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 "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/Analysis/LibCallSemantics.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/WinEHFuncInfo.h"
37 #include "llvm/IR/CallSite.h"
38 #include "llvm/IR/CallingConv.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DerivedTypes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/GlobalAlias.h"
43 #include "llvm/IR/GlobalVariable.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/Intrinsics.h"
46 #include "llvm/MC/MCAsmInfo.h"
47 #include "llvm/MC/MCContext.h"
48 #include "llvm/MC/MCExpr.h"
49 #include "llvm/MC/MCSymbol.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Target/TargetOptions.h"
55 #include "X86IntrinsicsInfo.h"
61 #define DEBUG_TYPE "x86-isel"
63 STATISTIC(NumTailCalls, "Number of tail calls");
65 static cl::opt<bool> ExperimentalVectorWideningLegalization(
66 "x86-experimental-vector-widening-legalization", cl::init(false),
67 cl::desc("Enable an experimental vector type legalization through widening "
68 "rather than promotion."),
71 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
72 const X86Subtarget &STI)
73 : TargetLowering(TM), Subtarget(&STI) {
74 X86ScalarSSEf64 = Subtarget->hasSSE2();
75 X86ScalarSSEf32 = Subtarget->hasSSE1();
76 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
78 // Set up the TargetLowering object.
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
119 if (Subtarget->isTargetDarwin()) {
120 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(false);
122 setUseUnderscoreLongJmp(false);
123 } else if (Subtarget->isTargetWindowsGNU()) {
124 // MS runtime is weird: it exports _setjmp, but longjmp!
125 setUseUnderscoreSetJmp(true);
126 setUseUnderscoreLongJmp(false);
128 setUseUnderscoreSetJmp(true);
129 setUseUnderscoreLongJmp(true);
132 // Set up the register classes.
133 addRegisterClass(MVT::i8, &X86::GR8RegClass);
134 addRegisterClass(MVT::i16, &X86::GR16RegClass);
135 addRegisterClass(MVT::i32, &X86::GR32RegClass);
136 if (Subtarget->is64Bit())
137 addRegisterClass(MVT::i64, &X86::GR64RegClass);
139 for (MVT VT : MVT::integer_valuetypes())
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
142 // We don't accept any truncstore of integer registers.
143 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
146 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
147 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
148 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
150 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
152 // SETOEQ and SETUNE require checking two conditions.
153 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
160 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
162 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
166 if (Subtarget->is64Bit()) {
167 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
168 // f32/f64 are legal, f80 is custom.
169 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 } else if (!Subtarget->useSoftFloat()) {
174 // We have an algorithm for SSE2->double, and we turn this into a
175 // 64-bit FILD followed by conditional FADD for other targets.
176 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
177 // We have an algorithm for SSE2, and we turn this into a 64-bit
178 // FILD or VCVTUSI2SS/SD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
182 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
184 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
187 if (!Subtarget->useSoftFloat()) {
188 // SSE has no i16 to fp conversion, only i32
189 if (X86ScalarSSEf32) {
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
191 // f32 and f64 cases are Legal, f80 case is not
192 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
202 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
204 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
205 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
207 if (!Subtarget->useSoftFloat()) {
208 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
209 // are Legal, f80 is custom lowered.
210 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
211 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
213 if (X86ScalarSSEf32) {
214 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
215 // f32 and f64 cases are Legal, f80 case is not
216 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
218 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
219 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
222 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
223 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Expand);
224 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Expand);
227 // Handle FP_TO_UINT by promoting the destination to a larger signed
229 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
233 if (Subtarget->is64Bit()) {
234 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
235 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
236 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
237 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
239 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
240 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
242 } else if (!Subtarget->useSoftFloat()) {
243 // Since AVX is a superset of SSE3, only check for SSE here.
244 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
245 // Expand FP_TO_UINT into a select.
246 // FIXME: We would like to use a Custom expander here eventually to do
247 // the optimal thing for SSE vs. the default expansion in the legalizer.
248 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
250 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
251 // With SSE3 we can use fisttpll to convert to a signed i64; without
252 // SSE, we're stuck with a fistpll.
253 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
255 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
258 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
259 if (!X86ScalarSSEf64) {
260 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
261 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
262 if (Subtarget->is64Bit()) {
263 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
264 // Without SSE, i64->f64 goes through memory.
265 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
269 // Scalar integer divide and remainder are lowered to use operations that
270 // produce two results, to match the available instructions. This exposes
271 // the two-result form to trivial CSE, which is able to combine x/y and x%y
272 // into a single instruction.
274 // Scalar integer multiply-high is also lowered to use two-result
275 // operations, to match the available instructions. However, plain multiply
276 // (low) operations are left as Legal, as there are single-result
277 // instructions for this in x86. Using the two-result multiply instructions
278 // when both high and low results are needed must be arranged by dagcombine.
279 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
280 setOperationAction(ISD::MULHS, VT, Expand);
281 setOperationAction(ISD::MULHU, VT, Expand);
282 setOperationAction(ISD::SDIV, VT, Expand);
283 setOperationAction(ISD::UDIV, VT, Expand);
284 setOperationAction(ISD::SREM, VT, Expand);
285 setOperationAction(ISD::UREM, VT, Expand);
287 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
288 setOperationAction(ISD::ADDC, VT, Custom);
289 setOperationAction(ISD::ADDE, VT, Custom);
290 setOperationAction(ISD::SUBC, VT, Custom);
291 setOperationAction(ISD::SUBE, VT, Custom);
294 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
295 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
296 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
298 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
299 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
300 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
301 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
302 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
303 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
305 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
306 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
307 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
308 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
309 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
310 if (Subtarget->is64Bit())
311 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
312 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
313 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
314 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
315 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
317 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
318 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
319 // is. We should promote the value to 64-bits to solve this.
320 // This is what the CRT headers do - `fmodf` is an inline header
321 // function casting to f64 and calling `fmod`.
322 setOperationAction(ISD::FREM , MVT::f32 , Promote);
324 setOperationAction(ISD::FREM , MVT::f32 , Expand);
327 setOperationAction(ISD::FREM , MVT::f64 , Expand);
328 setOperationAction(ISD::FREM , MVT::f80 , Expand);
329 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
331 // Promote the i8 variants and force them on up to i32 which has a shorter
333 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
334 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
335 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
336 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
337 if (Subtarget->hasBMI()) {
338 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
339 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
340 if (Subtarget->is64Bit())
341 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
343 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
344 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
345 if (Subtarget->is64Bit())
346 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
349 if (Subtarget->hasLZCNT()) {
350 // When promoting the i8 variants, force them to i32 for a shorter
352 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
353 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
355 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
358 if (Subtarget->is64Bit())
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
361 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
362 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
363 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
364 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
365 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
366 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
367 if (Subtarget->is64Bit()) {
368 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
369 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
373 // Special handling for half-precision floating point conversions.
374 // If we don't have F16C support, then lower half float conversions
375 // into library calls.
376 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
377 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
378 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
381 // There's never any support for operations beyond MVT::f32.
382 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
383 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
384 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
385 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
387 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
388 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
389 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
390 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
391 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
392 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
394 if (Subtarget->hasPOPCNT()) {
395 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
397 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
398 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
399 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
400 if (Subtarget->is64Bit())
401 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
404 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
406 if (!Subtarget->hasMOVBE())
407 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
409 // These should be promoted to a larger select which is supported.
410 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
411 // X86 wants to expand cmov itself.
412 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
413 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
414 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
415 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
416 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
417 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
418 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
420 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
421 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
422 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
423 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
424 if (Subtarget->is64Bit()) {
425 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
426 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
428 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
429 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
430 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
431 // support continuation, user-level threading, and etc.. As a result, no
432 // other SjLj exception interfaces are implemented and please don't build
433 // your own exception handling based on them.
434 // LLVM/Clang supports zero-cost DWARF exception handling.
435 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
436 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
439 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
440 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
441 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
442 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
443 if (Subtarget->is64Bit())
444 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
445 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
446 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
447 if (Subtarget->is64Bit()) {
448 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
449 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
450 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
451 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
452 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
454 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
455 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
456 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
457 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
458 if (Subtarget->is64Bit()) {
459 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
460 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
461 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
464 if (Subtarget->hasSSE1())
465 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
467 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
469 // Expand certain atomics
470 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
471 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
472 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
473 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
476 if (Subtarget->hasCmpxchg16b()) {
477 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
480 // FIXME - use subtarget debug flags
481 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
482 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
483 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
486 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
487 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
489 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
490 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
492 setOperationAction(ISD::TRAP, MVT::Other, Legal);
493 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
495 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
496 setOperationAction(ISD::VASTART , MVT::Other, Custom);
497 setOperationAction(ISD::VAEND , MVT::Other, Expand);
498 if (Subtarget->is64Bit()) {
499 setOperationAction(ISD::VAARG , MVT::Other, Custom);
500 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
502 // TargetInfo::CharPtrBuiltinVaList
503 setOperationAction(ISD::VAARG , MVT::Other, Expand);
504 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
507 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
508 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
510 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
512 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
513 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
514 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
516 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
517 // f32 and f64 use SSE.
518 // Set up the FP register classes.
519 addRegisterClass(MVT::f32, &X86::FR32RegClass);
520 addRegisterClass(MVT::f64, &X86::FR64RegClass);
522 // Use ANDPD to simulate FABS.
523 setOperationAction(ISD::FABS , MVT::f64, Custom);
524 setOperationAction(ISD::FABS , MVT::f32, Custom);
526 // Use XORP to simulate FNEG.
527 setOperationAction(ISD::FNEG , MVT::f64, Custom);
528 setOperationAction(ISD::FNEG , MVT::f32, Custom);
530 // Use ANDPD and ORPD to simulate FCOPYSIGN.
531 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
532 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
534 // Lower this to FGETSIGNx86 plus an AND.
535 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
536 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
538 // We don't support sin/cos/fmod
539 setOperationAction(ISD::FSIN , MVT::f64, Expand);
540 setOperationAction(ISD::FCOS , MVT::f64, Expand);
541 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
542 setOperationAction(ISD::FSIN , MVT::f32, Expand);
543 setOperationAction(ISD::FCOS , MVT::f32, Expand);
544 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
546 // Expand FP immediates into loads from the stack, except for the special
548 addLegalFPImmediate(APFloat(+0.0)); // xorpd
549 addLegalFPImmediate(APFloat(+0.0f)); // xorps
550 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
551 // Use SSE for f32, x87 for f64.
552 // Set up the FP register classes.
553 addRegisterClass(MVT::f32, &X86::FR32RegClass);
554 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
556 // Use ANDPS to simulate FABS.
557 setOperationAction(ISD::FABS , MVT::f32, Custom);
559 // Use XORP to simulate FNEG.
560 setOperationAction(ISD::FNEG , MVT::f32, Custom);
562 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
564 // Use ANDPS and ORPS to simulate FCOPYSIGN.
565 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
566 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
568 // We don't support sin/cos/fmod
569 setOperationAction(ISD::FSIN , MVT::f32, Expand);
570 setOperationAction(ISD::FCOS , MVT::f32, Expand);
571 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
573 // Special cases we handle for FP constants.
574 addLegalFPImmediate(APFloat(+0.0f)); // xorps
575 addLegalFPImmediate(APFloat(+0.0)); // FLD0
576 addLegalFPImmediate(APFloat(+1.0)); // FLD1
577 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
578 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
580 if (!TM.Options.UnsafeFPMath) {
581 setOperationAction(ISD::FSIN , MVT::f64, Expand);
582 setOperationAction(ISD::FCOS , MVT::f64, Expand);
583 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
585 } else if (!Subtarget->useSoftFloat()) {
586 // f32 and f64 in x87.
587 // Set up the FP register classes.
588 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
589 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
591 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
592 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
593 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
594 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
596 if (!TM.Options.UnsafeFPMath) {
597 setOperationAction(ISD::FSIN , MVT::f64, Expand);
598 setOperationAction(ISD::FSIN , MVT::f32, Expand);
599 setOperationAction(ISD::FCOS , MVT::f64, Expand);
600 setOperationAction(ISD::FCOS , MVT::f32, Expand);
601 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
602 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
604 addLegalFPImmediate(APFloat(+0.0)); // FLD0
605 addLegalFPImmediate(APFloat(+1.0)); // FLD1
606 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
607 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
608 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
614 // We don't support FMA.
615 setOperationAction(ISD::FMA, MVT::f64, Expand);
616 setOperationAction(ISD::FMA, MVT::f32, Expand);
618 // Long double always uses X87.
619 if (!Subtarget->useSoftFloat()) {
620 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
621 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
622 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
624 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
625 addLegalFPImmediate(TmpFlt); // FLD0
627 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
630 APFloat TmpFlt2(+1.0);
631 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
633 addLegalFPImmediate(TmpFlt2); // FLD1
634 TmpFlt2.changeSign();
635 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
638 if (!TM.Options.UnsafeFPMath) {
639 setOperationAction(ISD::FSIN , MVT::f80, Expand);
640 setOperationAction(ISD::FCOS , MVT::f80, Expand);
641 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
644 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
645 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
646 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
647 setOperationAction(ISD::FRINT, MVT::f80, Expand);
648 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
649 setOperationAction(ISD::FMA, MVT::f80, Expand);
652 // Always use a library call for pow.
653 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
654 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
655 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
657 setOperationAction(ISD::FLOG, MVT::f80, Expand);
658 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
659 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
660 setOperationAction(ISD::FEXP, MVT::f80, Expand);
661 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
662 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
663 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
665 // First set operation action for all vector types to either promote
666 // (for widening) or expand (for scalarization). Then we will selectively
667 // turn on ones that can be effectively codegen'd.
668 for (MVT VT : MVT::vector_valuetypes()) {
669 setOperationAction(ISD::ADD , VT, Expand);
670 setOperationAction(ISD::SUB , VT, Expand);
671 setOperationAction(ISD::FADD, VT, Expand);
672 setOperationAction(ISD::FNEG, VT, Expand);
673 setOperationAction(ISD::FSUB, VT, Expand);
674 setOperationAction(ISD::MUL , VT, Expand);
675 setOperationAction(ISD::FMUL, VT, Expand);
676 setOperationAction(ISD::SDIV, VT, Expand);
677 setOperationAction(ISD::UDIV, VT, Expand);
678 setOperationAction(ISD::FDIV, VT, Expand);
679 setOperationAction(ISD::SREM, VT, Expand);
680 setOperationAction(ISD::UREM, VT, Expand);
681 setOperationAction(ISD::LOAD, VT, Expand);
682 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
683 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
685 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
686 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
687 setOperationAction(ISD::FABS, VT, Expand);
688 setOperationAction(ISD::FSIN, VT, Expand);
689 setOperationAction(ISD::FSINCOS, VT, Expand);
690 setOperationAction(ISD::FCOS, VT, Expand);
691 setOperationAction(ISD::FSINCOS, VT, Expand);
692 setOperationAction(ISD::FREM, VT, Expand);
693 setOperationAction(ISD::FMA, VT, Expand);
694 setOperationAction(ISD::FPOWI, VT, Expand);
695 setOperationAction(ISD::FSQRT, VT, Expand);
696 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
697 setOperationAction(ISD::FFLOOR, VT, Expand);
698 setOperationAction(ISD::FCEIL, VT, Expand);
699 setOperationAction(ISD::FTRUNC, VT, Expand);
700 setOperationAction(ISD::FRINT, VT, Expand);
701 setOperationAction(ISD::FNEARBYINT, VT, Expand);
702 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
703 setOperationAction(ISD::MULHS, VT, Expand);
704 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
705 setOperationAction(ISD::MULHU, VT, Expand);
706 setOperationAction(ISD::SDIVREM, VT, Expand);
707 setOperationAction(ISD::UDIVREM, VT, Expand);
708 setOperationAction(ISD::FPOW, VT, Expand);
709 setOperationAction(ISD::CTPOP, VT, Expand);
710 setOperationAction(ISD::CTTZ, VT, Expand);
711 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
712 setOperationAction(ISD::CTLZ, VT, Expand);
713 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
714 setOperationAction(ISD::SHL, VT, Expand);
715 setOperationAction(ISD::SRA, VT, Expand);
716 setOperationAction(ISD::SRL, VT, Expand);
717 setOperationAction(ISD::ROTL, VT, Expand);
718 setOperationAction(ISD::ROTR, VT, Expand);
719 setOperationAction(ISD::BSWAP, VT, Expand);
720 setOperationAction(ISD::SETCC, VT, Expand);
721 setOperationAction(ISD::FLOG, VT, Expand);
722 setOperationAction(ISD::FLOG2, VT, Expand);
723 setOperationAction(ISD::FLOG10, VT, Expand);
724 setOperationAction(ISD::FEXP, VT, Expand);
725 setOperationAction(ISD::FEXP2, VT, Expand);
726 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
727 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
728 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
729 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
730 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
731 setOperationAction(ISD::TRUNCATE, VT, Expand);
732 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
733 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
734 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
735 setOperationAction(ISD::VSELECT, VT, Expand);
736 setOperationAction(ISD::SELECT_CC, VT, Expand);
737 for (MVT InnerVT : MVT::vector_valuetypes()) {
738 setTruncStoreAction(InnerVT, VT, Expand);
740 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
741 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
743 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
744 // types, we have to deal with them whether we ask for Expansion or not.
745 // Setting Expand causes its own optimisation problems though, so leave
747 if (VT.getVectorElementType() == MVT::i1)
748 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
750 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
751 // split/scalarized right now.
752 if (VT.getVectorElementType() == MVT::f16)
753 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
757 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
758 // with -msoft-float, disable use of MMX as well.
759 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
760 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
761 // No operations on x86mmx supported, everything uses intrinsics.
764 // MMX-sized vectors (other than x86mmx) are expected to be expanded
765 // into smaller operations.
766 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
767 setOperationAction(ISD::MULHS, MMXTy, Expand);
768 setOperationAction(ISD::AND, MMXTy, Expand);
769 setOperationAction(ISD::OR, MMXTy, Expand);
770 setOperationAction(ISD::XOR, MMXTy, Expand);
771 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
772 setOperationAction(ISD::SELECT, MMXTy, Expand);
773 setOperationAction(ISD::BITCAST, MMXTy, Expand);
775 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
777 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
778 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
780 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
781 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
782 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
783 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
784 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
785 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
786 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
787 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
788 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
789 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
790 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
792 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
793 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
796 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
797 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
799 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
800 // registers cannot be used even for integer operations.
801 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
802 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
803 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
804 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
806 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
807 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
808 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
809 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
810 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
811 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
812 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
813 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
814 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
815 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
816 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
817 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
818 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
819 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
820 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
821 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
822 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
823 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
824 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
825 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
826 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
827 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
828 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
830 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
831 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
832 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
833 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
835 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
836 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
837 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
838 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
840 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
841 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
846 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
847 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
848 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
849 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
851 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
852 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
853 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
854 // ISD::CTTZ v2i64 - scalarization is faster.
855 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
856 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
857 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
858 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
860 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
861 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
862 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
863 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
864 setOperationAction(ISD::VSELECT, VT, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
868 // We support custom legalizing of sext and anyext loads for specific
869 // memory vector types which we can load as a scalar (or sequence of
870 // scalars) and extend in-register to a legal 128-bit vector type. For sext
871 // loads these must work with a single scalar load.
872 for (MVT VT : MVT::integer_vector_valuetypes()) {
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
874 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
875 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
881 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
888 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
889 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
893 if (Subtarget->is64Bit()) {
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
895 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
898 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
899 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
900 setOperationAction(ISD::AND, VT, Promote);
901 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
902 setOperationAction(ISD::OR, VT, Promote);
903 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
904 setOperationAction(ISD::XOR, VT, Promote);
905 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
906 setOperationAction(ISD::LOAD, VT, Promote);
907 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
908 setOperationAction(ISD::SELECT, VT, Promote);
909 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
912 // Custom lower v2i64 and v2f64 selects.
913 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
914 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
915 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
916 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
918 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
919 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
921 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
923 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
924 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
925 // As there is no 64-bit GPR available, we need build a special custom
926 // sequence to convert from v2i32 to v2f32.
927 if (!Subtarget->is64Bit())
928 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
930 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
931 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
933 for (MVT VT : MVT::fp_vector_valuetypes())
934 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
936 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
937 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
938 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
941 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
942 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
943 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
944 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
945 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
946 setOperationAction(ISD::FRINT, RoundedTy, Legal);
947 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
950 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
951 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
952 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
953 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
954 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
955 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
956 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
957 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
959 // FIXME: Do we need to handle scalar-to-vector here?
960 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
962 // We directly match byte blends in the backend as they match the VSELECT
964 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
966 // SSE41 brings specific instructions for doing vector sign extend even in
967 // cases where we don't have SRA.
968 for (MVT VT : MVT::integer_vector_valuetypes()) {
969 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
970 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
971 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
974 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
976 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
977 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
978 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
979 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
980 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
983 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
984 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
985 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
986 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
987 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
989 // i8 and i16 vectors are custom because the source register and source
990 // source memory operand types are not the same width. f32 vectors are
991 // custom since the immediate controlling the insert encodes additional
993 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
994 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
995 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
996 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
998 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1000 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1001 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1003 // FIXME: these should be Legal, but that's only for the case where
1004 // the index is constant. For now custom expand to deal with that.
1005 if (Subtarget->is64Bit()) {
1006 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1007 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1011 if (Subtarget->hasSSE2()) {
1012 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1013 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1014 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1016 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1017 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1019 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1020 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1022 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1023 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1025 // In the customized shift lowering, the legal cases in AVX2 will be
1027 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1028 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1030 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1031 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1033 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1034 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1037 if (Subtarget->hasXOP()) {
1038 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1039 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1040 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1041 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1042 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1043 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1044 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1045 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1048 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1049 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1050 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1051 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1052 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1053 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1054 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1056 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1057 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1058 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1060 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1061 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1062 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1063 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1064 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1065 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1066 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1067 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1068 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1069 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1070 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1071 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1073 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1074 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1075 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1076 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1077 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1078 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1079 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1080 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1081 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1082 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1083 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1084 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1086 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1087 // even though v8i16 is a legal type.
1088 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1089 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1090 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1092 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1093 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1094 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1096 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1097 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1099 for (MVT VT : MVT::fp_vector_valuetypes())
1100 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1102 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1103 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1105 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1106 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1108 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1109 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1111 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1112 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1113 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1114 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1116 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1117 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1118 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1120 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1121 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1122 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1123 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1124 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1125 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1126 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1127 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1128 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1129 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1130 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1131 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1133 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1134 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1135 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1136 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1138 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1139 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1140 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1141 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1142 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1143 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1144 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1145 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1147 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1148 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1149 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1150 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1151 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1152 setOperationAction(ISD::FMA, MVT::f32, Legal);
1153 setOperationAction(ISD::FMA, MVT::f64, Legal);
1156 if (Subtarget->hasInt256()) {
1157 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1158 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1159 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1160 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1162 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1163 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1164 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1165 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1167 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1168 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1169 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1170 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1172 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1173 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1174 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1175 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1177 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1178 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1179 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1180 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1181 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1182 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1183 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1184 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1185 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1186 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1187 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1188 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1190 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1191 // when we have a 256bit-wide blend with immediate.
1192 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1194 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1195 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1196 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1197 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1198 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1199 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1202 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1203 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1204 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1205 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1206 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1207 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1209 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1210 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1211 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1212 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1214 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1215 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1216 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1217 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1219 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1220 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1221 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1222 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1224 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1225 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1226 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1227 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1228 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1229 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1230 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1231 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1232 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1233 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1234 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1235 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1238 // In the customized shift lowering, the legal cases in AVX2 will be
1240 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1241 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1243 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1244 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1246 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1247 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1249 // Custom lower several nodes for 256-bit types.
1250 for (MVT VT : MVT::vector_valuetypes()) {
1251 if (VT.getScalarSizeInBits() >= 32) {
1252 setOperationAction(ISD::MLOAD, VT, Legal);
1253 setOperationAction(ISD::MSTORE, VT, Legal);
1255 // Extract subvector is special because the value type
1256 // (result) is 128-bit but the source is 256-bit wide.
1257 if (VT.is128BitVector()) {
1258 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1260 // Do not attempt to custom lower other non-256-bit vectors
1261 if (!VT.is256BitVector())
1264 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1265 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1266 setOperationAction(ISD::VSELECT, VT, Custom);
1267 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1268 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1269 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1270 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1271 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1274 if (Subtarget->hasInt256())
1275 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1277 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1278 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1279 setOperationAction(ISD::AND, VT, Promote);
1280 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1281 setOperationAction(ISD::OR, VT, Promote);
1282 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1283 setOperationAction(ISD::XOR, VT, Promote);
1284 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1285 setOperationAction(ISD::LOAD, VT, Promote);
1286 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1287 setOperationAction(ISD::SELECT, VT, Promote);
1288 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1292 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1293 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1294 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1295 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1296 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1298 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1299 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1300 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1302 for (MVT VT : MVT::fp_vector_valuetypes())
1303 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1305 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1306 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1307 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1308 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1309 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1310 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1311 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1312 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1313 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1314 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1315 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1316 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1318 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1319 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1320 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1321 setOperationAction(ISD::XOR, MVT::i1, Legal);
1322 setOperationAction(ISD::OR, MVT::i1, Legal);
1323 setOperationAction(ISD::AND, MVT::i1, Legal);
1324 setOperationAction(ISD::SUB, MVT::i1, Custom);
1325 setOperationAction(ISD::ADD, MVT::i1, Custom);
1326 setOperationAction(ISD::MUL, MVT::i1, Custom);
1327 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1328 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1329 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1330 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1331 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1333 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1334 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1335 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1336 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1337 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1338 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1340 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1341 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1342 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1343 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1344 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1345 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1346 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1347 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1349 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1350 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1351 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1352 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1353 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1354 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1355 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1356 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1357 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1358 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1359 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1360 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1361 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1362 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1363 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1364 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1366 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1367 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1368 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1369 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1370 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1371 if (Subtarget->hasVLX()){
1372 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1373 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1374 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1375 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1376 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1378 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1379 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1380 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1381 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1382 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1384 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1385 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1386 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1387 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1388 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1389 if (Subtarget->hasDQI()) {
1390 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1391 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1393 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1394 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1395 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1396 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1397 if (Subtarget->hasVLX()) {
1398 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1399 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1400 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1401 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1402 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1403 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1404 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1405 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1408 if (Subtarget->hasVLX()) {
1409 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1410 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1411 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1412 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1413 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1414 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1415 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1416 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1418 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1419 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1420 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1421 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1422 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1423 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1424 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1425 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1426 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1427 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1428 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1429 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1430 if (Subtarget->hasDQI()) {
1431 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1432 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1434 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1436 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1437 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1438 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1439 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1440 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1441 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1442 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1443 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1445 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1446 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1447 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1448 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1449 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1451 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1452 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1454 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1456 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1457 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1458 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1459 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1460 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1461 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1462 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1463 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1464 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1465 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1466 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1468 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1469 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1470 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1471 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1472 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1473 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1474 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1475 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1477 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1478 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1480 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1481 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1483 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1485 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1486 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1488 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1489 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1491 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1492 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1494 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1495 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1496 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1497 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1498 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1499 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1501 if (Subtarget->hasCDI()) {
1502 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1503 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1504 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1505 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1507 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1508 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1509 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1510 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1511 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1512 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1513 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1514 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1516 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1517 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1519 if (Subtarget->hasVLX()) {
1520 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1521 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1522 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1523 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1524 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1525 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1526 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1527 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1529 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1530 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1531 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1532 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1535 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1536 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1537 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1539 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1540 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1541 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1543 } // Subtarget->hasCDI()
1545 if (Subtarget->hasDQI()) {
1546 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1547 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1548 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1550 // Custom lower several nodes.
1551 for (MVT VT : MVT::vector_valuetypes()) {
1552 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1554 setOperationAction(ISD::AND, VT, Legal);
1555 setOperationAction(ISD::OR, VT, Legal);
1556 setOperationAction(ISD::XOR, VT, Legal);
1558 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1559 setOperationAction(ISD::MGATHER, VT, Custom);
1560 setOperationAction(ISD::MSCATTER, VT, Custom);
1562 // Extract subvector is special because the value type
1563 // (result) is 256/128-bit but the source is 512-bit wide.
1564 if (VT.is128BitVector() || VT.is256BitVector()) {
1565 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1567 if (VT.getVectorElementType() == MVT::i1)
1568 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1570 // Do not attempt to custom lower other non-512-bit vectors
1571 if (!VT.is512BitVector())
1574 if (EltSize >= 32) {
1575 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1576 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1577 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1578 setOperationAction(ISD::VSELECT, VT, Legal);
1579 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1580 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1581 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1582 setOperationAction(ISD::MLOAD, VT, Legal);
1583 setOperationAction(ISD::MSTORE, VT, Legal);
1586 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1587 setOperationAction(ISD::SELECT, VT, Promote);
1588 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1592 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1593 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1594 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1596 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1597 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1599 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1600 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1601 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1602 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1603 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1604 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1605 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1606 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1607 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1608 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1609 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1610 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1611 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1612 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1613 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1614 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1615 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1616 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1617 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1618 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1619 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1620 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1621 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1622 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1623 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1624 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1625 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1626 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1628 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1629 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1630 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1631 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1632 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1633 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1634 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1635 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1636 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1637 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1638 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1639 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1640 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1642 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1643 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1644 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1645 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1646 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1647 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1648 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1649 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1651 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1652 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1653 if (Subtarget->hasVLX())
1654 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1656 if (Subtarget->hasCDI()) {
1657 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1658 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1659 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1660 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1663 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1664 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1665 setOperationAction(ISD::VSELECT, VT, Legal);
1669 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1670 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1671 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1673 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1674 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1675 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1676 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1677 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1678 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1679 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1680 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1681 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1682 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1686 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1687 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1688 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1689 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1690 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1691 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1692 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1693 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1695 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1696 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1697 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1698 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1699 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1700 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1701 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1702 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1705 // We want to custom lower some of our intrinsics.
1706 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1707 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1708 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1709 if (!Subtarget->is64Bit())
1710 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1712 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1713 // handle type legalization for these operations here.
1715 // FIXME: We really should do custom legalization for addition and
1716 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1717 // than generic legalization for 64-bit multiplication-with-overflow, though.
1718 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1719 if (VT == MVT::i64 && !Subtarget->is64Bit())
1721 // Add/Sub/Mul with overflow operations are custom lowered.
1722 setOperationAction(ISD::SADDO, VT, Custom);
1723 setOperationAction(ISD::UADDO, VT, Custom);
1724 setOperationAction(ISD::SSUBO, VT, Custom);
1725 setOperationAction(ISD::USUBO, VT, Custom);
1726 setOperationAction(ISD::SMULO, VT, Custom);
1727 setOperationAction(ISD::UMULO, VT, Custom);
1730 if (!Subtarget->is64Bit()) {
1731 // These libcalls are not available in 32-bit.
1732 setLibcallName(RTLIB::SHL_I128, nullptr);
1733 setLibcallName(RTLIB::SRL_I128, nullptr);
1734 setLibcallName(RTLIB::SRA_I128, nullptr);
1737 // Combine sin / cos into one node or libcall if possible.
1738 if (Subtarget->hasSinCos()) {
1739 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1740 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1741 if (Subtarget->isTargetDarwin()) {
1742 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1743 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1744 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1745 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1749 if (Subtarget->isTargetWin64()) {
1750 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1751 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1752 setOperationAction(ISD::SREM, MVT::i128, Custom);
1753 setOperationAction(ISD::UREM, MVT::i128, Custom);
1754 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1755 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1758 // We have target-specific dag combine patterns for the following nodes:
1759 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1760 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1761 setTargetDAGCombine(ISD::BITCAST);
1762 setTargetDAGCombine(ISD::VSELECT);
1763 setTargetDAGCombine(ISD::SELECT);
1764 setTargetDAGCombine(ISD::SHL);
1765 setTargetDAGCombine(ISD::SRA);
1766 setTargetDAGCombine(ISD::SRL);
1767 setTargetDAGCombine(ISD::OR);
1768 setTargetDAGCombine(ISD::AND);
1769 setTargetDAGCombine(ISD::ADD);
1770 setTargetDAGCombine(ISD::FADD);
1771 setTargetDAGCombine(ISD::FSUB);
1772 setTargetDAGCombine(ISD::FMA);
1773 setTargetDAGCombine(ISD::SUB);
1774 setTargetDAGCombine(ISD::LOAD);
1775 setTargetDAGCombine(ISD::MLOAD);
1776 setTargetDAGCombine(ISD::STORE);
1777 setTargetDAGCombine(ISD::MSTORE);
1778 setTargetDAGCombine(ISD::ZERO_EXTEND);
1779 setTargetDAGCombine(ISD::ANY_EXTEND);
1780 setTargetDAGCombine(ISD::SIGN_EXTEND);
1781 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1782 setTargetDAGCombine(ISD::SINT_TO_FP);
1783 setTargetDAGCombine(ISD::UINT_TO_FP);
1784 setTargetDAGCombine(ISD::SETCC);
1785 setTargetDAGCombine(ISD::BUILD_VECTOR);
1786 setTargetDAGCombine(ISD::MUL);
1787 setTargetDAGCombine(ISD::XOR);
1789 computeRegisterProperties(Subtarget->getRegisterInfo());
1791 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1792 MaxStoresPerMemsetOptSize = 8;
1793 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1794 MaxStoresPerMemcpyOptSize = 4;
1795 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1796 MaxStoresPerMemmoveOptSize = 4;
1797 setPrefLoopAlignment(4); // 2^4 bytes.
