DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT, N00, DAG.getValueType(SVT));
25269 SarConst = SarConst - (Size - ShiftSize);
25272 else if (SarConst.isNegative())
25273 return DAG.getNode(ISD::SHL, DL, VT, NN,
25274 DAG.getConstant(-SarConst, DL, CVT));
25276 return DAG.getNode(ISD::SRA, DL, VT, NN,
25277 DAG.getConstant(SarConst, DL, CVT));
25282 /// \brief Returns a vector of 0s if the node in input is a vector logical
25283 /// shift by a constant amount which is known to be bigger than or equal
25284 /// to the vector element size in bits.
25285 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
25286 const X86Subtarget *Subtarget) {
25287 EVT VT = N->getValueType(0);
25289 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
25290 (!Subtarget->hasInt256() ||
25291 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
25294 SDValue Amt = N->getOperand(1);
25296 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
25297 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
25298 APInt ShiftAmt = AmtSplat->getAPIntValue();
25299 unsigned MaxAmount =
25300 VT.getSimpleVT().getVectorElementType().getSizeInBits();
25302 // SSE2/AVX2 logical shifts always return a vector of 0s
25303 // if the shift amount is bigger than or equal to
25304 // the element size. The constant shift amount will be
25305 // encoded as a 8-bit immediate.
25306 if (ShiftAmt.trunc(8).uge(MaxAmount))
25307 return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, DL);
25313 /// PerformShiftCombine - Combine shifts.
25314 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
25315 TargetLowering::DAGCombinerInfo &DCI,
25316 const X86Subtarget *Subtarget) {
25317 if (N->getOpcode() == ISD::SHL)
25318 if (SDValue V = PerformSHLCombine(N, DAG))
25321 if (N->getOpcode() == ISD::SRA)
25322 if (SDValue V = PerformSRACombine(N, DAG))
25325 // Try to fold this logical shift into a zero vector.
25326 if (N->getOpcode() != ISD::SRA)
25327 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
25333 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
25334 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
25335 // and friends. Likewise for OR -> CMPNEQSS.
25336 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
25337 TargetLowering::DAGCombinerInfo &DCI,
25338 const X86Subtarget *Subtarget) {
25341 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
25342 // we're requiring SSE2 for both.
25343 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
25344 SDValue N0 = N->getOperand(0);
25345 SDValue N1 = N->getOperand(1);
25346 SDValue CMP0 = N0->getOperand(1);
25347 SDValue CMP1 = N1->getOperand(1);
25350 // The SETCCs should both refer to the same CMP.
25351 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
25354 SDValue CMP00 = CMP0->getOperand(0);
25355 SDValue CMP01 = CMP0->getOperand(1);
25356 EVT VT = CMP00.getValueType();
25358 if (VT == MVT::f32 || VT == MVT::f64) {
25359 bool ExpectingFlags = false;
25360 // Check for any users that want flags:
25361 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
25362 !ExpectingFlags && UI != UE; ++UI)
25363 switch (UI->getOpcode()) {
25368 ExpectingFlags = true;
25370 case ISD::CopyToReg:
25371 case ISD::SIGN_EXTEND:
25372 case ISD::ZERO_EXTEND:
25373 case ISD::ANY_EXTEND:
25377 if (!ExpectingFlags) {
25378 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
25379 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
25381 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
25382 X86::CondCode tmp = cc0;
25387 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
25388 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
25389 // FIXME: need symbolic constants for these magic numbers.
25390 // See X86ATTInstPrinter.cpp:printSSECC().
25391 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
25392 if (Subtarget->hasAVX512()) {
25393 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
25395 DAG.getConstant(x86cc, DL, MVT::i8));
25396 if (N->getValueType(0) != MVT::i1)
25397 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
25401 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
25402 CMP00.getValueType(), CMP00, CMP01,
25403 DAG.getConstant(x86cc, DL,
25406 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
25407 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
25409 if (is64BitFP && !Subtarget->is64Bit()) {
25410 // On a 32-bit target, we cannot bitcast the 64-bit float to a
25411 // 64-bit integer, since that's not a legal type. Since
25412 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
25413 // bits, but can do this little dance to extract the lowest 32 bits
25414 // and work with those going forward.
25415 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
25417 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
25418 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
25419 Vector32, DAG.getIntPtrConstant(0, DL));
25423 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
25424 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
25425 DAG.getConstant(1, DL, IntVT));
25426 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
25428 return OneBitOfTruth;
25436 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
25437 /// so it can be folded inside ANDNP.
25438 static bool CanFoldXORWithAllOnes(const SDNode *N) {
25439 EVT VT = N->getValueType(0);
25441 // Match direct AllOnes for 128 and 256-bit vectors
25442 if (ISD::isBuildVectorAllOnes(N))
25445 // Look through a bit convert.
25446 if (N->getOpcode() == ISD::BITCAST)
25447 N = N->getOperand(0).getNode();
25449 // Sometimes the operand may come from a insert_subvector building a 256-bit
25451 if (VT.is256BitVector() &&
25452 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
25453 SDValue V1 = N->getOperand(0);
25454 SDValue V2 = N->getOperand(1);
25456 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
25457 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
25458 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
25459 ISD::isBuildVectorAllOnes(V2.getNode()))
25466 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
25467 // register. In most cases we actually compare or select YMM-sized registers
25468 // and mixing the two types creates horrible code. This method optimizes
25469 // some of the transition sequences.
25470 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
25471 TargetLowering::DAGCombinerInfo &DCI,
25472 const X86Subtarget *Subtarget) {
25473 EVT VT = N->getValueType(0);
25474 if (!VT.is256BitVector())
25477 assert((N->getOpcode() == ISD::ANY_EXTEND ||
25478 N->getOpcode() == ISD::ZERO_EXTEND ||
25479 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
25481 SDValue Narrow = N->getOperand(0);
25482 EVT NarrowVT = Narrow->getValueType(0);
25483 if (!NarrowVT.is128BitVector())
25486 if (Narrow->getOpcode() != ISD::XOR &&
25487 Narrow->getOpcode() != ISD::AND &&
25488 Narrow->getOpcode() != ISD::OR)
25491 SDValue N0 = Narrow->getOperand(0);
25492 SDValue N1 = Narrow->getOperand(1);
25495 // The Left side has to be a trunc.
25496 if (N0.getOpcode() != ISD::TRUNCATE)
25499 // The type of the truncated inputs.
25500 EVT WideVT = N0->getOperand(0)->getValueType(0);
25504 // The right side has to be a 'trunc' or a constant vector.
25505 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
25506 ConstantSDNode *RHSConstSplat = nullptr;
25507 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
25508 RHSConstSplat = RHSBV->getConstantSplatNode();
25509 if (!RHSTrunc && !RHSConstSplat)
25512 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25514 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
25517 // Set N0 and N1 to hold the inputs to the new wide operation.
25518 N0 = N0->getOperand(0);
25519 if (RHSConstSplat) {
25520 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
25521 SDValue(RHSConstSplat, 0));
25522 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
25523 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
25524 } else if (RHSTrunc) {
25525 N1 = N1->getOperand(0);
25528 // Generate the wide operation.
25529 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
25530 unsigned Opcode = N->getOpcode();
25532 case ISD::ANY_EXTEND:
25534 case ISD::ZERO_EXTEND: {
25535 unsigned InBits = NarrowVT.getScalarSizeInBits();
25536 APInt Mask = APInt::getAllOnesValue(InBits);
25537 Mask = Mask.zext(VT.getScalarSizeInBits());
25538 return DAG.getNode(ISD::AND, DL, VT,
25539 Op, DAG.getConstant(Mask, DL, VT));
25541 case ISD::SIGN_EXTEND:
25542 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
25543 Op, DAG.getValueType(NarrowVT));
25545 llvm_unreachable("Unexpected opcode");
25549 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
25550 TargetLowering::DAGCombinerInfo &DCI,
25551 const X86Subtarget *Subtarget) {
25552 SDValue N0 = N->getOperand(0);
25553 SDValue N1 = N->getOperand(1);
25556 // A vector zext_in_reg may be represented as a shuffle,
25557 // feeding into a bitcast (this represents anyext) feeding into
25558 // an and with a mask.
25559 // We'd like to try to combine that into a shuffle with zero
25560 // plus a bitcast, removing the and.
25561 if (N0.getOpcode() != ISD::BITCAST ||
25562 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
25565 // The other side of the AND should be a splat of 2^C, where C
25566 // is the number of bits in the source type.
25567 if (N1.getOpcode() == ISD::BITCAST)
25568 N1 = N1.getOperand(0);
25569 if (N1.getOpcode() != ISD::BUILD_VECTOR)
25571 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
25573 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
25574 EVT SrcType = Shuffle->getValueType(0);
25576 // We expect a single-source shuffle
25577 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
25580 unsigned SrcSize = SrcType.getScalarSizeInBits();
25582 APInt SplatValue, SplatUndef;
25583 unsigned SplatBitSize;
25585 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
25586 SplatBitSize, HasAnyUndefs))
25589 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
25590 // Make sure the splat matches the mask we expect
25591 if (SplatBitSize > ResSize ||
25592 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
25595 // Make sure the input and output size make sense
25596 if (SrcSize >= ResSize || ResSize % SrcSize)
25599 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
25600 // The number of u's between each two values depends on the ratio between
25601 // the source and dest type.
25602 unsigned ZextRatio = ResSize / SrcSize;
25603 bool IsZext = true;
25604 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
25605 if (i % ZextRatio) {
25606 if (Shuffle->getMaskElt(i) > 0) {
25612 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
25613 // Expected element number
25623 // Ok, perform the transformation - replace the shuffle with
25624 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
25625 // (instead of undef) where the k elements come from the zero vector.
25626 SmallVector<int, 8> Mask;
25627 unsigned NumElems = SrcType.getVectorNumElements();
25628 for (unsigned i = 0; i < NumElems; ++i)
25630 Mask.push_back(NumElems);
25632 Mask.push_back(i / ZextRatio);
25634 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
25635 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
25636 return DAG.getBitcast(N0.getValueType(), NewShuffle);
25639 /// If both input operands of a logic op are being cast from floating point
25640 /// types, try to convert this into a floating point logic node to avoid
25641 /// unnecessary moves from SSE to integer registers.
25642 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
25643 const X86Subtarget *Subtarget) {
25644 unsigned FPOpcode = ISD::DELETED_NODE;
25645 if (N->getOpcode() == ISD::AND)
25646 FPOpcode = X86ISD::FAND;
25647 else if (N->getOpcode() == ISD::OR)
25648 FPOpcode = X86ISD::FOR;
25649 else if (N->getOpcode() == ISD::XOR)
25650 FPOpcode = X86ISD::FXOR;
25652 assert(FPOpcode != ISD::DELETED_NODE &&
25653 "Unexpected input node for FP logic conversion");
25655 EVT VT = N->getValueType(0);
25656 SDValue N0 = N->getOperand(0);
25657 SDValue N1 = N->getOperand(1);
25659 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
25660 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
25661 (Subtarget->hasSSE2() && VT == MVT::i64))) {
25662 SDValue N00 = N0.getOperand(0);
25663 SDValue N10 = N1.getOperand(0);
25664 EVT N00Type = N00.getValueType();
25665 EVT N10Type = N10.getValueType();
25666 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
25667 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
25668 return DAG.getBitcast(VT, FPLogic);
25674 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
25675 TargetLowering::DAGCombinerInfo &DCI,
25676 const X86Subtarget *Subtarget) {
25677 if (DCI.isBeforeLegalizeOps())
25680 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
25683 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25686 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25689 EVT VT = N->getValueType(0);
25690 SDValue N0 = N->getOperand(0);
25691 SDValue N1 = N->getOperand(1);
25694 // Create BEXTR instructions
25695 // BEXTR is ((X >> imm) & (2**size-1))
25696 if (VT == MVT::i32 || VT == MVT::i64) {
25697 // Check for BEXTR.
25698 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
25699 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
25700 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
25701 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25702 if (MaskNode && ShiftNode) {
25703 uint64_t Mask = MaskNode->getZExtValue();
25704 uint64_t Shift = ShiftNode->getZExtValue();
25705 if (isMask_64(Mask)) {
25706 uint64_t MaskSize = countPopulation(Mask);
25707 if (Shift + MaskSize <= VT.getSizeInBits())
25708 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
25709 DAG.getConstant(Shift | (MaskSize << 8), DL,
25718 // Want to form ANDNP nodes:
25719 // 1) In the hopes of then easily combining them with OR and AND nodes
25720 // to form PBLEND/PSIGN.