1799 // A predictable cmov does not hurt on an in-order CPU.
1800 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1801 PredictableSelectIsExpensive = !Subtarget->isAtom();
1802 EnableExtLdPromotion = true;
1803 setPrefFunctionAlignment(4); // 2^4 bytes.
1805 verifyIntrinsicTables();
1808 // This has so far only been implemented for 64-bit MachO.
1809 bool X86TargetLowering::useLoadStackGuardNode() const {
1810 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1813 TargetLoweringBase::LegalizeTypeAction
1814 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1815 if (ExperimentalVectorWideningLegalization &&
1816 VT.getVectorNumElements() != 1 &&
1817 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1818 return TypeWidenVector;
1820 return TargetLoweringBase::getPreferredVectorAction(VT);
1823 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1826 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1828 if (VT.isSimple()) {
1829 MVT VVT = VT.getSimpleVT();
1830 const unsigned NumElts = VVT.getVectorNumElements();
1831 const MVT EltVT = VVT.getVectorElementType();
1832 if (VVT.is512BitVector()) {
1833 if (Subtarget->hasAVX512())
1834 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1835 EltVT == MVT::f32 || EltVT == MVT::f64)
1837 case 8: return MVT::v8i1;
1838 case 16: return MVT::v16i1;
1840 if (Subtarget->hasBWI())
1841 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1843 case 32: return MVT::v32i1;
1844 case 64: return MVT::v64i1;
1848 if (VVT.is256BitVector() || VVT.is128BitVector()) {
1849 if (Subtarget->hasVLX())
1850 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1851 EltVT == MVT::f32 || EltVT == MVT::f64)
1853 case 2: return MVT::v2i1;
1854 case 4: return MVT::v4i1;
1855 case 8: return MVT::v8i1;
1857 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1858 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1860 case 8: return MVT::v8i1;
1861 case 16: return MVT::v16i1;
1862 case 32: return MVT::v32i1;
1867 return VT.changeVectorElementTypeToInteger();
1870 /// Helper for getByValTypeAlignment to determine
1871 /// the desired ByVal argument alignment.
1872 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1875 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1876 if (VTy->getBitWidth() == 128)
1878 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1879 unsigned EltAlign = 0;
1880 getMaxByValAlign(ATy->getElementType(), EltAlign);
1881 if (EltAlign > MaxAlign)
1882 MaxAlign = EltAlign;
1883 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1884 for (auto *EltTy : STy->elements()) {
1885 unsigned EltAlign = 0;
1886 getMaxByValAlign(EltTy, EltAlign);
1887 if (EltAlign > MaxAlign)
1888 MaxAlign = EltAlign;
1895 /// Return the desired alignment for ByVal aggregate
1896 /// function arguments in the caller parameter area. For X86, aggregates
1897 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1898 /// are at 4-byte boundaries.
1899 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1900 const DataLayout &DL) const {
1901 if (Subtarget->is64Bit()) {
1902 // Max of 8 and alignment of type.
1903 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1910 if (Subtarget->hasSSE1())
1911 getMaxByValAlign(Ty, Align);
1915 /// Returns the target specific optimal type for load
1916 /// and store operations as a result of memset, memcpy, and memmove
1917 /// lowering. If DstAlign is zero that means it's safe to destination
1918 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1919 /// means there isn't a need to check it against alignment requirement,
1920 /// probably because the source does not need to be loaded. If 'IsMemset' is
1921 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1922 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1923 /// source is constant so it does not need to be loaded.
1924 /// It returns EVT::Other if the type should be determined using generic
1925 /// target-independent logic.
1927 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1928 unsigned DstAlign, unsigned SrcAlign,
1929 bool IsMemset, bool ZeroMemset,
1931 MachineFunction &MF) const {
1932 const Function *F = MF.getFunction();
1933 if ((!IsMemset || ZeroMemset) &&
1934 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1936 (!Subtarget->isUnalignedMem16Slow() ||
1937 ((DstAlign == 0 || DstAlign >= 16) &&
1938 (SrcAlign == 0 || SrcAlign >= 16)))) {
1940 // FIXME: Check if unaligned 32-byte accesses are slow.
1941 if (Subtarget->hasInt256())
1943 if (Subtarget->hasFp256())
1946 if (Subtarget->hasSSE2())
1948 if (Subtarget->hasSSE1())
1950 } else if (!MemcpyStrSrc && Size >= 8 &&
1951 !Subtarget->is64Bit() &&
1952 Subtarget->hasSSE2()) {
1953 // Do not use f64 to lower memcpy if source is string constant. It's
1954 // better to use i32 to avoid the loads.
1958 // This is a compromise. If we reach here, unaligned accesses may be slow on
1959 // this target. However, creating smaller, aligned accesses could be even
1960 // slower and would certainly be a lot more code.
1961 if (Subtarget->is64Bit() && Size >= 8)
1966 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1968 return X86ScalarSSEf32;
1969 else if (VT == MVT::f64)
1970 return X86ScalarSSEf64;
1975 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1980 switch (VT.getSizeInBits()) {
1982 // 8-byte and under are always assumed to be fast.
1986 *Fast = !Subtarget->isUnalignedMem16Slow();
1989 *Fast = !Subtarget->isUnalignedMem32Slow();
1991 // TODO: What about AVX-512 (512-bit) accesses?
1994 // Misaligned accesses of any size are always allowed.
1998 /// Return the entry encoding for a jump table in the
1999 /// current function. The returned value is a member of the
2000 /// MachineJumpTableInfo::JTEntryKind enum.
2001 unsigned X86TargetLowering::getJumpTableEncoding() const {
2002 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2004 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2005 Subtarget->isPICStyleGOT())
2006 return MachineJumpTableInfo::EK_Custom32;
2008 // Otherwise, use the normal jump table encoding heuristics.
2009 return TargetLowering::getJumpTableEncoding();
2012 bool X86TargetLowering::useSoftFloat() const {
2013 return Subtarget->useSoftFloat();
2017 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2018 const MachineBasicBlock *MBB,
2019 unsigned uid,MCContext &Ctx) const{
2020 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2021 Subtarget->isPICStyleGOT());
2022 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2024 return MCSymbolRefExpr::create(MBB->getSymbol(),
2025 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2028 /// Returns relocation base for the given PIC jumptable.
2029 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2030 SelectionDAG &DAG) const {
2031 if (!Subtarget->is64Bit())
2032 // This doesn't have SDLoc associated with it, but is not really the
2033 // same as a Register.
2034 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2035 getPointerTy(DAG.getDataLayout()));
2039 /// This returns the relocation base for the given PIC jumptable,
2040 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2041 const MCExpr *X86TargetLowering::
2042 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2043 MCContext &Ctx) const {
2044 // X86-64 uses RIP relative addressing based on the jump table label.
2045 if (Subtarget->isPICStyleRIPRel())
2046 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2048 // Otherwise, the reference is relative to the PIC base.
2049 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2052 std::pair<const TargetRegisterClass *, uint8_t>
2053 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2055 const TargetRegisterClass *RRC = nullptr;
2057 switch (VT.SimpleTy) {
2059 return TargetLowering::findRepresentativeClass(TRI, VT);
2060 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2061 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2064 RRC = &X86::VR64RegClass;
2066 case MVT::f32: case MVT::f64:
2067 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2068 case MVT::v4f32: case MVT::v2f64:
2069 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2071 RRC = &X86::VR128RegClass;
2074 return std::make_pair(RRC, Cost);
2077 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2078 unsigned &Offset) const {
2079 if (!Subtarget->isTargetLinux())
2082 if (Subtarget->is64Bit()) {
2083 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2085 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2097 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2098 if (!Subtarget->isTargetAndroid())
2099 return TargetLowering::getSafeStackPointerLocation(IRB);
2101 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2102 // definition of TLS_SLOT_SAFESTACK in
2103 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2104 unsigned AddressSpace, Offset;
2105 if (Subtarget->is64Bit()) {
2106 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2108 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2118 return ConstantExpr::getIntToPtr(
2119 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2120 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2123 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2124 unsigned DestAS) const {
2125 assert(SrcAS != DestAS && "Expected different address spaces!");
2127 return SrcAS < 256 && DestAS < 256;
2130 //===----------------------------------------------------------------------===//
2131 // Return Value Calling Convention Implementation
2132 //===----------------------------------------------------------------------===//
2134 #include "X86GenCallingConv.inc"
2136 bool X86TargetLowering::CanLowerReturn(
2137 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2138 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2139 SmallVector<CCValAssign, 16> RVLocs;
2140 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2141 return CCInfo.CheckReturn(Outs, RetCC_X86);
2144 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2145 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2150 X86TargetLowering::LowerReturn(SDValue Chain,
2151 CallingConv::ID CallConv, bool isVarArg,
2152 const SmallVectorImpl<ISD::OutputArg> &Outs,
2153 const SmallVectorImpl<SDValue> &OutVals,
2154 SDLoc dl, SelectionDAG &DAG) const {
2155 MachineFunction &MF = DAG.getMachineFunction();
2156 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2158 SmallVector<CCValAssign, 16> RVLocs;
2159 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2160 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2163 SmallVector<SDValue, 6> RetOps;
2164 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2165 // Operand #1 = Bytes To Pop
2166 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2169 // Copy the result values into the output registers.
2170 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2171 CCValAssign &VA = RVLocs[i];
2172 assert(VA.isRegLoc() && "Can only return in registers!");
2173 SDValue ValToCopy = OutVals[i];
2174 EVT ValVT = ValToCopy.getValueType();
2176 // Promote values to the appropriate types.
2177 if (VA.getLocInfo() == CCValAssign::SExt)
2178 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2179 else if (VA.getLocInfo() == CCValAssign::ZExt)
2180 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2181 else if (VA.getLocInfo() == CCValAssign::AExt) {
2182 if (ValVT.isVector() && ValVT.getVectorElementType() == MVT::i1)
2183 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2185 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2187 else if (VA.getLocInfo() == CCValAssign::BCvt)
2188 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2190 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2191 "Unexpected FP-extend for return value.");
2193 // If this is x86-64, and we disabled SSE, we can't return FP values,
2194 // or SSE or MMX vectors.
2195 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2196 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2197 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2198 report_fatal_error("SSE register return with SSE disabled");
2200 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2201 // llvm-gcc has never done it right and no one has noticed, so this
2202 // should be OK for now.
2203 if (ValVT == MVT::f64 &&
2204 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2205 report_fatal_error("SSE2 register return with SSE2 disabled");
2207 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2208 // the RET instruction and handled by the FP Stackifier.
2209 if (VA.getLocReg() == X86::FP0 ||
2210 VA.getLocReg() == X86::FP1) {
2211 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2212 // change the value to the FP stack register class.
2213 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2214 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2215 RetOps.push_back(ValToCopy);
2216 // Don't emit a copytoreg.
2220 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2221 // which is returned in RAX / RDX.
2222 if (Subtarget->is64Bit()) {
2223 if (ValVT == MVT::x86mmx) {
2224 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2225 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2226 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2228 // If we don't have SSE2 available, convert to v4f32 so the generated
2229 // register is legal.
2230 if (!Subtarget->hasSSE2())
2231 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2236 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2237 Flag = Chain.getValue(1);
2238 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2241 // All x86 ABIs require that for returning structs by value we copy
2242 // the sret argument into %rax/%eax (depending on ABI) for the return.
2243 // We saved the argument into a virtual register in the entry block,
2244 // so now we copy the value out and into %rax/%eax.
2246 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2247 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2248 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2249 // either case FuncInfo->setSRetReturnReg() will have been called.
2250 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2251 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2252 getPointerTy(MF.getDataLayout()));
2255 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2256 X86::RAX : X86::EAX;
2257 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2258 Flag = Chain.getValue(1);
2260 // RAX/EAX now acts like a return value.
2262 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2265 RetOps[0] = Chain; // Update chain.
2267 // Add the flag if we have it.
2269 RetOps.push_back(Flag);
2271 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2274 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2275 if (N->getNumValues() != 1)
2277 if (!N->hasNUsesOfValue(1, 0))
2280 SDValue TCChain = Chain;
2281 SDNode *Copy = *N->use_begin();
2282 if (Copy->getOpcode() == ISD::CopyToReg) {
2283 // If the copy has a glue operand, we conservatively assume it isn't safe to
2284 // perform a tail call.
2285 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2287 TCChain = Copy->getOperand(0);
2288 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2291 bool HasRet = false;
2292 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2294 if (UI->getOpcode() != X86ISD::RET_FLAG)
2296 // If we are returning more than one value, we can definitely
2297 // not make a tail call see PR19530
2298 if (UI->getNumOperands() > 4)
2300 if (UI->getNumOperands() == 4 &&
2301 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2314 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2315 ISD::NodeType ExtendKind) const {
2317 // TODO: Is this also valid on 32-bit?
2318 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2319 ReturnMVT = MVT::i8;
2321 ReturnMVT = MVT::i32;
2323 EVT MinVT = getRegisterType(Context, ReturnMVT);
2324 return VT.bitsLT(MinVT) ? MinVT : VT;
2327 /// Lower the result values of a call into the
2328 /// appropriate copies out of appropriate physical registers.
2331 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2332 CallingConv::ID CallConv, bool isVarArg,
2333 const SmallVectorImpl<ISD::InputArg> &Ins,
2334 SDLoc dl, SelectionDAG &DAG,
2335 SmallVectorImpl<SDValue> &InVals) const {
2337 // Assign locations to each value returned by this call.
2338 SmallVector<CCValAssign, 16> RVLocs;
2339 bool Is64Bit = Subtarget->is64Bit();
2340 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2342 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2344 // Copy all of the result registers out of their specified physreg.
2345 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2346 CCValAssign &VA = RVLocs[i];
2347 EVT CopyVT = VA.getLocVT();
2349 // If this is x86-64, and we disabled SSE, we can't return FP values
2350 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2351 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2352 report_fatal_error("SSE register return with SSE disabled");
2355 // If we prefer to use the value in xmm registers, copy it out as f80 and
2356 // use a truncate to move it from fp stack reg to xmm reg.
2357 bool RoundAfterCopy = false;
2358 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2359 isScalarFPTypeInSSEReg(VA.getValVT())) {
2361 RoundAfterCopy = (CopyVT != VA.getLocVT());
2364 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2365 CopyVT, InFlag).getValue(1);
2366 SDValue Val = Chain.getValue(0);
2369 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2370 // This truncation won't change the value.
2371 DAG.getIntPtrConstant(1, dl));
2373 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2374 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2376 InFlag = Chain.getValue(2);
2377 InVals.push_back(Val);
2383 //===----------------------------------------------------------------------===//
2384 // C & StdCall & Fast Calling Convention implementation
2385 //===----------------------------------------------------------------------===//
2386 // StdCall calling convention seems to be standard for many Windows' API
2387 // routines and around. It differs from C calling convention just a little:
2388 // callee should clean up the stack, not caller. Symbols should be also
2389 // decorated in some fancy way :) It doesn't support any vector arguments.
2390 // For info on fast calling convention see Fast Calling Convention (tail call)
2391 // implementation LowerX86_32FastCCCallTo.
2393 /// CallIsStructReturn - Determines whether a call uses struct return
2395 enum StructReturnType {
2400 static StructReturnType
2401 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2403 return NotStructReturn;
2405 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2406 if (!Flags.isSRet())
2407 return NotStructReturn;
2408 if (Flags.isInReg())
2409 return RegStructReturn;
2410 return StackStructReturn;
2413 /// Determines whether a function uses struct return semantics.
2414 static StructReturnType
2415 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2417 return NotStructReturn;
2419 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2420 if (!Flags.isSRet())
2421 return NotStructReturn;
2422 if (Flags.isInReg())
2423 return RegStructReturn;
2424 return StackStructReturn;
2427 /// Make a copy of an aggregate at address specified by "Src" to address
2428 /// "Dst" with size and alignment information specified by the specific
2429 /// parameter attribute. The copy will be passed as a byval function parameter.
2431 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2432 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2434 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2436 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2437 /*isVolatile*/false, /*AlwaysInline=*/true,
2438 /*isTailCall*/false,
2439 MachinePointerInfo(), MachinePointerInfo());
2442 /// Return true if the calling convention is one that we can guarantee TCO for.
2443 static bool canGuaranteeTCO(CallingConv::ID CC) {
2444 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2445 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2448 /// Return true if we might ever do TCO for calls with this calling convention.
2449 static bool mayTailCallThisCC(CallingConv::ID CC) {
2451 // C calling conventions:
2452 case CallingConv::C:
2453 case CallingConv::X86_64_Win64:
2454 case CallingConv::X86_64_SysV:
2455 // Callee pop conventions:
2456 case CallingConv::X86_ThisCall:
2457 case CallingConv::X86_StdCall:
2458 case CallingConv::X86_VectorCall:
2459 case CallingConv::X86_FastCall:
2462 return canGuaranteeTCO(CC);
2466 /// Return true if the function is being made into a tailcall target by
2467 /// changing its ABI.
2468 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2469 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2472 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2474 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2475 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2479 CallingConv::ID CalleeCC = CS.getCallingConv();
2480 if (!mayTailCallThisCC(CalleeCC))
2487 X86TargetLowering::LowerMemArgument(SDValue Chain,
2488 CallingConv::ID CallConv,
2489 const SmallVectorImpl<ISD::InputArg> &Ins,
2490 SDLoc dl, SelectionDAG &DAG,
2491 const CCValAssign &VA,
2492 MachineFrameInfo *MFI,
2494 // Create the nodes corresponding to a load from this parameter slot.
2495 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2496 bool AlwaysUseMutable = shouldGuaranteeTCO(
2497 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2498 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2501 // If value is passed by pointer we have address passed instead of the value
2503 bool ExtendedInMem = VA.isExtInLoc() &&
2504 VA.getValVT().getScalarType() == MVT::i1;
2506 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2507 ValVT = VA.getLocVT();
2509 ValVT = VA.getValVT();
2511 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2512 // changed with more analysis.
2513 // In case of tail call optimization mark all arguments mutable. Since they
2514 // could be overwritten by lowering of arguments in case of a tail call.
2515 if (Flags.isByVal()) {
2516 unsigned Bytes = Flags.getByValSize();
2517 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2518 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2519 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2521 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2522 VA.getLocMemOffset(), isImmutable);
2523 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2524 SDValue Val = DAG.getLoad(
2525 ValVT, dl, Chain, FIN,
2526 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2528 return ExtendedInMem ?
2529 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2533 // FIXME: Get this from tablegen.
2534 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2535 const X86Subtarget *Subtarget) {
2536 assert(Subtarget->is64Bit());
2538 if (Subtarget->isCallingConvWin64(CallConv)) {
2539 static const MCPhysReg GPR64ArgRegsWin64[] = {
2540 X86::RCX, X86::RDX, X86::R8, X86::R9
2542 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2545 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2546 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2548 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2551 // FIXME: Get this from tablegen.
2552 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2553 CallingConv::ID CallConv,
2554 const X86Subtarget *Subtarget) {
2555 assert(Subtarget->is64Bit());
2556 if (Subtarget->isCallingConvWin64(CallConv)) {
2557 // The XMM registers which might contain var arg parameters are shadowed
2558 // in their paired GPR. So we only need to save the GPR to their home
2560 // TODO: __vectorcall will change this.
2564 const Function *Fn = MF.getFunction();
2565 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2566 bool isSoftFloat = Subtarget->useSoftFloat();
2567 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2568 "SSE register cannot be used when SSE is disabled!");
2569 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2570 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2574 static const MCPhysReg XMMArgRegs64Bit[] = {
2575 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2576 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2578 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2581 SDValue X86TargetLowering::LowerFormalArguments(
2582 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2583 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2584 SmallVectorImpl<SDValue> &InVals) const {
2585 MachineFunction &MF = DAG.getMachineFunction();
2586 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2587 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2589 const Function* Fn = MF.getFunction();
2590 if (Fn->hasExternalLinkage() &&
2591 Subtarget->isTargetCygMing() &&
2592 Fn->getName() == "main")
2593 FuncInfo->setForceFramePointer(true);
2595 MachineFrameInfo *MFI = MF.getFrameInfo();
2596 bool Is64Bit = Subtarget->is64Bit();
2597 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2599 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2600 "Var args not supported with calling convention fastcc, ghc or hipe");
2602 // Assign locations to all of the incoming arguments.
2603 SmallVector<CCValAssign, 16> ArgLocs;
2604 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2606 // Allocate shadow area for Win64
2608 CCInfo.AllocateStack(32, 8);
2610 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2612 unsigned LastVal = ~0U;
2614 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2615 CCValAssign &VA = ArgLocs[i];
2616 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2618 assert(VA.getValNo() != LastVal &&
2619 "Don't support value assigned to multiple locs yet");
2621 LastVal = VA.getValNo();
2623 if (VA.isRegLoc()) {
2624 EVT RegVT = VA.getLocVT();
2625 const TargetRegisterClass *RC;
2626 if (RegVT == MVT::i32)
2627 RC = &X86::GR32RegClass;
2628 else if (Is64Bit && RegVT == MVT::i64)
2629 RC = &X86::GR64RegClass;
2630 else if (RegVT == MVT::f32)
2631 RC = &X86::FR32RegClass;
2632 else if (RegVT == MVT::f64)
2633 RC = &X86::FR64RegClass;
2634 else if (RegVT.is512BitVector())
2635 RC = &X86::VR512RegClass;
2636 else if (RegVT.is256BitVector())
2637 RC = &X86::VR256RegClass;
2638 else if (RegVT.is128BitVector())
2639 RC = &X86::VR128RegClass;
2640 else if (RegVT == MVT::x86mmx)
2641 RC = &X86::VR64RegClass;
2642 else if (RegVT == MVT::i1)
2643 RC = &X86::VK1RegClass;
2644 else if (RegVT == MVT::v8i1)
2645 RC = &X86::VK8RegClass;
2646 else if (RegVT == MVT::v16i1)
2647 RC = &X86::VK16RegClass;
2648 else if (RegVT == MVT::v32i1)
2649 RC = &X86::VK32RegClass;
2650 else if (RegVT == MVT::v64i1)
2651 RC = &X86::VK64RegClass;
2653 llvm_unreachable("Unknown argument type!");
2655 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2656 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2658 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2659 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2661 if (VA.getLocInfo() == CCValAssign::SExt)
2662 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2663 DAG.getValueType(VA.getValVT()));
2664 else if (VA.getLocInfo() == CCValAssign::ZExt)
2665 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2666 DAG.getValueType(VA.getValVT()));
2667 else if (VA.getLocInfo() == CCValAssign::BCvt)
2668 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2670 if (VA.isExtInLoc()) {
2671 // Handle MMX values passed in XMM regs.
2672 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2673 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2675 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2678 assert(VA.isMemLoc());
2679 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2682 // If value is passed via pointer - do a load.
2683 if (VA.getLocInfo() == CCValAssign::Indirect)
2684 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2685 MachinePointerInfo(), false, false, false, 0);
2687 InVals.push_back(ArgValue);
2690 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2691 // All x86 ABIs require that for returning structs by value we copy the
2692 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2693 // the argument into a virtual register so that we can access it from the
2695 if (Ins[i].Flags.isSRet()) {
2696 unsigned Reg = FuncInfo->getSRetReturnReg();
2698 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2699 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2700 FuncInfo->setSRetReturnReg(Reg);
2702 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2703 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2708 unsigned StackSize = CCInfo.getNextStackOffset();
2709 // Align stack specially for tail calls.
2710 if (shouldGuaranteeTCO(CallConv,
2711 MF.getTarget().Options.GuaranteedTailCallOpt))
2712 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2714 // If the function takes variable number of arguments, make a frame index for
2715 // the start of the first vararg value... for expansion of llvm.va_start. We
2716 // can skip this if there are no va_start calls.
2717 if (MFI->hasVAStart() &&
2718 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2719 CallConv != CallingConv::X86_ThisCall))) {
2720 FuncInfo->setVarArgsFrameIndex(
2721 MFI->CreateFixedObject(1, StackSize, true));
2724 MachineModuleInfo &MMI = MF.getMMI();
2726 // Figure out if XMM registers are in use.
2727 assert(!(Subtarget->useSoftFloat() &&
2728 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2729 "SSE register cannot be used when SSE is disabled!");
2731 // 64-bit calling conventions support varargs and register parameters, so we
2732 // have to do extra work to spill them in the prologue.
2733 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2734 // Find the first unallocated argument registers.
2735 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2736 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2737 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2738 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2739 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2740 "SSE register cannot be used when SSE is disabled!");
2742 // Gather all the live in physical registers.
2743 SmallVector<SDValue, 6> LiveGPRs;
2744 SmallVector<SDValue, 8> LiveXMMRegs;
2746 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2747 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2749 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2751 if (!ArgXMMs.empty()) {
2752 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2753 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2754 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2755 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2756 LiveXMMRegs.push_back(
2757 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2762 // Get to the caller-allocated home save location. Add 8 to account
2763 // for the return address.
2764 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2765 FuncInfo->setRegSaveFrameIndex(
2766 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2767 // Fixup to set vararg frame on shadow area (4 x i64).
2769 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2771 // For X86-64, if there are vararg parameters that are passed via
2772 // registers, then we must store them to their spots on the stack so
2773 // they may be loaded by deferencing the result of va_next.
2774 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2775 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2776 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2777 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2780 // Store the integer parameter registers.
2781 SmallVector<SDValue, 8> MemOps;
2782 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2783 getPointerTy(DAG.getDataLayout()));
2784 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2785 for (SDValue Val : LiveGPRs) {
2786 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2787 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2789 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2790 MachinePointerInfo::getFixedStack(
2791 DAG.getMachineFunction(),
2792 FuncInfo->getRegSaveFrameIndex(), Offset),
2794 MemOps.push_back(Store);
2798 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2799 // Now store the XMM (fp + vector) parameter registers.
2800 SmallVector<SDValue, 12> SaveXMMOps;
2801 SaveXMMOps.push_back(Chain);
2802 SaveXMMOps.push_back(ALVal);
2803 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2804 FuncInfo->getRegSaveFrameIndex(), dl));
2805 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2806 FuncInfo->getVarArgsFPOffset(), dl));
2807 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2809 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2810 MVT::Other, SaveXMMOps));
2813 if (!MemOps.empty())
2814 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2817 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2818 // Find the largest legal vector type.
2819 MVT VecVT = MVT::Other;
2820 // FIXME: Only some x86_32 calling conventions support AVX512.
2821 if (Subtarget->hasAVX512() &&
2822 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2823 CallConv == CallingConv::Intel_OCL_BI)))
2824 VecVT = MVT::v16f32;
2825 else if (Subtarget->hasAVX())
2827 else if (Subtarget->hasSSE2())
2830 // We forward some GPRs and some vector types.
2831 SmallVector<MVT, 2> RegParmTypes;
2832 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2833 RegParmTypes.push_back(IntVT);
2834 if (VecVT != MVT::Other)
2835 RegParmTypes.push_back(VecVT);
2837 // Compute the set of forwarded registers. The rest are scratch.
2838 SmallVectorImpl<ForwardedRegister> &Forwards =
2839 FuncInfo->getForwardedMustTailRegParms();
2840 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2842 // Conservatively forward AL on x86_64, since it might be used for varargs.
2843 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2844 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2845 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2848 // Copy all forwards from physical to virtual registers.
2849 for (ForwardedRegister &F : Forwards) {
2850 // FIXME: Can we use a less constrained schedule?
2851 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2852 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2853 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2857 // Some CCs need callee pop.
2858 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2859 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2860 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2862 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2863 // If this is an sret function, the return should pop the hidden pointer.
2864 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2865 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2866 argsAreStructReturn(Ins) == StackStructReturn)
2867 FuncInfo->setBytesToPopOnReturn(4);
2871 // RegSaveFrameIndex is X86-64 only.
2872 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2873 if (CallConv == CallingConv::X86_FastCall ||
2874 CallConv == CallingConv::X86_ThisCall)
2875 // fastcc functions can't have varargs.
2876 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2879 FuncInfo->setArgumentStackSize(StackSize);
2881 if (MMI.hasWinEHFuncInfo(Fn)) {
2882 EHPersonality Personality = classifyEHPersonality(Fn->getPersonalityFn());
2883 if (Personality == EHPersonality::MSVC_CXX) {
2885 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2886 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2887 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx =
2889 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2890 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2891 MachinePointerInfo::getFixedStack(
2892 DAG.getMachineFunction(), UnwindHelpFI),
2893 /*isVolatile=*/true,
2894 /*isNonTemporal=*/false, /*Alignment=*/0);
2896 } else if (Personality == EHPersonality::CoreCLR) {
2898 // TODO: Add a mechanism to frame lowering that will allow us to indicate
2899 // that we'd prefer this slot be allocated towards the bottom of the frame
2900 // (i.e. near the stack pointer after allocating the frame). Every
2901 // funclet needs a copy of this slot in its (mostly empty) frame, and the
2902 // offset from the bottom of this and each funclet's frame must be the
2903 // same, so the size of funclets' (mostly empty) frames is dictated by
2904 // how far this slot is from the bottom (since they allocate just enough
2905 // space to accomodate holding this slot at the correct offset).
2906 int PSPSymFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2907 MMI.getWinEHFuncInfo(MF.getFunction()).PSPSymFrameIdx = PSPSymFI;
2915 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2916 SDValue StackPtr, SDValue Arg,
2917 SDLoc dl, SelectionDAG &DAG,
2918 const CCValAssign &VA,
2919 ISD::ArgFlagsTy Flags) const {
2920 unsigned LocMemOffset = VA.getLocMemOffset();
2921 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2922 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2924 if (Flags.isByVal())
2925 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2927 return DAG.getStore(
2928 Chain, dl, Arg, PtrOff,
2929 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2933 /// Emit a load of return address if tail call
2934 /// optimization is performed and it is required.
2936 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2937 SDValue &OutRetAddr, SDValue Chain,
2938 bool IsTailCall, bool Is64Bit,
2939 int FPDiff, SDLoc dl) const {
2940 // Adjust the Return address stack slot.
2941 EVT VT = getPointerTy(DAG.getDataLayout());
2942 OutRetAddr = getReturnAddressFrameIndex(DAG);
2944 // Load the "old" Return address.
2945 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2946 false, false, false, 0);
2947 return SDValue(OutRetAddr.getNode(), 1);
2950 /// Emit a store of the return address if tail call
2951 /// optimization is performed and it is required (FPDiff!=0).
2952 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2953 SDValue Chain, SDValue RetAddrFrIdx,
2954 EVT PtrVT, unsigned SlotSize,
2955 int FPDiff, SDLoc dl) {
2956 // Store the return address to the appropriate stack slot.
2957 if (!FPDiff) return Chain;
2958 // Calculate the new stack slot for the return address.
2959 int NewReturnAddrFI =
2960 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2962 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2963 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2964 MachinePointerInfo::getFixedStack(
2965 DAG.getMachineFunction(), NewReturnAddrFI),
2970 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2971 /// operation of specified width.
2972 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
2974 unsigned NumElems = VT.getVectorNumElements();
2975 SmallVector<int, 8> Mask;
2976 Mask.push_back(NumElems);
2977 for (unsigned i = 1; i != NumElems; ++i)
2979 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2983 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2984 SmallVectorImpl<SDValue> &InVals) const {
2985 SelectionDAG &DAG = CLI.DAG;
2987 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2988 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2989 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2990 SDValue Chain = CLI.Chain;
2991 SDValue Callee = CLI.Callee;
2992 CallingConv::ID CallConv = CLI.CallConv;
2993 bool &isTailCall = CLI.IsTailCall;
2994 bool isVarArg = CLI.IsVarArg;
2996 MachineFunction &MF = DAG.getMachineFunction();
2997 bool Is64Bit = Subtarget->is64Bit();
2998 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2999 StructReturnType SR = callIsStructReturn(Outs);
3000 bool IsSibcall = false;
3001 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3002 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3004 if (Attr.getValueAsString() == "true")
3007 if (Subtarget->isPICStyleGOT() &&
3008 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3009 // If we are using a GOT, disable tail calls to external symbols with
3010 // default visibility. Tail calling such a symbol requires using a GOT
3011 // relocation, which forces early binding of the symbol. This breaks code
3012 // that require lazy function symbol resolution. Using musttail or
3013 // GuaranteedTailCallOpt will override this.
3014 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3015 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3016 G->getGlobal()->hasDefaultVisibility()))
3020 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3022 // Force this to be a tail call. The verifier rules are enough to ensure
3023 // that we can lower this successfully without moving the return address
3026 } else if (isTailCall) {
3027 // Check if it's really possible to do a tail call.