25721 // 2) To match ANDN packed intrinsics
25722 if (VT != MVT::v2i64 && VT != MVT::v4i64)
25725 // Check LHS for vnot
25726 if (N0.getOpcode() == ISD::XOR &&
25727 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
25728 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
25729 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
25731 // Check RHS for vnot
25732 if (N1.getOpcode() == ISD::XOR &&
25733 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
25734 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
25735 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
25740 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
25741 TargetLowering::DAGCombinerInfo &DCI,
25742 const X86Subtarget *Subtarget) {
25743 if (DCI.isBeforeLegalizeOps())
25746 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
25749 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25752 SDValue N0 = N->getOperand(0);
25753 SDValue N1 = N->getOperand(1);
25754 EVT VT = N->getValueType(0);
25756 // look for psign/blend
25757 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
25758 if (!Subtarget->hasSSSE3() ||
25759 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
25762 // Canonicalize pandn to RHS
25763 if (N0.getOpcode() == X86ISD::ANDNP)
25765 // or (and (m, y), (pandn m, x))
25766 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
25767 SDValue Mask = N1.getOperand(0);
25768 SDValue X = N1.getOperand(1);
25770 if (N0.getOperand(0) == Mask)
25771 Y = N0.getOperand(1);
25772 if (N0.getOperand(1) == Mask)
25773 Y = N0.getOperand(0);
25775 // Check to see if the mask appeared in both the AND and ANDNP and
25779 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
25780 // Look through mask bitcast.
25781 if (Mask.getOpcode() == ISD::BITCAST)
25782 Mask = Mask.getOperand(0);
25783 if (X.getOpcode() == ISD::BITCAST)
25784 X = X.getOperand(0);
25785 if (Y.getOpcode() == ISD::BITCAST)
25786 Y = Y.getOperand(0);
25788 EVT MaskVT = Mask.getValueType();
25790 // Validate that the Mask operand is a vector sra node.
25791 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
25792 // there is no psrai.b
25793 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25794 unsigned SraAmt = ~0;
25795 if (Mask.getOpcode() == ISD::SRA) {
25796 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25797 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25798 SraAmt = AmtConst->getZExtValue();
25799 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25800 SDValue SraC = Mask.getOperand(1);
25801 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25803 if ((SraAmt + 1) != EltBits)
25808 // Now we know we at least have a plendvb with the mask val. See if
25809 // we can form a psignb/w/d.
25810 // psign = x.type == y.type == mask.type && y = sub(0, x);
25811 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25812 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25813 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25814 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25815 "Unsupported VT for PSIGN");
25816 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25817 return DAG.getBitcast(VT, Mask);
25819 // PBLENDVB only available on SSE 4.1
25820 if (!Subtarget->hasSSE41())
25823 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25825 X = DAG.getBitcast(BlendVT, X);
25826 Y = DAG.getBitcast(BlendVT, Y);
25827 Mask = DAG.getBitcast(BlendVT, Mask);
25828 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25829 return DAG.getBitcast(VT, Mask);
25833 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25836 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25837 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25839 // SHLD/SHRD instructions have lower register pressure, but on some
25840 // platforms they have higher latency than the equivalent
25841 // series of shifts/or that would otherwise be generated.
25842 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25843 // have higher latencies and we are not optimizing for size.
25844 if (!OptForSize && Subtarget->isSHLDSlow())
25847 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25849 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25851 if (!N0.hasOneUse() || !N1.hasOneUse())
25854 SDValue ShAmt0 = N0.getOperand(1);
25855 if (ShAmt0.getValueType() != MVT::i8)
25857 SDValue ShAmt1 = N1.getOperand(1);
25858 if (ShAmt1.getValueType() != MVT::i8)
25860 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25861 ShAmt0 = ShAmt0.getOperand(0);
25862 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25863 ShAmt1 = ShAmt1.getOperand(0);
25866 unsigned Opc = X86ISD::SHLD;
25867 SDValue Op0 = N0.getOperand(0);
25868 SDValue Op1 = N1.getOperand(0);
25869 if (ShAmt0.getOpcode() == ISD::SUB) {
25870 Opc = X86ISD::SHRD;
25871 std::swap(Op0, Op1);
25872 std::swap(ShAmt0, ShAmt1);
25875 unsigned Bits = VT.getSizeInBits();
25876 if (ShAmt1.getOpcode() == ISD::SUB) {
25877 SDValue Sum = ShAmt1.getOperand(0);
25878 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25879 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25880 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25881 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25882 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25883 return DAG.getNode(Opc, DL, VT,
25885 DAG.getNode(ISD::TRUNCATE, DL,
25888 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25889 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25891 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25892 return DAG.getNode(Opc, DL, VT,
25893 N0.getOperand(0), N1.getOperand(0),
25894 DAG.getNode(ISD::TRUNCATE, DL,
25901 // Generate NEG and CMOV for integer abs.
25902 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25903 EVT VT = N->getValueType(0);
25905 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25906 // 8-bit integer abs to NEG and CMOV.
25907 if (VT.isInteger() && VT.getSizeInBits() == 8)
25910 SDValue N0 = N->getOperand(0);
25911 SDValue N1 = N->getOperand(1);
25914 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25915 // and change it to SUB and CMOV.
25916 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25917 N0.getOpcode() == ISD::ADD &&
25918 N0.getOperand(1) == N1 &&
25919 N1.getOpcode() == ISD::SRA &&
25920 N1.getOperand(0) == N0.getOperand(0))
25921 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25922 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25923 // Generate SUB & CMOV.
25924 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25925 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25927 SDValue Ops[] = { N0.getOperand(0), Neg,
25928 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25929 SDValue(Neg.getNode(), 1) };
25930 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25935 // Try to turn tests against the signbit in the form of:
25936 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25939 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25940 // This is only worth doing if the output type is i8.
25941 if (N->getValueType(0) != MVT::i8)
25944 SDValue N0 = N->getOperand(0);
25945 SDValue N1 = N->getOperand(1);
25947 // We should be performing an xor against a truncated shift.
25948 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25951 // Make sure we are performing an xor against one.
25952 if (!isOneConstant(N1))
25955 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25956 SDValue Shift = N0.getOperand(0);
25957 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25960 // Make sure we are truncating from one of i16, i32 or i64.
25961 EVT ShiftTy = Shift.getValueType();
25962 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25965 // Make sure the shift amount extracts the sign bit.
25966 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25967 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25970 // Create a greater-than comparison against -1.
25971 // N.B. Using SETGE against 0 works but we want a canonical looking
25972 // comparison, using SETGT matches up with what TranslateX86CC.
25974 SDValue ShiftOp = Shift.getOperand(0);
25975 EVT ShiftOpTy = ShiftOp.getValueType();
25976 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25977 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25981 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25982 TargetLowering::DAGCombinerInfo &DCI,
25983 const X86Subtarget *Subtarget) {
25984 if (DCI.isBeforeLegalizeOps())
25987 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25990 if (Subtarget->hasCMov())
25991 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25994 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
26000 /// This function detects the AVG pattern between vectors of unsigned i8/i16,
26001 /// which is c = (a + b + 1) / 2, and replace this operation with the efficient
26002 /// X86ISD::AVG instruction.
26003 static SDValue detectAVGPattern(SDValue In, EVT VT, SelectionDAG &DAG,
26004 const X86Subtarget *Subtarget, SDLoc DL) {
26005 if (!VT.isVector() || !VT.isSimple())
26007 EVT InVT = In.getValueType();
26008 unsigned NumElems = VT.getVectorNumElements();
26010 EVT ScalarVT = VT.getVectorElementType();
26011 if (!((ScalarVT == MVT::i8 || ScalarVT == MVT::i16) &&
26012 isPowerOf2_32(NumElems)))
26015 // InScalarVT is the intermediate type in AVG pattern and it should be greater
26016 // than the original input type (i8/i16).
26017 EVT InScalarVT = InVT.getVectorElementType();
26018 if (InScalarVT.getSizeInBits() <= ScalarVT.getSizeInBits())
26021 if (Subtarget->hasAVX512()) {
26022 if (VT.getSizeInBits() > 512)
26024 } else if (Subtarget->hasAVX2()) {
26025 if (VT.getSizeInBits() > 256)
26028 if (VT.getSizeInBits() > 128)
26032 // Detect the following pattern:
26034 // %1 = zext <N x i8> %a to <N x i32>
26035 // %2 = zext <N x i8> %b to <N x i32>
26036 // %3 = add nuw nsw <N x i32> %1, <i32 1 x N>
26037 // %4 = add nuw nsw <N x i32> %3, %2
26038 // %5 = lshr <N x i32> %N, <i32 1 x N>
26039 // %6 = trunc <N x i32> %5 to <N x i8>
26041 // In AVX512, the last instruction can also be a trunc store.
26043 if (In.getOpcode() != ISD::SRL)
26046 // A lambda checking the given SDValue is a constant vector and each element
26047 // is in the range [Min, Max].
26048 auto IsConstVectorInRange = [](SDValue V, unsigned Min, unsigned Max) {
26049 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(V);
26050 if (!BV || !BV->isConstant())
26052 for (unsigned i = 0, e = V.getNumOperands(); i < e; i++) {
26053 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(i));
26056 uint64_t Val = C->getZExtValue();
26057 if (Val < Min || Val > Max)
26063 // Check if each element of the vector is left-shifted by one.
26064 auto LHS = In.getOperand(0);
26065 auto RHS = In.getOperand(1);
26066 if (!IsConstVectorInRange(RHS, 1, 1))
26068 if (LHS.getOpcode() != ISD::ADD)
26071 // Detect a pattern of a + b + 1 where the order doesn't matter.
26072 SDValue Operands[3];
26073 Operands[0] = LHS.getOperand(0);
26074 Operands[1] = LHS.getOperand(1);
26076 // Take care of the case when one of the operands is a constant vector whose
26077 // element is in the range [1, 256].
26078 if (IsConstVectorInRange(Operands[1], 1, ScalarVT == MVT::i8 ? 256 : 65536) &&
26079 Operands[0].getOpcode() == ISD::ZERO_EXTEND &&
26080 Operands[0].getOperand(0).getValueType() == VT) {
26081 // The pattern is detected. Subtract one from the constant vector, then
26082 // demote it and emit X86ISD::AVG instruction.
26083 SDValue One = DAG.getConstant(1, DL, InScalarVT);
26084 SDValue Ones = DAG.getNode(ISD::BUILD_VECTOR, DL, InVT,
26085 SmallVector<SDValue, 8>(NumElems, One));
26086 Operands[1] = DAG.getNode(ISD::SUB, DL, InVT, Operands[1], Ones);
26087 Operands[1] = DAG.getNode(ISD::TRUNCATE, DL, VT, Operands[1]);
26088 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26092 if (Operands[0].getOpcode() == ISD::ADD)
26093 std::swap(Operands[0], Operands[1]);
26094 else if (Operands[1].getOpcode() != ISD::ADD)
26096 Operands[2] = Operands[1].getOperand(0);
26097 Operands[1] = Operands[1].getOperand(1);
26099 // Now we have three operands of two additions. Check that one of them is a
26100 // constant vector with ones, and the other two are promoted from i8/i16.
26101 for (int i = 0; i < 3; ++i) {
26102 if (!IsConstVectorInRange(Operands[i], 1, 1))
26104 std::swap(Operands[i], Operands[2]);
26106 // Check if Operands[0] and Operands[1] are results of type promotion.
26107 for (int j = 0; j < 2; ++j)
26108 if (Operands[j].getOpcode() != ISD::ZERO_EXTEND ||
26109 Operands[j].getOperand(0).getValueType() != VT)
26112 // The pattern is detected, emit X86ISD::AVG instruction.
26113 return DAG.getNode(X86ISD::AVG, DL, VT, Operands[0].getOperand(0),
26114 Operands[1].getOperand(0));
26120 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
26121 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
26122 TargetLowering::DAGCombinerInfo &DCI,
26123 const X86Subtarget *Subtarget) {
26124 LoadSDNode *Ld = cast<LoadSDNode>(N);
26125 EVT RegVT = Ld->getValueType(0);
26126 EVT MemVT = Ld->getMemoryVT();
26128 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26130 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
26131 // into two 16-byte operations.
26132 ISD::LoadExtType Ext = Ld->getExtensionType();
26134 unsigned AddressSpace = Ld->getAddressSpace();
26135 unsigned Alignment = Ld->getAlignment();
26136 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
26137 Ext == ISD::NON_EXTLOAD &&
26138 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
26139 AddressSpace, Alignment, &Fast) && !Fast) {
26140 unsigned NumElems = RegVT.getVectorNumElements();
26144 SDValue Ptr = Ld->getBasePtr();
26145 SDValue Increment =
26146 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26148 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
26150 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26151 Ld->getPointerInfo(), Ld->isVolatile(),
26152 Ld->isNonTemporal(), Ld->isInvariant(),
26154 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26155 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
26156 Ld->getPointerInfo(), Ld->isVolatile(),
26157 Ld->isNonTemporal(), Ld->isInvariant(),
26158 std::min(16U, Alignment));
26159 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
26161 Load2.getValue(1));
26163 SDValue NewVec = DAG.getUNDEF(RegVT);
26164 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
26165 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
26166 return DCI.CombineTo(N, NewVec, TF, true);
26172 /// PerformMLOADCombine - Resolve extending loads
26173 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
26174 TargetLowering::DAGCombinerInfo &DCI,
26175 const X86Subtarget *Subtarget) {
26176 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
26177 if (Mld->getExtensionType() != ISD::SEXTLOAD)
26180 EVT VT = Mld->getValueType(0);
26181 unsigned NumElems = VT.getVectorNumElements();
26182 EVT LdVT = Mld->getMemoryVT();
26185 assert(LdVT != VT && "Cannot extend to the same type");
26186 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
26187 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
26188 // From, To sizes and ElemCount must be pow of two
26189 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26190 "Unexpected size for extending masked load");
26192 unsigned SizeRatio = ToSz / FromSz;
26193 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
26195 // Create a type on which we perform the shuffle
26196 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26197 LdVT.getScalarType(), NumElems*SizeRatio);
26198 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26200 // Convert Src0 value
26201 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
26202 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
26203 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26204 for (unsigned i = 0; i != NumElems; ++i)
26205 ShuffleVec[i] = i * SizeRatio;
26207 // Can't shuffle using an illegal type.
26208 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26209 "WideVecVT should be legal");
26210 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
26211 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
26213 // Prepare the new mask
26215 SDValue Mask = Mld->getMask();
26216 if (Mask.getValueType() == VT) {
26217 // Mask and original value have the same type
26218 NewMask = DAG.getBitcast(WideVecVT, Mask);
26219 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26220 for (unsigned i = 0; i != NumElems; ++i)
26221 ShuffleVec[i] = i * SizeRatio;
26222 for (unsigned i = NumElems; i != NumElems * SizeRatio; ++i)
26223 ShuffleVec[i] = NumElems * SizeRatio;
26224 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26225 DAG.getConstant(0, dl, WideVecVT),
26229 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26230 unsigned WidenNumElts = NumElems*SizeRatio;
26231 unsigned MaskNumElts = VT.getVectorNumElements();
26232 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26235 unsigned NumConcat = WidenNumElts / MaskNumElts;
26236 SmallVector<SDValue, 16> Ops(NumConcat);
26237 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26239 for (unsigned i = 1; i != NumConcat; ++i)
26242 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26245 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
26246 Mld->getBasePtr(), NewMask, WideSrc0,
26247 Mld->getMemoryVT(), Mld->getMemOperand(),
26249 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
26250 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
26252 /// PerformMSTORECombine - Resolve truncating stores
26253 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
26254 const X86Subtarget *Subtarget) {
26255 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
26256 if (!Mst->isTruncatingStore())
26259 EVT VT = Mst->getValue().getValueType();
26260 unsigned NumElems = VT.getVectorNumElements();
26261 EVT StVT = Mst->getMemoryVT();
26264 assert(StVT != VT && "Cannot truncate to the same type");
26265 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26266 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26268 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26270 // The truncating store is legal in some cases. For example
26271 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26272 // are designated for truncate store.
26273 // In this case we don't need any further transformations.
26274 if (TLI.isTruncStoreLegal(VT, StVT))
26277 // From, To sizes and ElemCount must be pow of two
26278 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
26279 "Unexpected size for truncating masked store");
26280 // We are going to use the original vector elt for storing.
26281 // Accumulated smaller vector elements must be a multiple of the store size.
26282 assert (((NumElems * FromSz) % ToSz) == 0 &&
26283 "Unexpected ratio for truncating masked store");
26285 unsigned SizeRatio = FromSz / ToSz;
26286 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26288 // Create a type on which we perform the shuffle
26289 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26290 StVT.getScalarType(), NumElems*SizeRatio);
26292 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26294 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
26295 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
26296 for (unsigned i = 0; i != NumElems; ++i)
26297 ShuffleVec[i] = i * SizeRatio;
26299 // Can't shuffle using an illegal type.
26300 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
26301 "WideVecVT should be legal");
26303 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26304 DAG.getUNDEF(WideVecVT),
26308 SDValue Mask = Mst->getMask();
26309 if (Mask.getValueType() == VT) {
26310 // Mask and original value have the same type
26311 NewMask = DAG.getBitcast(WideVecVT, Mask);
26312 for (unsigned i = 0; i != NumElems; ++i)
26313 ShuffleVec[i] = i * SizeRatio;
26314 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
26315 ShuffleVec[i] = NumElems*SizeRatio;
26316 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
26317 DAG.getConstant(0, dl, WideVecVT),
26321 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
26322 unsigned WidenNumElts = NumElems*SizeRatio;
26323 unsigned MaskNumElts = VT.getVectorNumElements();
26324 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
26327 unsigned NumConcat = WidenNumElts / MaskNumElts;
26328 SmallVector<SDValue, 16> Ops(NumConcat);
26329 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
26331 for (unsigned i = 1; i != NumConcat; ++i)
26334 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
26337 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal,
26338 Mst->getBasePtr(), NewMask, StVT,
26339 Mst->getMemOperand(), false);
26341 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
26342 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
26343 const X86Subtarget *Subtarget) {
26344 StoreSDNode *St = cast<StoreSDNode>(N);
26345 EVT VT = St->getValue().getValueType();
26346 EVT StVT = St->getMemoryVT();
26348 SDValue StoredVal = St->getOperand(1);
26349 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26351 // If we are saving a concatenation of two XMM registers and 32-byte stores
26352 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
26354 unsigned AddressSpace = St->getAddressSpace();
26355 unsigned Alignment = St->getAlignment();
26356 if (VT.is256BitVector() && StVT == VT &&
26357 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
26358 AddressSpace, Alignment, &Fast) && !Fast) {
26359 unsigned NumElems = VT.getVectorNumElements();
26363 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
26364 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
26367 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
26368 SDValue Ptr0 = St->getBasePtr();
26369 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
26371 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
26372 St->getPointerInfo(), St->isVolatile(),
26373 St->isNonTemporal(), Alignment);
26374 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
26375 St->getPointerInfo(), St->isVolatile(),
26376 St->isNonTemporal(),
26377 std::min(16U, Alignment));
26378 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
26381 // Optimize trunc store (of multiple scalars) to shuffle and store.
26382 // First, pack all of the elements in one place. Next, store to memory
26383 // in fewer chunks.
26384 if (St->isTruncatingStore() && VT.isVector()) {
26385 // Check if we can detect an AVG pattern from the truncation. If yes,
26386 // replace the trunc store by a normal store with the result of X86ISD::AVG
26389 detectAVGPattern(St->getValue(), St->getMemoryVT(), DAG, Subtarget, dl);
26391 return DAG.getStore(St->getChain(), dl, Avg, St->getBasePtr(),
26392 St->getPointerInfo(), St->isVolatile(),
26393 St->isNonTemporal(), St->getAlignment());
26395 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26396 unsigned NumElems = VT.getVectorNumElements();
26397 assert(StVT != VT && "Cannot truncate to the same type");
26398 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
26399 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
26401 // The truncating store is legal in some cases. For example
26402 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
26403 // are designated for truncate store.
26404 // In this case we don't need any further transformations.
26405 if (TLI.isTruncStoreLegal(VT, StVT))
26408 // From, To sizes and ElemCount must be pow of two
26409 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
26410 // We are going to use the original vector elt for storing.
26411 // Accumulated smaller vector elements must be a multiple of the store size.
26412 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
26414 unsigned SizeRatio = FromSz / ToSz;
26416 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
26418 // Create a type on which we perform the shuffle
26419 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
26420 StVT.getScalarType(), NumElems*SizeRatio);
26422 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
26424 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
26425 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
26426 for (unsigned i = 0; i != NumElems; ++i)
26427 ShuffleVec[i] = i * SizeRatio;
26429 // Can't shuffle using an illegal type.
26430 if (!TLI.isTypeLegal(WideVecVT))
26433 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
26434 DAG.getUNDEF(WideVecVT),
26436 // At this point all of the data is stored at the bottom of the
26437 // register. We now need to save it to mem.
26439 // Find the largest store unit
26440 MVT StoreType = MVT::i8;
26441 for (MVT Tp : MVT::integer_valuetypes()) {
26442 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
26446 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
26447 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
26448 (64 <= NumElems * ToSz))
26449 StoreType = MVT::f64;
26451 // Bitcast the original vector into a vector of store-size units
26452 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
26453 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
26454 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
26455 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
26456 SmallVector<SDValue, 8> Chains;
26457 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
26458 TLI.getPointerTy(DAG.getDataLayout()));
26459 SDValue Ptr = St->getBasePtr();
26461 // Perform one or more big stores into memory.
26462 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
26463 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
26464 StoreType, ShuffWide,
26465 DAG.getIntPtrConstant(i, dl));
26466 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
26467 St->getPointerInfo(), St->isVolatile(),
26468 St->isNonTemporal(), St->getAlignment());
26469 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
26470 Chains.push_back(Ch);
26473 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
26476 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
26477 // the FP state in cases where an emms may be missing.
26478 // A preferable solution to the general problem is to figure out the right
26479 // places to insert EMMS. This qualifies as a quick hack.
26481 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
26482 if (VT.getSizeInBits() != 64)
26485 const Function *F = DAG.getMachineFunction().getFunction();
26486 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
26488 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
26489 if ((VT.isVector() ||
26490 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
26491 isa<LoadSDNode>(St->getValue()) &&
26492 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
26493 St->getChain().hasOneUse() && !St->isVolatile()) {
26494 SDNode* LdVal = St->getValue().getNode();
26495 LoadSDNode *Ld = nullptr;
26496 int TokenFactorIndex = -1;
26497 SmallVector<SDValue, 8> Ops;
26498 SDNode* ChainVal = St->getChain().getNode();
26499 // Must be a store of a load. We currently handle two cases: the load
26500 // is a direct child, and it's under an intervening TokenFactor. It is
26501 // possible to dig deeper under nested TokenFactors.
26502 if (ChainVal == LdVal)
26503 Ld = cast<LoadSDNode>(St->getChain());
26504 else if (St->getValue().hasOneUse() &&
26505 ChainVal->getOpcode() == ISD::TokenFactor) {
26506 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
26507 if (ChainVal->getOperand(i).getNode() == LdVal) {
26508 TokenFactorIndex = i;
26509 Ld = cast<LoadSDNode>(St->getValue());
26511 Ops.push_back(ChainVal->getOperand(i));
26515 if (!Ld || !ISD::isNormalLoad(Ld))
26518 // If this is not the MMX case, i.e. we are just turning i64 load/store
26519 // into f64 load/store, avoid the transformation if there are multiple
26520 // uses of the loaded value.
26521 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
26526 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
26527 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
26529 if (Subtarget->is64Bit() || F64IsLegal) {
26530 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
26531 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
26532 Ld->getPointerInfo(), Ld->isVolatile(),
26533 Ld->isNonTemporal(), Ld->isInvariant(),
26534 Ld->getAlignment());
26535 SDValue NewChain = NewLd.getValue(1);
26536 if (TokenFactorIndex != -1) {
26537 Ops.push_back(NewChain);
26538 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26540 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
26541 St->getPointerInfo(),
26542 St->isVolatile(), St->isNonTemporal(),
26543 St->getAlignment());
26546 // Otherwise, lower to two pairs of 32-bit loads / stores.