3028 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3029 isVarArg, SR != NotStructReturn,
3030 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3031 Outs, OutVals, Ins, DAG);
3033 // Sibcalls are automatically detected tailcalls which do not require
3035 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3042 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3043 "Var args not supported with calling convention fastcc, ghc or hipe");
3045 // Analyze operands of the call, assigning locations to each operand.
3046 SmallVector<CCValAssign, 16> ArgLocs;
3047 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3049 // Allocate shadow area for Win64
3051 CCInfo.AllocateStack(32, 8);
3053 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3055 // Get a count of how many bytes are to be pushed on the stack.
3056 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3058 // This is a sibcall. The memory operands are available in caller's
3059 // own caller's stack.
3061 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3062 canGuaranteeTCO(CallConv))
3063 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3066 if (isTailCall && !IsSibcall && !IsMustTail) {
3067 // Lower arguments at fp - stackoffset + fpdiff.
3068 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3070 FPDiff = NumBytesCallerPushed - NumBytes;
3072 // Set the delta of movement of the returnaddr stackslot.
3073 // But only set if delta is greater than previous delta.
3074 if (FPDiff < X86Info->getTCReturnAddrDelta())
3075 X86Info->setTCReturnAddrDelta(FPDiff);
3078 unsigned NumBytesToPush = NumBytes;
3079 unsigned NumBytesToPop = NumBytes;
3081 // If we have an inalloca argument, all stack space has already been allocated
3082 // for us and be right at the top of the stack. We don't support multiple
3083 // arguments passed in memory when using inalloca.
3084 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3086 if (!ArgLocs.back().isMemLoc())
3087 report_fatal_error("cannot use inalloca attribute on a register "
3089 if (ArgLocs.back().getLocMemOffset() != 0)
3090 report_fatal_error("any parameter with the inalloca attribute must be "
3091 "the only memory argument");
3095 Chain = DAG.getCALLSEQ_START(
3096 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3098 SDValue RetAddrFrIdx;
3099 // Load return address for tail calls.
3100 if (isTailCall && FPDiff)
3101 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3102 Is64Bit, FPDiff, dl);
3104 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3105 SmallVector<SDValue, 8> MemOpChains;
3108 // Walk the register/memloc assignments, inserting copies/loads. In the case
3109 // of tail call optimization arguments are handle later.
3110 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3111 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3112 // Skip inalloca arguments, they have already been written.
3113 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3114 if (Flags.isInAlloca())
3117 CCValAssign &VA = ArgLocs[i];
3118 EVT RegVT = VA.getLocVT();
3119 SDValue Arg = OutVals[i];
3120 bool isByVal = Flags.isByVal();
3122 // Promote the value if needed.
3123 switch (VA.getLocInfo()) {
3124 default: llvm_unreachable("Unknown loc info!");
3125 case CCValAssign::Full: break;
3126 case CCValAssign::SExt:
3127 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3129 case CCValAssign::ZExt:
3130 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3132 case CCValAssign::AExt:
3133 if (Arg.getValueType().isVector() &&
3134 Arg.getValueType().getVectorElementType() == MVT::i1)
3135 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3136 else if (RegVT.is128BitVector()) {
3137 // Special case: passing MMX values in XMM registers.
3138 Arg = DAG.getBitcast(MVT::i64, Arg);
3139 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3140 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3142 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3144 case CCValAssign::BCvt:
3145 Arg = DAG.getBitcast(RegVT, Arg);
3147 case CCValAssign::Indirect: {
3148 // Store the argument.
3149 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3150 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3151 Chain = DAG.getStore(
3152 Chain, dl, Arg, SpillSlot,
3153 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3160 if (VA.isRegLoc()) {
3161 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3162 if (isVarArg && IsWin64) {
3163 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3164 // shadow reg if callee is a varargs function.
3165 unsigned ShadowReg = 0;
3166 switch (VA.getLocReg()) {
3167 case X86::XMM0: ShadowReg = X86::RCX; break;
3168 case X86::XMM1: ShadowReg = X86::RDX; break;
3169 case X86::XMM2: ShadowReg = X86::R8; break;
3170 case X86::XMM3: ShadowReg = X86::R9; break;
3173 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3175 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3176 assert(VA.isMemLoc());
3177 if (!StackPtr.getNode())
3178 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3179 getPointerTy(DAG.getDataLayout()));
3180 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3181 dl, DAG, VA, Flags));
3185 if (!MemOpChains.empty())
3186 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3188 if (Subtarget->isPICStyleGOT()) {
3189 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3192 RegsToPass.push_back(std::make_pair(
3193 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3194 getPointerTy(DAG.getDataLayout()))));
3196 // If we are tail calling and generating PIC/GOT style code load the
3197 // address of the callee into ECX. The value in ecx is used as target of
3198 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3199 // for tail calls on PIC/GOT architectures. Normally we would just put the
3200 // address of GOT into ebx and then call target@PLT. But for tail calls
3201 // ebx would be restored (since ebx is callee saved) before jumping to the
3204 // Note: The actual moving to ECX is done further down.
3205 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3206 if (G && !G->getGlobal()->hasLocalLinkage() &&
3207 G->getGlobal()->hasDefaultVisibility())
3208 Callee = LowerGlobalAddress(Callee, DAG);
3209 else if (isa<ExternalSymbolSDNode>(Callee))
3210 Callee = LowerExternalSymbol(Callee, DAG);
3214 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3215 // From AMD64 ABI document:
3216 // For calls that may call functions that use varargs or stdargs
3217 // (prototype-less calls or calls to functions containing ellipsis (...) in
3218 // the declaration) %al is used as hidden argument to specify the number
3219 // of SSE registers used. The contents of %al do not need to match exactly
3220 // the number of registers, but must be an ubound on the number of SSE
3221 // registers used and is in the range 0 - 8 inclusive.
3223 // Count the number of XMM registers allocated.
3224 static const MCPhysReg XMMArgRegs[] = {
3225 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3226 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3228 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3229 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3230 && "SSE registers cannot be used when SSE is disabled");
3232 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3233 DAG.getConstant(NumXMMRegs, dl,
3237 if (isVarArg && IsMustTail) {
3238 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3239 for (const auto &F : Forwards) {
3240 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3241 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3245 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3246 // don't need this because the eligibility check rejects calls that require
3247 // shuffling arguments passed in memory.
3248 if (!IsSibcall && isTailCall) {
3249 // Force all the incoming stack arguments to be loaded from the stack
3250 // before any new outgoing arguments are stored to the stack, because the
3251 // outgoing stack slots may alias the incoming argument stack slots, and
3252 // the alias isn't otherwise explicit. This is slightly more conservative
3253 // than necessary, because it means that each store effectively depends
3254 // on every argument instead of just those arguments it would clobber.
3255 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3257 SmallVector<SDValue, 8> MemOpChains2;
3260 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3261 CCValAssign &VA = ArgLocs[i];
3264 assert(VA.isMemLoc());
3265 SDValue Arg = OutVals[i];
3266 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3267 // Skip inalloca arguments. They don't require any work.
3268 if (Flags.isInAlloca())
3270 // Create frame index.
3271 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3272 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3273 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3274 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3276 if (Flags.isByVal()) {
3277 // Copy relative to framepointer.
3278 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3279 if (!StackPtr.getNode())
3280 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3281 getPointerTy(DAG.getDataLayout()));
3282 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3285 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3289 // Store relative to framepointer.
3290 MemOpChains2.push_back(DAG.getStore(
3291 ArgChain, dl, Arg, FIN,
3292 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3297 if (!MemOpChains2.empty())
3298 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3300 // Store the return address to the appropriate stack slot.
3301 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3302 getPointerTy(DAG.getDataLayout()),
3303 RegInfo->getSlotSize(), FPDiff, dl);
3306 // Build a sequence of copy-to-reg nodes chained together with token chain
3307 // and flag operands which copy the outgoing args into registers.
3309 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3310 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3311 RegsToPass[i].second, InFlag);
3312 InFlag = Chain.getValue(1);
3315 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3316 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3317 // In the 64-bit large code model, we have to make all calls
3318 // through a register, since the call instruction's 32-bit
3319 // pc-relative offset may not be large enough to hold the whole
3321 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3322 // If the callee is a GlobalAddress node (quite common, every direct call
3323 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3325 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3327 // We should use extra load for direct calls to dllimported functions in
3329 const GlobalValue *GV = G->getGlobal();
3330 if (!GV->hasDLLImportStorageClass()) {
3331 unsigned char OpFlags = 0;
3332 bool ExtraLoad = false;
3333 unsigned WrapperKind = ISD::DELETED_NODE;
3335 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3336 // external symbols most go through the PLT in PIC mode. If the symbol
3337 // has hidden or protected visibility, or if it is static or local, then
3338 // we don't need to use the PLT - we can directly call it.
3339 if (Subtarget->isTargetELF() &&
3340 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3341 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3342 OpFlags = X86II::MO_PLT;
3343 } else if (Subtarget->isPICStyleStubAny() &&
3344 !GV->isStrongDefinitionForLinker() &&
3345 (!Subtarget->getTargetTriple().isMacOSX() ||
3346 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3347 // PC-relative references to external symbols should go through $stub,
3348 // unless we're building with the leopard linker or later, which
3349 // automatically synthesizes these stubs.
3350 OpFlags = X86II::MO_DARWIN_STUB;
3351 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3352 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3353 // If the function is marked as non-lazy, generate an indirect call
3354 // which loads from the GOT directly. This avoids runtime overhead
3355 // at the cost of eager binding (and one extra byte of encoding).
3356 OpFlags = X86II::MO_GOTPCREL;
3357 WrapperKind = X86ISD::WrapperRIP;
3361 Callee = DAG.getTargetGlobalAddress(
3362 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3364 // Add a wrapper if needed.
3365 if (WrapperKind != ISD::DELETED_NODE)
3366 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3367 getPointerTy(DAG.getDataLayout()), Callee);
3368 // Add extra indirection if needed.
3370 Callee = DAG.getLoad(
3371 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3372 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3375 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3376 unsigned char OpFlags = 0;
3378 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3379 // external symbols should go through the PLT.
3380 if (Subtarget->isTargetELF() &&
3381 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3382 OpFlags = X86II::MO_PLT;
3383 } else if (Subtarget->isPICStyleStubAny() &&
3384 (!Subtarget->getTargetTriple().isMacOSX() ||
3385 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3386 // PC-relative references to external symbols should go through $stub,
3387 // unless we're building with the leopard linker or later, which
3388 // automatically synthesizes these stubs.
3389 OpFlags = X86II::MO_DARWIN_STUB;
3392 Callee = DAG.getTargetExternalSymbol(
3393 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3394 } else if (Subtarget->isTarget64BitILP32() &&
3395 Callee->getValueType(0) == MVT::i32) {
3396 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3397 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3400 // Returns a chain & a flag for retval copy to use.
3401 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3402 SmallVector<SDValue, 8> Ops;
3404 if (!IsSibcall && isTailCall) {
3405 Chain = DAG.getCALLSEQ_END(Chain,
3406 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3407 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3408 InFlag = Chain.getValue(1);
3411 Ops.push_back(Chain);
3412 Ops.push_back(Callee);
3415 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3417 // Add argument registers to the end of the list so that they are known live
3419 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3420 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3421 RegsToPass[i].second.getValueType()));
3423 // Add a register mask operand representing the call-preserved registers.
3424 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3425 assert(Mask && "Missing call preserved mask for calling convention");
3427 // If this is an invoke in a 32-bit function using a funclet-based
3428 // personality, assume the function clobbers all registers. If an exception
3429 // is thrown, the runtime will not restore CSRs.
3430 // FIXME: Model this more precisely so that we can register allocate across
3431 // the normal edge and spill and fill across the exceptional edge.
3432 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3433 const Function *CallerFn = MF.getFunction();
3434 EHPersonality Pers =
3435 CallerFn->hasPersonalityFn()
3436 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3437 : EHPersonality::Unknown;
3438 if (isFuncletEHPersonality(Pers))
3439 Mask = RegInfo->getNoPreservedMask();
3442 Ops.push_back(DAG.getRegisterMask(Mask));
3444 if (InFlag.getNode())
3445 Ops.push_back(InFlag);
3449 //// If this is the first return lowered for this function, add the regs
3450 //// to the liveout set for the function.
3451 // This isn't right, although it's probably harmless on x86; liveouts
3452 // should be computed from returns not tail calls. Consider a void
3453 // function making a tail call to a function returning int.
3454 MF.getFrameInfo()->setHasTailCall();
3455 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3458 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3459 InFlag = Chain.getValue(1);
3461 // Create the CALLSEQ_END node.
3462 unsigned NumBytesForCalleeToPop;
3463 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3464 DAG.getTarget().Options.GuaranteedTailCallOpt))
3465 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3466 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3467 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3468 SR == StackStructReturn)
3469 // If this is a call to a struct-return function, the callee
3470 // pops the hidden struct pointer, so we have to push it back.
3471 // This is common for Darwin/X86, Linux & Mingw32 targets.
3472 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3473 NumBytesForCalleeToPop = 4;
3475 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3477 // Returns a flag for retval copy to use.
3479 Chain = DAG.getCALLSEQ_END(Chain,
3480 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3481 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3484 InFlag = Chain.getValue(1);
3487 // Handle result values, copying them out of physregs into vregs that we
3489 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3490 Ins, dl, DAG, InVals);
3493 //===----------------------------------------------------------------------===//
3494 // Fast Calling Convention (tail call) implementation
3495 //===----------------------------------------------------------------------===//
3497 // Like std call, callee cleans arguments, convention except that ECX is
3498 // reserved for storing the tail called function address. Only 2 registers are
3499 // free for argument passing (inreg). Tail call optimization is performed
3501 // * tailcallopt is enabled
3502 // * caller/callee are fastcc
3503 // On X86_64 architecture with GOT-style position independent code only local
3504 // (within module) calls are supported at the moment.
3505 // To keep the stack aligned according to platform abi the function
3506 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3507 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3508 // If a tail called function callee has more arguments than the caller the
3509 // caller needs to make sure that there is room to move the RETADDR to. This is
3510 // achieved by reserving an area the size of the argument delta right after the
3511 // original RETADDR, but before the saved framepointer or the spilled registers
3512 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3524 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3527 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3528 SelectionDAG& DAG) const {
3529 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3530 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3531 unsigned StackAlignment = TFI.getStackAlignment();
3532 uint64_t AlignMask = StackAlignment - 1;
3533 int64_t Offset = StackSize;
3534 unsigned SlotSize = RegInfo->getSlotSize();
3535 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3536 // Number smaller than 12 so just add the difference.
3537 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3539 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3540 Offset = ((~AlignMask) & Offset) + StackAlignment +
3541 (StackAlignment-SlotSize);
3546 /// Return true if the given stack call argument is already available in the
3547 /// same position (relatively) of the caller's incoming argument stack.
3549 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3550 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3551 const X86InstrInfo *TII) {
3552 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3554 if (Arg.getOpcode() == ISD::CopyFromReg) {
3555 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3556 if (!TargetRegisterInfo::isVirtualRegister(VR))
3558 MachineInstr *Def = MRI->getVRegDef(VR);
3561 if (!Flags.isByVal()) {
3562 if (!TII->isLoadFromStackSlot(Def, FI))
3565 unsigned Opcode = Def->getOpcode();
3566 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3567 Opcode == X86::LEA64_32r) &&
3568 Def->getOperand(1).isFI()) {
3569 FI = Def->getOperand(1).getIndex();
3570 Bytes = Flags.getByValSize();
3574 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3575 if (Flags.isByVal())
3576 // ByVal argument is passed in as a pointer but it's now being
3577 // dereferenced. e.g.
3578 // define @foo(%struct.X* %A) {
3579 // tail call @bar(%struct.X* byval %A)
3582 SDValue Ptr = Ld->getBasePtr();
3583 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3586 FI = FINode->getIndex();
3587 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3588 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3589 FI = FINode->getIndex();
3590 Bytes = Flags.getByValSize();
3594 assert(FI != INT_MAX);
3595 if (!MFI->isFixedObjectIndex(FI))
3597 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3600 /// Check whether the call is eligible for tail call optimization. Targets
3601 /// that want to do tail call optimization should implement this function.
3602 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3603 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3604 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3605 const SmallVectorImpl<ISD::OutputArg> &Outs,
3606 const SmallVectorImpl<SDValue> &OutVals,
3607 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3608 if (!mayTailCallThisCC(CalleeCC))
3611 // If -tailcallopt is specified, make fastcc functions tail-callable.
3612 MachineFunction &MF = DAG.getMachineFunction();
3613 const Function *CallerF = MF.getFunction();
3615 // If the function return type is x86_fp80 and the callee return type is not,
3616 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3617 // perform a tailcall optimization here.
3618 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3621 CallingConv::ID CallerCC = CallerF->getCallingConv();
3622 bool CCMatch = CallerCC == CalleeCC;
3623 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3624 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3626 // Win64 functions have extra shadow space for argument homing. Don't do the
3627 // sibcall if the caller and callee have mismatched expectations for this
3629 if (IsCalleeWin64 != IsCallerWin64)
3632 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3633 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3638 // Look for obvious safe cases to perform tail call optimization that do not
3639 // require ABI changes. This is what gcc calls sibcall.
3641 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3642 // emit a special epilogue.
3643 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3644 if (RegInfo->needsStackRealignment(MF))
3647 // Also avoid sibcall optimization if either caller or callee uses struct
3648 // return semantics.
3649 if (isCalleeStructRet || isCallerStructRet)
3652 // Do not sibcall optimize vararg calls unless all arguments are passed via
3654 if (isVarArg && !Outs.empty()) {
3655 // Optimizing for varargs on Win64 is unlikely to be safe without
3656 // additional testing.
3657 if (IsCalleeWin64 || IsCallerWin64)
3660 SmallVector<CCValAssign, 16> ArgLocs;
3661 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3664 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3665 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3666 if (!ArgLocs[i].isRegLoc())
3670 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3671 // stack. Therefore, if it's not used by the call it is not safe to optimize
3672 // this into a sibcall.
3673 bool Unused = false;
3674 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3681 SmallVector<CCValAssign, 16> RVLocs;
3682 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3684 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3685 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3686 CCValAssign &VA = RVLocs[i];
3687 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3692 // If the calling conventions do not match, then we'd better make sure the
3693 // results are returned in the same way as what the caller expects.
3695 SmallVector<CCValAssign, 16> RVLocs1;
3696 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3698 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3700 SmallVector<CCValAssign, 16> RVLocs2;
3701 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3703 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3705 if (RVLocs1.size() != RVLocs2.size())
3707 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3708 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3710 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3712 if (RVLocs1[i].isRegLoc()) {
3713 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3716 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3722 unsigned StackArgsSize = 0;
3724 // If the callee takes no arguments then go on to check the results of the
3726 if (!Outs.empty()) {
3727 // Check if stack adjustment is needed. For now, do not do this if any
3728 // argument is passed on the stack.
3729 SmallVector<CCValAssign, 16> ArgLocs;
3730 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3733 // Allocate shadow area for Win64
3735 CCInfo.AllocateStack(32, 8);
3737 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3738 StackArgsSize = CCInfo.getNextStackOffset();
3740 if (CCInfo.getNextStackOffset()) {
3741 // Check if the arguments are already laid out in the right way as
3742 // the caller's fixed stack objects.
3743 MachineFrameInfo *MFI = MF.getFrameInfo();
3744 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3745 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3746 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3747 CCValAssign &VA = ArgLocs[i];
3748 SDValue Arg = OutVals[i];
3749 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3750 if (VA.getLocInfo() == CCValAssign::Indirect)
3752 if (!VA.isRegLoc()) {
3753 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3760 // If the tailcall address may be in a register, then make sure it's
3761 // possible to register allocate for it. In 32-bit, the call address can
3762 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3763 // callee-saved registers are restored. These happen to be the same
3764 // registers used to pass 'inreg' arguments so watch out for those.
3765 if (!Subtarget->is64Bit() &&
3766 ((!isa<GlobalAddressSDNode>(Callee) &&
3767 !isa<ExternalSymbolSDNode>(Callee)) ||
3768 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3769 unsigned NumInRegs = 0;
3770 // In PIC we need an extra register to formulate the address computation
3772 unsigned MaxInRegs =
3773 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3775 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3776 CCValAssign &VA = ArgLocs[i];
3779 unsigned Reg = VA.getLocReg();
3782 case X86::EAX: case X86::EDX: case X86::ECX:
3783 if (++NumInRegs == MaxInRegs)
3791 bool CalleeWillPop =
3792 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3793 MF.getTarget().Options.GuaranteedTailCallOpt);
3795 if (unsigned BytesToPop =
3796 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3797 // If we have bytes to pop, the callee must pop them.
3798 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3799 if (!CalleePopMatches)
3801 } else if (CalleeWillPop && StackArgsSize > 0) {
3802 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3810 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3811 const TargetLibraryInfo *libInfo) const {
3812 return X86::createFastISel(funcInfo, libInfo);
3815 //===----------------------------------------------------------------------===//
3816 // Other Lowering Hooks
3817 //===----------------------------------------------------------------------===//
3819 static bool MayFoldLoad(SDValue Op) {
3820 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3823 static bool MayFoldIntoStore(SDValue Op) {
3824 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3827 static bool isTargetShuffle(unsigned Opcode) {
3829 default: return false;
3830 case X86ISD::BLENDI:
3831 case X86ISD::PSHUFB:
3832 case X86ISD::PSHUFD:
3833 case X86ISD::PSHUFHW:
3834 case X86ISD::PSHUFLW:
3836 case X86ISD::PALIGNR:
3837 case X86ISD::MOVLHPS:
3838 case X86ISD::MOVLHPD:
3839 case X86ISD::MOVHLPS:
3840 case X86ISD::MOVLPS:
3841 case X86ISD::MOVLPD:
3842 case X86ISD::MOVSHDUP:
3843 case X86ISD::MOVSLDUP:
3844 case X86ISD::MOVDDUP:
3847 case X86ISD::UNPCKL:
3848 case X86ISD::UNPCKH:
3849 case X86ISD::VPERMILPI:
3850 case X86ISD::VPERM2X128:
3851 case X86ISD::VPERMI:
3852 case X86ISD::VPERMV:
3853 case X86ISD::VPERMV3:
3858 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3859 SDValue V1, unsigned TargetMask,
3860 SelectionDAG &DAG) {
3862 default: llvm_unreachable("Unknown x86 shuffle node");
3863 case X86ISD::PSHUFD:
3864 case X86ISD::PSHUFHW:
3865 case X86ISD::PSHUFLW:
3866 case X86ISD::VPERMILPI:
3867 case X86ISD::VPERMI:
3868 return DAG.getNode(Opc, dl, VT, V1,
3869 DAG.getConstant(TargetMask, dl, MVT::i8));
3873 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3874 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3876 default: llvm_unreachable("Unknown x86 shuffle node");
3877 case X86ISD::MOVLHPS:
3878 case X86ISD::MOVLHPD:
3879 case X86ISD::MOVHLPS:
3880 case X86ISD::MOVLPS:
3881 case X86ISD::MOVLPD:
3884 case X86ISD::UNPCKL:
3885 case X86ISD::UNPCKH:
3886 return DAG.getNode(Opc, dl, VT, V1, V2);
3890 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3891 MachineFunction &MF = DAG.getMachineFunction();
3892 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3893 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3894 int ReturnAddrIndex = FuncInfo->getRAIndex();
3896 if (ReturnAddrIndex == 0) {
3897 // Set up a frame object for the return address.
3898 unsigned SlotSize = RegInfo->getSlotSize();
3899 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3902 FuncInfo->setRAIndex(ReturnAddrIndex);
3905 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3908 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3909 bool hasSymbolicDisplacement) {
3910 // Offset should fit into 32 bit immediate field.
3911 if (!isInt<32>(Offset))
3914 // If we don't have a symbolic displacement - we don't have any extra
3916 if (!hasSymbolicDisplacement)
3919 // FIXME: Some tweaks might be needed for medium code model.
3920 if (M != CodeModel::Small && M != CodeModel::Kernel)
3923 // For small code model we assume that latest object is 16MB before end of 31
3924 // bits boundary. We may also accept pretty large negative constants knowing
3925 // that all objects are in the positive half of address space.
3926 if (M == CodeModel::Small && Offset < 16*1024*1024)
3929 // For kernel code model we know that all object resist in the negative half
3930 // of 32bits address space. We may not accept negative offsets, since they may
3931 // be just off and we may accept pretty large positive ones.
3932 if (M == CodeModel::Kernel && Offset >= 0)
3938 /// Determines whether the callee is required to pop its own arguments.
3939 /// Callee pop is necessary to support tail calls.
3940 bool X86::isCalleePop(CallingConv::ID CallingConv,
3941 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3942 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3943 // can guarantee TCO.
3944 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3947 switch (CallingConv) {
3950 case CallingConv::X86_StdCall:
3951 case CallingConv::X86_FastCall:
3952 case CallingConv::X86_ThisCall:
3953 case CallingConv::X86_VectorCall:
3958 /// \brief Return true if the condition is an unsigned comparison operation.
3959 static bool isX86CCUnsigned(unsigned X86CC) {
3961 default: llvm_unreachable("Invalid integer condition!");
3962 case X86::COND_E: return true;
3963 case X86::COND_G: return false;
3964 case X86::COND_GE: return false;
3965 case X86::COND_L: return false;
3966 case X86::COND_LE: return false;
3967 case X86::COND_NE: return true;
3968 case X86::COND_B: return true;
3969 case X86::COND_A: return true;
3970 case X86::COND_BE: return true;
3971 case X86::COND_AE: return true;
3975 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3976 /// condition code, returning the condition code and the LHS/RHS of the
3977 /// comparison to make.
3978 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3979 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3981 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3982 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3983 // X > -1 -> X == 0, jump !sign.
3984 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3985 return X86::COND_NS;
3987 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3988 // X < 0 -> X == 0, jump on sign.
3991 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3993 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3994 return X86::COND_LE;
3998 switch (SetCCOpcode) {
3999 default: llvm_unreachable("Invalid integer condition!");
4000 case ISD::SETEQ: return X86::COND_E;
4001 case ISD::SETGT: return X86::COND_G;
4002 case ISD::SETGE: return X86::COND_GE;
4003 case ISD::SETLT: return X86::COND_L;
4004 case ISD::SETLE: return X86::COND_LE;
4005 case ISD::SETNE: return X86::COND_NE;
4006 case ISD::SETULT: return X86::COND_B;
4007 case ISD::SETUGT: return X86::COND_A;
4008 case ISD::SETULE: return X86::COND_BE;
4009 case ISD::SETUGE: return X86::COND_AE;
4013 // First determine if it is required or is profitable to flip the operands.
4015 // If LHS is a foldable load, but RHS is not, flip the condition.
4016 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4017 !ISD::isNON_EXTLoad(RHS.getNode())) {
4018 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4019 std::swap(LHS, RHS);
4022 switch (SetCCOpcode) {
4028 std::swap(LHS, RHS);
4032 // On a floating point condition, the flags are set as follows:
4034 // 0 | 0 | 0 | X > Y
4035 // 0 | 0 | 1 | X < Y
4036 // 1 | 0 | 0 | X == Y
4037 // 1 | 1 | 1 | unordered
4038 switch (SetCCOpcode) {
4039 default: llvm_unreachable("Condcode should be pre-legalized away");
4041 case ISD::SETEQ: return X86::COND_E;
4042 case ISD::SETOLT: // flipped
4044 case ISD::SETGT: return X86::COND_A;
4045 case ISD::SETOLE: // flipped
4047 case ISD::SETGE: return X86::COND_AE;
4048 case ISD::SETUGT: // flipped
4050 case ISD::SETLT: return X86::COND_B;
4051 case ISD::SETUGE: // flipped
4053 case ISD::SETLE: return X86::COND_BE;
4055 case ISD::SETNE: return X86::COND_NE;
4056 case ISD::SETUO: return X86::COND_P;
4057 case ISD::SETO: return X86::COND_NP;
4059 case ISD::SETUNE: return X86::COND_INVALID;
4063 /// Is there a floating point cmov for the specific X86 condition code?
4064 /// Current x86 isa includes the following FP cmov instructions:
4065 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4066 static bool hasFPCMov(unsigned X86CC) {
4082 /// Returns true if the target can instruction select the
4083 /// specified FP immediate natively. If false, the legalizer will
4084 /// materialize the FP immediate as a load from a constant pool.
4085 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4086 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4087 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4093 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4094 ISD::LoadExtType ExtTy,
4096 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4097 // relocation target a movq or addq instruction: don't let the load shrink.
4098 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4099 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4100 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4101 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4105 /// \brief Returns true if it is beneficial to convert a load of a constant
4106 /// to just the constant itself.
4107 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4109 assert(Ty->isIntegerTy());
4111 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4112 if (BitSize == 0 || BitSize > 64)
4117 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4118 unsigned Index) const {
4119 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4122 return (Index == 0 || Index == ResVT.getVectorNumElements());
4125 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4126 // Speculate cttz only if we can directly use TZCNT.
4127 return Subtarget->hasBMI();
4130 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4131 // Speculate ctlz only if we can directly use LZCNT.
4132 return Subtarget->hasLZCNT();
4135 /// Return true if every element in Mask, beginning
4136 /// from position Pos and ending in Pos+Size is undef.
4137 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4138 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4144 /// Return true if Val is undef or if its value falls within the
4145 /// specified range (L, H].
4146 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4147 return (Val < 0) || (Val >= Low && Val < Hi);
4150 /// Val is either less than zero (undef) or equal to the specified value.
4151 static bool isUndefOrEqual(int Val, int CmpVal) {
4152 return (Val < 0 || Val == CmpVal);
4155 /// Return true if every element in Mask, beginning
4156 /// from position Pos and ending in Pos+Size, falls within the specified
4157 /// sequential range (Low, Low+Size]. or is undef.
4158 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4159 unsigned Pos, unsigned Size, int Low) {
4160 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4161 if (!isUndefOrEqual(Mask[i], Low))
4166 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4167 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4168 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4169 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4170 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4173 // The index should be aligned on a vecWidth-bit boundary.
4175 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4177 MVT VT = N->getSimpleValueType(0);
4178 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4179 bool Result = (Index * ElSize) % vecWidth == 0;
4184 /// Return true if the specified INSERT_SUBVECTOR
4185 /// operand specifies a subvector insert that is suitable for input to
4186 /// insertion of 128 or 256-bit subvectors
4187 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4188 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4189 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4191 // The index should be aligned on a vecWidth-bit boundary.
4193 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4195 MVT VT = N->getSimpleValueType(0);
4196 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4197 bool Result = (Index * ElSize) % vecWidth == 0;
4202 bool X86::isVINSERT128Index(SDNode *N) {
4203 return isVINSERTIndex(N, 128);
4206 bool X86::isVINSERT256Index(SDNode *N) {
4207 return isVINSERTIndex(N, 256);
4210 bool X86::isVEXTRACT128Index(SDNode *N) {
4211 return isVEXTRACTIndex(N, 128);
4214 bool X86::isVEXTRACT256Index(SDNode *N) {
4215 return isVEXTRACTIndex(N, 256);
4218 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4219 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4220 assert(isa<ConstantSDNode>(N->getOperand(1).getNode()) &&
4221 "Illegal extract subvector for VEXTRACT");
4224 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4226 MVT VecVT = N->getOperand(0).getSimpleValueType();
4227 MVT ElVT = VecVT.getVectorElementType();
4229 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4230 return Index / NumElemsPerChunk;
4233 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4234 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4235 assert(isa<ConstantSDNode>(N->getOperand(2).getNode()) &&
4236 "Illegal insert subvector for VINSERT");
4239 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4241 MVT VecVT = N->getSimpleValueType(0);
4242 MVT ElVT = VecVT.getVectorElementType();
4244 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4245 return Index / NumElemsPerChunk;
4248 /// Return the appropriate immediate to extract the specified
4249 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4250 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4251 return getExtractVEXTRACTImmediate(N, 128);
4254 /// Return the appropriate immediate to extract the specified
4255 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4256 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4257 return getExtractVEXTRACTImmediate(N, 256);
4260 /// Return the appropriate immediate to insert at the specified
4261 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4262 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4263 return getInsertVINSERTImmediate(N, 128);
4266 /// Return the appropriate immediate to insert at the specified
4267 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4268 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4269 return getInsertVINSERTImmediate(N, 256);
4272 /// Returns true if V is a constant integer zero.
4273 static bool isZero(SDValue V) {
4274 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4275 return C && C->isNullValue();
4278 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4279 bool X86::isZeroNode(SDValue Elt) {
4282 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4283 return CFP->getValueAPF().isPosZero();
4287 // Build a vector of constants
4288 // Use an UNDEF node if MaskElt == -1.
4289 // Spilt 64-bit constants in the 32-bit mode.