26547 SDValue LoAddr = Ld->getBasePtr();
26548 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
26549 DAG.getConstant(4, LdDL, MVT::i32));
26551 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
26552 Ld->getPointerInfo(),
26553 Ld->isVolatile(), Ld->isNonTemporal(),
26554 Ld->isInvariant(), Ld->getAlignment());
26555 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
26556 Ld->getPointerInfo().getWithOffset(4),
26557 Ld->isVolatile(), Ld->isNonTemporal(),
26559 MinAlign(Ld->getAlignment(), 4));
26561 SDValue NewChain = LoLd.getValue(1);
26562 if (TokenFactorIndex != -1) {
26563 Ops.push_back(LoLd);
26564 Ops.push_back(HiLd);
26565 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
26568 LoAddr = St->getBasePtr();
26569 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
26570 DAG.getConstant(4, StDL, MVT::i32));
26572 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
26573 St->getPointerInfo(),
26574 St->isVolatile(), St->isNonTemporal(),
26575 St->getAlignment());
26576 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
26577 St->getPointerInfo().getWithOffset(4),
26579 St->isNonTemporal(),
26580 MinAlign(St->getAlignment(), 4));
26581 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
26584 // This is similar to the above case, but here we handle a scalar 64-bit
26585 // integer store that is extracted from a vector on a 32-bit target.
26586 // If we have SSE2, then we can treat it like a floating-point double
26587 // to get past legalization. The execution dependencies fixup pass will
26588 // choose the optimal machine instruction for the store if this really is
26589 // an integer or v2f32 rather than an f64.
26590 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
26591 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
26592 SDValue OldExtract = St->getOperand(1);
26593 SDValue ExtOp0 = OldExtract.getOperand(0);
26594 unsigned VecSize = ExtOp0.getValueSizeInBits();
26595 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
26596 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
26597 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
26598 BitCast, OldExtract.getOperand(1));
26599 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
26600 St->getPointerInfo(), St->isVolatile(),
26601 St->isNonTemporal(), St->getAlignment());
26607 /// Return 'true' if this vector operation is "horizontal"
26608 /// and return the operands for the horizontal operation in LHS and RHS. A
26609 /// horizontal operation performs the binary operation on successive elements
26610 /// of its first operand, then on successive elements of its second operand,
26611 /// returning the resulting values in a vector. For example, if
26612 /// A = < float a0, float a1, float a2, float a3 >
26614 /// B = < float b0, float b1, float b2, float b3 >
26615 /// then the result of doing a horizontal operation on A and B is
26616 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
26617 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
26618 /// A horizontal-op B, for some already available A and B, and if so then LHS is
26619 /// set to A, RHS to B, and the routine returns 'true'.
26620 /// Note that the binary operation should have the property that if one of the
26621 /// operands is UNDEF then the result is UNDEF.
26622 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
26623 // Look for the following pattern: if
26624 // A = < float a0, float a1, float a2, float a3 >
26625 // B = < float b0, float b1, float b2, float b3 >
26627 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
26628 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
26629 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
26630 // which is A horizontal-op B.
26632 // At least one of the operands should be a vector shuffle.
26633 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
26634 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
26637 MVT VT = LHS.getSimpleValueType();
26639 assert((VT.is128BitVector() || VT.is256BitVector()) &&
26640 "Unsupported vector type for horizontal add/sub");
26642 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
26643 // operate independently on 128-bit lanes.
26644 unsigned NumElts = VT.getVectorNumElements();
26645 unsigned NumLanes = VT.getSizeInBits()/128;
26646 unsigned NumLaneElts = NumElts / NumLanes;
26647 assert((NumLaneElts % 2 == 0) &&
26648 "Vector type should have an even number of elements in each lane");
26649 unsigned HalfLaneElts = NumLaneElts/2;
26651 // View LHS in the form
26652 // LHS = VECTOR_SHUFFLE A, B, LMask
26653 // If LHS is not a shuffle then pretend it is the shuffle
26654 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
26655 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
26658 SmallVector<int, 16> LMask(NumElts);
26659 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26660 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
26661 A = LHS.getOperand(0);
26662 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
26663 B = LHS.getOperand(1);
26664 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
26665 std::copy(Mask.begin(), Mask.end(), LMask.begin());
26667 if (LHS.getOpcode() != ISD::UNDEF)
26669 for (unsigned i = 0; i != NumElts; ++i)
26673 // Likewise, view RHS in the form
26674 // RHS = VECTOR_SHUFFLE C, D, RMask
26676 SmallVector<int, 16> RMask(NumElts);
26677 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
26678 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
26679 C = RHS.getOperand(0);
26680 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
26681 D = RHS.getOperand(1);
26682 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
26683 std::copy(Mask.begin(), Mask.end(), RMask.begin());
26685 if (RHS.getOpcode() != ISD::UNDEF)
26687 for (unsigned i = 0; i != NumElts; ++i)
26691 // Check that the shuffles are both shuffling the same vectors.
26692 if (!(A == C && B == D) && !(A == D && B == C))
26695 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
26696 if (!A.getNode() && !B.getNode())
26699 // If A and B occur in reverse order in RHS, then "swap" them (which means
26700 // rewriting the mask).
26702 ShuffleVectorSDNode::commuteMask(RMask);
26704 // At this point LHS and RHS are equivalent to
26705 // LHS = VECTOR_SHUFFLE A, B, LMask
26706 // RHS = VECTOR_SHUFFLE A, B, RMask
26707 // Check that the masks correspond to performing a horizontal operation.
26708 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
26709 for (unsigned i = 0; i != NumLaneElts; ++i) {
26710 int LIdx = LMask[i+l], RIdx = RMask[i+l];
26712 // Ignore any UNDEF components.
26713 if (LIdx < 0 || RIdx < 0 ||
26714 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
26715 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
26718 // Check that successive elements are being operated on. If not, this is
26719 // not a horizontal operation.
26720 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
26721 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
26722 if (!(LIdx == Index && RIdx == Index + 1) &&
26723 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
26728 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
26729 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
26733 /// Do target-specific dag combines on floating point adds.
26734 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
26735 const X86Subtarget *Subtarget) {
26736 EVT VT = N->getValueType(0);
26737 SDValue LHS = N->getOperand(0);
26738 SDValue RHS = N->getOperand(1);
26740 // Try to synthesize horizontal adds from adds of shuffles.
26741 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26742 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26743 isHorizontalBinOp(LHS, RHS, true))
26744 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
26748 /// Do target-specific dag combines on floating point subs.
26749 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
26750 const X86Subtarget *Subtarget) {
26751 EVT VT = N->getValueType(0);
26752 SDValue LHS = N->getOperand(0);
26753 SDValue RHS = N->getOperand(1);
26755 // Try to synthesize horizontal subs from subs of shuffles.
26756 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
26757 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
26758 isHorizontalBinOp(LHS, RHS, false))
26759 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
26763 /// Truncate a group of v4i32 into v16i8/v8i16 using X86ISD::PACKUS.
26765 combineVectorTruncationWithPACKUS(SDNode *N, SelectionDAG &DAG,
26766 SmallVector<SDValue, 8> &Regs) {
26767 assert(Regs.size() > 0 && (Regs[0].getValueType() == MVT::v4i32 ||
26768 Regs[0].getValueType() == MVT::v2i64));
26769 EVT OutVT = N->getValueType(0);
26770 EVT OutSVT = OutVT.getVectorElementType();
26771 EVT InVT = Regs[0].getValueType();
26772 EVT InSVT = InVT.getVectorElementType();
26775 // First, use mask to unset all bits that won't appear in the result.
26776 assert((OutSVT == MVT::i8 || OutSVT == MVT::i16) &&
26777 "OutSVT can only be either i8 or i16.");
26779 DAG.getConstant(OutSVT == MVT::i8 ? 0xFF : 0xFFFF, DL, InSVT);
26780 SDValue MaskVec = DAG.getNode(
26781 ISD::BUILD_VECTOR, DL, InVT,
26782 SmallVector<SDValue, 8>(InVT.getVectorNumElements(), MaskVal));
26783 for (auto &Reg : Regs)
26784 Reg = DAG.getNode(ISD::AND, DL, InVT, MaskVec, Reg);
26786 MVT UnpackedVT, PackedVT;
26787 if (OutSVT == MVT::i8) {
26788 UnpackedVT = MVT::v8i16;
26789 PackedVT = MVT::v16i8;
26791 UnpackedVT = MVT::v4i32;
26792 PackedVT = MVT::v8i16;
26795 // In each iteration, truncate the type by a half size.
26796 auto RegNum = Regs.size();
26797 for (unsigned j = 1, e = InSVT.getSizeInBits() / OutSVT.getSizeInBits();
26798 j < e; j *= 2, RegNum /= 2) {
26799 for (unsigned i = 0; i < RegNum; i++)
26800 Regs[i] = DAG.getNode(ISD::BITCAST, DL, UnpackedVT, Regs[i]);
26801 for (unsigned i = 0; i < RegNum / 2; i++)
26802 Regs[i] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[i * 2],
26806 // If the type of the result is v8i8, we need do one more X86ISD::PACKUS, and
26807 // then extract a subvector as the result since v8i8 is not a legal type.
26808 if (OutVT == MVT::v8i8) {
26809 Regs[0] = DAG.getNode(X86ISD::PACKUS, DL, PackedVT, Regs[0], Regs[0]);
26810 Regs[0] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OutVT, Regs[0],
26811 DAG.getIntPtrConstant(0, DL));
26813 } else if (RegNum > 1) {
26814 Regs.resize(RegNum);
26815 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26820 /// Truncate a group of v4i32 into v8i16 using X86ISD::PACKSS.
26822 combineVectorTruncationWithPACKSS(SDNode *N, SelectionDAG &DAG,
26823 SmallVector<SDValue, 8> &Regs) {
26824 assert(Regs.size() > 0 && Regs[0].getValueType() == MVT::v4i32);
26825 EVT OutVT = N->getValueType(0);
26828 // Shift left by 16 bits, then arithmetic-shift right by 16 bits.
26829 SDValue ShAmt = DAG.getConstant(16, DL, MVT::i32);
26830 for (auto &Reg : Regs) {
26831 Reg = getTargetVShiftNode(X86ISD::VSHLI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26832 Reg = getTargetVShiftNode(X86ISD::VSRAI, DL, MVT::v4i32, Reg, ShAmt, DAG);
26835 for (unsigned i = 0, e = Regs.size() / 2; i < e; i++)
26836 Regs[i] = DAG.getNode(X86ISD::PACKSS, DL, MVT::v8i16, Regs[i * 2],
26839 if (Regs.size() > 2) {
26840 Regs.resize(Regs.size() / 2);
26841 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Regs);
26846 /// This function transforms truncation from vXi32/vXi64 to vXi8/vXi16 into
26847 /// X86ISD::PACKUS/X86ISD::PACKSS operations. We do it here because after type
26848 /// legalization the truncation will be translated into a BUILD_VECTOR with each
26849 /// element that is extracted from a vector and then truncated, and it is
26850 /// diffcult to do this optimization based on them.
26851 static SDValue combineVectorTruncation(SDNode *N, SelectionDAG &DAG,
26852 const X86Subtarget *Subtarget) {
26853 EVT OutVT = N->getValueType(0);
26854 if (!OutVT.isVector())
26857 SDValue In = N->getOperand(0);
26858 if (!In.getValueType().isSimple())
26861 EVT InVT = In.getValueType();
26862 unsigned NumElems = OutVT.getVectorNumElements();
26864 // TODO: On AVX2, the behavior of X86ISD::PACKUS is different from that on
26865 // SSE2, and we need to take care of it specially.
26866 // AVX512 provides vpmovdb.
26867 if (!Subtarget->hasSSE2() || Subtarget->hasAVX2())
26870 EVT OutSVT = OutVT.getVectorElementType();
26871 EVT InSVT = InVT.getVectorElementType();
26872 if (!((InSVT == MVT::i32 || InSVT == MVT::i64) &&
26873 (OutSVT == MVT::i8 || OutSVT == MVT::i16) && isPowerOf2_32(NumElems) &&
26877 // SSSE3's pshufb results in less instructions in the cases below.
26878 if (Subtarget->hasSSSE3() && NumElems == 8 &&
26879 ((OutSVT == MVT::i8 && InSVT != MVT::i64) ||
26880 (InSVT == MVT::i32 && OutSVT == MVT::i16)))
26885 // Split a long vector into vectors of legal type.
26886 unsigned RegNum = InVT.getSizeInBits() / 128;
26887 SmallVector<SDValue, 8> SubVec(RegNum);
26888 if (InSVT == MVT::i32) {
26889 for (unsigned i = 0; i < RegNum; i++)
26890 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
26891 DAG.getIntPtrConstant(i * 4, DL));
26893 for (unsigned i = 0; i < RegNum; i++)
26894 SubVec[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
26895 DAG.getIntPtrConstant(i * 2, DL));
26898 // SSE2 provides PACKUS for only 2 x v8i16 -> v16i8 and SSE4.1 provides PAKCUS
26899 // for 2 x v4i32 -> v8i16. For SSSE3 and below, we need to use PACKSS to
26900 // truncate 2 x v4i32 to v8i16.