4290 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4292 SDLoc dl, bool IsMask = false) {
4294 SmallVector<SDValue, 32> Ops;
4297 MVT ConstVecVT = VT;
4298 unsigned NumElts = VT.getVectorNumElements();
4299 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4300 if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
4301 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4305 MVT EltVT = ConstVecVT.getVectorElementType();
4306 for (unsigned i = 0; i < NumElts; ++i) {
4307 bool IsUndef = Values[i] < 0 && IsMask;
4308 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4309 DAG.getConstant(Values[i], dl, EltVT);
4310 Ops.push_back(OpNode);
4312 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4313 DAG.getConstant(0, dl, EltVT));
4315 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4317 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4321 /// Returns a vector of specified type with all zero elements.
4322 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4323 SelectionDAG &DAG, SDLoc dl) {
4324 assert(VT.isVector() && "Expected a vector type");
4326 // Always build SSE zero vectors as <4 x i32> bitcasted
4327 // to their dest type. This ensures they get CSE'd.
4329 if (VT.is128BitVector()) { // SSE
4330 if (Subtarget->hasSSE2()) { // SSE2
4331 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4332 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4334 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4335 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4337 } else if (VT.is256BitVector()) { // AVX
4338 if (Subtarget->hasInt256()) { // AVX2
4339 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4340 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4341 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4343 // 256-bit logic and arithmetic instructions in AVX are all
4344 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4345 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4346 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4347 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4349 } else if (VT.is512BitVector()) { // AVX-512
4350 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4351 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4352 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4353 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4354 } else if (VT.getVectorElementType() == MVT::i1) {
4356 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4357 && "Unexpected vector type");
4358 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4359 && "Unexpected vector type");
4360 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4361 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4362 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4364 llvm_unreachable("Unexpected vector type");
4366 return DAG.getBitcast(VT, Vec);
4369 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4370 SelectionDAG &DAG, SDLoc dl,
4371 unsigned vectorWidth) {
4372 assert((vectorWidth == 128 || vectorWidth == 256) &&
4373 "Unsupported vector width");
4374 EVT VT = Vec.getValueType();
4375 EVT ElVT = VT.getVectorElementType();
4376 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4377 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4378 VT.getVectorNumElements()/Factor);
4380 // Extract from UNDEF is UNDEF.
4381 if (Vec.getOpcode() == ISD::UNDEF)
4382 return DAG.getUNDEF(ResultVT);
4384 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4385 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4386 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4388 // This is the index of the first element of the vectorWidth-bit chunk
4389 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4390 IdxVal &= ~(ElemsPerChunk - 1);
4392 // If the input is a buildvector just emit a smaller one.
4393 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4394 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4395 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4397 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4398 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4401 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4402 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4403 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4404 /// instructions or a simple subregister reference. Idx is an index in the
4405 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4406 /// lowering EXTRACT_VECTOR_ELT operations easier.
4407 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4408 SelectionDAG &DAG, SDLoc dl) {
4409 assert((Vec.getValueType().is256BitVector() ||
4410 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4411 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4414 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4415 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4416 SelectionDAG &DAG, SDLoc dl) {
4417 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4418 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4421 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4422 unsigned IdxVal, SelectionDAG &DAG,
4423 SDLoc dl, unsigned vectorWidth) {
4424 assert((vectorWidth == 128 || vectorWidth == 256) &&
4425 "Unsupported vector width");
4426 // Inserting UNDEF is Result
4427 if (Vec.getOpcode() == ISD::UNDEF)
4429 EVT VT = Vec.getValueType();
4430 EVT ElVT = VT.getVectorElementType();
4431 EVT ResultVT = Result.getValueType();
4433 // Insert the relevant vectorWidth bits.
4434 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4435 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4437 // This is the index of the first element of the vectorWidth-bit chunk
4438 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4439 IdxVal &= ~(ElemsPerChunk - 1);
4441 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4442 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4445 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4446 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4447 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4448 /// simple superregister reference. Idx is an index in the 128 bits
4449 /// we want. It need not be aligned to a 128-bit boundary. That makes
4450 /// lowering INSERT_VECTOR_ELT operations easier.
4451 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4452 SelectionDAG &DAG, SDLoc dl) {
4453 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4455 // For insertion into the zero index (low half) of a 256-bit vector, it is
4456 // more efficient to generate a blend with immediate instead of an insert*128.
4457 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4458 // extend the subvector to the size of the result vector. Make sure that
4459 // we are not recursing on that node by checking for undef here.
4460 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4461 Result.getOpcode() != ISD::UNDEF) {
4462 EVT ResultVT = Result.getValueType();
4463 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4464 SDValue Undef = DAG.getUNDEF(ResultVT);
4465 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4468 // The blend instruction, and therefore its mask, depend on the data type.
4469 MVT ScalarType = ResultVT.getVectorElementType().getSimpleVT();
4470 if (ScalarType.isFloatingPoint()) {
4471 // Choose either vblendps (float) or vblendpd (double).
4472 unsigned ScalarSize = ScalarType.getSizeInBits();
4473 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4474 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4475 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4476 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4479 const X86Subtarget &Subtarget =
4480 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4482 // AVX2 is needed for 256-bit integer blend support.
4483 // Integers must be cast to 32-bit because there is only vpblendd;
4484 // vpblendw can't be used for this because it has a handicapped mask.
4486 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4487 // is still more efficient than using the wrong domain vinsertf128 that
4488 // will be created by InsertSubVector().
4489 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4491 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4492 Vec256 = DAG.getBitcast(CastVT, Vec256);
4493 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4494 return DAG.getBitcast(ResultVT, Vec256);
4497 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4500 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4501 SelectionDAG &DAG, SDLoc dl) {
4502 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4503 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4506 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4507 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4508 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4509 /// large BUILD_VECTORS.
4510 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4511 unsigned NumElems, SelectionDAG &DAG,
4513 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4514 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4517 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4518 unsigned NumElems, SelectionDAG &DAG,
4520 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4521 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4524 /// Returns a vector of specified type with all bits set.
4525 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4526 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4527 /// Then bitcast to their original type, ensuring they get CSE'd.
4528 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4529 SelectionDAG &DAG, SDLoc dl) {
4530 assert(VT.isVector() && "Expected a vector type");
4532 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4534 if (VT.is512BitVector()) {
4535 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4536 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4537 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4538 } else if (VT.is256BitVector()) {
4539 if (Subtarget->hasInt256()) { // AVX2
4540 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4541 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4543 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4544 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4546 } else if (VT.is128BitVector()) {
4547 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4549 llvm_unreachable("Unexpected vector type");
4551 return DAG.getBitcast(VT, Vec);
4554 /// Returns a vector_shuffle node for an unpackl operation.
4555 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4557 unsigned NumElems = VT.getVectorNumElements();
4558 SmallVector<int, 8> Mask;
4559 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4561 Mask.push_back(i + NumElems);
4563 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4566 /// Returns a vector_shuffle node for an unpackh operation.
4567 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4569 unsigned NumElems = VT.getVectorNumElements();
4570 SmallVector<int, 8> Mask;
4571 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4572 Mask.push_back(i + Half);
4573 Mask.push_back(i + NumElems + Half);
4575 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4578 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4579 /// This produces a shuffle where the low element of V2 is swizzled into the
4580 /// zero/undef vector, landing at element Idx.
4581 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4582 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4584 const X86Subtarget *Subtarget,
4585 SelectionDAG &DAG) {
4586 MVT VT = V2.getSimpleValueType();
4588 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4589 unsigned NumElems = VT.getVectorNumElements();
4590 SmallVector<int, 16> MaskVec;
4591 for (unsigned i = 0; i != NumElems; ++i)
4592 // If this is the insertion idx, put the low elt of V2 here.
4593 MaskVec.push_back(i == Idx ? NumElems : i);
4594 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4597 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4598 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4599 /// uses one source. Note that this will set IsUnary for shuffles which use a
4600 /// single input multiple times, and in those cases it will
4601 /// adjust the mask to only have indices within that single input.
4602 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4603 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4604 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4605 unsigned NumElems = VT.getVectorNumElements();
4609 bool IsFakeUnary = false;
4610 switch(N->getOpcode()) {
4611 case X86ISD::BLENDI:
4612 ImmN = N->getOperand(N->getNumOperands()-1);
4613 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4616 ImmN = N->getOperand(N->getNumOperands()-1);
4617 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4618 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4620 case X86ISD::UNPCKH:
4621 DecodeUNPCKHMask(VT, Mask);
4622 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4624 case X86ISD::UNPCKL:
4625 DecodeUNPCKLMask(VT, Mask);
4626 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4628 case X86ISD::MOVHLPS:
4629 DecodeMOVHLPSMask(NumElems, Mask);
4630 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4632 case X86ISD::MOVLHPS:
4633 DecodeMOVLHPSMask(NumElems, Mask);
4634 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4636 case X86ISD::PALIGNR:
4637 ImmN = N->getOperand(N->getNumOperands()-1);
4638 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4640 case X86ISD::PSHUFD:
4641 case X86ISD::VPERMILPI:
4642 ImmN = N->getOperand(N->getNumOperands()-1);
4643 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4646 case X86ISD::PSHUFHW:
4647 ImmN = N->getOperand(N->getNumOperands()-1);
4648 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4651 case X86ISD::PSHUFLW:
4652 ImmN = N->getOperand(N->getNumOperands()-1);
4653 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4656 case X86ISD::PSHUFB: {
4658 SDValue MaskNode = N->getOperand(1);
4659 while (MaskNode->getOpcode() == ISD::BITCAST)
4660 MaskNode = MaskNode->getOperand(0);
4662 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4663 // If we have a build-vector, then things are easy.
4664 MVT VT = MaskNode.getSimpleValueType();
4665 assert(VT.isVector() &&
4666 "Can't produce a non-vector with a build_vector!");
4667 if (!VT.isInteger())
4670 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4672 SmallVector<uint64_t, 32> RawMask;
4673 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4674 SDValue Op = MaskNode->getOperand(i);
4675 if (Op->getOpcode() == ISD::UNDEF) {
4676 RawMask.push_back((uint64_t)SM_SentinelUndef);
4679 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4682 APInt MaskElement = CN->getAPIntValue();
4684 // We now have to decode the element which could be any integer size and
4685 // extract each byte of it.
4686 for (int j = 0; j < NumBytesPerElement; ++j) {
4687 // Note that this is x86 and so always little endian: the low byte is
4688 // the first byte of the mask.
4689 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4690 MaskElement = MaskElement.lshr(8);
4693 DecodePSHUFBMask(RawMask, Mask);
4697 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4701 SDValue Ptr = MaskLoad->getBasePtr();
4702 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4703 Ptr->getOpcode() == X86ISD::WrapperRIP)
4704 Ptr = Ptr->getOperand(0);
4706 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4707 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4710 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4711 DecodePSHUFBMask(C, Mask);
4719 case X86ISD::VPERMI:
4720 ImmN = N->getOperand(N->getNumOperands()-1);
4721 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4726 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4728 case X86ISD::VPERM2X128:
4729 ImmN = N->getOperand(N->getNumOperands()-1);
4730 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4731 if (Mask.empty()) return false;
4732 // Mask only contains negative index if an element is zero.
4733 if (std::any_of(Mask.begin(), Mask.end(),
4734 [](int M){ return M == SM_SentinelZero; }))
4737 case X86ISD::MOVSLDUP:
4738 DecodeMOVSLDUPMask(VT, Mask);
4741 case X86ISD::MOVSHDUP:
4742 DecodeMOVSHDUPMask(VT, Mask);
4745 case X86ISD::MOVDDUP:
4746 DecodeMOVDDUPMask(VT, Mask);
4749 case X86ISD::MOVLHPD:
4750 case X86ISD::MOVLPD:
4751 case X86ISD::MOVLPS:
4752 // Not yet implemented
4754 case X86ISD::VPERMV: {
4756 SDValue MaskNode = N->getOperand(0);
4757 while (MaskNode->getOpcode() == ISD::BITCAST)
4758 MaskNode = MaskNode->getOperand(0);
4760 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4761 SmallVector<uint64_t, 32> RawMask;
4762 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4763 // If we have a build-vector, then things are easy.
4764 assert(MaskNode.getSimpleValueType().isInteger() &&
4765 MaskNode.getSimpleValueType().getVectorNumElements() ==
4766 VT.getVectorNumElements());
4768 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4769 SDValue Op = MaskNode->getOperand(i);
4770 if (Op->getOpcode() == ISD::UNDEF)
4771 RawMask.push_back((uint64_t)SM_SentinelUndef);
4772 else if (isa<ConstantSDNode>(Op)) {
4773 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4774 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4778 DecodeVPERMVMask(RawMask, Mask);
4781 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4782 unsigned NumEltsInMask = MaskNode->getNumOperands();
4783 MaskNode = MaskNode->getOperand(0);
4784 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4786 APInt MaskEltValue = CN->getAPIntValue();
4787 for (unsigned i = 0; i < NumEltsInMask; ++i)
4788 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4789 DecodeVPERMVMask(RawMask, Mask);
4792 // It may be a scalar load
4795 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4799 SDValue Ptr = MaskLoad->getBasePtr();
4800 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4801 Ptr->getOpcode() == X86ISD::WrapperRIP)
4802 Ptr = Ptr->getOperand(0);
4804 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4805 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4808 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4810 DecodeVPERMVMask(C, VT, Mask);
4817 case X86ISD::VPERMV3: {
4819 SDValue MaskNode = N->getOperand(1);
4820 while (MaskNode->getOpcode() == ISD::BITCAST)
4821 MaskNode = MaskNode->getOperand(1);
4823 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4824 // If we have a build-vector, then things are easy.
4825 assert(MaskNode.getSimpleValueType().isInteger() &&
4826 MaskNode.getSimpleValueType().getVectorNumElements() ==
4827 VT.getVectorNumElements());
4829 SmallVector<uint64_t, 32> RawMask;
4830 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4832 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4833 SDValue Op = MaskNode->getOperand(i);
4834 if (Op->getOpcode() == ISD::UNDEF)
4835 RawMask.push_back((uint64_t)SM_SentinelUndef);
4837 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4840 APInt MaskElement = CN->getAPIntValue();
4841 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4844 DecodeVPERMV3Mask(RawMask, Mask);
4848 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4852 SDValue Ptr = MaskLoad->getBasePtr();
4853 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4854 Ptr->getOpcode() == X86ISD::WrapperRIP)
4855 Ptr = Ptr->getOperand(0);
4857 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4858 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4861 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4863 DecodeVPERMV3Mask(C, VT, Mask);
4870 default: llvm_unreachable("unknown target shuffle node");
4873 // If we have a fake unary shuffle, the shuffle mask is spread across two
4874 // inputs that are actually the same node. Re-map the mask to always point
4875 // into the first input.
4878 if (M >= (int)Mask.size())
4884 /// Returns the scalar element that will make up the ith
4885 /// element of the result of the vector shuffle.
4886 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4889 return SDValue(); // Limit search depth.
4891 SDValue V = SDValue(N, 0);
4892 EVT VT = V.getValueType();
4893 unsigned Opcode = V.getOpcode();
4895 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4896 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4897 int Elt = SV->getMaskElt(Index);
4900 return DAG.getUNDEF(VT.getVectorElementType());
4902 unsigned NumElems = VT.getVectorNumElements();
4903 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4904 : SV->getOperand(1);
4905 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4908 // Recurse into target specific vector shuffles to find scalars.
4909 if (isTargetShuffle(Opcode)) {
4910 MVT ShufVT = V.getSimpleValueType();
4911 unsigned NumElems = ShufVT.getVectorNumElements();
4912 SmallVector<int, 16> ShuffleMask;
4915 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4918 int Elt = ShuffleMask[Index];
4920 return DAG.getUNDEF(ShufVT.getVectorElementType());
4922 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4924 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4928 // Actual nodes that may contain scalar elements
4929 if (Opcode == ISD::BITCAST) {
4930 V = V.getOperand(0);
4931 EVT SrcVT = V.getValueType();
4932 unsigned NumElems = VT.getVectorNumElements();
4934 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4938 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4939 return (Index == 0) ? V.getOperand(0)
4940 : DAG.getUNDEF(VT.getVectorElementType());
4942 if (V.getOpcode() == ISD::BUILD_VECTOR)
4943 return V.getOperand(Index);
4948 /// Custom lower build_vector of v16i8.
4949 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4950 unsigned NumNonZero, unsigned NumZero,
4952 const X86Subtarget* Subtarget,
4953 const TargetLowering &TLI) {
4961 // SSE4.1 - use PINSRB to insert each byte directly.
4962 if (Subtarget->hasSSE41()) {
4963 for (unsigned i = 0; i < 16; ++i) {
4964 bool isNonZero = (NonZeros & (1 << i)) != 0;
4968 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4970 V = DAG.getUNDEF(MVT::v16i8);
4973 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4974 MVT::v16i8, V, Op.getOperand(i),
4975 DAG.getIntPtrConstant(i, dl));
4982 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4983 for (unsigned i = 0; i < 16; ++i) {
4984 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4985 if (ThisIsNonZero && First) {
4987 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4989 V = DAG.getUNDEF(MVT::v8i16);
4994 SDValue ThisElt, LastElt;
4995 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4996 if (LastIsNonZero) {
4997 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4998 MVT::i16, Op.getOperand(i-1));
5000 if (ThisIsNonZero) {
5001 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5002 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5003 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5005 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5009 if (ThisElt.getNode())
5010 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5011 DAG.getIntPtrConstant(i/2, dl));
5015 return DAG.getBitcast(MVT::v16i8, V);
5018 /// Custom lower build_vector of v8i16.
5019 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5020 unsigned NumNonZero, unsigned NumZero,
5022 const X86Subtarget* Subtarget,
5023 const TargetLowering &TLI) {
5030 for (unsigned i = 0; i < 8; ++i) {
5031 bool isNonZero = (NonZeros & (1 << i)) != 0;
5035 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5037 V = DAG.getUNDEF(MVT::v8i16);
5040 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5041 MVT::v8i16, V, Op.getOperand(i),
5042 DAG.getIntPtrConstant(i, dl));
5049 /// Custom lower build_vector of v4i32 or v4f32.
5050 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5051 const X86Subtarget *Subtarget,
5052 const TargetLowering &TLI) {
5053 // Find all zeroable elements.
5054 std::bitset<4> Zeroable;
5055 for (int i=0; i < 4; ++i) {
5056 SDValue Elt = Op->getOperand(i);
5057 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5059 assert(Zeroable.size() - Zeroable.count() > 1 &&
5060 "We expect at least two non-zero elements!");
5062 // We only know how to deal with build_vector nodes where elements are either
5063 // zeroable or extract_vector_elt with constant index.
5064 SDValue FirstNonZero;
5065 unsigned FirstNonZeroIdx;
5066 for (unsigned i=0; i < 4; ++i) {
5069 SDValue Elt = Op->getOperand(i);
5070 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5071 !isa<ConstantSDNode>(Elt.getOperand(1)))
5073 // Make sure that this node is extracting from a 128-bit vector.
5074 MVT VT = Elt.getOperand(0).getSimpleValueType();
5075 if (!VT.is128BitVector())
5077 if (!FirstNonZero.getNode()) {
5079 FirstNonZeroIdx = i;
5083 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5084 SDValue V1 = FirstNonZero.getOperand(0);
5085 MVT VT = V1.getSimpleValueType();
5087 // See if this build_vector can be lowered as a blend with zero.
5089 unsigned EltMaskIdx, EltIdx;
5091 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5092 if (Zeroable[EltIdx]) {
5093 // The zero vector will be on the right hand side.
5094 Mask[EltIdx] = EltIdx+4;
5098 Elt = Op->getOperand(EltIdx);
5099 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5100 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5101 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5103 Mask[EltIdx] = EltIdx;
5107 // Let the shuffle legalizer deal with blend operations.
5108 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5109 if (V1.getSimpleValueType() != VT)
5110 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5111 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5114 // See if we can lower this build_vector to a INSERTPS.
5115 if (!Subtarget->hasSSE41())
5118 SDValue V2 = Elt.getOperand(0);
5119 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5122 bool CanFold = true;
5123 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5127 SDValue Current = Op->getOperand(i);
5128 SDValue SrcVector = Current->getOperand(0);
5131 CanFold = SrcVector == V1 &&
5132 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5138 assert(V1.getNode() && "Expected at least two non-zero elements!");
5139 if (V1.getSimpleValueType() != MVT::v4f32)
5140 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5141 if (V2.getSimpleValueType() != MVT::v4f32)
5142 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5144 // Ok, we can emit an INSERTPS instruction.
5145 unsigned ZMask = Zeroable.to_ulong();
5147 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5148 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5150 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5151 DAG.getIntPtrConstant(InsertPSMask, DL));
5152 return DAG.getBitcast(VT, Result);
5155 /// Return a vector logical shift node.
5156 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5157 unsigned NumBits, SelectionDAG &DAG,
5158 const TargetLowering &TLI, SDLoc dl) {
5159 assert(VT.is128BitVector() && "Unknown type for VShift");
5160 MVT ShVT = MVT::v2i64;
5161 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5162 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5163 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5164 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5165 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5166 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5170 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5172 // Check if the scalar load can be widened into a vector load. And if
5173 // the address is "base + cst" see if the cst can be "absorbed" into
5174 // the shuffle mask.
5175 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5176 SDValue Ptr = LD->getBasePtr();
5177 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5179 EVT PVT = LD->getValueType(0);
5180 if (PVT != MVT::i32 && PVT != MVT::f32)
5185 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5186 FI = FINode->getIndex();
5188 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5189 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5190 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5191 Offset = Ptr.getConstantOperandVal(1);
5192 Ptr = Ptr.getOperand(0);
5197 // FIXME: 256-bit vector instructions don't require a strict alignment,
5198 // improve this code to support it better.
5199 unsigned RequiredAlign = VT.getSizeInBits()/8;
5200 SDValue Chain = LD->getChain();
5201 // Make sure the stack object alignment is at least 16 or 32.
5202 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5203 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5204 if (MFI->isFixedObjectIndex(FI)) {
5205 // Can't change the alignment. FIXME: It's possible to compute
5206 // the exact stack offset and reference FI + adjust offset instead.
5207 // If someone *really* cares about this. That's the way to implement it.
5210 MFI->setObjectAlignment(FI, RequiredAlign);
5214 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5215 // Ptr + (Offset & ~15).
5218 if ((Offset % RequiredAlign) & 3)
5220 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5223 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5224 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5227 int EltNo = (Offset - StartOffset) >> 2;
5228 unsigned NumElems = VT.getVectorNumElements();
5230 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5231 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5232 LD->getPointerInfo().getWithOffset(StartOffset),
5233 false, false, false, 0);
5235 SmallVector<int, 8> Mask(NumElems, EltNo);
5237 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5243 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5244 /// elements can be replaced by a single large load which has the same value as
5245 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5247 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5249 /// FIXME: we'd also like to handle the case where the last elements are zero
5250 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5251 /// There's even a handy isZeroNode for that purpose.
5252 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5253 SDLoc &DL, SelectionDAG &DAG,
5254 bool isAfterLegalize) {
5255 unsigned NumElems = Elts.size();
5257 LoadSDNode *LDBase = nullptr;
5258 unsigned LastLoadedElt = -1U;
5260 // For each element in the initializer, see if we've found a load or an undef.
5261 // If we don't find an initial load element, or later load elements are
5262 // non-consecutive, bail out.
5263 for (unsigned i = 0; i < NumElems; ++i) {
5264 SDValue Elt = Elts[i];
5265 // Look through a bitcast.
5266 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5267 Elt = Elt.getOperand(0);
5268 if (!Elt.getNode() ||
5269 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5272 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5274 LDBase = cast<LoadSDNode>(Elt.getNode());
5278 if (Elt.getOpcode() == ISD::UNDEF)
5281 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5282 EVT LdVT = Elt.getValueType();
5283 // Each loaded element must be the correct fractional portion of the
5284 // requested vector load.
5285 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5287 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5292 // If we have found an entire vector of loads and undefs, then return a large
5293 // load of the entire vector width starting at the base pointer. If we found
5294 // consecutive loads for the low half, generate a vzext_load node.
5295 if (LastLoadedElt == NumElems - 1) {
5296 assert(LDBase && "Did not find base load for merging consecutive loads");
5297 EVT EltVT = LDBase->getValueType(0);
5298 // Ensure that the input vector size for the merged loads matches the
5299 // cumulative size of the input elements.
5300 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5303 if (isAfterLegalize &&
5304 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5307 SDValue NewLd = SDValue();
5309 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5310 LDBase->getPointerInfo(), LDBase->isVolatile(),
5311 LDBase->isNonTemporal(), LDBase->isInvariant(),
5312 LDBase->getAlignment());
5314 if (LDBase->hasAnyUseOfValue(1)) {
5315 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5317 SDValue(NewLd.getNode(), 1));
5318 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5319 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5320 SDValue(NewLd.getNode(), 1));
5326 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5327 //of a v4i32 / v4f32. It's probably worth generalizing.
5328 EVT EltVT = VT.getVectorElementType();
5329 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5330 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5331 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5332 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5334 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5335 LDBase->getPointerInfo(),
5336 LDBase->getAlignment(),
5337 false/*isVolatile*/, true/*ReadMem*/,
5340 // Make sure the newly-created LOAD is in the same position as LDBase in
5341 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5342 // update uses of LDBase's output chain to use the TokenFactor.
5343 if (LDBase->hasAnyUseOfValue(1)) {
5344 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5345 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5346 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5347 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5348 SDValue(ResNode.getNode(), 1));
5351 return DAG.getBitcast(VT, ResNode);
5356 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5357 /// to generate a splat value for the following cases:
5358 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5359 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5360 /// a scalar load, or a constant.
5361 /// The VBROADCAST node is returned when a pattern is found,
5362 /// or SDValue() otherwise.
5363 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5364 SelectionDAG &DAG) {
5365 // VBROADCAST requires AVX.
5366 // TODO: Splats could be generated for non-AVX CPUs using SSE
5367 // instructions, but there's less potential gain for only 128-bit vectors.
5368 if (!Subtarget->hasAVX())
5371 MVT VT = Op.getSimpleValueType();
5374 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5375 "Unsupported vector type for broadcast.");
5380 switch (Op.getOpcode()) {
5382 // Unknown pattern found.
5385 case ISD::BUILD_VECTOR: {
5386 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5387 BitVector UndefElements;
5388 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5390 // We need a splat of a single value to use broadcast, and it doesn't
5391 // make any sense if the value is only in one element of the vector.
5392 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5396 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5397 Ld.getOpcode() == ISD::ConstantFP);
5399 // Make sure that all of the users of a non-constant load are from the
5400 // BUILD_VECTOR node.
5401 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5406 case ISD::VECTOR_SHUFFLE: {
5407 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5409 // Shuffles must have a splat mask where the first element is
5411 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5414 SDValue Sc = Op.getOperand(0);
5415 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5416 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5418 if (!Subtarget->hasInt256())
5421 // Use the register form of the broadcast instruction available on AVX2.
5422 if (VT.getSizeInBits() >= 256)
5423 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5424 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5427 Ld = Sc.getOperand(0);
5428 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5429 Ld.getOpcode() == ISD::ConstantFP);
5431 // The scalar_to_vector node and the suspected
5432 // load node must have exactly one user.
5433 // Constants may have multiple users.
5435 // AVX-512 has register version of the broadcast
5436 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5437 Ld.getValueType().getSizeInBits() >= 32;
5438 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5445 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5446 bool IsGE256 = (VT.getSizeInBits() >= 256);
5448 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5449 // instruction to save 8 or more bytes of constant pool data.
5450 // TODO: If multiple splats are generated to load the same constant,
5451 // it may be detrimental to overall size. There needs to be a way to detect
5452 // that condition to know if this is truly a size win.
5453 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5455 // Handle broadcasting a single constant scalar from the constant pool
5457 // On Sandybridge (no AVX2), it is still better to load a constant vector
5458 // from the constant pool and not to broadcast it from a scalar.
5459 // But override that restriction when optimizing for size.
5460 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5461 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5462 EVT CVT = Ld.getValueType();
5463 assert(!CVT.isVector() && "Must not broadcast a vector type");
5465 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5466 // For size optimization, also splat v2f64 and v2i64, and for size opt
5467 // with AVX2, also splat i8 and i16.
5468 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5469 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5470 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5471 const Constant *C = nullptr;
5472 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5473 C = CI->getConstantIntValue();
5474 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5475 C = CF->getConstantFPValue();
5477 assert(C && "Invalid constant type");
5479 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5481 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5482 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5484 CVT, dl, DAG.getEntryNode(), CP,
5485 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5486 false, false, Alignment);
5488 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5492 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5494 // Handle AVX2 in-register broadcasts.
5495 if (!IsLoad && Subtarget->hasInt256() &&
5496 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5497 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5499 // The scalar source must be a normal load.
5503 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5504 (Subtarget->hasVLX() && ScalarSize == 64))
5505 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5507 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5508 // double since there is no vbroadcastsd xmm
5509 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5510 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5511 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5514 // Unsupported broadcast.
5518 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5519 /// underlying vector and index.
5521 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5523 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5525 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5526 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5529 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5531 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5533 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5534 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5537 // In this case the vector is the extract_subvector expression and the index
5538 // is 2, as specified by the shuffle.
5539 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5540 SDValue ShuffleVec = SVOp->getOperand(0);
5541 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5542 assert(ShuffleVecVT.getVectorElementType() ==
5543 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5545 int ShuffleIdx = SVOp->getMaskElt(Idx);
5546 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5547 ExtractedFromVec = ShuffleVec;
5553 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5554 MVT VT = Op.getSimpleValueType();
5556 // Skip if insert_vec_elt is not supported.
5557 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5558 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5562 unsigned NumElems = Op.getNumOperands();
5566 SmallVector<unsigned, 4> InsertIndices;
5567 SmallVector<int, 8> Mask(NumElems, -1);
5569 for (unsigned i = 0; i != NumElems; ++i) {
5570 unsigned Opc = Op.getOperand(i).getOpcode();
5572 if (Opc == ISD::UNDEF)
5575 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5576 // Quit if more than 1 elements need inserting.
5577 if (InsertIndices.size() > 1)
5580 InsertIndices.push_back(i);
5584 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5585 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5586 // Quit if non-constant index.
5587 if (!isa<ConstantSDNode>(ExtIdx))
5589 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5591 // Quit if extracted from vector of different type.
5592 if (ExtractedFromVec.getValueType() != VT)
5595 if (!VecIn1.getNode())
5596 VecIn1 = ExtractedFromVec;
5597 else if (VecIn1 != ExtractedFromVec) {
5598 if (!VecIn2.getNode())
5599 VecIn2 = ExtractedFromVec;
5600 else if (VecIn2 != ExtractedFromVec)
5601 // Quit if more than 2 vectors to shuffle
5605 if (ExtractedFromVec == VecIn1)
5607 else if (ExtractedFromVec == VecIn2)
5608 Mask[i] = Idx + NumElems;
5611 if (!VecIn1.getNode())
5614 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5615 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5616 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5617 unsigned Idx = InsertIndices[i];
5618 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5619 DAG.getIntPtrConstant(Idx, DL));
5625 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5626 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5627 Op.getScalarValueSizeInBits() == 1 &&
5628 "Can not convert non-constant vector");
5629 uint64_t Immediate = 0;
5630 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5631 SDValue In = Op.getOperand(idx);
5632 if (In.getOpcode() != ISD::UNDEF)
5633 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5637 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5638 return DAG.getConstant(Immediate, dl, VT);
5640 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5642 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5644 MVT VT = Op.getSimpleValueType();
5645 assert((VT.getVectorElementType() == MVT::i1) &&
5646 "Unexpected type in LowerBUILD_VECTORvXi1!");
5649 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5650 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5651 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5652 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5655 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5656 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5657 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5658 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5661 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5662 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5663 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5664 return DAG.getBitcast(VT, Imm);
5665 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5666 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5667 DAG.getIntPtrConstant(0, dl));
5670 // Vector has one or more non-const elements
5671 uint64_t Immediate = 0;
5672 SmallVector<unsigned, 16> NonConstIdx;
5673 bool IsSplat = true;
5674 bool HasConstElts = false;
5676 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5677 SDValue In = Op.getOperand(idx);
5678 if (In.getOpcode() == ISD::UNDEF)
5680 if (!isa<ConstantSDNode>(In))
5681 NonConstIdx.push_back(idx);
5683 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5684 HasConstElts = true;
5688 else if (In != Op.getOperand(SplatIdx))
5692 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5694 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5695 DAG.getConstant(1, dl, VT),
5696 DAG.getConstant(0, dl, VT));
5698 // insert elements one by one
5702 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5703 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5705 else if (HasConstElts)
5706 Imm = DAG.getConstant(0, dl, VT);
5708 Imm = DAG.getUNDEF(VT);
5709 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5710 DstVec = DAG.getBitcast(VT, Imm);
5712 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5713 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5714 DAG.getIntPtrConstant(0, dl));
5717 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5718 unsigned InsertIdx = NonConstIdx[i];
5719 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5720 Op.getOperand(InsertIdx),
5721 DAG.getIntPtrConstant(InsertIdx, dl));
5726 /// \brief Return true if \p N implements a horizontal binop and return the
5727 /// operands for the horizontal binop into V0 and V1.
5729 /// This is a helper function of LowerToHorizontalOp().
5730 /// This function checks that the build_vector \p N in input implements a
5731 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5732 /// operation to match.
5733 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5734 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5735 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5738 /// This function only analyzes elements of \p N whose indices are
5739 /// in range [BaseIdx, LastIdx).
5740 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5742 unsigned BaseIdx, unsigned LastIdx,
5743 SDValue &V0, SDValue &V1) {
5744 EVT VT = N->getValueType(0);
5746 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5747 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5748 "Invalid Vector in input!");
5750 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5751 bool CanFold = true;
5752 unsigned ExpectedVExtractIdx = BaseIdx;
5753 unsigned NumElts = LastIdx - BaseIdx;
5754 V0 = DAG.getUNDEF(VT);
5755 V1 = DAG.getUNDEF(VT);
5757 // Check if N implements a horizontal binop.
5758 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5759 SDValue Op = N->getOperand(i + BaseIdx);
5762 if (Op->getOpcode() == ISD::UNDEF) {
5763 // Update the expected vector extract index.
5764 if (i * 2 == NumElts)
5765 ExpectedVExtractIdx = BaseIdx;
5766 ExpectedVExtractIdx += 2;
5770 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5775 SDValue Op0 = Op.getOperand(0);
5776 SDValue Op1 = Op.getOperand(1);
5778 // Try to match the following pattern:
5779 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5780 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5781 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5782 Op0.getOperand(0) == Op1.getOperand(0) &&
5783 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5784 isa<ConstantSDNode>(Op1.getOperand(1)));
5788 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5789 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5791 if (i * 2 < NumElts) {
5792 if (V0.getOpcode() == ISD::UNDEF) {
5793 V0 = Op0.getOperand(0);
5794 if (V0.getValueType() != VT)
5798 if (V1.getOpcode() == ISD::UNDEF) {
5799 V1 = Op0.getOperand(0);
5800 if (V1.getValueType() != VT)
5803 if (i * 2 == NumElts)
5804 ExpectedVExtractIdx = BaseIdx;
5807 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5808 if (I0 == ExpectedVExtractIdx)
5809 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5810 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5811 // Try to match the following dag sequence:
5812 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5813 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5817 ExpectedVExtractIdx += 2;
5823 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5824 /// a concat_vector.
5826 /// This is a helper function of LowerToHorizontalOp().
5827 /// This function expects two 256-bit vectors called V0 and V1.
5828 /// At first, each vector is split into two separate 128-bit vectors.
5829 /// Then, the resulting 128-bit vectors are used to implement two
5830 /// horizontal binary operations.
5832 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5834 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5835 /// the two new horizontal binop.
5836 /// When Mode is set, the first horizontal binop dag node would take as input
5837 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5838 /// horizontal binop dag node would take as input the lower 128-bit of V1
5839 /// and the upper 128-bit of V1.
5841 /// HADD V0_LO, V0_HI
5842 /// HADD V1_LO, V1_HI
5844 /// Otherwise, the first horizontal binop dag node takes as input the lower
5845 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5846 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5848 /// HADD V0_LO, V1_LO
5849 /// HADD V0_HI, V1_HI
5851 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5852 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5853 /// the upper 128-bits of the result.
5854 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5855 SDLoc DL, SelectionDAG &DAG,
5856 unsigned X86Opcode, bool Mode,
5857 bool isUndefLO, bool isUndefHI) {
5858 EVT VT = V0.getValueType();
5859 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5860 "Invalid nodes in input!");
5862 unsigned NumElts = VT.getVectorNumElements();
5863 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5864 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5865 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5866 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5867 EVT NewVT = V0_LO.getValueType();
5869 SDValue LO = DAG.getUNDEF(NewVT);
5870 SDValue HI = DAG.getUNDEF(NewVT);
5873 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5874 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5875 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5876 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5877 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5879 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5880 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5881 V1_LO->getOpcode() != ISD::UNDEF))
5882 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5884 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5885 V1_HI->getOpcode() != ISD::UNDEF))
5886 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5889 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5892 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5894 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5895 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5896 MVT VT = BV->getSimpleValueType(0);
5897 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5898 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5902 unsigned NumElts = VT.getVectorNumElements();
5903 SDValue InVec0 = DAG.getUNDEF(VT);
5904 SDValue InVec1 = DAG.getUNDEF(VT);
5906 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5907 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5909 // Odd-numbered elements in the input build vector are obtained from
5910 // adding two integer/float elements.
5911 // Even-numbered elements in the input build vector are obtained from
5912 // subtracting two integer/float elements.
5913 unsigned ExpectedOpcode = ISD::FSUB;
5914 unsigned NextExpectedOpcode = ISD::FADD;
5915 bool AddFound = false;
5916 bool SubFound = false;
5918 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5919 SDValue Op = BV->getOperand(i);
5921 // Skip 'undef' values.
5922 unsigned Opcode = Op.getOpcode();
5923 if (Opcode == ISD::UNDEF) {
5924 std::swap(ExpectedOpcode, NextExpectedOpcode);
5928 // Early exit if we found an unexpected opcode.
5929 if (Opcode != ExpectedOpcode)
5932 SDValue Op0 = Op.getOperand(0);
5933 SDValue Op1 = Op.getOperand(1);
5935 // Try to match the following pattern:
5936 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5937 // Early exit if we cannot match that sequence.
5938 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5939 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5940 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5941 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5942 Op0.getOperand(1) != Op1.getOperand(1))
5945 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5949 // We found a valid add/sub node. Update the information accordingly.
5955 // Update InVec0 and InVec1.
5956 if (InVec0.getOpcode() == ISD::UNDEF) {
5957 InVec0 = Op0.getOperand(0);
5958 if (InVec0.getSimpleValueType() != VT)
5961 if (InVec1.getOpcode() == ISD::UNDEF) {
5962 InVec1 = Op1.getOperand(0);
5963 if (InVec1.getSimpleValueType() != VT)
5967 // Make sure that operands in input to each add/sub node always
5968 // come from a same pair of vectors.
5969 if (InVec0 != Op0.getOperand(0)) {
5970 if (ExpectedOpcode == ISD::FSUB)
5973 // FADD is commutable. Try to commute the operands
5974 // and then test again.
5975 std::swap(Op0, Op1);
5976 if (InVec0 != Op0.getOperand(0))
5980 if (InVec1 != Op1.getOperand(0))
5983 // Update the pair of expected opcodes.
5984 std::swap(ExpectedOpcode, NextExpectedOpcode);
5987 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5988 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5989 InVec1.getOpcode() != ISD::UNDEF)
5990 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5995 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5996 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5997 const X86Subtarget *Subtarget,
5998 SelectionDAG &DAG) {
5999 MVT VT = BV->getSimpleValueType(0);
6000 unsigned NumElts = VT.getVectorNumElements();
6001 unsigned NumUndefsLO = 0;
6002 unsigned NumUndefsHI = 0;
6003 unsigned Half = NumElts/2;
6005 // Count the number of UNDEF operands in the build_vector in input.
6006 for (unsigned i = 0, e = Half; i != e; ++i)
6007 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6010 for (unsigned i = Half, e = NumElts; i != e; ++i)
6011 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6014 // Early exit if this is either a build_vector of all UNDEFs or all the
6015 // operands but one are UNDEF.
6016 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6020 SDValue InVec0, InVec1;
6021 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6022 // Try to match an SSE3 float HADD/HSUB.
6023 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6024 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6026 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6027 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6028 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6029 // Try to match an SSSE3 integer HADD/HSUB.
6030 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6031 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6033 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6034 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6037 if (!Subtarget->hasAVX())
6040 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6041 // Try to match an AVX horizontal add/sub of packed single/double
6042 // precision floating point values from 256-bit vectors.
6043 SDValue InVec2, InVec3;
6044 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6045 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6046 ((InVec0.getOpcode() == ISD::UNDEF ||
6047 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6048 ((InVec1.getOpcode() == ISD::UNDEF ||
6049 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6050 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6052 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6053 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6054 ((InVec0.getOpcode() == ISD::UNDEF ||
6055 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6056 ((InVec1.getOpcode() == ISD::UNDEF ||
6057 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6058 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6059 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6060 // Try to match an AVX2 horizontal add/sub of signed integers.
6061 SDValue InVec2, InVec3;
6063 bool CanFold = true;
6065 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6066 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6067 ((InVec0.getOpcode() == ISD::UNDEF ||
6068 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6069 ((InVec1.getOpcode() == ISD::UNDEF ||
6070 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6071 X86Opcode = X86ISD::HADD;
6072 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6073 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6074 ((InVec0.getOpcode() == ISD::UNDEF ||
6075 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6076 ((InVec1.getOpcode() == ISD::UNDEF ||
6077 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6078 X86Opcode = X86ISD::HSUB;
6083 // Fold this build_vector into a single horizontal add/sub.
6084 // Do this only if the target has AVX2.
6085 if (Subtarget->hasAVX2())
6086 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6088 // Do not try to expand this build_vector into a pair of horizontal
6089 // add/sub if we can emit a pair of scalar add/sub.
6090 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6093 // Convert this build_vector into a pair of horizontal binop followed by
6095 bool isUndefLO = NumUndefsLO == Half;
6096 bool isUndefHI = NumUndefsHI == Half;
6097 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6098 isUndefLO, isUndefHI);
6102 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6103 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6105 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6106 X86Opcode = X86ISD::HADD;
6107 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6108 X86Opcode = X86ISD::HSUB;
6109 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6110 X86Opcode = X86ISD::FHADD;
6111 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6112 X86Opcode = X86ISD::FHSUB;
6116 // Don't try to expand this build_vector into a pair of horizontal add/sub
6117 // if we can simply emit a pair of scalar add/sub.
6118 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6121 // Convert this build_vector into two horizontal add/sub followed by
6123 bool isUndefLO = NumUndefsLO == Half;
6124 bool isUndefHI = NumUndefsHI == Half;
6125 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6126 isUndefLO, isUndefHI);
6133 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6136 MVT VT = Op.getSimpleValueType();
6137 MVT ExtVT = VT.getVectorElementType();
6138 unsigned NumElems = Op.getNumOperands();
6140 // Generate vectors for predicate vectors.
6141 if (VT.getVectorElementType() == MVT::i1 && Subtarget->hasAVX512())
6142 return LowerBUILD_VECTORvXi1(Op, DAG);
6144 // Vectors containing all zeros can be matched by pxor and xorps later
6145 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6146 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6147 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6148 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6151 return getZeroVector(VT, Subtarget, DAG, dl);
6154 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6155 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6156 // vpcmpeqd on 256-bit vectors.
6157 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6158 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6161 if (!VT.is512BitVector())
6162 return getOnesVector(VT, Subtarget, DAG, dl);
6165 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6166 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6168 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6169 return HorizontalOp;
6170 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6173 unsigned EVTBits = ExtVT.getSizeInBits();
6175 unsigned NumZero = 0;
6176 unsigned NumNonZero = 0;
6177 unsigned NonZeros = 0;
6178 bool IsAllConstants = true;
6179 SmallSet<SDValue, 8> Values;
6180 for (unsigned i = 0; i < NumElems; ++i) {
6181 SDValue Elt = Op.getOperand(i);
6182 if (Elt.getOpcode() == ISD::UNDEF)
6185 if (Elt.getOpcode() != ISD::Constant &&
6186 Elt.getOpcode() != ISD::ConstantFP)
6187 IsAllConstants = false;
6188 if (X86::isZeroNode(Elt))
6191 NonZeros |= (1 << i);
6196 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6197 if (NumNonZero == 0)
6198 return DAG.getUNDEF(VT);
6200 // Special case for single non-zero, non-undef, element.
6201 if (NumNonZero == 1) {
6202 unsigned Idx = countTrailingZeros(NonZeros);
6203 SDValue Item = Op.getOperand(Idx);
6205 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6206 // the value are obviously zero, truncate the value to i32 and do the
6207 // insertion that way. Only do this if the value is non-constant or if the
6208 // value is a constant being inserted into element 0. It is cheaper to do
6209 // a constant pool load than it is to do a movd + shuffle.
6210 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6211 (!IsAllConstants || Idx == 0)) {
6212 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6214 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6215 MVT VecVT = MVT::v4i32;
6217 // Truncate the value (which may itself be a constant) to i32, and
6218 // convert it to a vector with movd (S2V+shuffle to zero extend).
6219 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6220 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6221 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6222 Item, Idx * 2, true, Subtarget, DAG));
6226 // If we have a constant or non-constant insertion into the low element of
6227 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6228 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6229 // depending on what the source datatype is.
6232 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6234 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6235 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6236 if (VT.is512BitVector()) {
6237 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6238 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6239 Item, DAG.getIntPtrConstant(0, dl));
6241 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6242 "Expected an SSE value type!");
6243 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6244 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6245 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6248 // We can't directly insert an i8 or i16 into a vector, so zero extend
6250 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6251 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6252 if (VT.is256BitVector()) {
6253 if (Subtarget->hasAVX()) {
6254 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6255 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6257 // Without AVX, we need to extend to a 128-bit vector and then
6258 // insert into the 256-bit vector.
6259 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6260 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6261 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6264 assert(VT.is128BitVector() && "Expected an SSE value type!");
6265 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6266 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6268 return DAG.getBitcast(VT, Item);
6272 // Is it a vector logical left shift?
6273 if (NumElems == 2 && Idx == 1 &&
6274 X86::isZeroNode(Op.getOperand(0)) &&
6275 !X86::isZeroNode(Op.getOperand(1))) {
6276 unsigned NumBits = VT.getSizeInBits();
6277 return getVShift(true, VT,
6278 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6279 VT, Op.getOperand(1)),
6280 NumBits/2, DAG, *this, dl);
6283 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6286 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6287 // is a non-constant being inserted into an element other than the low one,
6288 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6289 // movd/movss) to move this into the low element, then shuffle it into
6291 if (EVTBits == 32) {
6292 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6293 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6297 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6298 if (Values.size() == 1) {
6299 if (EVTBits == 32) {
6300 // Instead of a shuffle like this:
6301 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6302 // Check if it's possible to issue this instead.
6303 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6304 unsigned Idx = countTrailingZeros(NonZeros);
6305 SDValue Item = Op.getOperand(Idx);
6306 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6307 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6312 // A vector full of immediates; various special cases are already
6313 // handled, so this is best done with a single constant-pool load.
6317 // For AVX-length vectors, see if we can use a vector load to get all of the
6318 // elements, otherwise build the individual 128-bit pieces and use
6319 // shuffles to put them in place.
6320 if (VT.is256BitVector() || VT.is512BitVector()) {
6321 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6323 // Check for a build vector of consecutive loads.
6324 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6327 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6329 // Build both the lower and upper subvector.
6330 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6331 makeArrayRef(&V[0], NumElems/2));
6332 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6333 makeArrayRef(&V[NumElems / 2], NumElems/2));
6335 // Recreate the wider vector with the lower and upper part.
6336 if (VT.is256BitVector())
6337 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6338 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6341 // Let legalizer expand 2-wide build_vectors.
6342 if (EVTBits == 64) {
6343 if (NumNonZero == 1) {
6344 // One half is zero or undef.
6345 unsigned Idx = countTrailingZeros(NonZeros);
6346 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6347 Op.getOperand(Idx));
6348 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6353 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6354 if (EVTBits == 8 && NumElems == 16)
6355 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6359 if (EVTBits == 16 && NumElems == 8)
6360 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6364 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6365 if (EVTBits == 32 && NumElems == 4)
6366 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6369 // If element VT is == 32 bits, turn it into a number of shuffles.
6370 SmallVector<SDValue, 8> V(NumElems);
6371 if (NumElems == 4 && NumZero > 0) {
6372 for (unsigned i = 0; i < 4; ++i) {
6373 bool isZero = !(NonZeros & (1 << i));
6375 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6377 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6380 for (unsigned i = 0; i < 2; ++i) {
6381 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6384 V[i] = V[i*2]; // Must be a zero vector.
6387 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6390 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6393 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6398 bool Reverse1 = (NonZeros & 0x3) == 2;
6399 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6403 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6404 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6406 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6409 if (Values.size() > 1 && VT.is128BitVector()) {
6410 // Check for a build vector of consecutive loads.
6411 for (unsigned i = 0; i < NumElems; ++i)
6412 V[i] = Op.getOperand(i);
6414 // Check for elements which are consecutive loads.
6415 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6418 // Check for a build vector from mostly shuffle plus few inserting.
6419 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6422 // For SSE 4.1, use insertps to put the high elements into the low element.
6423 if (Subtarget->hasSSE41()) {
6425 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6426 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6428 Result = DAG.getUNDEF(VT);
6430 for (unsigned i = 1; i < NumElems; ++i) {
6431 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6432 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6433 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6438 // Otherwise, expand into a number of unpckl*, start by extending each of
6439 // our (non-undef) elements to the full vector width with the element in the
6440 // bottom slot of the vector (which generates no code for SSE).
6441 for (unsigned i = 0; i < NumElems; ++i) {
6442 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6443 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6445 V[i] = DAG.getUNDEF(VT);
6448 // Next, we iteratively mix elements, e.g. for v4f32:
6449 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6450 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6451 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6452 unsigned EltStride = NumElems >> 1;
6453 while (EltStride != 0) {
6454 for (unsigned i = 0; i < EltStride; ++i) {
6455 // If V[i+EltStride] is undef and this is the first round of mixing,
6456 // then it is safe to just drop this shuffle: V[i] is already in the
6457 // right place, the one element (since it's the first round) being
6458 // inserted as undef can be dropped. This isn't safe for successive
6459 // rounds because they will permute elements within both vectors.
6460 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6461 EltStride == NumElems/2)
6464 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6473 // 256-bit AVX can use the vinsertf128 instruction
6474 // to create 256-bit vectors from two other 128-bit ones.
6475 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6477 MVT ResVT = Op.getSimpleValueType();
6479 assert((ResVT.is256BitVector() ||
6480 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6482 SDValue V1 = Op.getOperand(0);
6483 SDValue V2 = Op.getOperand(1);
6484 unsigned NumElems = ResVT.getVectorNumElements();
6485 if (ResVT.is256BitVector())
6486 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6488 if (Op.getNumOperands() == 4) {
6489 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6490 ResVT.getVectorNumElements()/2);
6491 SDValue V3 = Op.getOperand(2);
6492 SDValue V4 = Op.getOperand(3);
6493 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6494 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6496 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6499 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6500 const X86Subtarget *Subtarget,
6501 SelectionDAG & DAG) {
6503 MVT ResVT = Op.getSimpleValueType();
6504 unsigned NumOfOperands = Op.getNumOperands();
6506 assert(isPowerOf2_32(NumOfOperands) &&
6507 "Unexpected number of operands in CONCAT_VECTORS");
6509 if (NumOfOperands > 2) {
6510 MVT HalfVT = MVT::getVectorVT(ResVT.getVectorElementType(),
6511 ResVT.getVectorNumElements()/2);
6512 SmallVector<SDValue, 2> Ops;
6513 for (unsigned i = 0; i < NumOfOperands/2; i++)
6514 Ops.push_back(Op.getOperand(i));
6515 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6517 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6518 Ops.push_back(Op.getOperand(i));
6519 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6520 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6523 SDValue V1 = Op.getOperand(0);
6524 SDValue V2 = Op.getOperand(1);
6525 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6526 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6528 if (IsZeroV1 && IsZeroV2)
6529 return getZeroVector(ResVT, Subtarget, DAG, dl);
6531 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6532 SDValue Undef = DAG.getUNDEF(ResVT);
6533 unsigned NumElems = ResVT.getVectorNumElements();
6534 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6536 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6537 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6541 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6542 // Zero the upper bits of V1
6543 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6544 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6547 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6550 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6551 const X86Subtarget *Subtarget,
6552 SelectionDAG &DAG) {
6553 MVT VT = Op.getSimpleValueType();
6554 if (VT.getVectorElementType() == MVT::i1)
6555 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6557 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6558 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6559 Op.getNumOperands() == 4)));
6561 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6562 // from two other 128-bit ones.
6564 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6565 return LowerAVXCONCAT_VECTORS(Op, DAG);
6568 //===----------------------------------------------------------------------===//
6569 // Vector shuffle lowering
6571 // This is an experimental code path for lowering vector shuffles on x86. It is
6572 // designed to handle arbitrary vector shuffles and blends, gracefully
6573 // degrading performance as necessary. It works hard to recognize idiomatic
6574 // shuffles and lower them to optimal instruction patterns without leaving
6575 // a framework that allows reasonably efficient handling of all vector shuffle
6577 //===----------------------------------------------------------------------===//
6579 /// \brief Tiny helper function to identify a no-op mask.
6581 /// This is a somewhat boring predicate function. It checks whether the mask
6582 /// array input, which is assumed to be a single-input shuffle mask of the kind
6583 /// used by the X86 shuffle instructions (not a fully general
6584 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6585 /// in-place shuffle are 'no-op's.
6586 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6587 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6588 if (Mask[i] != -1 && Mask[i] != i)
6593 /// \brief Helper function to classify a mask as a single-input mask.
6595 /// This isn't a generic single-input test because in the vector shuffle
6596 /// lowering we canonicalize single inputs to be the first input operand. This
6597 /// means we can more quickly test for a single input by only checking whether
6598 /// an input from the second operand exists. We also assume that the size of
6599 /// mask corresponds to the size of the input vectors which isn't true in the
6600 /// fully general case.
6601 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6603 if (M >= (int)Mask.size())
6608 /// \brief Test whether there are elements crossing 128-bit lanes in this
6611 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6612 /// and we routinely test for these.
6613 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6614 int LaneSize = 128 / VT.getScalarSizeInBits();
6615 int Size = Mask.size();
6616 for (int i = 0; i < Size; ++i)
6617 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6622 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6624 /// This checks a shuffle mask to see if it is performing the same
6625 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6626 /// that it is also not lane-crossing. It may however involve a blend from the
6627 /// same lane of a second vector.
6629 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6630 /// non-trivial to compute in the face of undef lanes. The representation is
6631 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6632 /// entries from both V1 and V2 inputs to the wider mask.
6634 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6635 SmallVectorImpl<int> &RepeatedMask) {
6636 int LaneSize = 128 / VT.getScalarSizeInBits();
6637 RepeatedMask.resize(LaneSize, -1);
6638 int Size = Mask.size();
6639 for (int i = 0; i < Size; ++i) {
6642 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6643 // This entry crosses lanes, so there is no way to model this shuffle.
6646 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6647 if (RepeatedMask[i % LaneSize] == -1)
6648 // This is the first non-undef entry in this slot of a 128-bit lane.
6649 RepeatedMask[i % LaneSize] =
6650 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6651 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6652 // Found a mismatch with the repeated mask.
6658 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6661 /// This is a fast way to test a shuffle mask against a fixed pattern:
6663 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6665 /// It returns true if the mask is exactly as wide as the argument list, and
6666 /// each element of the mask is either -1 (signifying undef) or the value given
6667 /// in the argument.
6668 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6669 ArrayRef<int> ExpectedMask) {
6670 if (Mask.size() != ExpectedMask.size())
6673 int Size = Mask.size();
6675 // If the values are build vectors, we can look through them to find
6676 // equivalent inputs that make the shuffles equivalent.
6677 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6678 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6680 for (int i = 0; i < Size; ++i)
6681 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6682 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6683 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6684 if (!MaskBV || !ExpectedBV ||
6685 MaskBV->getOperand(Mask[i] % Size) !=
6686 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6693 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6695 /// This helper function produces an 8-bit shuffle immediate corresponding to
6696 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6697 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6700 /// NB: We rely heavily on "undef" masks preserving the input lane.
6701 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6702 SelectionDAG &DAG) {
6703 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6704 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6705 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6706 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6707 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6710 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6711 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6712 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6713 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6714 return DAG.getConstant(Imm, DL, MVT::i8);
6717 /// \brief Compute whether each element of a shuffle is zeroable.
6719 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6720 /// Either it is an undef element in the shuffle mask, the element of the input
6721 /// referenced is undef, or the element of the input referenced is known to be
6722 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6723 /// as many lanes with this technique as possible to simplify the remaining
6725 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6726 SDValue V1, SDValue V2) {
6727 SmallBitVector Zeroable(Mask.size(), false);
6729 while (V1.getOpcode() == ISD::BITCAST)
6730 V1 = V1->getOperand(0);
6731 while (V2.getOpcode() == ISD::BITCAST)
6732 V2 = V2->getOperand(0);
6734 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6735 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6737 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6739 // Handle the easy cases.
6740 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6745 // If this is an index into a build_vector node (which has the same number
6746 // of elements), dig out the input value and use it.
6747 SDValue V = M < Size ? V1 : V2;
6748 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6751 SDValue Input = V.getOperand(M % Size);
6752 // The UNDEF opcode check really should be dead code here, but not quite
6753 // worth asserting on (it isn't invalid, just unexpected).
6754 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6761 // X86 has dedicated unpack instructions that can handle specific blend
6762 // operations: UNPCKH and UNPCKL.
6763 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6764 SDValue V1, SDValue V2,
6765 SelectionDAG &DAG) {
6766 int NumElts = VT.getVectorNumElements();
6767 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6768 SmallVector<int, 8> Unpckl;
6769 SmallVector<int, 8> Unpckh;
6771 for (int i = 0; i < NumElts; ++i) {
6772 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6773 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6774 int HiPos = LoPos + NumEltsInLane / 2;
6775 Unpckl.push_back(LoPos);
6776 Unpckh.push_back(HiPos);
6779 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6780 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6781 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6782 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6784 // Commute and try again.
6785 ShuffleVectorSDNode::commuteMask(Unpckl);
6786 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6787 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6789 ShuffleVectorSDNode::commuteMask(Unpckh);
6790 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6791 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6796 /// \brief Try to emit a bitmask instruction for a shuffle.
6798 /// This handles cases where we can model a blend exactly as a bitmask due to
6799 /// one of the inputs being zeroable.
6800 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6801 SDValue V2, ArrayRef<int> Mask,
6802 SelectionDAG &DAG) {
6803 MVT EltVT = VT.getVectorElementType();
6804 int NumEltBits = EltVT.getSizeInBits();
6805 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6806 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6807 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6809 if (EltVT.isFloatingPoint()) {
6810 Zero = DAG.getBitcast(EltVT, Zero);
6811 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6813 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6814 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6816 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6819 if (Mask[i] % Size != i)
6820 return SDValue(); // Not a blend.
6822 V = Mask[i] < Size ? V1 : V2;
6823 else if (V != (Mask[i] < Size ? V1 : V2))
6824 return SDValue(); // Can only let one input through the mask.
6826 VMaskOps[i] = AllOnes;
6829 return SDValue(); // No non-zeroable elements!
6831 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6832 V = DAG.getNode(VT.isFloatingPoint()
6833 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6838 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6840 /// This is used as a fallback approach when first class blend instructions are
6841 /// unavailable. Currently it is only suitable for integer vectors, but could
6842 /// be generalized for floating point vectors if desirable.
6843 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6844 SDValue V2, ArrayRef<int> Mask,
6845 SelectionDAG &DAG) {
6846 assert(VT.isInteger() && "Only supports integer vector types!");
6847 MVT EltVT = VT.getVectorElementType();
6848 int NumEltBits = EltVT.getSizeInBits();
6849 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6850 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6852 SmallVector<SDValue, 16> MaskOps;
6853 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6854 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6855 return SDValue(); // Shuffled input!
6856 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6859 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6860 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6861 // We have to cast V2 around.
6862 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6863 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6864 DAG.getBitcast(MaskVT, V1Mask),
6865 DAG.getBitcast(MaskVT, V2)));
6866 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6869 /// \brief Try to emit a blend instruction for a shuffle.
6871 /// This doesn't do any checks for the availability of instructions for blending
6872 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6873 /// be matched in the backend with the type given. What it does check for is
6874 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6875 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6876 SDValue V2, ArrayRef<int> Original,
6877 const X86Subtarget *Subtarget,
6878 SelectionDAG &DAG) {
6879 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6880 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6881 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6882 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6883 bool ForceV1Zero = false, ForceV2Zero = false;
6885 // Attempt to generate the binary blend mask. If an input is zero then
6886 // we can use any lane.
6887 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
6888 unsigned BlendMask = 0;
6889 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6895 if (M == i + Size) {
6896 BlendMask |= 1u << i;
6907 BlendMask |= 1u << i;
6912 return SDValue(); // Shuffled input!
6915 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
6917 V1 = getZeroVector(VT, Subtarget, DAG, DL);
6919 V2 = getZeroVector(VT, Subtarget, DAG, DL);
6921 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
6922 unsigned ScaledMask = 0;
6923 for (int i = 0; i != Size; ++i)
6924 if (BlendMask & (1u << i))
6925 for (int j = 0; j != Scale; ++j)
6926 ScaledMask |= 1u << (i * Scale + j);
6930 switch (VT.SimpleTy) {
6935 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6936 DAG.getConstant(BlendMask, DL, MVT::i8));
6940 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6944 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6945 // that instruction.
6946 if (Subtarget->hasAVX2()) {
6947 // Scale the blend by the number of 32-bit dwords per element.
6948 int Scale = VT.getScalarSizeInBits() / 32;
6949 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6950 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6951 V1 = DAG.getBitcast(BlendVT, V1);
6952 V2 = DAG.getBitcast(BlendVT, V2);
6953 return DAG.getBitcast(
6954 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6955 DAG.getConstant(BlendMask, DL, MVT::i8)));
6959 // For integer shuffles we need to expand the mask and cast the inputs to
6960 // v8i16s prior to blending.
6961 int Scale = 8 / VT.getVectorNumElements();
6962 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6963 V1 = DAG.getBitcast(MVT::v8i16, V1);
6964 V2 = DAG.getBitcast(MVT::v8i16, V2);
6965 return DAG.getBitcast(VT,
6966 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6967 DAG.getConstant(BlendMask, DL, MVT::i8)));
6971 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6972 SmallVector<int, 8> RepeatedMask;
6973 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6974 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6975 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6977 for (int i = 0; i < 8; ++i)
6978 if (RepeatedMask[i] >= 16)
6979 BlendMask |= 1u << i;
6980 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6981 DAG.getConstant(BlendMask, DL, MVT::i8));
6987 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
6988 "256-bit byte-blends require AVX2 support!");
6990 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6991 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6994 // Scale the blend by the number of bytes per element.
6995 int Scale = VT.getScalarSizeInBits() / 8;
6997 // This form of blend is always done on bytes. Compute the byte vector
6999 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7001 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7002 // mix of LLVM's code generator and the x86 backend. We tell the code
7003 // generator that boolean values in the elements of an x86 vector register
7004 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7005 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7006 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7007 // of the element (the remaining are ignored) and 0 in that high bit would
7008 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7009 // the LLVM model for boolean values in vector elements gets the relevant
7010 // bit set, it is set backwards and over constrained relative to x86's
7012 SmallVector<SDValue, 32> VSELECTMask;
7013 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7014 for (int j = 0; j < Scale; ++j)
7015 VSELECTMask.push_back(
7016 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7017 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7020 V1 = DAG.getBitcast(BlendVT, V1);
7021 V2 = DAG.getBitcast(BlendVT, V2);
7022 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7023 DAG.getNode(ISD::BUILD_VECTOR, DL,
7024 BlendVT, VSELECTMask),
7029 llvm_unreachable("Not a supported integer vector type!");
7033 /// \brief Try to lower as a blend of elements from two inputs followed by
7034 /// a single-input permutation.
7036 /// This matches the pattern where we can blend elements from two inputs and
7037 /// then reduce the shuffle to a single-input permutation.
7038 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7041 SelectionDAG &DAG) {
7042 // We build up the blend mask while checking whether a blend is a viable way
7043 // to reduce the shuffle.
7044 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7045 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7047 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7051 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7053 if (BlendMask[Mask[i] % Size] == -1)
7054 BlendMask[Mask[i] % Size] = Mask[i];
7055 else if (BlendMask[Mask[i] % Size] != Mask[i])
7056 return SDValue(); // Can't blend in the needed input!
7058 PermuteMask[i] = Mask[i] % Size;
7061 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7062 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7065 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7066 /// blends and permutes.
7068 /// This matches the extremely common pattern for handling combined
7069 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7070 /// operations. It will try to pick the best arrangement of shuffles and
7072 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7076 SelectionDAG &DAG) {
7077 // Shuffle the input elements into the desired positions in V1 and V2 and
7078 // blend them together.
7079 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7080 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7081 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7082 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7083 if (Mask[i] >= 0 && Mask[i] < Size) {
7084 V1Mask[i] = Mask[i];
7086 } else if (Mask[i] >= Size) {
7087 V2Mask[i] = Mask[i] - Size;
7088 BlendMask[i] = i + Size;
7091 // Try to lower with the simpler initial blend strategy unless one of the
7092 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7093 // shuffle may be able to fold with a load or other benefit. However, when
7094 // we'll have to do 2x as many shuffles in order to achieve this, blending
7095 // first is a better strategy.
7096 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7097 if (SDValue BlendPerm =
7098 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7101 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7102 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7103 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7106 /// \brief Try to lower a vector shuffle as a byte rotation.