26901 if (Subtarget->hasSSE41() || OutSVT == MVT::i8)
26902 return combineVectorTruncationWithPACKUS(N, DAG, SubVec);
26903 else if (InSVT == MVT::i32)
26904 return combineVectorTruncationWithPACKSS(N, DAG, SubVec);
26909 static SDValue PerformTRUNCATECombine(SDNode *N, SelectionDAG &DAG,
26910 const X86Subtarget *Subtarget) {
26911 // Try to detect AVG pattern first.
26912 SDValue Avg = detectAVGPattern(N->getOperand(0), N->getValueType(0), DAG,
26913 Subtarget, SDLoc(N));
26917 return combineVectorTruncation(N, DAG, Subtarget);
26920 /// Do target-specific dag combines on floating point negations.
26921 static SDValue PerformFNEGCombine(SDNode *N, SelectionDAG &DAG,
26922 const X86Subtarget *Subtarget) {
26923 EVT VT = N->getValueType(0);
26924 EVT SVT = VT.getScalarType();
26925 SDValue Arg = N->getOperand(0);
26928 // Let legalize expand this if it isn't a legal type yet.
26929 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26932 // If we're negating a FMUL node on a target with FMA, then we can avoid the
26933 // use of a constant by performing (-0 - A*B) instead.
26934 // FIXME: Check rounding control flags as well once it becomes available.
26935 if (Arg.getOpcode() == ISD::FMUL && (SVT == MVT::f32 || SVT == MVT::f64) &&
26936 Arg->getFlags()->hasNoSignedZeros() && Subtarget->hasAnyFMA()) {
26937 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
26938 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26939 Arg.getOperand(1), Zero);
26942 // If we're negating a FMA node, then we can adjust the
26943 // instruction to include the extra negation.
26944 if (Arg.hasOneUse()) {
26945 switch (Arg.getOpcode()) {
26946 case X86ISD::FMADD:
26947 return DAG.getNode(X86ISD::FNMSUB, DL, VT, Arg.getOperand(0),
26948 Arg.getOperand(1), Arg.getOperand(2));
26949 case X86ISD::FMSUB:
26950 return DAG.getNode(X86ISD::FNMADD, DL, VT, Arg.getOperand(0),
26951 Arg.getOperand(1), Arg.getOperand(2));
26952 case X86ISD::FNMADD:
26953 return DAG.getNode(X86ISD::FMSUB, DL, VT, Arg.getOperand(0),
26954 Arg.getOperand(1), Arg.getOperand(2));
26955 case X86ISD::FNMSUB:
26956 return DAG.getNode(X86ISD::FMADD, DL, VT, Arg.getOperand(0),
26957 Arg.getOperand(1), Arg.getOperand(2));
26963 static SDValue lowerX86FPLogicOp(SDNode *N, SelectionDAG &DAG,
26964 const X86Subtarget *Subtarget) {
26965 EVT VT = N->getValueType(0);
26966 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
26967 // VXORPS, VORPS, VANDPS, VANDNPS are supported only under DQ extention.
26968 // These logic operations may be executed in the integer domain.
26970 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
26971 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
26973 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
26974 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
26975 unsigned IntOpcode = 0;
26976 switch (N->getOpcode()) {
26977 default: llvm_unreachable("Unexpected FP logic op");
26978 case X86ISD::FOR: IntOpcode = ISD::OR; break;
26979 case X86ISD::FXOR: IntOpcode = ISD::XOR; break;
26980 case X86ISD::FAND: IntOpcode = ISD::AND; break;
26981 case X86ISD::FANDN: IntOpcode = X86ISD::ANDNP; break;
26983 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
26984 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
26988 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
26989 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
26990 const X86Subtarget *Subtarget) {
26991 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
26993 // F[X]OR(0.0, x) -> x
26994 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
26995 if (C->getValueAPF().isPosZero())
26996 return N->getOperand(1);
26998 // F[X]OR(x, 0.0) -> x
26999 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27000 if (C->getValueAPF().isPosZero())
27001 return N->getOperand(0);
27003 return lowerX86FPLogicOp(N, DAG, Subtarget);
27006 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
27007 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
27008 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
27010 // Only perform optimizations if UnsafeMath is used.
27011 if (!DAG.getTarget().Options.UnsafeFPMath)
27014 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
27015 // into FMINC and FMAXC, which are Commutative operations.
27016 unsigned NewOp = 0;
27017 switch (N->getOpcode()) {
27018 default: llvm_unreachable("unknown opcode");
27019 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
27020 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
27023 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
27024 N->getOperand(0), N->getOperand(1));
27027 static SDValue performFMinNumFMaxNumCombine(SDNode *N, SelectionDAG &DAG,
27028 const X86Subtarget *Subtarget) {
27029 if (Subtarget->useSoftFloat())
27032 // TODO: Check for global or instruction-level "nnan". In that case, we
27033 // should be able to lower to FMAX/FMIN alone.
27034 // TODO: If an operand is already known to be a NaN or not a NaN, this
27035 // should be an optional swap and FMAX/FMIN.
27037 EVT VT = N->getValueType(0);
27038 if (!((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
27039 (Subtarget->hasSSE2() && (VT == MVT::f64 || VT == MVT::v2f64)) ||
27040 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))))
27043 // This takes at least 3 instructions, so favor a library call when operating
27044 // on a scalar and minimizing code size.
27045 if (!VT.isVector() && DAG.getMachineFunction().getFunction()->optForMinSize())
27048 SDValue Op0 = N->getOperand(0);
27049 SDValue Op1 = N->getOperand(1);
27051 EVT SetCCType = DAG.getTargetLoweringInfo().getSetCCResultType(
27052 DAG.getDataLayout(), *DAG.getContext(), VT);
27054 // There are 4 possibilities involving NaN inputs, and these are the required
27058 // ----------------
27059 // Num | Max | Op0 |
27060 // Op0 ----------------
27061 // NaN | Op1 | NaN |
27062 // ----------------
27064 // The SSE FP max/min instructions were not designed for this case, but rather
27066 // Min = Op1 < Op0 ? Op1 : Op0
27067 // Max = Op1 > Op0 ? Op1 : Op0
27069 // So they always return Op0 if either input is a NaN. However, we can still
27070 // use those instructions for fmaxnum by selecting away a NaN input.
27072 // If either operand is NaN, the 2nd source operand (Op0) is passed through.
27073 auto MinMaxOp = N->getOpcode() == ISD::FMAXNUM ? X86ISD::FMAX : X86ISD::FMIN;
27074 SDValue MinOrMax = DAG.getNode(MinMaxOp, DL, VT, Op1, Op0);
27075 SDValue IsOp0Nan = DAG.getSetCC(DL, SetCCType , Op0, Op0, ISD::SETUO);
27077 // If Op0 is a NaN, select Op1. Otherwise, select the max. If both operands
27078 // are NaN, the NaN value of Op1 is the result.
27079 auto SelectOpcode = VT.isVector() ? ISD::VSELECT : ISD::SELECT;
27080 return DAG.getNode(SelectOpcode, DL, VT, IsOp0Nan, Op1, MinOrMax);
27083 /// Do target-specific dag combines on X86ISD::FAND nodes.
27084 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG,
27085 const X86Subtarget *Subtarget) {
27086 // FAND(0.0, x) -> 0.0
27087 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27088 if (C->getValueAPF().isPosZero())
27089 return N->getOperand(0);
27091 // FAND(x, 0.0) -> 0.0
27092 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27093 if (C->getValueAPF().isPosZero())
27094 return N->getOperand(1);
27096 return lowerX86FPLogicOp(N, DAG, Subtarget);
27099 /// Do target-specific dag combines on X86ISD::FANDN nodes
27100 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG,
27101 const X86Subtarget *Subtarget) {
27102 // FANDN(0.0, x) -> x
27103 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
27104 if (C->getValueAPF().isPosZero())
27105 return N->getOperand(1);
27107 // FANDN(x, 0.0) -> 0.0
27108 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
27109 if (C->getValueAPF().isPosZero())
27110 return N->getOperand(1);
27112 return lowerX86FPLogicOp(N, DAG, Subtarget);
27115 static SDValue PerformBTCombine(SDNode *N,
27117 TargetLowering::DAGCombinerInfo &DCI) {
27118 // BT ignores high bits in the bit index operand.
27119 SDValue Op1 = N->getOperand(1);
27120 if (Op1.hasOneUse()) {
27121 unsigned BitWidth = Op1.getValueSizeInBits();
27122 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
27123 APInt KnownZero, KnownOne;
27124 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
27125 !DCI.isBeforeLegalizeOps());
27126 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27127 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
27128 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
27129 DCI.CommitTargetLoweringOpt(TLO);
27134 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
27135 SDValue Op = N->getOperand(0);
27136 if (Op.getOpcode() == ISD::BITCAST)
27137 Op = Op.getOperand(0);
27138 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
27139 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
27140 VT.getVectorElementType().getSizeInBits() ==
27141 OpVT.getVectorElementType().getSizeInBits()) {
27142 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
27147 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
27148 const X86Subtarget *Subtarget) {
27149 EVT VT = N->getValueType(0);
27150 if (!VT.isVector())
27153 SDValue N0 = N->getOperand(0);
27154 SDValue N1 = N->getOperand(1);
27155 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
27158 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
27159 // both SSE and AVX2 since there is no sign-extended shift right
27160 // operation on a vector with 64-bit elements.
27161 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
27162 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
27163 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
27164 N0.getOpcode() == ISD::SIGN_EXTEND)) {
27165 SDValue N00 = N0.getOperand(0);
27167 // EXTLOAD has a better solution on AVX2,
27168 // it may be replaced with X86ISD::VSEXT node.
27169 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
27170 if (!ISD::isNormalLoad(N00.getNode()))
27173 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
27174 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
27176 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
27182 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
27183 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
27184 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
27185 /// eliminate extend, add, and shift instructions.
27186 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
27187 const X86Subtarget *Subtarget) {
27188 // TODO: This should be valid for other integer types.
27189 EVT VT = Sext->getValueType(0);
27190 if (VT != MVT::i64)
27193 // We need an 'add nsw' feeding into the 'sext'.
27194 SDValue Add = Sext->getOperand(0);
27195 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
27198 // Having a constant operand to the 'add' ensures that we are not increasing
27199 // the instruction count because the constant is extended for free below.
27200 // A constant operand can also become the displacement field of an LEA.
27201 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
27205 // Don't make the 'add' bigger if there's no hope of combining it with some
27206 // other 'add' or 'shl' instruction.
27207 // TODO: It may be profitable to generate simpler LEA instructions in place
27208 // of single 'add' instructions, but the cost model for selecting an LEA
27209 // currently has a high threshold.
27210 bool HasLEAPotential = false;
27211 for (auto *User : Sext->uses()) {
27212 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
27213 HasLEAPotential = true;
27217 if (!HasLEAPotential)
27220 // Everything looks good, so pull the 'sext' ahead of the 'add'.
27221 int64_t AddConstant = AddOp1->getSExtValue();
27222 SDValue AddOp0 = Add.getOperand(0);
27223 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
27224 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
27226 // The wider add is guaranteed to not wrap because both operands are
27229 Flags.setNoSignedWrap(true);
27230 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
27233 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
27234 TargetLowering::DAGCombinerInfo &DCI,
27235 const X86Subtarget *Subtarget) {
27236 SDValue N0 = N->getOperand(0);
27237 EVT VT = N->getValueType(0);
27238 EVT SVT = VT.getScalarType();
27239 EVT InVT = N0.getValueType();
27240 EVT InSVT = InVT.getScalarType();
27243 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
27244 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
27245 // This exposes the sext to the sdivrem lowering, so that it directly extends
27246 // from AH (which we otherwise need to do contortions to access).
27247 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
27248 InVT == MVT::i8 && VT == MVT::i32) {
27249 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
27250 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
27251 N0.getOperand(0), N0.getOperand(1));
27252 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
27253 return R.getValue(1);
27256 if (!DCI.isBeforeLegalizeOps()) {
27257 if (InVT == MVT::i1) {
27258 SDValue Zero = DAG.getConstant(0, DL, VT);
27260 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
27261 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
27266 if (VT.isVector() && Subtarget->hasSSE2()) {
27267 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
27268 EVT InVT = N.getValueType();
27269 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
27270 Size / InVT.getScalarSizeInBits());
27271 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
27272 DAG.getUNDEF(InVT));
27274 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
27277 // If target-size is less than 128-bits, extend to a type that would extend
27278 // to 128 bits, extend that and extract the original target vector.
27279 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
27280 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27281 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27282 unsigned Scale = 128 / VT.getSizeInBits();
27284 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
27285 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
27286 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
27287 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
27288 DAG.getIntPtrConstant(0, DL));
27291 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
27292 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
27293 if (VT.getSizeInBits() == 128 &&
27294 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27295 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27296 SDValue ExOp = ExtendVecSize(DL, N0, 128);
27297 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
27300 // On pre-AVX2 targets, split into 128-bit nodes of
27301 // ISD::SIGN_EXTEND_VECTOR_INREG.
27302 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
27303 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
27304 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
27305 unsigned NumVecs = VT.getSizeInBits() / 128;
27306 unsigned NumSubElts = 128 / SVT.getSizeInBits();
27307 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
27308 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
27310 SmallVector<SDValue, 8> Opnds;
27311 for (unsigned i = 0, Offset = 0; i != NumVecs;
27312 ++i, Offset += NumSubElts) {
27313 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
27314 DAG.getIntPtrConstant(Offset, DL));
27315 SrcVec = ExtendVecSize(DL, SrcVec, 128);
27316 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
27317 Opnds.push_back(SrcVec);
27319 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
27323 if (Subtarget->hasAVX() && VT.is256BitVector())
27324 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27327 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
27333 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
27334 const X86Subtarget* Subtarget) {
27336 EVT VT = N->getValueType(0);
27338 // Let legalize expand this if it isn't a legal type yet.
27339 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
27342 EVT ScalarVT = VT.getScalarType();
27343 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) || !Subtarget->hasAnyFMA())
27346 SDValue A = N->getOperand(0);
27347 SDValue B = N->getOperand(1);
27348 SDValue C = N->getOperand(2);
27350 bool NegA = (A.getOpcode() == ISD::FNEG);
27351 bool NegB = (B.getOpcode() == ISD::FNEG);
27352 bool NegC = (C.getOpcode() == ISD::FNEG);
27354 // Negative multiplication when NegA xor NegB
27355 bool NegMul = (NegA != NegB);
27357 A = A.getOperand(0);
27359 B = B.getOperand(0);
27361 C = C.getOperand(0);
27365 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
27367 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
27369 return DAG.getNode(Opcode, dl, VT, A, B, C);
27372 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
27373 TargetLowering::DAGCombinerInfo &DCI,
27374 const X86Subtarget *Subtarget) {
27375 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
27376 // (and (i32 x86isd::setcc_carry), 1)
27377 // This eliminates the zext. This transformation is necessary because
27378 // ISD::SETCC is always legalized to i8.
27380 SDValue N0 = N->getOperand(0);
27381 EVT VT = N->getValueType(0);
27383 if (N0.getOpcode() == ISD::AND &&
27385 N0.getOperand(0).hasOneUse()) {
27386 SDValue N00 = N0.getOperand(0);
27387 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27388 if (!isOneConstant(N0.getOperand(1)))
27390 return DAG.getNode(ISD::AND, dl, VT,
27391 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27392 N00.getOperand(0), N00.getOperand(1)),
27393 DAG.getConstant(1, dl, VT));
27397 if (N0.getOpcode() == ISD::TRUNCATE &&
27399 N0.getOperand(0).hasOneUse()) {
27400 SDValue N00 = N0.getOperand(0);
27401 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
27402 return DAG.getNode(ISD::AND, dl, VT,
27403 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
27404 N00.getOperand(0), N00.getOperand(1)),
27405 DAG.getConstant(1, dl, VT));
27409 if (VT.is256BitVector())
27410 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
27413 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
27414 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
27415 // This exposes the zext to the udivrem lowering, so that it directly extends
27416 // from AH (which we otherwise need to do contortions to access).
27417 if (N0.getOpcode() == ISD::UDIVREM &&
27418 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
27419 (VT == MVT::i32 || VT == MVT::i64)) {
27420 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
27421 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
27422 N0.getOperand(0), N0.getOperand(1));
27423 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
27424 return R.getValue(1);
27430 // Optimize x == -y --> x+y == 0
27431 // x != -y --> x+y != 0
27432 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
27433 const X86Subtarget* Subtarget) {
27434 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
27435 SDValue LHS = N->getOperand(0);
27436 SDValue RHS = N->getOperand(1);
27437 EVT VT = N->getValueType(0);
27440 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
27441 if (isNullConstant(LHS.getOperand(0)) && LHS.hasOneUse()) {
27442 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
27443 LHS.getOperand(1));
27444 return DAG.getSetCC(DL, N->getValueType(0), addV,
27445 DAG.getConstant(0, DL, addV.getValueType()), CC);
27447 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
27448 if (isNullConstant(RHS.getOperand(0)) && RHS.hasOneUse()) {
27449 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
27450 RHS.getOperand(1));
27451 return DAG.getSetCC(DL, N->getValueType(0), addV,
27452 DAG.getConstant(0, DL, addV.getValueType()), CC);
27455 if (VT.getScalarType() == MVT::i1 &&
27456 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
27458 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27459 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27460 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27462 if (!IsSEXT0 || !IsVZero1) {
27463 // Swap the operands and update the condition code.
27464 std::swap(LHS, RHS);
27465 CC = ISD::getSetCCSwappedOperands(CC);
27467 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
27468 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
27469 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
27472 if (IsSEXT0 && IsVZero1) {
27473 assert(VT == LHS.getOperand(0).getValueType() &&
27474 "Uexpected operand type");
27475 if (CC == ISD::SETGT)
27476 return DAG.getConstant(0, DL, VT);
27477 if (CC == ISD::SETLE)
27478 return DAG.getConstant(1, DL, VT);
27479 if (CC == ISD::SETEQ || CC == ISD::SETGE)
27480 return DAG.getNOT(DL, LHS.getOperand(0), VT);
27482 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
27483 "Unexpected condition code!");
27484 return LHS.getOperand(0);
27491 static SDValue PerformGatherScatterCombine(SDNode *N, SelectionDAG &DAG) {
27493 // Gather and Scatter instructions use k-registers for masks. The type of
27494 // the masks is v*i1. So the mask will be truncated anyway.
27495 // The SIGN_EXTEND_INREG my be dropped.
27496 SDValue Mask = N->getOperand(2);
27497 if (Mask.getOpcode() == ISD::SIGN_EXTEND_INREG) {
27498 SmallVector<SDValue, 5> NewOps(N->op_begin(), N->op_end());
27499 NewOps[2] = Mask.getOperand(0);
27500 DAG.UpdateNodeOperands(N, NewOps);
27505 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
27506 // as "sbb reg,reg", since it can be extended without zext and produces
27507 // an all-ones bit which is more useful than 0/1 in some cases.
27508 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
27511 return DAG.getNode(ISD::AND, DL, VT,
27512 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27513 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27515 DAG.getConstant(1, DL, VT));
27516 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
27517 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
27518 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
27519 DAG.getConstant(X86::COND_B, DL, MVT::i8),
27523 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
27524 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
27525 TargetLowering::DAGCombinerInfo &DCI,
27526 const X86Subtarget *Subtarget) {
27528 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
27529 SDValue EFLAGS = N->getOperand(1);
27531 if (CC == X86::COND_A) {
27532 // Try to convert COND_A into COND_B in an attempt to facilitate
27533 // materializing "setb reg".
27535 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
27536 // cannot take an immediate as its first operand.
27538 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
27539 EFLAGS.getValueType().isInteger() &&
27540 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
27541 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
27542 EFLAGS.getNode()->getVTList(),
27543 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
27544 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
27545 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
27549 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
27550 // a zext and produces an all-ones bit which is more useful than 0/1 in some
27552 if (CC == X86::COND_B)
27553 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
27555 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27556 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27557 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
27563 // Optimize branch condition evaluation.
27565 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
27566 TargetLowering::DAGCombinerInfo &DCI,
27567 const X86Subtarget *Subtarget) {
27569 SDValue Chain = N->getOperand(0);
27570 SDValue Dest = N->getOperand(1);
27571 SDValue EFLAGS = N->getOperand(3);
27572 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
27574 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
27575 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
27576 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
27583 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
27584 SelectionDAG &DAG) {
27585 // Take advantage of vector comparisons producing 0 or -1 in each lane to
27586 // optimize away operation when it's from a constant.
27588 // The general transformation is:
27589 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
27590 // AND(VECTOR_CMP(x,y), constant2)
27591 // constant2 = UNARYOP(constant)
27593 // Early exit if this isn't a vector operation, the operand of the
27594 // unary operation isn't a bitwise AND, or if the sizes of the operations
27595 // aren't the same.
27596 EVT VT = N->getValueType(0);
27597 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
27598 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
27599 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
27602 // Now check that the other operand of the AND is a constant. We could
27603 // make the transformation for non-constant splats as well, but it's unclear
27604 // that would be a benefit as it would not eliminate any operations, just
27605 // perform one more step in scalar code before moving to the vector unit.
27606 if (BuildVectorSDNode *BV =
27607 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
27608 // Bail out if the vector isn't a constant.
27609 if (!BV->isConstant())
27612 // Everything checks out. Build up the new and improved node.
27614 EVT IntVT = BV->getValueType(0);
27615 // Create a new constant of the appropriate type for the transformed
27617 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
27618 // The AND node needs bitcasts to/from an integer vector type around it.
27619 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
27620 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
27621 N->getOperand(0)->getOperand(0), MaskConst);
27622 SDValue Res = DAG.getBitcast(VT, NewAnd);
27629 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27630 const X86Subtarget *Subtarget) {
27631 SDValue Op0 = N->getOperand(0);
27632 EVT VT = N->getValueType(0);
27633 EVT InVT = Op0.getValueType();
27634 EVT InSVT = InVT.getScalarType();
27635 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
27637 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
27638 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
27639 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27641 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27642 InVT.getVectorNumElements());
27643 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
27645 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
27646 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
27648 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27654 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
27655 const X86Subtarget *Subtarget) {
27656 // First try to optimize away the conversion entirely when it's
27657 // conditionally from a constant. Vectors only.
27658 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
27661 // Now move on to more general possibilities.
27662 SDValue Op0 = N->getOperand(0);
27663 EVT VT = N->getValueType(0);
27664 EVT InVT = Op0.getValueType();
27665 EVT InSVT = InVT.getScalarType();
27667 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
27668 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
27669 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
27671 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
27672 InVT.getVectorNumElements());
27673 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
27674 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
27677 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
27678 // a 32-bit target where SSE doesn't support i64->FP operations.
27679 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
27680 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
27681 EVT LdVT = Ld->getValueType(0);
27683 // This transformation is not supported if the result type is f16
27684 if (VT == MVT::f16)
27687 if (!Ld->isVolatile() && !VT.isVector() &&
27688 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
27689 !Subtarget->is64Bit() && LdVT == MVT::i64) {
27690 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
27691 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
27692 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
27699 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
27700 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
27701 X86TargetLowering::DAGCombinerInfo &DCI) {
27702 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
27703 // the result is either zero or one (depending on the input carry bit).
27704 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
27705 if (X86::isZeroNode(N->getOperand(0)) &&
27706 X86::isZeroNode(N->getOperand(1)) &&
27707 // We don't have a good way to replace an EFLAGS use, so only do this when
27709 SDValue(N, 1).use_empty()) {
27711 EVT VT = N->getValueType(0);
27712 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
27713 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
27714 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
27715 DAG.getConstant(X86::COND_B, DL,
27718 DAG.getConstant(1, DL, VT));
27719 return DCI.CombineTo(N, Res1, CarryOut);
27725 // fold (add Y, (sete X, 0)) -> adc 0, Y
27726 // (add Y, (setne X, 0)) -> sbb -1, Y
27727 // (sub (sete X, 0), Y) -> sbb 0, Y
27728 // (sub (setne X, 0), Y) -> adc -1, Y
27729 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
27732 // Look through ZExts.