7108 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7109 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7110 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7111 /// try to generically lower a vector shuffle through such an pattern. It
7112 /// does not check for the profitability of lowering either as PALIGNR or
7113 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7114 /// This matches shuffle vectors that look like:
7116 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7118 /// Essentially it concatenates V1 and V2, shifts right by some number of
7119 /// elements, and takes the low elements as the result. Note that while this is
7120 /// specified as a *right shift* because x86 is little-endian, it is a *left
7121 /// rotate* of the vector lanes.
7122 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7125 const X86Subtarget *Subtarget,
7126 SelectionDAG &DAG) {
7127 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7129 int NumElts = Mask.size();
7130 int NumLanes = VT.getSizeInBits() / 128;
7131 int NumLaneElts = NumElts / NumLanes;
7133 // We need to detect various ways of spelling a rotation:
7134 // [11, 12, 13, 14, 15, 0, 1, 2]
7135 // [-1, 12, 13, 14, -1, -1, 1, -1]
7136 // [-1, -1, -1, -1, -1, -1, 1, 2]
7137 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7138 // [-1, 4, 5, 6, -1, -1, 9, -1]
7139 // [-1, 4, 5, 6, -1, -1, -1, -1]
7142 for (int l = 0; l < NumElts; l += NumLaneElts) {
7143 for (int i = 0; i < NumLaneElts; ++i) {
7144 if (Mask[l + i] == -1)
7146 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7148 // Get the mod-Size index and lane correct it.
7149 int LaneIdx = (Mask[l + i] % NumElts) - l;
7150 // Make sure it was in this lane.
7151 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7154 // Determine where a rotated vector would have started.
7155 int StartIdx = i - LaneIdx;
7157 // The identity rotation isn't interesting, stop.
7160 // If we found the tail of a vector the rotation must be the missing
7161 // front. If we found the head of a vector, it must be how much of the
7163 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7166 Rotation = CandidateRotation;
7167 else if (Rotation != CandidateRotation)
7168 // The rotations don't match, so we can't match this mask.
7171 // Compute which value this mask is pointing at.
7172 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7174 // Compute which of the two target values this index should be assigned
7175 // to. This reflects whether the high elements are remaining or the low
7176 // elements are remaining.
7177 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7179 // Either set up this value if we've not encountered it before, or check
7180 // that it remains consistent.
7183 else if (TargetV != MaskV)
7184 // This may be a rotation, but it pulls from the inputs in some
7185 // unsupported interleaving.
7190 // Check that we successfully analyzed the mask, and normalize the results.
7191 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7192 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7198 // The actual rotate instruction rotates bytes, so we need to scale the
7199 // rotation based on how many bytes are in the vector lane.
7200 int Scale = 16 / NumLaneElts;
7202 // SSSE3 targets can use the palignr instruction.
7203 if (Subtarget->hasSSSE3()) {
7204 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7205 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7206 Lo = DAG.getBitcast(AlignVT, Lo);
7207 Hi = DAG.getBitcast(AlignVT, Hi);
7209 return DAG.getBitcast(
7210 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7211 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7214 assert(VT.is128BitVector() &&
7215 "Rotate-based lowering only supports 128-bit lowering!");
7216 assert(Mask.size() <= 16 &&
7217 "Can shuffle at most 16 bytes in a 128-bit vector!");
7219 // Default SSE2 implementation
7220 int LoByteShift = 16 - Rotation * Scale;
7221 int HiByteShift = Rotation * Scale;
7223 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7224 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7225 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7227 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7228 DAG.getConstant(LoByteShift, DL, MVT::i8));
7229 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7230 DAG.getConstant(HiByteShift, DL, MVT::i8));
7231 return DAG.getBitcast(VT,
7232 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7235 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7237 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7238 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7239 /// matches elements from one of the input vectors shuffled to the left or
7240 /// right with zeroable elements 'shifted in'. It handles both the strictly
7241 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7244 /// PSHL : (little-endian) left bit shift.
7245 /// [ zz, 0, zz, 2 ]
7246 /// [ -1, 4, zz, -1 ]
7247 /// PSRL : (little-endian) right bit shift.
7249 /// [ -1, -1, 7, zz]
7250 /// PSLLDQ : (little-endian) left byte shift
7251 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7252 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7253 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7254 /// PSRLDQ : (little-endian) right byte shift
7255 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7256 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7257 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7258 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7259 SDValue V2, ArrayRef<int> Mask,
7260 SelectionDAG &DAG) {
7261 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7263 int Size = Mask.size();
7264 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7266 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7267 for (int i = 0; i < Size; i += Scale)
7268 for (int j = 0; j < Shift; ++j)
7269 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7275 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7276 for (int i = 0; i != Size; i += Scale) {
7277 unsigned Pos = Left ? i + Shift : i;
7278 unsigned Low = Left ? i : i + Shift;
7279 unsigned Len = Scale - Shift;
7280 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7281 Low + (V == V1 ? 0 : Size)))
7285 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7286 bool ByteShift = ShiftEltBits > 64;
7287 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7288 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7289 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7291 // Normalize the scale for byte shifts to still produce an i64 element
7293 Scale = ByteShift ? Scale / 2 : Scale;
7295 // We need to round trip through the appropriate type for the shift.
7296 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7297 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7298 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7299 "Illegal integer vector type");
7300 V = DAG.getBitcast(ShiftVT, V);
7302 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7303 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7304 return DAG.getBitcast(VT, V);
7307 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7308 // keep doubling the size of the integer elements up to that. We can
7309 // then shift the elements of the integer vector by whole multiples of
7310 // their width within the elements of the larger integer vector. Test each
7311 // multiple to see if we can find a match with the moved element indices
7312 // and that the shifted in elements are all zeroable.
7313 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7314 for (int Shift = 1; Shift != Scale; ++Shift)
7315 for (bool Left : {true, false})
7316 if (CheckZeros(Shift, Scale, Left))
7317 for (SDValue V : {V1, V2})
7318 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7325 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7326 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7327 SDValue V2, ArrayRef<int> Mask,
7328 SelectionDAG &DAG) {
7329 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7330 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7332 int Size = Mask.size();
7333 int HalfSize = Size / 2;
7334 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7336 // Upper half must be undefined.
7337 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7340 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7341 // Remainder of lower half result is zero and upper half is all undef.
7342 auto LowerAsEXTRQ = [&]() {
7343 // Determine the extraction length from the part of the
7344 // lower half that isn't zeroable.
7346 for (; Len > 0; --Len)
7347 if (!Zeroable[Len - 1])
7349 assert(Len > 0 && "Zeroable shuffle mask");
7351 // Attempt to match first Len sequential elements from the lower half.
7354 for (int i = 0; i != Len; ++i) {
7358 SDValue &V = (M < Size ? V1 : V2);
7361 // All mask elements must be in the lower half.
7365 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7376 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7377 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7378 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7379 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7380 DAG.getConstant(BitLen, DL, MVT::i8),
7381 DAG.getConstant(BitIdx, DL, MVT::i8));
7384 if (SDValue ExtrQ = LowerAsEXTRQ())
7387 // INSERTQ: Extract lowest Len elements from lower half of second source and
7388 // insert over first source, starting at Idx.
7389 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7390 auto LowerAsInsertQ = [&]() {
7391 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7394 // Attempt to match first source from mask before insertion point.
7395 if (isUndefInRange(Mask, 0, Idx)) {
7397 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7399 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7405 // Extend the extraction length looking to match both the insertion of
7406 // the second source and the remaining elements of the first.
7407 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7412 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7414 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7420 // Match the remaining elements of the lower half.
7421 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7423 } else if ((!Base || (Base == V1)) &&
7424 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7426 } else if ((!Base || (Base == V2)) &&
7427 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7434 // We may not have a base (first source) - this can safely be undefined.
7436 Base = DAG.getUNDEF(VT);
7438 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7439 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7440 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7441 DAG.getConstant(BitLen, DL, MVT::i8),
7442 DAG.getConstant(BitIdx, DL, MVT::i8));
7449 if (SDValue InsertQ = LowerAsInsertQ())
7455 /// \brief Lower a vector shuffle as a zero or any extension.
7457 /// Given a specific number of elements, element bit width, and extension
7458 /// stride, produce either a zero or any extension based on the available
7459 /// features of the subtarget. The extended elements are consecutive and
7460 /// begin and can start from an offseted element index in the input; to
7461 /// avoid excess shuffling the offset must either being in the bottom lane
7462 /// or at the start of a higher lane. All extended elements must be from
7464 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7465 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7466 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7467 assert(Scale > 1 && "Need a scale to extend.");
7468 int EltBits = VT.getScalarSizeInBits();
7469 int NumElements = VT.getVectorNumElements();
7470 int NumEltsPerLane = 128 / EltBits;
7471 int OffsetLane = Offset / NumEltsPerLane;
7472 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7473 "Only 8, 16, and 32 bit elements can be extended.");
7474 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7475 assert(0 <= Offset && "Extension offset must be positive.");
7476 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7477 "Extension offset must be in the first lane or start an upper lane.");
7479 // Check that an index is in same lane as the base offset.
7480 auto SafeOffset = [&](int Idx) {
7481 return OffsetLane == (Idx / NumEltsPerLane);
7484 // Shift along an input so that the offset base moves to the first element.
7485 auto ShuffleOffset = [&](SDValue V) {
7489 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7490 for (int i = 0; i * Scale < NumElements; ++i) {
7491 int SrcIdx = i + Offset;
7492 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7494 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7497 // Found a valid zext mask! Try various lowering strategies based on the
7498 // input type and available ISA extensions.
7499 if (Subtarget->hasSSE41()) {
7500 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7501 // PUNPCK will catch this in a later shuffle match.
7502 if (Offset && Scale == 2 && VT.is128BitVector())
7504 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7505 NumElements / Scale);
7506 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7507 return DAG.getBitcast(VT, InputV);
7510 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7512 // For any extends we can cheat for larger element sizes and use shuffle
7513 // instructions that can fold with a load and/or copy.
7514 if (AnyExt && EltBits == 32) {
7515 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7517 return DAG.getBitcast(
7518 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7519 DAG.getBitcast(MVT::v4i32, InputV),
7520 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7522 if (AnyExt && EltBits == 16 && Scale > 2) {
7523 int PSHUFDMask[4] = {Offset / 2, -1,
7524 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7525 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7526 DAG.getBitcast(MVT::v4i32, InputV),
7527 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7528 int PSHUFWMask[4] = {1, -1, -1, -1};
7529 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7530 return DAG.getBitcast(
7531 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7532 DAG.getBitcast(MVT::v8i16, InputV),
7533 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7536 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7538 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7539 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7540 assert(VT.is128BitVector() && "Unexpected vector width!");
7542 int LoIdx = Offset * EltBits;
7543 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7544 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7545 DAG.getConstant(EltBits, DL, MVT::i8),
7546 DAG.getConstant(LoIdx, DL, MVT::i8)));
7548 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7549 !SafeOffset(Offset + 1))
7550 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7552 int HiIdx = (Offset + 1) * EltBits;
7553 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7554 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7555 DAG.getConstant(EltBits, DL, MVT::i8),
7556 DAG.getConstant(HiIdx, DL, MVT::i8)));
7557 return DAG.getNode(ISD::BITCAST, DL, VT,
7558 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7561 // If this would require more than 2 unpack instructions to expand, use
7562 // pshufb when available. We can only use more than 2 unpack instructions
7563 // when zero extending i8 elements which also makes it easier to use pshufb.
7564 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7565 assert(NumElements == 16 && "Unexpected byte vector width!");
7566 SDValue PSHUFBMask[16];
7567 for (int i = 0; i < 16; ++i) {
7568 int Idx = Offset + (i / Scale);
7569 PSHUFBMask[i] = DAG.getConstant(
7570 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7572 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7573 return DAG.getBitcast(VT,
7574 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7575 DAG.getNode(ISD::BUILD_VECTOR, DL,
7576 MVT::v16i8, PSHUFBMask)));
7579 // If we are extending from an offset, ensure we start on a boundary that
7580 // we can unpack from.
7581 int AlignToUnpack = Offset % (NumElements / Scale);
7582 if (AlignToUnpack) {
7583 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7584 for (int i = AlignToUnpack; i < NumElements; ++i)
7585 ShMask[i - AlignToUnpack] = i;
7586 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7587 Offset -= AlignToUnpack;
7590 // Otherwise emit a sequence of unpacks.
7592 unsigned UnpackLoHi = X86ISD::UNPCKL;
7593 if (Offset >= (NumElements / 2)) {
7594 UnpackLoHi = X86ISD::UNPCKH;
7595 Offset -= (NumElements / 2);
7598 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7599 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7600 : getZeroVector(InputVT, Subtarget, DAG, DL);
7601 InputV = DAG.getBitcast(InputVT, InputV);
7602 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7606 } while (Scale > 1);
7607 return DAG.getBitcast(VT, InputV);
7610 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7612 /// This routine will try to do everything in its power to cleverly lower
7613 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7614 /// check for the profitability of this lowering, it tries to aggressively
7615 /// match this pattern. It will use all of the micro-architectural details it
7616 /// can to emit an efficient lowering. It handles both blends with all-zero
7617 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7618 /// masking out later).
7620 /// The reason we have dedicated lowering for zext-style shuffles is that they
7621 /// are both incredibly common and often quite performance sensitive.
7622 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7623 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7624 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7625 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7627 int Bits = VT.getSizeInBits();
7628 int NumLanes = Bits / 128;
7629 int NumElements = VT.getVectorNumElements();
7630 int NumEltsPerLane = NumElements / NumLanes;
7631 assert(VT.getScalarSizeInBits() <= 32 &&
7632 "Exceeds 32-bit integer zero extension limit");
7633 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7635 // Define a helper function to check a particular ext-scale and lower to it if
7637 auto Lower = [&](int Scale) -> SDValue {
7642 for (int i = 0; i < NumElements; ++i) {
7645 continue; // Valid anywhere but doesn't tell us anything.
7646 if (i % Scale != 0) {
7647 // Each of the extended elements need to be zeroable.
7651 // We no longer are in the anyext case.
7656 // Each of the base elements needs to be consecutive indices into the
7657 // same input vector.
7658 SDValue V = M < NumElements ? V1 : V2;
7659 M = M % NumElements;
7662 Offset = M - (i / Scale);
7663 } else if (InputV != V)
7664 return SDValue(); // Flip-flopping inputs.
7666 // Offset must start in the lowest 128-bit lane or at the start of an
7668 // FIXME: Is it ever worth allowing a negative base offset?
7669 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7670 (Offset % NumEltsPerLane) == 0))
7673 // If we are offsetting, all referenced entries must come from the same
7675 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7678 if ((M % NumElements) != (Offset + (i / Scale)))
7679 return SDValue(); // Non-consecutive strided elements.
7683 // If we fail to find an input, we have a zero-shuffle which should always
7684 // have already been handled.
7685 // FIXME: Maybe handle this here in case during blending we end up with one?
7689 // If we are offsetting, don't extend if we only match a single input, we
7690 // can always do better by using a basic PSHUF or PUNPCK.
7691 if (Offset != 0 && Matches < 2)
7694 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7695 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7698 // The widest scale possible for extending is to a 64-bit integer.
7699 assert(Bits % 64 == 0 &&
7700 "The number of bits in a vector must be divisible by 64 on x86!");
7701 int NumExtElements = Bits / 64;
7703 // Each iteration, try extending the elements half as much, but into twice as
7705 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7706 assert(NumElements % NumExtElements == 0 &&
7707 "The input vector size must be divisible by the extended size.");
7708 if (SDValue V = Lower(NumElements / NumExtElements))
7712 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7716 // Returns one of the source operands if the shuffle can be reduced to a
7717 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7718 auto CanZExtLowHalf = [&]() {
7719 for (int i = NumElements / 2; i != NumElements; ++i)
7722 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7724 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7729 if (SDValue V = CanZExtLowHalf()) {
7730 V = DAG.getBitcast(MVT::v2i64, V);
7731 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7732 return DAG.getBitcast(VT, V);
7735 // No viable ext lowering found.
7739 /// \brief Try to get a scalar value for a specific element of a vector.
7741 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7742 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7743 SelectionDAG &DAG) {
7744 MVT VT = V.getSimpleValueType();
7745 MVT EltVT = VT.getVectorElementType();
7746 while (V.getOpcode() == ISD::BITCAST)
7747 V = V.getOperand(0);
7748 // If the bitcasts shift the element size, we can't extract an equivalent
7750 MVT NewVT = V.getSimpleValueType();
7751 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7754 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7755 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7756 // Ensure the scalar operand is the same size as the destination.
7757 // FIXME: Add support for scalar truncation where possible.
7758 SDValue S = V.getOperand(Idx);
7759 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7760 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7766 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7768 /// This is particularly important because the set of instructions varies
7769 /// significantly based on whether the operand is a load or not.
7770 static bool isShuffleFoldableLoad(SDValue V) {
7771 while (V.getOpcode() == ISD::BITCAST)
7772 V = V.getOperand(0);
7774 return ISD::isNON_EXTLoad(V.getNode());
7777 /// \brief Try to lower insertion of a single element into a zero vector.
7779 /// This is a common pattern that we have especially efficient patterns to lower
7780 /// across all subtarget feature sets.
7781 static SDValue lowerVectorShuffleAsElementInsertion(
7782 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7783 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7784 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7786 MVT EltVT = VT.getVectorElementType();
7788 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7789 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7791 bool IsV1Zeroable = true;
7792 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7793 if (i != V2Index && !Zeroable[i]) {
7794 IsV1Zeroable = false;
7798 // Check for a single input from a SCALAR_TO_VECTOR node.
7799 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7800 // all the smarts here sunk into that routine. However, the current
7801 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7802 // vector shuffle lowering is dead.
7803 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7805 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7806 // We need to zext the scalar if it is smaller than an i32.
7807 V2S = DAG.getBitcast(EltVT, V2S);
7808 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7809 // Using zext to expand a narrow element won't work for non-zero
7814 // Zero-extend directly to i32.
7816 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7818 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7819 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7820 EltVT == MVT::i16) {
7821 // Either not inserting from the low element of the input or the input
7822 // element size is too small to use VZEXT_MOVL to clear the high bits.
7826 if (!IsV1Zeroable) {
7827 // If V1 can't be treated as a zero vector we have fewer options to lower
7828 // this. We can't support integer vectors or non-zero targets cheaply, and
7829 // the V1 elements can't be permuted in any way.
7830 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7831 if (!VT.isFloatingPoint() || V2Index != 0)
7833 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7834 V1Mask[V2Index] = -1;
7835 if (!isNoopShuffleMask(V1Mask))
7837 // This is essentially a special case blend operation, but if we have
7838 // general purpose blend operations, they are always faster. Bail and let
7839 // the rest of the lowering handle these as blends.
7840 if (Subtarget->hasSSE41())
7843 // Otherwise, use MOVSD or MOVSS.
7844 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7845 "Only two types of floating point element types to handle!");
7846 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7850 // This lowering only works for the low element with floating point vectors.
7851 if (VT.isFloatingPoint() && V2Index != 0)
7854 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7856 V2 = DAG.getBitcast(VT, V2);
7859 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7860 // the desired position. Otherwise it is more efficient to do a vector
7861 // shift left. We know that we can do a vector shift left because all
7862 // the inputs are zero.
7863 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7864 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7865 V2Shuffle[V2Index] = 0;
7866 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7868 V2 = DAG.getBitcast(MVT::v2i64, V2);
7870 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7871 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7872 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7873 DAG.getDataLayout(), VT)));
7874 V2 = DAG.getBitcast(VT, V2);
7880 /// \brief Try to lower broadcast of a single - truncated - integer element,
7881 /// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
7883 /// This assumes we have AVX2.
7884 static SDValue lowerVectorShuffleAsTruncBroadcast(SDLoc DL, MVT VT, SDValue V0,
7886 const X86Subtarget *Subtarget,
7887 SelectionDAG &DAG) {
7888 assert(Subtarget->hasAVX2() &&
7889 "We can only lower integer broadcasts with AVX2!");
7891 EVT EltVT = VT.getVectorElementType();
7892 EVT V0VT = V0.getValueType();
7894 assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
7895 assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
7897 EVT V0EltVT = V0VT.getVectorElementType();
7898 if (!V0EltVT.isInteger())
7901 const unsigned EltSize = EltVT.getSizeInBits();
7902 const unsigned V0EltSize = V0EltVT.getSizeInBits();
7904 // This is only a truncation if the original element type is larger.
7905 if (V0EltSize <= EltSize)
7908 assert(((V0EltSize % EltSize) == 0) &&
7909 "Scalar type sizes must all be powers of 2 on x86!");
7911 const unsigned V0Opc = V0.getOpcode();
7912 const unsigned Scale = V0EltSize / EltSize;
7913 const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
7915 if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
7916 V0Opc != ISD::BUILD_VECTOR)
7919 SDValue Scalar = V0.getOperand(V0BroadcastIdx);
7921 // If we're extracting non-least-significant bits, shift so we can truncate.
7922 // Hopefully, we can fold away the trunc/srl/load into the broadcast.
7923 // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
7924 // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
7925 if (const int OffsetIdx = BroadcastIdx % Scale)
7926 Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
7927 DAG.getConstant(OffsetIdx * EltSize, DL, Scalar.getValueType()));
7929 return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
7930 DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
7933 /// \brief Try to lower broadcast of a single element.
7935 /// For convenience, this code also bundles all of the subtarget feature set
7936 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7937 /// a convenient way to factor it out.
7938 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7940 const X86Subtarget *Subtarget,
7941 SelectionDAG &DAG) {
7942 if (!Subtarget->hasAVX())
7944 if (VT.isInteger() && !Subtarget->hasAVX2())
7947 // Check that the mask is a broadcast.
7948 int BroadcastIdx = -1;
7950 if (M >= 0 && BroadcastIdx == -1)
7952 else if (M >= 0 && M != BroadcastIdx)
7955 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7956 "a sorted mask where the broadcast "
7959 // Go up the chain of (vector) values to find a scalar load that we can
7960 // combine with the broadcast.
7962 switch (V.getOpcode()) {
7963 case ISD::CONCAT_VECTORS: {
7964 int OperandSize = Mask.size() / V.getNumOperands();
7965 V = V.getOperand(BroadcastIdx / OperandSize);
7966 BroadcastIdx %= OperandSize;
7970 case ISD::INSERT_SUBVECTOR: {
7971 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7972 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7976 int BeginIdx = (int)ConstantIdx->getZExtValue();
7978 BeginIdx + (int)VInner.getSimpleValueType().getVectorNumElements();
7979 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7980 BroadcastIdx -= BeginIdx;
7991 // Check if this is a broadcast of a scalar. We special case lowering
7992 // for scalars so that we can more effectively fold with loads.
7993 // First, look through bitcast: if the original value has a larger element
7994 // type than the shuffle, the broadcast element is in essence truncated.
7995 // Make that explicit to ease folding.
7996 if (V.getOpcode() == ISD::BITCAST && VT.isInteger())
7997 if (SDValue TruncBroadcast = lowerVectorShuffleAsTruncBroadcast(
7998 DL, VT, V.getOperand(0), BroadcastIdx, Subtarget, DAG))
7999 return TruncBroadcast;
8001 // Also check the simpler case, where we can directly reuse the scalar.
8002 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8003 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8004 V = V.getOperand(BroadcastIdx);
8006 // If the scalar isn't a load, we can't broadcast from it in AVX1.
8007 // Only AVX2 has register broadcasts.
8008 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8010 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8011 // We can't broadcast from a vector register without AVX2, and we can only
8012 // broadcast from the zero-element of a vector register.
8016 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8019 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8020 // INSERTPS when the V1 elements are already in the correct locations
8021 // because otherwise we can just always use two SHUFPS instructions which
8022 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8023 // perform INSERTPS if a single V1 element is out of place and all V2
8024 // elements are zeroable.
8025 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8027 SelectionDAG &DAG) {
8028 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8029 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8030 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8031 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8033 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8036 int V1DstIndex = -1;
8037 int V2DstIndex = -1;
8038 bool V1UsedInPlace = false;
8040 for (int i = 0; i < 4; ++i) {
8041 // Synthesize a zero mask from the zeroable elements (includes undefs).
8047 // Flag if we use any V1 inputs in place.
8049 V1UsedInPlace = true;
8053 // We can only insert a single non-zeroable element.
8054 if (V1DstIndex != -1 || V2DstIndex != -1)
8058 // V1 input out of place for insertion.
8061 // V2 input for insertion.
8066 // Don't bother if we have no (non-zeroable) element for insertion.
8067 if (V1DstIndex == -1 && V2DstIndex == -1)
8070 // Determine element insertion src/dst indices. The src index is from the
8071 // start of the inserted vector, not the start of the concatenated vector.
8072 unsigned V2SrcIndex = 0;
8073 if (V1DstIndex != -1) {
8074 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8075 // and don't use the original V2 at all.
8076 V2SrcIndex = Mask[V1DstIndex];
8077 V2DstIndex = V1DstIndex;
8080 V2SrcIndex = Mask[V2DstIndex] - 4;
8083 // If no V1 inputs are used in place, then the result is created only from
8084 // the zero mask and the V2 insertion - so remove V1 dependency.
8086 V1 = DAG.getUNDEF(MVT::v4f32);
8088 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8089 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8091 // Insert the V2 element into the desired position.
8093 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8094 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8097 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8098 /// UNPCK instruction.
8100 /// This specifically targets cases where we end up with alternating between
8101 /// the two inputs, and so can permute them into something that feeds a single
8102 /// UNPCK instruction. Note that this routine only targets integer vectors
8103 /// because for floating point vectors we have a generalized SHUFPS lowering
8104 /// strategy that handles everything that doesn't *exactly* match an unpack,
8105 /// making this clever lowering unnecessary.
8106 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8107 SDValue V1, SDValue V2,
8109 SelectionDAG &DAG) {
8110 assert(!VT.isFloatingPoint() &&
8111 "This routine only supports integer vectors.");
8112 assert(!isSingleInputShuffleMask(Mask) &&
8113 "This routine should only be used when blending two inputs.");
8114 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8116 int Size = Mask.size();
8118 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8119 return M >= 0 && M % Size < Size / 2;
8121 int NumHiInputs = std::count_if(
8122 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8124 bool UnpackLo = NumLoInputs >= NumHiInputs;
8126 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8127 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8128 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8130 for (int i = 0; i < Size; ++i) {
8134 // Each element of the unpack contains Scale elements from this mask.
8135 int UnpackIdx = i / Scale;
8137 // We only handle the case where V1 feeds the first slots of the unpack.
8138 // We rely on canonicalization to ensure this is the case.
8139 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8142 // Setup the mask for this input. The indexing is tricky as we have to
8143 // handle the unpack stride.
8144 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8145 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8149 // If we will have to shuffle both inputs to use the unpack, check whether
8150 // we can just unpack first and shuffle the result. If so, skip this unpack.
8151 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8152 !isNoopShuffleMask(V2Mask))
8155 // Shuffle the inputs into place.
8156 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8157 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8159 // Cast the inputs to the type we will use to unpack them.
8160 V1 = DAG.getBitcast(UnpackVT, V1);
8161 V2 = DAG.getBitcast(UnpackVT, V2);
8163 // Unpack the inputs and cast the result back to the desired type.
8164 return DAG.getBitcast(
8165 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8169 // We try each unpack from the largest to the smallest to try and find one
8170 // that fits this mask.
8171 int OrigNumElements = VT.getVectorNumElements();
8172 int OrigScalarSize = VT.getScalarSizeInBits();
8173 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8174 int Scale = ScalarSize / OrigScalarSize;
8175 int NumElements = OrigNumElements / Scale;
8176 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8177 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8181 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8183 if (NumLoInputs == 0 || NumHiInputs == 0) {
8184 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8185 "We have to have *some* inputs!");
8186 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8188 // FIXME: We could consider the total complexity of the permute of each
8189 // possible unpacking. Or at the least we should consider how many
8190 // half-crossings are created.
8191 // FIXME: We could consider commuting the unpacks.
8193 SmallVector<int, 32> PermMask;
8194 PermMask.assign(Size, -1);
8195 for (int i = 0; i < Size; ++i) {
8199 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8202 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8204 return DAG.getVectorShuffle(
8205 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8207 DAG.getUNDEF(VT), PermMask);
8213 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8215 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8216 /// support for floating point shuffles but not integer shuffles. These
8217 /// instructions will incur a domain crossing penalty on some chips though so
8218 /// it is better to avoid lowering through this for integer vectors where
8220 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8221 const X86Subtarget *Subtarget,
8222 SelectionDAG &DAG) {
8224 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8225 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8226 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8227 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8228 ArrayRef<int> Mask = SVOp->getMask();
8229 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8231 if (isSingleInputShuffleMask(Mask)) {
8232 // Use low duplicate instructions for masks that match their pattern.
8233 if (Subtarget->hasSSE3())
8234 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8235 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8237 // Straight shuffle of a single input vector. Simulate this by using the
8238 // single input as both of the "inputs" to this instruction..
8239 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8241 if (Subtarget->hasAVX()) {
8242 // If we have AVX, we can use VPERMILPS which will allow folding a load
8243 // into the shuffle.
8244 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8245 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8248 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8249 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8251 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8252 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8254 // If we have a single input, insert that into V1 if we can do so cheaply.
8255 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8256 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8257 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8259 // Try inverting the insertion since for v2 masks it is easy to do and we
8260 // can't reliably sort the mask one way or the other.
8261 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8262 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8263 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8264 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8268 // Try to use one of the special instruction patterns to handle two common
8269 // blend patterns if a zero-blend above didn't work.
8270 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8271 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8272 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8273 // We can either use a special instruction to load over the low double or
8274 // to move just the low double.
8276 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8278 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8280 if (Subtarget->hasSSE41())
8281 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8285 // Use dedicated unpack instructions for masks that match their pattern.
8287 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8290 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8291 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8292 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8295 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8297 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8298 /// the integer unit to minimize domain crossing penalties. However, for blends
8299 /// it falls back to the floating point shuffle operation with appropriate bit
8301 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8302 const X86Subtarget *Subtarget,
8303 SelectionDAG &DAG) {
8305 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8306 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8307 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8308 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8309 ArrayRef<int> Mask = SVOp->getMask();
8310 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8312 if (isSingleInputShuffleMask(Mask)) {
8313 // Check for being able to broadcast a single element.
8314 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8315 Mask, Subtarget, DAG))
8318 // Straight shuffle of a single input vector. For everything from SSE2
8319 // onward this has a single fast instruction with no scary immediates.
8320 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8321 V1 = DAG.getBitcast(MVT::v4i32, V1);
8322 int WidenedMask[4] = {
8323 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8324 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8325 return DAG.getBitcast(
8327 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8328 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8330 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8331 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8332 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8333 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8335 // If we have a blend of two PACKUS operations an the blend aligns with the
8336 // low and half halves, we can just merge the PACKUS operations. This is
8337 // particularly important as it lets us merge shuffles that this routine itself
8339 auto GetPackNode = [](SDValue V) {
8340 while (V.getOpcode() == ISD::BITCAST)
8341 V = V.getOperand(0);
8343 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8345 if (SDValue V1Pack = GetPackNode(V1))
8346 if (SDValue V2Pack = GetPackNode(V2))
8347 return DAG.getBitcast(MVT::v2i64,
8348 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8349 Mask[0] == 0 ? V1Pack.getOperand(0)
8350 : V1Pack.getOperand(1),
8351 Mask[1] == 2 ? V2Pack.getOperand(0)
8352 : V2Pack.getOperand(1)));
8354 // Try to use shift instructions.
8356 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8359 // When loading a scalar and then shuffling it into a vector we can often do
8360 // the insertion cheaply.
8361 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8362 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8364 // Try inverting the insertion since for v2 masks it is easy to do and we
8365 // can't reliably sort the mask one way or the other.
8366 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8367 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8368 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8371 // We have different paths for blend lowering, but they all must use the
8372 // *exact* same predicate.
8373 bool IsBlendSupported = Subtarget->hasSSE41();
8374 if (IsBlendSupported)
8375 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8379 // Use dedicated unpack instructions for masks that match their pattern.
8381 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8384 // Try to use byte rotation instructions.
8385 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8386 if (Subtarget->hasSSSE3())
8387 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8388 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8391 // If we have direct support for blends, we should lower by decomposing into
8392 // a permute. That will be faster than the domain cross.
8393 if (IsBlendSupported)
8394 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8397 // We implement this with SHUFPD which is pretty lame because it will likely
8398 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8399 // However, all the alternatives are still more cycles and newer chips don't
8400 // have this problem. It would be really nice if x86 had better shuffles here.
8401 V1 = DAG.getBitcast(MVT::v2f64, V1);
8402 V2 = DAG.getBitcast(MVT::v2f64, V2);
8403 return DAG.getBitcast(MVT::v2i64,
8404 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8407 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8409 /// This is used to disable more specialized lowerings when the shufps lowering
8410 /// will happen to be efficient.
8411 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8412 // This routine only handles 128-bit shufps.