27733 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
27734 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
27737 SDValue SetCC = Ext.getOperand(0);
27738 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
27741 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
27742 if (CC != X86::COND_E && CC != X86::COND_NE)
27745 SDValue Cmp = SetCC.getOperand(1);
27746 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
27747 !X86::isZeroNode(Cmp.getOperand(1)) ||
27748 !Cmp.getOperand(0).getValueType().isInteger())
27751 SDValue CmpOp0 = Cmp.getOperand(0);
27752 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
27753 DAG.getConstant(1, DL, CmpOp0.getValueType()));
27755 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
27756 if (CC == X86::COND_NE)
27757 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
27758 DL, OtherVal.getValueType(), OtherVal,
27759 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
27761 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
27762 DL, OtherVal.getValueType(), OtherVal,
27763 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
27766 /// PerformADDCombine - Do target-specific dag combines on integer adds.
27767 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
27768 const X86Subtarget *Subtarget) {
27769 EVT VT = N->getValueType(0);
27770 SDValue Op0 = N->getOperand(0);
27771 SDValue Op1 = N->getOperand(1);
27773 // Try to synthesize horizontal adds from adds of shuffles.
27774 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27775 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27776 isHorizontalBinOp(Op0, Op1, true))
27777 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
27779 return OptimizeConditionalInDecrement(N, DAG);
27782 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
27783 const X86Subtarget *Subtarget) {
27784 SDValue Op0 = N->getOperand(0);
27785 SDValue Op1 = N->getOperand(1);
27787 // X86 can't encode an immediate LHS of a sub. See if we can push the
27788 // negation into a preceding instruction.
27789 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
27790 // If the RHS of the sub is a XOR with one use and a constant, invert the
27791 // immediate. Then add one to the LHS of the sub so we can turn
27792 // X-Y -> X+~Y+1, saving one register.
27793 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
27794 isa<ConstantSDNode>(Op1.getOperand(1))) {
27795 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
27796 EVT VT = Op0.getValueType();
27797 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
27799 DAG.getConstant(~XorC, SDLoc(Op1), VT));
27800 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
27801 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
27805 // Try to synthesize horizontal adds from adds of shuffles.
27806 EVT VT = N->getValueType(0);
27807 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
27808 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
27809 isHorizontalBinOp(Op0, Op1, true))
27810 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
27812 return OptimizeConditionalInDecrement(N, DAG);
27815 /// performVZEXTCombine - Performs build vector combines
27816 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
27817 TargetLowering::DAGCombinerInfo &DCI,
27818 const X86Subtarget *Subtarget) {
27820 MVT VT = N->getSimpleValueType(0);
27821 SDValue Op = N->getOperand(0);
27822 MVT OpVT = Op.getSimpleValueType();
27823 MVT OpEltVT = OpVT.getVectorElementType();
27824 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
27826 // (vzext (bitcast (vzext (x)) -> (vzext x)
27828 while (V.getOpcode() == ISD::BITCAST)
27829 V = V.getOperand(0);
27831 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
27832 MVT InnerVT = V.getSimpleValueType();
27833 MVT InnerEltVT = InnerVT.getVectorElementType();
27835 // If the element sizes match exactly, we can just do one larger vzext. This
27836 // is always an exact type match as vzext operates on integer types.
27837 if (OpEltVT == InnerEltVT) {
27838 assert(OpVT == InnerVT && "Types must match for vzext!");
27839 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
27842 // The only other way we can combine them is if only a single element of the
27843 // inner vzext is used in the input to the outer vzext.
27844 if (InnerEltVT.getSizeInBits() < InputBits)
27847 // In this case, the inner vzext is completely dead because we're going to
27848 // only look at bits inside of the low element. Just do the outer vzext on
27849 // a bitcast of the input to the inner.
27850 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
27853 // Check if we can bypass extracting and re-inserting an element of an input
27854 // vector. Essentially:
27855 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
27856 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
27857 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
27858 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
27859 SDValue ExtractedV = V.getOperand(0);
27860 SDValue OrigV = ExtractedV.getOperand(0);
27861 if (isNullConstant(ExtractedV.getOperand(1))) {
27862 MVT OrigVT = OrigV.getSimpleValueType();
27863 // Extract a subvector if necessary...
27864 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
27865 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
27866 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
27867 OrigVT.getVectorNumElements() / Ratio);
27868 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
27869 DAG.getIntPtrConstant(0, DL));
27871 Op = DAG.getBitcast(OpVT, OrigV);
27872 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
27879 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
27880 DAGCombinerInfo &DCI) const {
27881 SelectionDAG &DAG = DCI.DAG;
27882 switch (N->getOpcode()) {
27884 case ISD::EXTRACT_VECTOR_ELT:
27885 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
27888 case X86ISD::SHRUNKBLEND:
27889 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
27890 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
27891 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
27892 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
27893 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
27894 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
27895 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
27898 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
27899 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
27900 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
27901 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
27902 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
27903 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
27904 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
27905 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
27906 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
27907 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
27908 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
27909 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
27910 case ISD::FNEG: return PerformFNEGCombine(N, DAG, Subtarget);
27911 case ISD::TRUNCATE: return PerformTRUNCATECombine(N, DAG, Subtarget);
27913 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
27915 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
27917 case ISD::FMAXNUM: return performFMinNumFMaxNumCombine(N, DAG,
27919 case X86ISD::FAND: return PerformFANDCombine(N, DAG, Subtarget);
27920 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG, Subtarget);
27921 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
27922 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
27923 case ISD::ANY_EXTEND:
27924 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
27925 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
27926 case ISD::SIGN_EXTEND_INREG:
27927 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
27928 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
27929 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
27930 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
27931 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
27932 case X86ISD::SHUFP: // Handle all target specific shuffles
27933 case X86ISD::PALIGNR:
27934 case X86ISD::BLENDI:
27935 case X86ISD::UNPCKH:
27936 case X86ISD::UNPCKL:
27937 case X86ISD::MOVHLPS:
27938 case X86ISD::MOVLHPS:
27939 case X86ISD::PSHUFB:
27940 case X86ISD::PSHUFD:
27941 case X86ISD::PSHUFHW:
27942 case X86ISD::PSHUFLW:
27943 case X86ISD::MOVSS:
27944 case X86ISD::MOVSD:
27945 case X86ISD::VPERMILPI:
27946 case X86ISD::VPERM2X128:
27947 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
27948 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
27950 case ISD::MSCATTER: return PerformGatherScatterCombine(N, DAG);
27956 /// isTypeDesirableForOp - Return true if the target has native support for
27957 /// the specified value type and it is 'desirable' to use the type for the
27958 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
27959 /// instruction encodings are longer and some i16 instructions are slow.
27960 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
27961 if (!isTypeLegal(VT))
27963 if (VT != MVT::i16)
27970 case ISD::SIGN_EXTEND:
27971 case ISD::ZERO_EXTEND:
27972 case ISD::ANY_EXTEND:
27985 /// This function checks if any of the users of EFLAGS copies the EFLAGS. We
27986 /// know that the code that lowers COPY of EFLAGS has to use the stack, and if
27987 /// we don't adjust the stack we clobber the first frame index.
27988 /// See X86InstrInfo::copyPhysReg.
27989 bool X86TargetLowering::hasCopyImplyingStackAdjustment(
27990 MachineFunction *MF) const {
27991 const MachineRegisterInfo &MRI = MF->getRegInfo();
27993 return any_of(MRI.reg_instructions(X86::EFLAGS),
27994 [](const MachineInstr &RI) { return RI.isCopy(); });
27997 /// IsDesirableToPromoteOp - This method query the target whether it is
27998 /// beneficial for dag combiner to promote the specified node. If true, it
27999 /// should return the desired promotion type by reference.
28000 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
28001 EVT VT = Op.getValueType();
28002 if (VT != MVT::i16)
28005 bool Promote = false;
28006 bool Commute = false;
28007 switch (Op.getOpcode()) {
28010 LoadSDNode *LD = cast<LoadSDNode>(Op);
28011 // If the non-extending load has a single use and it's not live out, then it
28012 // might be folded.
28013 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
28014 Op.hasOneUse()*/) {
28015 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
28016 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
28017 // The only case where we'd want to promote LOAD (rather then it being
28018 // promoted as an operand is when it's only use is liveout.
28019 if (UI->getOpcode() != ISD::CopyToReg)
28026 case ISD::SIGN_EXTEND:
28027 case ISD::ZERO_EXTEND:
28028 case ISD::ANY_EXTEND:
28033 SDValue N0 = Op.getOperand(0);
28034 // Look out for (store (shl (load), x)).
28035 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
28048 SDValue N0 = Op.getOperand(0);
28049 SDValue N1 = Op.getOperand(1);
28050 if (!Commute && MayFoldLoad(N1))
28052 // Avoid disabling potential load folding opportunities.
28053 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
28055 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
28065 //===----------------------------------------------------------------------===//
28066 // X86 Inline Assembly Support
28067 //===----------------------------------------------------------------------===//
28069 // Helper to match a string separated by whitespace.
28070 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
28071 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
28073 for (StringRef Piece : Pieces) {
28074 if (!S.startswith(Piece)) // Check if the piece matches.
28077 S = S.substr(Piece.size());
28078 StringRef::size_type Pos = S.find_first_not_of(" \t");
28079 if (Pos == 0) // We matched a prefix.
28088 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
28090 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
28091 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
28092 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
28093 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
28095 if (AsmPieces.size() == 3)
28097 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
28104 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
28105 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
28107 std::string AsmStr = IA->getAsmString();
28109 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
28110 if (!Ty || Ty->getBitWidth() % 16 != 0)
28113 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
28114 SmallVector<StringRef, 4> AsmPieces;
28115 SplitString(AsmStr, AsmPieces, ";\n");
28117 switch (AsmPieces.size()) {
28118 default: return false;
28120 // FIXME: this should verify that we are targeting a 486 or better. If not,
28121 // we will turn this bswap into something that will be lowered to logical
28122 // ops instead of emitting the bswap asm. For now, we don't support 486 or
28123 // lower so don't worry about this.
28125 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
28126 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
28127 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
28128 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
28129 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
28130 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
28131 // No need to check constraints, nothing other than the equivalent of
28132 // "=r,0" would be valid here.
28133 return IntrinsicLowering::LowerToByteSwap(CI);
28136 // rorw $$8, ${0:w} --> llvm.bswap.i16
28137 if (CI->getType()->isIntegerTy(16) &&
28138 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28139 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
28140 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
28142 StringRef ConstraintsStr = IA->getConstraintString();
28143 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28144 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28145 if (clobbersFlagRegisters(AsmPieces))
28146 return IntrinsicLowering::LowerToByteSwap(CI);
28150 if (CI->getType()->isIntegerTy(32) &&
28151 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
28152 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
28153 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
28154 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
28156 StringRef ConstraintsStr = IA->getConstraintString();
28157 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
28158 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
28159 if (clobbersFlagRegisters(AsmPieces))
28160 return IntrinsicLowering::LowerToByteSwap(CI);
28163 if (CI->getType()->isIntegerTy(64)) {
28164 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
28165 if (Constraints.size() >= 2 &&
28166 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
28167 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
28168 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
28169 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
28170 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
28171 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
28172 return IntrinsicLowering::LowerToByteSwap(CI);
28180 /// getConstraintType - Given a constraint letter, return the type of
28181 /// constraint it is for this target.
28182 X86TargetLowering::ConstraintType
28183 X86TargetLowering::getConstraintType(StringRef Constraint) const {
28184 if (Constraint.size() == 1) {
28185 switch (Constraint[0]) {
28196 return C_RegisterClass;
28220 return TargetLowering::getConstraintType(Constraint);
28223 /// Examine constraint type and operand type and determine a weight value.
28224 /// This object must already have been set up with the operand type
28225 /// and the current alternative constraint selected.
28226 TargetLowering::ConstraintWeight
28227 X86TargetLowering::getSingleConstraintMatchWeight(
28228 AsmOperandInfo &info, const char *constraint) const {
28229 ConstraintWeight weight = CW_Invalid;
28230 Value *CallOperandVal = info.CallOperandVal;
28231 // If we don't have a value, we can't do a match,
28232 // but allow it at the lowest weight.
28233 if (!CallOperandVal)
28235 Type *type = CallOperandVal->getType();
28236 // Look at the constraint type.