8413 assert(Mask.size() == 4 && "Unsupported mask size!");
8415 // To lower with a single SHUFPS we need to have the low half and high half
8416 // each requiring a single input.
8417 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8419 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8425 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8427 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8428 /// It makes no assumptions about whether this is the *best* lowering, it simply
8430 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8431 ArrayRef<int> Mask, SDValue V1,
8432 SDValue V2, SelectionDAG &DAG) {
8433 SDValue LowV = V1, HighV = V2;
8434 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8437 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8439 if (NumV2Elements == 1) {
8441 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8444 // Compute the index adjacent to V2Index and in the same half by toggling
8446 int V2AdjIndex = V2Index ^ 1;
8448 if (Mask[V2AdjIndex] == -1) {
8449 // Handles all the cases where we have a single V2 element and an undef.
8450 // This will only ever happen in the high lanes because we commute the
8451 // vector otherwise.
8453 std::swap(LowV, HighV);
8454 NewMask[V2Index] -= 4;
8456 // Handle the case where the V2 element ends up adjacent to a V1 element.
8457 // To make this work, blend them together as the first step.
8458 int V1Index = V2AdjIndex;
8459 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8460 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8461 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8463 // Now proceed to reconstruct the final blend as we have the necessary
8464 // high or low half formed.
8471 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8472 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8474 } else if (NumV2Elements == 2) {
8475 if (Mask[0] < 4 && Mask[1] < 4) {
8476 // Handle the easy case where we have V1 in the low lanes and V2 in the
8480 } else if (Mask[2] < 4 && Mask[3] < 4) {
8481 // We also handle the reversed case because this utility may get called
8482 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8483 // arrange things in the right direction.
8489 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8490 // trying to place elements directly, just blend them and set up the final
8491 // shuffle to place them.
8493 // The first two blend mask elements are for V1, the second two are for
8495 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8496 Mask[2] < 4 ? Mask[2] : Mask[3],
8497 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8498 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8499 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8500 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8502 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8505 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8506 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8507 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8508 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8511 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8512 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8515 /// \brief Lower 4-lane 32-bit floating point shuffles.
8517 /// Uses instructions exclusively from the floating point unit to minimize
8518 /// domain crossing penalties, as these are sufficient to implement all v4f32
8520 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8521 const X86Subtarget *Subtarget,
8522 SelectionDAG &DAG) {
8524 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8525 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8526 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8527 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8528 ArrayRef<int> Mask = SVOp->getMask();
8529 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8532 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8534 if (NumV2Elements == 0) {
8535 // Check for being able to broadcast a single element.
8536 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8537 Mask, Subtarget, DAG))
8540 // Use even/odd duplicate instructions for masks that match their pattern.
8541 if (Subtarget->hasSSE3()) {
8542 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8543 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8544 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8545 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8548 if (Subtarget->hasAVX()) {
8549 // If we have AVX, we can use VPERMILPS which will allow folding a load
8550 // into the shuffle.
8551 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8552 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8555 // Otherwise, use a straight shuffle of a single input vector. We pass the
8556 // input vector to both operands to simulate this with a SHUFPS.
8557 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8558 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8561 // There are special ways we can lower some single-element blends. However, we
8562 // have custom ways we can lower more complex single-element blends below that
8563 // we defer to if both this and BLENDPS fail to match, so restrict this to
8564 // when the V2 input is targeting element 0 of the mask -- that is the fast
8566 if (NumV2Elements == 1 && Mask[0] >= 4)
8567 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8568 Mask, Subtarget, DAG))
8571 if (Subtarget->hasSSE41()) {
8572 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8576 // Use INSERTPS if we can complete the shuffle efficiently.
8577 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8580 if (!isSingleSHUFPSMask(Mask))
8581 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8582 DL, MVT::v4f32, V1, V2, Mask, DAG))
8586 // Use dedicated unpack instructions for masks that match their pattern.
8588 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8591 // Otherwise fall back to a SHUFPS lowering strategy.
8592 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8595 /// \brief Lower 4-lane i32 vector shuffles.
8597 /// We try to handle these with integer-domain shuffles where we can, but for
8598 /// blends we use the floating point domain blend instructions.
8599 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8600 const X86Subtarget *Subtarget,
8601 SelectionDAG &DAG) {
8603 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8604 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8605 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8606 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8607 ArrayRef<int> Mask = SVOp->getMask();
8608 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8610 // Whenever we can lower this as a zext, that instruction is strictly faster
8611 // than any alternative. It also allows us to fold memory operands into the
8612 // shuffle in many cases.
8613 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8614 Mask, Subtarget, DAG))
8618 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8620 if (NumV2Elements == 0) {
8621 // Check for being able to broadcast a single element.
8622 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8623 Mask, Subtarget, DAG))
8626 // Straight shuffle of a single input vector. For everything from SSE2
8627 // onward this has a single fast instruction with no scary immediates.
8628 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8629 // but we aren't actually going to use the UNPCK instruction because doing
8630 // so prevents folding a load into this instruction or making a copy.
8631 const int UnpackLoMask[] = {0, 0, 1, 1};
8632 const int UnpackHiMask[] = {2, 2, 3, 3};
8633 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8634 Mask = UnpackLoMask;
8635 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8636 Mask = UnpackHiMask;
8638 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8639 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8642 // Try to use shift instructions.
8644 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8647 // There are special ways we can lower some single-element blends.
8648 if (NumV2Elements == 1)
8649 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8650 Mask, Subtarget, DAG))
8653 // We have different paths for blend lowering, but they all must use the
8654 // *exact* same predicate.
8655 bool IsBlendSupported = Subtarget->hasSSE41();
8656 if (IsBlendSupported)
8657 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8661 if (SDValue Masked =
8662 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8665 // Use dedicated unpack instructions for masks that match their pattern.
8667 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8670 // Try to use byte rotation instructions.
8671 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8672 if (Subtarget->hasSSSE3())
8673 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8674 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8677 // If we have direct support for blends, we should lower by decomposing into
8678 // a permute. That will be faster than the domain cross.
8679 if (IsBlendSupported)
8680 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8683 // Try to lower by permuting the inputs into an unpack instruction.
8684 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8688 // We implement this with SHUFPS because it can blend from two vectors.
8689 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8690 // up the inputs, bypassing domain shift penalties that we would encur if we
8691 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8693 return DAG.getBitcast(
8695 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8696 DAG.getBitcast(MVT::v4f32, V2), Mask));
8699 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8700 /// shuffle lowering, and the most complex part.
8702 /// The lowering strategy is to try to form pairs of input lanes which are
8703 /// targeted at the same half of the final vector, and then use a dword shuffle
8704 /// to place them onto the right half, and finally unpack the paired lanes into
8705 /// their final position.
8707 /// The exact breakdown of how to form these dword pairs and align them on the
8708 /// correct sides is really tricky. See the comments within the function for
8709 /// more of the details.
8711 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8712 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8713 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8714 /// vector, form the analogous 128-bit 8-element Mask.
8715 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8716 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8717 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8718 assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
8719 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8721 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8722 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8723 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8725 SmallVector<int, 4> LoInputs;
8726 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8727 [](int M) { return M >= 0; });
8728 std::sort(LoInputs.begin(), LoInputs.end());
8729 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8730 SmallVector<int, 4> HiInputs;
8731 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8732 [](int M) { return M >= 0; });
8733 std::sort(HiInputs.begin(), HiInputs.end());
8734 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8736 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8737 int NumHToL = LoInputs.size() - NumLToL;
8739 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8740 int NumHToH = HiInputs.size() - NumLToH;
8741 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8742 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8743 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8744 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8746 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8747 // such inputs we can swap two of the dwords across the half mark and end up
8748 // with <=2 inputs to each half in each half. Once there, we can fall through
8749 // to the generic code below. For example:
8751 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8752 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8754 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8755 // and an existing 2-into-2 on the other half. In this case we may have to
8756 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8757 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8758 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8759 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8760 // half than the one we target for fixing) will be fixed when we re-enter this
8761 // path. We will also combine away any sequence of PSHUFD instructions that
8762 // result into a single instruction. Here is an example of the tricky case:
8764 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8765 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8767 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8769 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8770 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8772 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8773 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8775 // The result is fine to be handled by the generic logic.
8776 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8777 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8778 int AOffset, int BOffset) {
8779 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8780 "Must call this with A having 3 or 1 inputs from the A half.");
8781 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8782 "Must call this with B having 1 or 3 inputs from the B half.");
8783 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8784 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8786 bool ThreeAInputs = AToAInputs.size() == 3;
8788 // Compute the index of dword with only one word among the three inputs in
8789 // a half by taking the sum of the half with three inputs and subtracting
8790 // the sum of the actual three inputs. The difference is the remaining
8793 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8794 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8795 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8796 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8797 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8798 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8799 int TripleNonInputIdx =
8800 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8801 TripleDWord = TripleNonInputIdx / 2;
8803 // We use xor with one to compute the adjacent DWord to whichever one the
8805 OneInputDWord = (OneInput / 2) ^ 1;
8807 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8808 // and BToA inputs. If there is also such a problem with the BToB and AToB
8809 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8810 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8811 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8812 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8813 // Compute how many inputs will be flipped by swapping these DWords. We
8815 // to balance this to ensure we don't form a 3-1 shuffle in the other
8817 int NumFlippedAToBInputs =
8818 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8819 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8820 int NumFlippedBToBInputs =
8821 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8822 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8823 if ((NumFlippedAToBInputs == 1 &&
8824 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8825 (NumFlippedBToBInputs == 1 &&
8826 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8827 // We choose whether to fix the A half or B half based on whether that
8828 // half has zero flipped inputs. At zero, we may not be able to fix it
8829 // with that half. We also bias towards fixing the B half because that
8830 // will more commonly be the high half, and we have to bias one way.
8831 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8832 ArrayRef<int> Inputs) {
8833 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8834 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8835 PinnedIdx ^ 1) != Inputs.end();
8836 // Determine whether the free index is in the flipped dword or the
8837 // unflipped dword based on where the pinned index is. We use this bit
8838 // in an xor to conditionally select the adjacent dword.
8839 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8840 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8841 FixFreeIdx) != Inputs.end();
8842 if (IsFixIdxInput == IsFixFreeIdxInput)
8844 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8845 FixFreeIdx) != Inputs.end();
8846 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8847 "We need to be changing the number of flipped inputs!");
8848 int PSHUFHalfMask[] = {0, 1, 2, 3};
8849 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8850 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8852 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8855 if (M != -1 && M == FixIdx)
8857 else if (M != -1 && M == FixFreeIdx)
8860 if (NumFlippedBToBInputs != 0) {
8862 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8863 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8865 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8866 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8867 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8872 int PSHUFDMask[] = {0, 1, 2, 3};
8873 PSHUFDMask[ADWord] = BDWord;
8874 PSHUFDMask[BDWord] = ADWord;
8877 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8878 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8880 // Adjust the mask to match the new locations of A and B.
8882 if (M != -1 && M/2 == ADWord)
8883 M = 2 * BDWord + M % 2;
8884 else if (M != -1 && M/2 == BDWord)
8885 M = 2 * ADWord + M % 2;
8887 // Recurse back into this routine to re-compute state now that this isn't
8888 // a 3 and 1 problem.
8889 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8892 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8893 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8894 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8895 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8897 // At this point there are at most two inputs to the low and high halves from
8898 // each half. That means the inputs can always be grouped into dwords and
8899 // those dwords can then be moved to the correct half with a dword shuffle.
8900 // We use at most one low and one high word shuffle to collect these paired
8901 // inputs into dwords, and finally a dword shuffle to place them.
8902 int PSHUFLMask[4] = {-1, -1, -1, -1};
8903 int PSHUFHMask[4] = {-1, -1, -1, -1};
8904 int PSHUFDMask[4] = {-1, -1, -1, -1};
8906 // First fix the masks for all the inputs that are staying in their
8907 // original halves. This will then dictate the targets of the cross-half
8909 auto fixInPlaceInputs =
8910 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8911 MutableArrayRef<int> SourceHalfMask,
8912 MutableArrayRef<int> HalfMask, int HalfOffset) {
8913 if (InPlaceInputs.empty())
8915 if (InPlaceInputs.size() == 1) {
8916 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8917 InPlaceInputs[0] - HalfOffset;
8918 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8921 if (IncomingInputs.empty()) {
8922 // Just fix all of the in place inputs.
8923 for (int Input : InPlaceInputs) {
8924 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8925 PSHUFDMask[Input / 2] = Input / 2;
8930 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8931 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8932 InPlaceInputs[0] - HalfOffset;
8933 // Put the second input next to the first so that they are packed into
8934 // a dword. We find the adjacent index by toggling the low bit.
8935 int AdjIndex = InPlaceInputs[0] ^ 1;
8936 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8937 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8938 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8940 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8941 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8943 // Now gather the cross-half inputs and place them into a free dword of
8944 // their target half.
8945 // FIXME: This operation could almost certainly be simplified dramatically to
8946 // look more like the 3-1 fixing operation.
8947 auto moveInputsToRightHalf = [&PSHUFDMask](
8948 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8949 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8950 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8952 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8953 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8955 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8957 int LowWord = Word & ~1;
8958 int HighWord = Word | 1;
8959 return isWordClobbered(SourceHalfMask, LowWord) ||
8960 isWordClobbered(SourceHalfMask, HighWord);
8963 if (IncomingInputs.empty())
8966 if (ExistingInputs.empty()) {
8967 // Map any dwords with inputs from them into the right half.
8968 for (int Input : IncomingInputs) {
8969 // If the source half mask maps over the inputs, turn those into
8970 // swaps and use the swapped lane.
8971 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8972 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8973 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8974 Input - SourceOffset;
8975 // We have to swap the uses in our half mask in one sweep.
8976 for (int &M : HalfMask)
8977 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8979 else if (M == Input)
8980 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8982 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8983 Input - SourceOffset &&
8984 "Previous placement doesn't match!");
8986 // Note that this correctly re-maps both when we do a swap and when
8987 // we observe the other side of the swap above. We rely on that to
8988 // avoid swapping the members of the input list directly.
8989 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8992 // Map the input's dword into the correct half.
8993 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8994 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8996 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8998 "Previous placement doesn't match!");
9001 // And just directly shift any other-half mask elements to be same-half
9002 // as we will have mirrored the dword containing the element into the
9003 // same position within that half.
9004 for (int &M : HalfMask)
9005 if (M >= SourceOffset && M < SourceOffset + 4) {
9006 M = M - SourceOffset + DestOffset;
9007 assert(M >= 0 && "This should never wrap below zero!");
9012 // Ensure we have the input in a viable dword of its current half. This
9013 // is particularly tricky because the original position may be clobbered
9014 // by inputs being moved and *staying* in that half.
9015 if (IncomingInputs.size() == 1) {
9016 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9017 int InputFixed = std::find(std::begin(SourceHalfMask),
9018 std::end(SourceHalfMask), -1) -
9019 std::begin(SourceHalfMask) + SourceOffset;
9020 SourceHalfMask[InputFixed - SourceOffset] =
9021 IncomingInputs[0] - SourceOffset;
9022 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9024 IncomingInputs[0] = InputFixed;
9026 } else if (IncomingInputs.size() == 2) {
9027 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9028 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9029 // We have two non-adjacent or clobbered inputs we need to extract from
9030 // the source half. To do this, we need to map them into some adjacent
9031 // dword slot in the source mask.
9032 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9033 IncomingInputs[1] - SourceOffset};
9035 // If there is a free slot in the source half mask adjacent to one of
9036 // the inputs, place the other input in it. We use (Index XOR 1) to
9037 // compute an adjacent index.
9038 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9039 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9040 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9041 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9042 InputsFixed[1] = InputsFixed[0] ^ 1;
9043 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9044 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9045 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9046 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9047 InputsFixed[0] = InputsFixed[1] ^ 1;
9048 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9049 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9050 // The two inputs are in the same DWord but it is clobbered and the
9051 // adjacent DWord isn't used at all. Move both inputs to the free
9053 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9054 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9055 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9056 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9058 // The only way we hit this point is if there is no clobbering
9059 // (because there are no off-half inputs to this half) and there is no
9060 // free slot adjacent to one of the inputs. In this case, we have to
9061 // swap an input with a non-input.
9062 for (int i = 0; i < 4; ++i)
9063 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9064 "We can't handle any clobbers here!");
9065 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9066 "Cannot have adjacent inputs here!");
9068 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9069 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9071 // We also have to update the final source mask in this case because
9072 // it may need to undo the above swap.
9073 for (int &M : FinalSourceHalfMask)
9074 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9075 M = InputsFixed[1] + SourceOffset;
9076 else if (M == InputsFixed[1] + SourceOffset)
9077 M = (InputsFixed[0] ^ 1) + SourceOffset;
9079 InputsFixed[1] = InputsFixed[0] ^ 1;
9082 // Point everything at the fixed inputs.
9083 for (int &M : HalfMask)
9084 if (M == IncomingInputs[0])
9085 M = InputsFixed[0] + SourceOffset;
9086 else if (M == IncomingInputs[1])
9087 M = InputsFixed[1] + SourceOffset;
9089 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9090 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9093 llvm_unreachable("Unhandled input size!");
9096 // Now hoist the DWord down to the right half.
9097 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9098 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9099 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9100 for (int &M : HalfMask)
9101 for (int Input : IncomingInputs)
9103 M = FreeDWord * 2 + Input % 2;
9105 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9106 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9107 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9108 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9110 // Now enact all the shuffles we've computed to move the inputs into their
9112 if (!isNoopShuffleMask(PSHUFLMask))
9113 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9114 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9115 if (!isNoopShuffleMask(PSHUFHMask))
9116 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9117 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9118 if (!isNoopShuffleMask(PSHUFDMask))
9121 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9122 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9124 // At this point, each half should contain all its inputs, and we can then
9125 // just shuffle them into their final position.
9126 assert(std::count_if(LoMask.begin(), LoMask.end(),
9127 [](int M) { return M >= 4; }) == 0 &&
9128 "Failed to lift all the high half inputs to the low mask!");
9129 assert(std::count_if(HiMask.begin(), HiMask.end(),
9130 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9131 "Failed to lift all the low half inputs to the high mask!");
9133 // Do a half shuffle for the low mask.
9134 if (!isNoopShuffleMask(LoMask))
9135 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9136 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9138 // Do a half shuffle with the high mask after shifting its values down.
9139 for (int &M : HiMask)
9142 if (!isNoopShuffleMask(HiMask))
9143 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9144 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9149 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9150 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9151 SDValue V2, ArrayRef<int> Mask,
9152 SelectionDAG &DAG, bool &V1InUse,
9154 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9160 int Size = Mask.size();
9161 int Scale = 16 / Size;
9162 for (int i = 0; i < 16; ++i) {
9163 if (Mask[i / Scale] == -1) {
9164 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9166 const int ZeroMask = 0x80;
9167 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9169 int V2Idx = Mask[i / Scale] < Size
9171 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9172 if (Zeroable[i / Scale])
9173 V1Idx = V2Idx = ZeroMask;
9174 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9175 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9176 V1InUse |= (ZeroMask != V1Idx);
9177 V2InUse |= (ZeroMask != V2Idx);
9182 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9183 DAG.getBitcast(MVT::v16i8, V1),
9184 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9186 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9187 DAG.getBitcast(MVT::v16i8, V2),
9188 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9190 // If we need shuffled inputs from both, blend the two.
9192 if (V1InUse && V2InUse)
9193 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9195 V = V1InUse ? V1 : V2;
9197 // Cast the result back to the correct type.
9198 return DAG.getBitcast(VT, V);
9201 /// \brief Generic lowering of 8-lane i16 shuffles.
9203 /// This handles both single-input shuffles and combined shuffle/blends with
9204 /// two inputs. The single input shuffles are immediately delegated to
9205 /// a dedicated lowering routine.
9207 /// The blends are lowered in one of three fundamental ways. If there are few
9208 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9209 /// of the input is significantly cheaper when lowered as an interleaving of
9210 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9211 /// halves of the inputs separately (making them have relatively few inputs)
9212 /// and then concatenate them.
9213 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9214 const X86Subtarget *Subtarget,
9215 SelectionDAG &DAG) {
9217 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9218 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9219 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9220 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9221 ArrayRef<int> OrigMask = SVOp->getMask();
9222 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9223 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9224 MutableArrayRef<int> Mask(MaskStorage);
9226 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9228 // Whenever we can lower this as a zext, that instruction is strictly faster
9229 // than any alternative.
9230 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9231 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9234 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9236 auto isV2 = [](int M) { return M >= 8; };
9238 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9240 if (NumV2Inputs == 0) {
9241 // Check for being able to broadcast a single element.
9242 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9243 Mask, Subtarget, DAG))
9246 // Try to use shift instructions.
9248 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9251 // Use dedicated unpack instructions for masks that match their pattern.
9253 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9256 // Try to use byte rotation instructions.
9257 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9258 Mask, Subtarget, DAG))
9261 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9265 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9266 "All single-input shuffles should be canonicalized to be V1-input "
9269 // Try to use shift instructions.
9271 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9274 // See if we can use SSE4A Extraction / Insertion.
9275 if (Subtarget->hasSSE4A())
9276 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9279 // There are special ways we can lower some single-element blends.
9280 if (NumV2Inputs == 1)
9281 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9282 Mask, Subtarget, DAG))
9285 // We have different paths for blend lowering, but they all must use the
9286 // *exact* same predicate.
9287 bool IsBlendSupported = Subtarget->hasSSE41();
9288 if (IsBlendSupported)
9289 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9293 if (SDValue Masked =
9294 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9297 // Use dedicated unpack instructions for masks that match their pattern.
9299 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9302 // Try to use byte rotation instructions.
9303 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9304 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9307 if (SDValue BitBlend =
9308 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9311 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9315 // If we can't directly blend but can use PSHUFB, that will be better as it
9316 // can both shuffle and set up the inefficient blend.
9317 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9318 bool V1InUse, V2InUse;
9319 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9323 // We can always bit-blend if we have to so the fallback strategy is to
9324 // decompose into single-input permutes and blends.
9325 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9329 /// \brief Check whether a compaction lowering can be done by dropping even
9330 /// elements and compute how many times even elements must be dropped.
9332 /// This handles shuffles which take every Nth element where N is a power of
9333 /// two. Example shuffle masks:
9335 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9336 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9337 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9338 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9339 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9340 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9342 /// Any of these lanes can of course be undef.
9344 /// This routine only supports N <= 3.
9345 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9348 /// \returns N above, or the number of times even elements must be dropped if
9349 /// there is such a number. Otherwise returns zero.
9350 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9351 // Figure out whether we're looping over two inputs or just one.
9352 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9354 // The modulus for the shuffle vector entries is based on whether this is
9355 // a single input or not.
9356 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9357 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9358 "We should only be called with masks with a power-of-2 size!");
9360 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9362 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9363 // and 2^3 simultaneously. This is because we may have ambiguity with
9364 // partially undef inputs.
9365 bool ViableForN[3] = {true, true, true};
9367 for (int i = 0, e = Mask.size(); i < e; ++i) {
9368 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9373 bool IsAnyViable = false;
9374 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9375 if (ViableForN[j]) {
9378 // The shuffle mask must be equal to (i * 2^N) % M.
9379 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9382 ViableForN[j] = false;
9384 // Early exit if we exhaust the possible powers of two.
9389 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9393 // Return 0 as there is no viable power of two.
9397 /// \brief Generic lowering of v16i8 shuffles.
9399 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9400 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9401 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9402 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9404 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9405 const X86Subtarget *Subtarget,
9406 SelectionDAG &DAG) {
9408 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9409 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9410 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9411 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9412 ArrayRef<int> Mask = SVOp->getMask();
9413 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9415 // Try to use shift instructions.
9417 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9420 // Try to use byte rotation instructions.
9421 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9422 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9425 // Try to use a zext lowering.
9426 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9427 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9430 // See if we can use SSE4A Extraction / Insertion.
9431 if (Subtarget->hasSSE4A())
9432 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9436 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9438 // For single-input shuffles, there are some nicer lowering tricks we can use.
9439 if (NumV2Elements == 0) {
9440 // Check for being able to broadcast a single element.
9441 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9442 Mask, Subtarget, DAG))
9445 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9446 // Notably, this handles splat and partial-splat shuffles more efficiently.
9447 // However, it only makes sense if the pre-duplication shuffle simplifies
9448 // things significantly. Currently, this means we need to be able to
9449 // express the pre-duplication shuffle as an i16 shuffle.
9451 // FIXME: We should check for other patterns which can be widened into an
9452 // i16 shuffle as well.
9453 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9454 for (int i = 0; i < 16; i += 2)
9455 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9460 auto tryToWidenViaDuplication = [&]() -> SDValue {
9461 if (!canWidenViaDuplication(Mask))
9463 SmallVector<int, 4> LoInputs;
9464 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9465 [](int M) { return M >= 0 && M < 8; });
9466 std::sort(LoInputs.begin(), LoInputs.end());
9467 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9469 SmallVector<int, 4> HiInputs;
9470 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9471 [](int M) { return M >= 8; });
9472 std::sort(HiInputs.begin(), HiInputs.end());
9473 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9476 bool TargetLo = LoInputs.size() >= HiInputs.size();
9477 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9478 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9480 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9481 SmallDenseMap<int, int, 8> LaneMap;
9482 for (int I : InPlaceInputs) {
9483 PreDupI16Shuffle[I/2] = I/2;
9486 int j = TargetLo ? 0 : 4, je = j + 4;
9487 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9488 // Check if j is already a shuffle of this input. This happens when
9489 // there are two adjacent bytes after we move the low one.
9490 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9491 // If we haven't yet mapped the input, search for a slot into which
9493 while (j < je && PreDupI16Shuffle[j] != -1)
9497 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9500 // Map this input with the i16 shuffle.
9501 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9504 // Update the lane map based on the mapping we ended up with.
9505 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9507 V1 = DAG.getBitcast(
9509 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9510 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9512 // Unpack the bytes to form the i16s that will be shuffled into place.
9513 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9514 MVT::v16i8, V1, V1);
9516 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9517 for (int i = 0; i < 16; ++i)
9518 if (Mask[i] != -1) {
9519 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9520 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9521 if (PostDupI16Shuffle[i / 2] == -1)
9522 PostDupI16Shuffle[i / 2] = MappedMask;
9524 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9525 "Conflicting entrties in the original shuffle!");
9527 return DAG.getBitcast(
9529 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9530 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9532 if (SDValue V = tryToWidenViaDuplication())
9536 if (SDValue Masked =
9537 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9540 // Use dedicated unpack instructions for masks that match their pattern.
9542 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9545 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9546 // with PSHUFB. It is important to do this before we attempt to generate any
9547 // blends but after all of the single-input lowerings. If the single input
9548 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9549 // want to preserve that and we can DAG combine any longer sequences into
9550 // a PSHUFB in the end. But once we start blending from multiple inputs,
9551 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9552 // and there are *very* few patterns that would actually be faster than the
9553 // PSHUFB approach because of its ability to zero lanes.
9555 // FIXME: The only exceptions to the above are blends which are exact
9556 // interleavings with direct instructions supporting them. We currently don't
9557 // handle those well here.
9558 if (Subtarget->hasSSSE3()) {
9559 bool V1InUse = false;
9560 bool V2InUse = false;
9562 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9563 DAG, V1InUse, V2InUse);
9565 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9566 // do so. This avoids using them to handle blends-with-zero which is
9567 // important as a single pshufb is significantly faster for that.
9568 if (V1InUse && V2InUse) {
9569 if (Subtarget->hasSSE41())
9570 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9571 Mask, Subtarget, DAG))
9574 // We can use an unpack to do the blending rather than an or in some
9575 // cases. Even though the or may be (very minorly) more efficient, we
9576 // preference this lowering because there are common cases where part of
9577 // the complexity of the shuffles goes away when we do the final blend as
9579 // FIXME: It might be worth trying to detect if the unpack-feeding
9580 // shuffles will both be pshufb, in which case we shouldn't bother with
9582 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9583 DL, MVT::v16i8, V1, V2, Mask, DAG))
9590 // There are special ways we can lower some single-element blends.
9591 if (NumV2Elements == 1)
9592 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9593 Mask, Subtarget, DAG))
9596 if (SDValue BitBlend =
9597 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9600 // Check whether a compaction lowering can be done. This handles shuffles
9601 // which take every Nth element for some even N. See the helper function for
9604 // We special case these as they can be particularly efficiently handled with
9605 // the PACKUSB instruction on x86 and they show up in common patterns of
9606 // rearranging bytes to truncate wide elements.
9607 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9608 // NumEvenDrops is the power of two stride of the elements. Another way of
9609 // thinking about it is that we need to drop the even elements this many
9610 // times to get the original input.
9611 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9613 // First we need to zero all the dropped bytes.
9614 assert(NumEvenDrops <= 3 &&
9615 "No support for dropping even elements more than 3 times.");
9616 // We use the mask type to pick which bytes are preserved based on how many
9617 // elements are dropped.
9618 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9619 SDValue ByteClearMask = DAG.getBitcast(
9620 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9621 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9623 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9625 // Now pack things back together.
9626 V1 = DAG.getBitcast(MVT::v8i16, V1);
9627 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9628 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9629 for (int i = 1; i < NumEvenDrops; ++i) {
9630 Result = DAG.getBitcast(MVT::v8i16, Result);
9631 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9637 // Handle multi-input cases by blending single-input shuffles.
9638 if (NumV2Elements > 0)
9639 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9642 // The fallback path for single-input shuffles widens this into two v8i16
9643 // vectors with unpacks, shuffles those, and then pulls them back together
9647 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9648 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9649 for (int i = 0; i < 16; ++i)
9651 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9653 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9655 SDValue VLoHalf, VHiHalf;
9656 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9657 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9659 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9660 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9661 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9662 [](int M) { return M >= 0 && M % 2 == 1; })) {
9663 // Use a mask to drop the high bytes.
9664 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9665 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9666 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9668 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9669 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9671 // Squash the masks to point directly into VLoHalf.
9672 for (int &M : LoBlendMask)
9675 for (int &M : HiBlendMask)
9679 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9680 // VHiHalf so that we can blend them as i16s.
9681 VLoHalf = DAG.getBitcast(
9682 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9683 VHiHalf = DAG.getBitcast(
9684 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9687 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9688 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9690 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9693 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9695 /// This routine breaks down the specific type of 128-bit shuffle and
9696 /// dispatches to the lowering routines accordingly.
9697 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9698 MVT VT, const X86Subtarget *Subtarget,
9699 SelectionDAG &DAG) {
9700 switch (VT.SimpleTy) {
9702 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9704 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9706 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9708 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9710 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9712 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9715 llvm_unreachable("Unimplemented!");
9719 /// \brief Helper function to test whether a shuffle mask could be
9720 /// simplified by widening the elements being shuffled.
9722 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9723 /// leaves it in an unspecified state.
9725 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9726 /// shuffle masks. The latter have the special property of a '-2' representing
9727 /// a zero-ed lane of a vector.
9728 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9729 SmallVectorImpl<int> &WidenedMask) {
9730 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9731 // If both elements are undef, its trivial.
9732 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9733 WidenedMask.push_back(SM_SentinelUndef);
9737 // Check for an undef mask and a mask value properly aligned to fit with
9738 // a pair of values. If we find such a case, use the non-undef mask's value.
9739 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9740 WidenedMask.push_back(Mask[i + 1] / 2);
9743 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9744 WidenedMask.push_back(Mask[i] / 2);
9748 // When zeroing, we need to spread the zeroing across both lanes to widen.
9749 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9750 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9751 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9752 WidenedMask.push_back(SM_SentinelZero);
9758 // Finally check if the two mask values are adjacent and aligned with
9760 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9761 WidenedMask.push_back(Mask[i] / 2);
9765 // Otherwise we can't safely widen the elements used in this shuffle.
9768 assert(WidenedMask.size() == Mask.size() / 2 &&
9769 "Incorrect size of mask after widening the elements!");
9774 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9776 /// This routine just extracts two subvectors, shuffles them independently, and
9777 /// then concatenates them back together. This should work effectively with all
9778 /// AVX vector shuffle types.
9779 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9780 SDValue V2, ArrayRef<int> Mask,
9781 SelectionDAG &DAG) {
9782 assert(VT.getSizeInBits() >= 256 &&
9783 "Only for 256-bit or wider vector shuffles!");
9784 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9785 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9787 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9788 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9790 int NumElements = VT.getVectorNumElements();
9791 int SplitNumElements = NumElements / 2;
9792 MVT ScalarVT = VT.getVectorElementType();
9793 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9795 // Rather than splitting build-vectors, just build two narrower build
9796 // vectors. This helps shuffling with splats and zeros.