28237 switch (*constraint) {
28239 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
28250 if (CallOperandVal->getType()->isIntegerTy())
28251 weight = CW_SpecificReg;
28256 if (type->isFloatingPointTy())
28257 weight = CW_SpecificReg;
28260 if (type->isX86_MMXTy() && Subtarget->hasMMX())
28261 weight = CW_SpecificReg;
28265 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
28266 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
28267 weight = CW_Register;
28270 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
28271 if (C->getZExtValue() <= 31)
28272 weight = CW_Constant;
28276 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28277 if (C->getZExtValue() <= 63)
28278 weight = CW_Constant;
28282 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28283 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
28284 weight = CW_Constant;
28288 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28289 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
28290 weight = CW_Constant;
28294 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28295 if (C->getZExtValue() <= 3)
28296 weight = CW_Constant;
28300 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28301 if (C->getZExtValue() <= 0xff)
28302 weight = CW_Constant;
28307 if (isa<ConstantFP>(CallOperandVal)) {
28308 weight = CW_Constant;
28312 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28313 if ((C->getSExtValue() >= -0x80000000LL) &&
28314 (C->getSExtValue() <= 0x7fffffffLL))
28315 weight = CW_Constant;
28319 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
28320 if (C->getZExtValue() <= 0xffffffff)
28321 weight = CW_Constant;
28328 /// LowerXConstraint - try to replace an X constraint, which matches anything,
28329 /// with another that has more specific requirements based on the type of the
28330 /// corresponding operand.
28331 const char *X86TargetLowering::
28332 LowerXConstraint(EVT ConstraintVT) const {
28333 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
28334 // 'f' like normal targets.
28335 if (ConstraintVT.isFloatingPoint()) {
28336 if (Subtarget->hasSSE2())
28338 if (Subtarget->hasSSE1())
28342 return TargetLowering::LowerXConstraint(ConstraintVT);
28345 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
28346 /// vector. If it is invalid, don't add anything to Ops.
28347 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
28348 std::string &Constraint,
28349 std::vector<SDValue>&Ops,
28350 SelectionDAG &DAG) const {
28353 // Only support length 1 constraints for now.
28354 if (Constraint.length() > 1) return;
28356 char ConstraintLetter = Constraint[0];
28357 switch (ConstraintLetter) {
28360 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28361 if (C->getZExtValue() <= 31) {
28362 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28363 Op.getValueType());
28369 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28370 if (C->getZExtValue() <= 63) {
28371 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28372 Op.getValueType());
28378 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28379 if (isInt<8>(C->getSExtValue())) {
28380 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28381 Op.getValueType());
28387 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28388 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
28389 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
28390 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
28391 Op.getValueType());
28397 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28398 if (C->getZExtValue() <= 3) {
28399 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28400 Op.getValueType());
28406 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28407 if (C->getZExtValue() <= 255) {
28408 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28409 Op.getValueType());
28415 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28416 if (C->getZExtValue() <= 127) {
28417 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28418 Op.getValueType());
28424 // 32-bit signed value
28425 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28426 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28427 C->getSExtValue())) {
28428 // Widen to 64 bits here to get it sign extended.
28429 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
28432 // FIXME gcc accepts some relocatable values here too, but only in certain
28433 // memory models; it's complicated.
28438 // 32-bit unsigned value
28439 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
28440 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
28441 C->getZExtValue())) {
28442 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
28443 Op.getValueType());
28447 // FIXME gcc accepts some relocatable values here too, but only in certain
28448 // memory models; it's complicated.
28452 // Literal immediates are always ok.
28453 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
28454 // Widen to 64 bits here to get it sign extended.
28455 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
28459 // In any sort of PIC mode addresses need to be computed at runtime by
28460 // adding in a register or some sort of table lookup. These can't
28461 // be used as immediates.
28462 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
28465 // If we are in non-pic codegen mode, we allow the address of a global (with
28466 // an optional displacement) to be used with 'i'.
28467 GlobalAddressSDNode *GA = nullptr;
28468 int64_t Offset = 0;
28470 // Match either (GA), (GA+C), (GA+C1+C2), etc.
28472 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
28473 Offset += GA->getOffset();
28475 } else if (Op.getOpcode() == ISD::ADD) {
28476 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28477 Offset += C->getZExtValue();
28478 Op = Op.getOperand(0);
28481 } else if (Op.getOpcode() == ISD::SUB) {
28482 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
28483 Offset += -C->getZExtValue();
28484 Op = Op.getOperand(0);
28489 // Otherwise, this isn't something we can handle, reject it.
28493 const GlobalValue *GV = GA->getGlobal();
28494 // If we require an extra load to get this address, as in PIC mode, we
28495 // can't accept it.
28496 if (isGlobalStubReference(
28497 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
28500 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
28501 GA->getValueType(0), Offset);
28506 if (Result.getNode()) {
28507 Ops.push_back(Result);
28510 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
28513 std::pair<unsigned, const TargetRegisterClass *>
28514 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
28515 StringRef Constraint,
28517 // First, see if this is a constraint that directly corresponds to an LLVM
28519 if (Constraint.size() == 1) {
28520 // GCC Constraint Letters
28521 switch (Constraint[0]) {
28523 // TODO: Slight differences here in allocation order and leaving
28524 // RIP in the class. Do they matter any more here than they do
28525 // in the normal allocation?
28526 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
28527 if (Subtarget->is64Bit()) {
28528 if (VT == MVT::i32 || VT == MVT::f32)
28529 return std::make_pair(0U, &X86::GR32RegClass);
28530 if (VT == MVT::i16)
28531 return std::make_pair(0U, &X86::GR16RegClass);
28532 if (VT == MVT::i8 || VT == MVT::i1)
28533 return std::make_pair(0U, &X86::GR8RegClass);
28534 if (VT == MVT::i64 || VT == MVT::f64)
28535 return std::make_pair(0U, &X86::GR64RegClass);
28538 // 32-bit fallthrough
28539 case 'Q': // Q_REGS
28540 if (VT == MVT::i32 || VT == MVT::f32)
28541 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
28542 if (VT == MVT::i16)
28543 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
28544 if (VT == MVT::i8 || VT == MVT::i1)
28545 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
28546 if (VT == MVT::i64)
28547 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
28549 case 'r': // GENERAL_REGS
28550 case 'l': // INDEX_REGS
28551 if (VT == MVT::i8 || VT == MVT::i1)
28552 return std::make_pair(0U, &X86::GR8RegClass);
28553 if (VT == MVT::i16)
28554 return std::make_pair(0U, &X86::GR16RegClass);
28555 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
28556 return std::make_pair(0U, &X86::GR32RegClass);
28557 return std::make_pair(0U, &X86::GR64RegClass);
28558 case 'R': // LEGACY_REGS
28559 if (VT == MVT::i8 || VT == MVT::i1)
28560 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
28561 if (VT == MVT::i16)
28562 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
28563 if (VT == MVT::i32 || !Subtarget->is64Bit())
28564 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
28565 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
28566 case 'f': // FP Stack registers.
28567 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
28568 // value to the correct fpstack register class.
28569 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
28570 return std::make_pair(0U, &X86::RFP32RegClass);
28571 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
28572 return std::make_pair(0U, &X86::RFP64RegClass);
28573 return std::make_pair(0U, &X86::RFP80RegClass);
28574 case 'y': // MMX_REGS if MMX allowed.
28575 if (!Subtarget->hasMMX()) break;
28576 return std::make_pair(0U, &X86::VR64RegClass);
28577 case 'Y': // SSE_REGS if SSE2 allowed
28578 if (!Subtarget->hasSSE2()) break;
28580 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
28581 if (!Subtarget->hasSSE1()) break;
28583 switch (VT.SimpleTy) {
28585 // Scalar SSE types.
28588 return std::make_pair(0U, &X86::FR32RegClass);
28591 return std::make_pair(0U, &X86::FR64RegClass);
28592 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28600 return std::make_pair(0U, &X86::VR128RegClass);
28608 return std::make_pair(0U, &X86::VR256RegClass);
28613 return std::make_pair(0U, &X86::VR512RegClass);
28619 // Use the default implementation in TargetLowering to convert the register
28620 // constraint into a member of a register class.
28621 std::pair<unsigned, const TargetRegisterClass*> Res;
28622 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
28624 // Not found as a standard register?
28626 // Map st(0) -> st(7) -> ST0
28627 if (Constraint.size() == 7 && Constraint[0] == '{' &&
28628 tolower(Constraint[1]) == 's' &&
28629 tolower(Constraint[2]) == 't' &&
28630 Constraint[3] == '(' &&
28631 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
28632 Constraint[5] == ')' &&
28633 Constraint[6] == '}') {
28635 Res.first = X86::FP0+Constraint[4]-'0';
28636 Res.second = &X86::RFP80RegClass;
28640 // GCC allows "st(0)" to be called just plain "st".
28641 if (StringRef("{st}").equals_lower(Constraint)) {
28642 Res.first = X86::FP0;
28643 Res.second = &X86::RFP80RegClass;
28648 if (StringRef("{flags}").equals_lower(Constraint)) {
28649 Res.first = X86::EFLAGS;
28650 Res.second = &X86::CCRRegClass;
28654 // 'A' means EAX + EDX.
28655 if (Constraint == "A") {
28656 Res.first = X86::EAX;
28657 Res.second = &X86::GR32_ADRegClass;
28663 // Otherwise, check to see if this is a register class of the wrong value
28664 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
28665 // turn into {ax},{dx}.
28666 // MVT::Other is used to specify clobber names.
28667 if (Res.second->hasType(VT) || VT == MVT::Other)
28668 return Res; // Correct type already, nothing to do.
28670 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
28671 // return "eax". This should even work for things like getting 64bit integer
28672 // registers when given an f64 type.
28673 const TargetRegisterClass *Class = Res.second;
28674 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
28675 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
28676 unsigned Size = VT.getSizeInBits();
28677 if (Size == 1) Size = 8;
28678 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, Size);
28680 Res.first = DestReg;
28681 Res.second = Size == 8 ? &X86::GR8RegClass
28682 : Size == 16 ? &X86::GR16RegClass
28683 : Size == 32 ? &X86::GR32RegClass
28684 : &X86::GR64RegClass;
28685 assert(Res.second->contains(Res.first) && "Register in register class");
28687 // No register found/type mismatch.
28689 Res.second = nullptr;
28691 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
28692 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
28693 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
28694 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
28695 Class == &X86::VR512RegClass) {
28696 // Handle references to XMM physical registers that got mapped into the
28697 // wrong class. This can happen with constraints like {xmm0} where the
28698 // target independent register mapper will just pick the first match it can
28699 // find, ignoring the required type.
28701 // TODO: Handle f128 and i128 in FR128RegClass after it is tested well.
28702 if (VT == MVT::f32 || VT == MVT::i32)
28703 Res.second = &X86::FR32RegClass;
28704 else if (VT == MVT::f64 || VT == MVT::i64)
28705 Res.second = &X86::FR64RegClass;
28706 else if (X86::VR128RegClass.hasType(VT))
28707 Res.second = &X86::VR128RegClass;
28708 else if (X86::VR256RegClass.hasType(VT))
28709 Res.second = &X86::VR256RegClass;
28710 else if (X86::VR512RegClass.hasType(VT))
28711 Res.second = &X86::VR512RegClass;
28713 // Type mismatch and not a clobber: Return an error;
28715 Res.second = nullptr;
28722 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
28723 const AddrMode &AM, Type *Ty,
28724 unsigned AS) const {
28725 // Scaling factors are not free at all.
28726 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
28727 // will take 2 allocations in the out of order engine instead of 1
28728 // for plain addressing mode, i.e. inst (reg1).
28730 // vaddps (%rsi,%drx), %ymm0, %ymm1
28731 // Requires two allocations (one for the load, one for the computation)
28733 // vaddps (%rsi), %ymm0, %ymm1
28734 // Requires just 1 allocation, i.e., freeing allocations for other operations
28735 // and having less micro operations to execute.
28737 // For some X86 architectures, this is even worse because for instance for
28738 // stores, the complex addressing mode forces the instruction to use the
28739 // "load" ports instead of the dedicated "store" port.
28740 // E.g., on Haswell:
28741 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
28742 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
28743 if (isLegalAddressingMode(DL, AM, Ty, AS))
28744 // Scale represents reg2 * scale, thus account for 1
28745 // as soon as we use a second register.
28746 return AM.Scale != 0;
28750 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
28751 // Integer division on x86 is expensive. However, when aggressively optimizing
28752 // for code size, we prefer to use a div instruction, as it is usually smaller
28753 // than the alternative sequence.
28754 // The exception to this is vector division. Since x86 doesn't have vector
28755 // integer division, leaving the division as-is is a loss even in terms of
28756 // size, because it will have to be scalarized, while the alternative code
28757 // sequence can be performed in vector form.
28758 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
28759 Attribute::MinSize);
28760 return OptSize && !VT.isVector();