9797 auto SplitVector = [&](SDValue V) {
9798 while (V.getOpcode() == ISD::BITCAST)
9799 V = V->getOperand(0);
9801 MVT OrigVT = V.getSimpleValueType();
9802 int OrigNumElements = OrigVT.getVectorNumElements();
9803 int OrigSplitNumElements = OrigNumElements / 2;
9804 MVT OrigScalarVT = OrigVT.getVectorElementType();
9805 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9809 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9811 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9812 DAG.getIntPtrConstant(0, DL));
9813 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9814 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9817 SmallVector<SDValue, 16> LoOps, HiOps;
9818 for (int i = 0; i < OrigSplitNumElements; ++i) {
9819 LoOps.push_back(BV->getOperand(i));
9820 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9822 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9823 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9825 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9826 DAG.getBitcast(SplitVT, HiV));
9829 SDValue LoV1, HiV1, LoV2, HiV2;
9830 std::tie(LoV1, HiV1) = SplitVector(V1);
9831 std::tie(LoV2, HiV2) = SplitVector(V2);
9833 // Now create two 4-way blends of these half-width vectors.
9834 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9835 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9836 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9837 for (int i = 0; i < SplitNumElements; ++i) {
9838 int M = HalfMask[i];
9839 if (M >= NumElements) {
9840 if (M >= NumElements + SplitNumElements)
9844 V2BlendMask.push_back(M - NumElements);
9845 V1BlendMask.push_back(-1);
9846 BlendMask.push_back(SplitNumElements + i);
9847 } else if (M >= 0) {
9848 if (M >= SplitNumElements)
9852 V2BlendMask.push_back(-1);
9853 V1BlendMask.push_back(M);
9854 BlendMask.push_back(i);
9856 V2BlendMask.push_back(-1);
9857 V1BlendMask.push_back(-1);
9858 BlendMask.push_back(-1);
9862 // Because the lowering happens after all combining takes place, we need to
9863 // manually combine these blend masks as much as possible so that we create
9864 // a minimal number of high-level vector shuffle nodes.
9866 // First try just blending the halves of V1 or V2.
9867 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9868 return DAG.getUNDEF(SplitVT);
9869 if (!UseLoV2 && !UseHiV2)
9870 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9871 if (!UseLoV1 && !UseHiV1)
9872 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9874 SDValue V1Blend, V2Blend;
9875 if (UseLoV1 && UseHiV1) {
9877 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9879 // We only use half of V1 so map the usage down into the final blend mask.
9880 V1Blend = UseLoV1 ? LoV1 : HiV1;
9881 for (int i = 0; i < SplitNumElements; ++i)
9882 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9883 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9885 if (UseLoV2 && UseHiV2) {
9887 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9889 // We only use half of V2 so map the usage down into the final blend mask.
9890 V2Blend = UseLoV2 ? LoV2 : HiV2;
9891 for (int i = 0; i < SplitNumElements; ++i)
9892 if (BlendMask[i] >= SplitNumElements)
9893 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9895 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9897 SDValue Lo = HalfBlend(LoMask);
9898 SDValue Hi = HalfBlend(HiMask);
9899 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9902 /// \brief Either split a vector in halves or decompose the shuffles and the
9905 /// This is provided as a good fallback for many lowerings of non-single-input
9906 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9907 /// between splitting the shuffle into 128-bit components and stitching those
9908 /// back together vs. extracting the single-input shuffles and blending those
9910 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9911 SDValue V2, ArrayRef<int> Mask,
9912 SelectionDAG &DAG) {
9913 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9914 "lower single-input shuffles as it "
9915 "could then recurse on itself.");
9916 int Size = Mask.size();
9918 // If this can be modeled as a broadcast of two elements followed by a blend,
9919 // prefer that lowering. This is especially important because broadcasts can
9920 // often fold with memory operands.
9921 auto DoBothBroadcast = [&] {
9922 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9925 if (V2BroadcastIdx == -1)
9926 V2BroadcastIdx = M - Size;
9927 else if (M - Size != V2BroadcastIdx)
9929 } else if (M >= 0) {
9930 if (V1BroadcastIdx == -1)
9932 else if (M != V1BroadcastIdx)
9937 if (DoBothBroadcast())
9938 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9941 // If the inputs all stem from a single 128-bit lane of each input, then we
9942 // split them rather than blending because the split will decompose to
9943 // unusually few instructions.
9944 int LaneCount = VT.getSizeInBits() / 128;
9945 int LaneSize = Size / LaneCount;
9946 SmallBitVector LaneInputs[2];
9947 LaneInputs[0].resize(LaneCount, false);
9948 LaneInputs[1].resize(LaneCount, false);
9949 for (int i = 0; i < Size; ++i)
9951 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9952 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9953 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9955 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9956 // that the decomposed single-input shuffles don't end up here.
9957 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9960 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9961 /// a permutation and blend of those lanes.
9963 /// This essentially blends the out-of-lane inputs to each lane into the lane
9964 /// from a permuted copy of the vector. This lowering strategy results in four
9965 /// instructions in the worst case for a single-input cross lane shuffle which
9966 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9967 /// of. Special cases for each particular shuffle pattern should be handled
9968 /// prior to trying this lowering.
9969 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9970 SDValue V1, SDValue V2,
9972 SelectionDAG &DAG) {
9973 // FIXME: This should probably be generalized for 512-bit vectors as well.
9974 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
9975 int LaneSize = Mask.size() / 2;
9977 // If there are only inputs from one 128-bit lane, splitting will in fact be
9978 // less expensive. The flags track whether the given lane contains an element
9979 // that crosses to another lane.
9980 bool LaneCrossing[2] = {false, false};
9981 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9982 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9983 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9984 if (!LaneCrossing[0] || !LaneCrossing[1])
9985 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9987 if (isSingleInputShuffleMask(Mask)) {
9988 SmallVector<int, 32> FlippedBlendMask;
9989 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9990 FlippedBlendMask.push_back(
9991 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9993 : Mask[i] % LaneSize +
9994 (i / LaneSize) * LaneSize + Size));
9996 // Flip the vector, and blend the results which should now be in-lane. The
9997 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9998 // 5 for the high source. The value 3 selects the high half of source 2 and
9999 // the value 2 selects the low half of source 2. We only use source 2 to
10000 // allow folding it into a memory operand.
10001 unsigned PERMMask = 3 | 2 << 4;
10002 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10003 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
10004 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10007 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10008 // will be handled by the above logic and a blend of the results, much like
10009 // other patterns in AVX.
10010 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10013 /// \brief Handle lowering 2-lane 128-bit shuffles.
10014 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10015 SDValue V2, ArrayRef<int> Mask,
10016 const X86Subtarget *Subtarget,
10017 SelectionDAG &DAG) {
10018 // TODO: If minimizing size and one of the inputs is a zero vector and the
10019 // the zero vector has only one use, we could use a VPERM2X128 to save the
10020 // instruction bytes needed to explicitly generate the zero vector.
10022 // Blends are faster and handle all the non-lane-crossing cases.
10023 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10027 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10028 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10030 // If either input operand is a zero vector, use VPERM2X128 because its mask
10031 // allows us to replace the zero input with an implicit zero.
10032 if (!IsV1Zero && !IsV2Zero) {
10033 // Check for patterns which can be matched with a single insert of a 128-bit
10035 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10036 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10037 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10038 VT.getVectorNumElements() / 2);
10039 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10040 DAG.getIntPtrConstant(0, DL));
10041 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10042 OnlyUsesV1 ? V1 : V2,
10043 DAG.getIntPtrConstant(0, DL));
10044 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10048 // Otherwise form a 128-bit permutation. After accounting for undefs,
10049 // convert the 64-bit shuffle mask selection values into 128-bit
10050 // selection bits by dividing the indexes by 2 and shifting into positions
10051 // defined by a vperm2*128 instruction's immediate control byte.
10053 // The immediate permute control byte looks like this:
10054 // [1:0] - select 128 bits from sources for low half of destination
10056 // [3] - zero low half of destination
10057 // [5:4] - select 128 bits from sources for high half of destination
10059 // [7] - zero high half of destination
10061 int MaskLO = Mask[0];
10062 if (MaskLO == SM_SentinelUndef)
10063 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10065 int MaskHI = Mask[2];
10066 if (MaskHI == SM_SentinelUndef)
10067 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10069 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10071 // If either input is a zero vector, replace it with an undef input.
10072 // Shuffle mask values < 4 are selecting elements of V1.
10073 // Shuffle mask values >= 4 are selecting elements of V2.
10074 // Adjust each half of the permute mask by clearing the half that was
10075 // selecting the zero vector and setting the zero mask bit.
10077 V1 = DAG.getUNDEF(VT);
10079 PermMask = (PermMask & 0xf0) | 0x08;
10081 PermMask = (PermMask & 0x0f) | 0x80;
10084 V2 = DAG.getUNDEF(VT);
10086 PermMask = (PermMask & 0xf0) | 0x08;
10088 PermMask = (PermMask & 0x0f) | 0x80;
10091 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10092 DAG.getConstant(PermMask, DL, MVT::i8));
10095 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10096 /// shuffling each lane.
10098 /// This will only succeed when the result of fixing the 128-bit lanes results
10099 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10100 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10101 /// the lane crosses early and then use simpler shuffles within each lane.
10103 /// FIXME: It might be worthwhile at some point to support this without
10104 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10105 /// in x86 only floating point has interesting non-repeating shuffles, and even
10106 /// those are still *marginally* more expensive.
10107 static SDValue lowerVectorShuffleByMerging128BitLanes(
10108 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10109 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10110 assert(!isSingleInputShuffleMask(Mask) &&
10111 "This is only useful with multiple inputs.");
10113 int Size = Mask.size();
10114 int LaneSize = 128 / VT.getScalarSizeInBits();
10115 int NumLanes = Size / LaneSize;
10116 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10118 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10119 // check whether the in-128-bit lane shuffles share a repeating pattern.
10120 SmallVector<int, 4> Lanes;
10121 Lanes.resize(NumLanes, -1);
10122 SmallVector<int, 4> InLaneMask;
10123 InLaneMask.resize(LaneSize, -1);
10124 for (int i = 0; i < Size; ++i) {
10128 int j = i / LaneSize;
10130 if (Lanes[j] < 0) {
10131 // First entry we've seen for this lane.
10132 Lanes[j] = Mask[i] / LaneSize;
10133 } else if (Lanes[j] != Mask[i] / LaneSize) {
10134 // This doesn't match the lane selected previously!
10138 // Check that within each lane we have a consistent shuffle mask.
10139 int k = i % LaneSize;
10140 if (InLaneMask[k] < 0) {
10141 InLaneMask[k] = Mask[i] % LaneSize;
10142 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10143 // This doesn't fit a repeating in-lane mask.
10148 // First shuffle the lanes into place.
10149 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10150 VT.getSizeInBits() / 64);
10151 SmallVector<int, 8> LaneMask;
10152 LaneMask.resize(NumLanes * 2, -1);
10153 for (int i = 0; i < NumLanes; ++i)
10154 if (Lanes[i] >= 0) {
10155 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10156 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10159 V1 = DAG.getBitcast(LaneVT, V1);
10160 V2 = DAG.getBitcast(LaneVT, V2);
10161 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10163 // Cast it back to the type we actually want.
10164 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10166 // Now do a simple shuffle that isn't lane crossing.
10167 SmallVector<int, 8> NewMask;
10168 NewMask.resize(Size, -1);
10169 for (int i = 0; i < Size; ++i)
10171 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10172 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10173 "Must not introduce lane crosses at this point!");
10175 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10178 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10181 /// This returns true if the elements from a particular input are already in the
10182 /// slot required by the given mask and require no permutation.
10183 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10184 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10185 int Size = Mask.size();
10186 for (int i = 0; i < Size; ++i)
10187 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10193 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10194 ArrayRef<int> Mask, SDValue V1,
10195 SDValue V2, SelectionDAG &DAG) {
10197 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10198 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10199 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10200 int NumElts = VT.getVectorNumElements();
10201 bool ShufpdMask = true;
10202 bool CommutableMask = true;
10203 unsigned Immediate = 0;
10204 for (int i = 0; i < NumElts; ++i) {
10207 int Val = (i & 6) + NumElts * (i & 1);
10208 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10209 if (Mask[i] < Val || Mask[i] > Val + 1)
10210 ShufpdMask = false;
10211 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10212 CommutableMask = false;
10213 Immediate |= (Mask[i] % 2) << i;
10216 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10217 DAG.getConstant(Immediate, DL, MVT::i8));
10218 if (CommutableMask)
10219 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10220 DAG.getConstant(Immediate, DL, MVT::i8));
10224 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10226 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10227 /// isn't available.
10228 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10229 const X86Subtarget *Subtarget,
10230 SelectionDAG &DAG) {
10232 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10233 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10234 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10235 ArrayRef<int> Mask = SVOp->getMask();
10236 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10238 SmallVector<int, 4> WidenedMask;
10239 if (canWidenShuffleElements(Mask, WidenedMask))
10240 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10243 if (isSingleInputShuffleMask(Mask)) {
10244 // Check for being able to broadcast a single element.
10245 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10246 Mask, Subtarget, DAG))
10249 // Use low duplicate instructions for masks that match their pattern.
10250 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10251 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10253 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10254 // Non-half-crossing single input shuffles can be lowerid with an
10255 // interleaved permutation.
10256 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10257 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10258 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10259 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10262 // With AVX2 we have direct support for this permutation.
10263 if (Subtarget->hasAVX2())
10264 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10265 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10267 // Otherwise, fall back.
10268 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10272 // Use dedicated unpack instructions for masks that match their pattern.
10274 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10277 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10281 // Check if the blend happens to exactly fit that of SHUFPD.
10283 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10286 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10287 // shuffle. However, if we have AVX2 and either inputs are already in place,
10288 // we will be able to shuffle even across lanes the other input in a single
10289 // instruction so skip this pattern.
10290 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10291 isShuffleMaskInputInPlace(1, Mask))))
10292 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10293 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10296 // If we have AVX2 then we always want to lower with a blend because an v4 we
10297 // can fully permute the elements.
10298 if (Subtarget->hasAVX2())
10299 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10302 // Otherwise fall back on generic lowering.
10303 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10306 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10308 /// This routine is only called when we have AVX2 and thus a reasonable
10309 /// instruction set for v4i64 shuffling..
10310 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10311 const X86Subtarget *Subtarget,
10312 SelectionDAG &DAG) {
10314 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10315 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10316 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10317 ArrayRef<int> Mask = SVOp->getMask();
10318 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10319 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10321 SmallVector<int, 4> WidenedMask;
10322 if (canWidenShuffleElements(Mask, WidenedMask))
10323 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10326 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10330 // Check for being able to broadcast a single element.
10331 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10332 Mask, Subtarget, DAG))
10335 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10336 // use lower latency instructions that will operate on both 128-bit lanes.
10337 SmallVector<int, 2> RepeatedMask;
10338 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10339 if (isSingleInputShuffleMask(Mask)) {
10340 int PSHUFDMask[] = {-1, -1, -1, -1};
10341 for (int i = 0; i < 2; ++i)
10342 if (RepeatedMask[i] >= 0) {
10343 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10344 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10346 return DAG.getBitcast(
10348 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10349 DAG.getBitcast(MVT::v8i32, V1),
10350 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10354 // AVX2 provides a direct instruction for permuting a single input across
10356 if (isSingleInputShuffleMask(Mask))
10357 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10358 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10360 // Try to use shift instructions.
10361 if (SDValue Shift =
10362 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10365 // Use dedicated unpack instructions for masks that match their pattern.
10367 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10370 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10371 // shuffle. However, if we have AVX2 and either inputs are already in place,
10372 // we will be able to shuffle even across lanes the other input in a single
10373 // instruction so skip this pattern.
10374 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10375 isShuffleMaskInputInPlace(1, Mask))))
10376 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10377 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10380 // Otherwise fall back on generic blend lowering.
10381 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10385 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10387 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10388 /// isn't available.
10389 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10390 const X86Subtarget *Subtarget,
10391 SelectionDAG &DAG) {
10393 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10394 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10395 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10396 ArrayRef<int> Mask = SVOp->getMask();
10397 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10399 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10403 // Check for being able to broadcast a single element.
10404 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10405 Mask, Subtarget, DAG))
10408 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10409 // options to efficiently lower the shuffle.
10410 SmallVector<int, 4> RepeatedMask;
10411 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10412 assert(RepeatedMask.size() == 4 &&
10413 "Repeated masks must be half the mask width!");
10415 // Use even/odd duplicate instructions for masks that match their pattern.
10416 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10417 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10418 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10419 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10421 if (isSingleInputShuffleMask(Mask))
10422 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10423 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10425 // Use dedicated unpack instructions for masks that match their pattern.
10427 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10430 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10431 // have already handled any direct blends. We also need to squash the
10432 // repeated mask into a simulated v4f32 mask.
10433 for (int i = 0; i < 4; ++i)
10434 if (RepeatedMask[i] >= 8)
10435 RepeatedMask[i] -= 4;
10436 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10439 // If we have a single input shuffle with different shuffle patterns in the
10440 // two 128-bit lanes use the variable mask to VPERMILPS.
10441 if (isSingleInputShuffleMask(Mask)) {
10442 SDValue VPermMask[8];
10443 for (int i = 0; i < 8; ++i)
10444 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10445 : DAG.getConstant(Mask[i], DL, MVT::i32);
10446 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10447 return DAG.getNode(
10448 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10449 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10451 if (Subtarget->hasAVX2())
10452 return DAG.getNode(
10453 X86ISD::VPERMV, DL, MVT::v8f32,
10454 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10455 MVT::v8i32, VPermMask)),
10458 // Otherwise, fall back.
10459 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10463 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10465 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10466 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10469 // If we have AVX2 then we always want to lower with a blend because at v8 we
10470 // can fully permute the elements.
10471 if (Subtarget->hasAVX2())
10472 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10475 // Otherwise fall back on generic lowering.
10476 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10479 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10481 /// This routine is only called when we have AVX2 and thus a reasonable
10482 /// instruction set for v8i32 shuffling..
10483 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10484 const X86Subtarget *Subtarget,
10485 SelectionDAG &DAG) {
10487 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10488 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10489 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10490 ArrayRef<int> Mask = SVOp->getMask();
10491 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10492 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10494 // Whenever we can lower this as a zext, that instruction is strictly faster
10495 // than any alternative. It also allows us to fold memory operands into the
10496 // shuffle in many cases.
10497 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10498 Mask, Subtarget, DAG))
10501 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10505 // Check for being able to broadcast a single element.
10506 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10507 Mask, Subtarget, DAG))
10510 // If the shuffle mask is repeated in each 128-bit lane we can use more
10511 // efficient instructions that mirror the shuffles across the two 128-bit
10513 SmallVector<int, 4> RepeatedMask;
10514 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10515 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10516 if (isSingleInputShuffleMask(Mask))
10517 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10518 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10520 // Use dedicated unpack instructions for masks that match their pattern.
10522 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10526 // Try to use shift instructions.
10527 if (SDValue Shift =
10528 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10531 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10532 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10535 // If the shuffle patterns aren't repeated but it is a single input, directly
10536 // generate a cross-lane VPERMD instruction.
10537 if (isSingleInputShuffleMask(Mask)) {
10538 SDValue VPermMask[8];
10539 for (int i = 0; i < 8; ++i)
10540 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10541 : DAG.getConstant(Mask[i], DL, MVT::i32);
10542 return DAG.getNode(
10543 X86ISD::VPERMV, DL, MVT::v8i32,
10544 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10547 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10549 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10550 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10553 // Otherwise fall back on generic blend lowering.
10554 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10558 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10560 /// This routine is only called when we have AVX2 and thus a reasonable
10561 /// instruction set for v16i16 shuffling..
10562 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10563 const X86Subtarget *Subtarget,
10564 SelectionDAG &DAG) {
10566 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10567 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10568 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10569 ArrayRef<int> Mask = SVOp->getMask();
10570 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10571 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10573 // Whenever we can lower this as a zext, that instruction is strictly faster
10574 // than any alternative. It also allows us to fold memory operands into the
10575 // shuffle in many cases.
10576 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10577 Mask, Subtarget, DAG))
10580 // Check for being able to broadcast a single element.
10581 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10582 Mask, Subtarget, DAG))
10585 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10589 // Use dedicated unpack instructions for masks that match their pattern.
10591 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10594 // Try to use shift instructions.
10595 if (SDValue Shift =
10596 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10599 // Try to use byte rotation instructions.
10600 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10601 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10604 if (isSingleInputShuffleMask(Mask)) {
10605 // There are no generalized cross-lane shuffle operations available on i16
10607 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10608 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10611 SmallVector<int, 8> RepeatedMask;
10612 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10613 // As this is a single-input shuffle, the repeated mask should be
10614 // a strictly valid v8i16 mask that we can pass through to the v8i16
10615 // lowering to handle even the v16 case.
10616 return lowerV8I16GeneralSingleInputVectorShuffle(
10617 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10620 SDValue PSHUFBMask[32];
10621 for (int i = 0; i < 16; ++i) {
10622 if (Mask[i] == -1) {
10623 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10627 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10628 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10629 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10630 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10632 return DAG.getBitcast(MVT::v16i16,
10633 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10634 DAG.getBitcast(MVT::v32i8, V1),
10635 DAG.getNode(ISD::BUILD_VECTOR, DL,
10636 MVT::v32i8, PSHUFBMask)));
10639 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10641 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10642 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10645 // Otherwise fall back on generic lowering.
10646 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10649 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10651 /// This routine is only called when we have AVX2 and thus a reasonable
10652 /// instruction set for v32i8 shuffling..
10653 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10654 const X86Subtarget *Subtarget,
10655 SelectionDAG &DAG) {
10657 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10658 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10660 ArrayRef<int> Mask = SVOp->getMask();
10661 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10662 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10664 // Whenever we can lower this as a zext, that instruction is strictly faster
10665 // than any alternative. It also allows us to fold memory operands into the
10666 // shuffle in many cases.
10667 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10668 Mask, Subtarget, DAG))
10671 // Check for being able to broadcast a single element.
10672 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10673 Mask, Subtarget, DAG))
10676 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10680 // Use dedicated unpack instructions for masks that match their pattern.
10682 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10685 // Try to use shift instructions.
10686 if (SDValue Shift =
10687 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10690 // Try to use byte rotation instructions.
10691 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10692 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10695 if (isSingleInputShuffleMask(Mask)) {
10696 // There are no generalized cross-lane shuffle operations available on i8
10698 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10699 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10702 SDValue PSHUFBMask[32];
10703 for (int i = 0; i < 32; ++i)
10706 ? DAG.getUNDEF(MVT::i8)
10707 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10710 return DAG.getNode(
10711 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10712 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10715 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10717 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10718 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10721 // Otherwise fall back on generic lowering.
10722 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10725 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10727 /// This routine either breaks down the specific type of a 256-bit x86 vector
10728 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10729 /// together based on the available instructions.
10730 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10731 MVT VT, const X86Subtarget *Subtarget,
10732 SelectionDAG &DAG) {
10734 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10735 ArrayRef<int> Mask = SVOp->getMask();
10737 // If we have a single input to the zero element, insert that into V1 if we
10738 // can do so cheaply.
10739 int NumElts = VT.getVectorNumElements();
10740 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10741 return M >= NumElts;
10744 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10745 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10746 DL, VT, V1, V2, Mask, Subtarget, DAG))
10749 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10750 // can check for those subtargets here and avoid much of the subtarget
10751 // querying in the per-vector-type lowering routines. With AVX1 we have
10752 // essentially *zero* ability to manipulate a 256-bit vector with integer
10753 // types. Since we'll use floating point types there eventually, just
10754 // immediately cast everything to a float and operate entirely in that domain.
10755 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10756 int ElementBits = VT.getScalarSizeInBits();
10757 if (ElementBits < 32)
10758 // No floating point type available, decompose into 128-bit vectors.
10759 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10761 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10762 VT.getVectorNumElements());
10763 V1 = DAG.getBitcast(FpVT, V1);
10764 V2 = DAG.getBitcast(FpVT, V2);
10765 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10768 switch (VT.SimpleTy) {
10770 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10772 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10774 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10776 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10778 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10780 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10783 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10787 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10788 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10789 ArrayRef<int> Mask,
10790 SDValue V1, SDValue V2,
10791 SelectionDAG &DAG) {
10792 assert(VT.getScalarSizeInBits() == 64 &&
10793 "Unexpected element type size for 128bit shuffle.");
10795 // To handle 256 bit vector requires VLX and most probably
10796 // function lowerV2X128VectorShuffle() is better solution.
10797 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10799 SmallVector<int, 4> WidenedMask;
10800 if (!canWidenShuffleElements(Mask, WidenedMask))
10803 // Form a 128-bit permutation.
10804 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10805 // bits defined by a vshuf64x2 instruction's immediate control byte.
10806 unsigned PermMask = 0, Imm = 0;
10807 unsigned ControlBitsNum = WidenedMask.size() / 2;
10809 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10810 if (WidenedMask[i] == SM_SentinelZero)
10813 // Use first element in place of undef mask.
10814 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10815 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10818 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10819 DAG.getConstant(PermMask, DL, MVT::i8));
10822 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10823 ArrayRef<int> Mask, SDValue V1,
10824 SDValue V2, SelectionDAG &DAG) {
10826 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10828 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10829 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10831 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10832 if (isSingleInputShuffleMask(Mask))
10833 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10835 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10838 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10839 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10840 const X86Subtarget *Subtarget,
10841 SelectionDAG &DAG) {
10843 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10844 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10845 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10846 ArrayRef<int> Mask = SVOp->getMask();
10847 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10849 if (SDValue Shuf128 =
10850 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10853 if (SDValue Unpck =
10854 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10857 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10860 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10861 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10862 const X86Subtarget *Subtarget,
10863 SelectionDAG &DAG) {
10865 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10866 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10867 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10868 ArrayRef<int> Mask = SVOp->getMask();
10869 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10871 if (SDValue Unpck =
10872 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10875 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10878 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10879 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10880 const X86Subtarget *Subtarget,
10881 SelectionDAG &DAG) {
10883 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10884 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10885 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10886 ArrayRef<int> Mask = SVOp->getMask();
10887 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10889 if (SDValue Shuf128 =
10890 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10893 if (SDValue Unpck =
10894 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10897 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10900 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10901 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10902 const X86Subtarget *Subtarget,
10903 SelectionDAG &DAG) {
10905 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10906 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10907 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10908 ArrayRef<int> Mask = SVOp->getMask();
10909 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10911 if (SDValue Unpck =
10912 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10915 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10918 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10919 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10920 const X86Subtarget *Subtarget,
10921 SelectionDAG &DAG) {
10923 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10924 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10925 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10926 ArrayRef<int> Mask = SVOp->getMask();
10927 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10928 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10930 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10933 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10934 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10935 const X86Subtarget *Subtarget,
10936 SelectionDAG &DAG) {
10938 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10939 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10940 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10941 ArrayRef<int> Mask = SVOp->getMask();
10942 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10943 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10945 // FIXME: Implement direct support for this type!
10946 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10949 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10951 /// This routine either breaks down the specific type of a 512-bit x86 vector
10952 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10953 /// together based on the available instructions.
10954 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10955 MVT VT, const X86Subtarget *Subtarget,
10956 SelectionDAG &DAG) {
10958 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10959 ArrayRef<int> Mask = SVOp->getMask();
10960 assert(Subtarget->hasAVX512() &&
10961 "Cannot lower 512-bit vectors w/ basic ISA!");
10963 // Check for being able to broadcast a single element.
10964 if (SDValue Broadcast =
10965 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10968 // Dispatch to each element type for lowering. If we don't have supprot for
10969 // specific element type shuffles at 512 bits, immediately split them and
10970 // lower them. Each lowering routine of a given type is allowed to assume that
10971 // the requisite ISA extensions for that element type are available.
10972 switch (VT.SimpleTy) {
10974 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10976 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10978 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10980 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10982 if (Subtarget->hasBWI())
10983 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10986 if (Subtarget->hasBWI())
10987 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10991 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10994 // Otherwise fall back on splitting.
10995 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10998 // Lower vXi1 vector shuffles.
10999 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11000 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11001 // vector, shuffle and then truncate it back.
11002 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11003 MVT VT, const X86Subtarget *Subtarget,
11004 SelectionDAG &DAG) {
11006 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11007 ArrayRef<int> Mask = SVOp->getMask();
11008 assert(Subtarget->hasAVX512() &&
11009 "Cannot lower 512-bit vectors w/o basic ISA!");
11011 switch (VT.SimpleTy) {
11013 llvm_unreachable("Expected a vector of i1 elements");
11015 ExtVT = MVT::v2i64;
11018 ExtVT = MVT::v4i32;
11021 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11024 ExtVT = MVT::v16i32;
11027 ExtVT = MVT::v32i16;
11030 ExtVT = MVT::v64i8;
11034 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11035 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11036 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11037 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11039 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11042 V2 = DAG.getUNDEF(ExtVT);
11043 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11044 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11045 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11046 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11048 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11049 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11050 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11052 /// \brief Top-level lowering for x86 vector shuffles.
11054 /// This handles decomposition, canonicalization, and lowering of all x86
11055 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11056 /// above in helper routines. The canonicalization attempts to widen shuffles
11057 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11058 /// s.t. only one of the two inputs needs to be tested, etc.
11059 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11060 SelectionDAG &DAG) {
11061 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11062 ArrayRef<int> Mask = SVOp->getMask();
11063 SDValue V1 = Op.getOperand(0);
11064 SDValue V2 = Op.getOperand(1);
11065 MVT VT = Op.getSimpleValueType();
11066 int NumElements = VT.getVectorNumElements();
11068 bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
11070 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11071 "Can't lower MMX shuffles");
11073 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11074 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11075 if (V1IsUndef && V2IsUndef)
11076 return DAG.getUNDEF(VT);
11078 // When we create a shuffle node we put the UNDEF node to second operand,
11079 // but in some cases the first operand may be transformed to UNDEF.
11080 // In this case we should just commute the node.
11082 return DAG.getCommutedVectorShuffle(*SVOp);
11084 // Check for non-undef masks pointing at an undef vector and make the masks
11085 // undef as well. This makes it easier to match the shuffle based solely on
11089 if (M >= NumElements) {
11090 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11091 for (int &M : NewMask)
11092 if (M >= NumElements)
11094 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11097 // We actually see shuffles that are entirely re-arrangements of a set of
11098 // zero inputs. This mostly happens while decomposing complex shuffles into
11099 // simple ones. Directly lower these as a buildvector of zeros.
11100 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11101 if (Zeroable.all())
11102 return getZeroVector(VT, Subtarget, DAG, dl);
11104 // Try to collapse shuffles into using a vector type with fewer elements but
11105 // wider element types. We cap this to not form integers or floating point
11106 // elements wider than 64 bits, but it might be interesting to form i128
11107 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11108 SmallVector<int, 16> WidenedMask;
11109 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11110 canWidenShuffleElements(Mask, WidenedMask)) {
11111 MVT NewEltVT = VT.isFloatingPoint()
11112 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11113 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11114 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11115 // Make sure that the new vector type is legal. For example, v2f64 isn't
11117 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11118 V1 = DAG.getBitcast(NewVT, V1);
11119 V2 = DAG.getBitcast(NewVT, V2);
11120 return DAG.getBitcast(
11121 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11125 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11126 for (int M : SVOp->getMask())
11128 ++NumUndefElements;
11129 else if (M < NumElements)
11134 // Commute the shuffle as needed such that more elements come from V1 than
11135 // V2. This allows us to match the shuffle pattern strictly on how many
11136 // elements come from V1 without handling the symmetric cases.
11137 if (NumV2Elements > NumV1Elements)
11138 return DAG.getCommutedVectorShuffle(*SVOp);
11140 // When the number of V1 and V2 elements are the same, try to minimize the
11141 // number of uses of V2 in the low half of the vector. When that is tied,
11142 // ensure that the sum of indices for V1 is equal to or lower than the sum
11143 // indices for V2. When those are equal, try to ensure that the number of odd
11144 // indices for V1 is lower than the number of odd indices for V2.
11145 if (NumV1Elements == NumV2Elements) {
11146 int LowV1Elements = 0, LowV2Elements = 0;
11147 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11148 if (M >= NumElements)
11152 if (LowV2Elements > LowV1Elements) {
11153 return DAG.getCommutedVectorShuffle(*SVOp);
11154 } else if (LowV2Elements == LowV1Elements) {
11155 int SumV1Indices = 0, SumV2Indices = 0;
11156 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11157 if (SVOp->getMask()[i] >= NumElements)
11159 else if (SVOp->getMask()[i] >= 0)
11161 if (SumV2Indices < SumV1Indices) {
11162 return DAG.getCommutedVectorShuffle(*SVOp);
11163 } else if (SumV2Indices == SumV1Indices) {
11164 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11165 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11166 if (SVOp->getMask()[i] >= NumElements)
11167 NumV2OddIndices += i % 2;
11168 else if (SVOp->getMask()[i] >= 0)
11169 NumV1OddIndices += i % 2;
11170 if (NumV2OddIndices < NumV1OddIndices)
11171 return DAG.getCommutedVectorShuffle(*SVOp);
11176 // For each vector width, delegate to a specialized lowering routine.
11177 if (VT.is128BitVector())
11178 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11180 if (VT.is256BitVector())
11181 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11183 if (VT.is512BitVector())
11184 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11187 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11188 llvm_unreachable("Unimplemented!");
11191 // This function assumes its argument is a BUILD_VECTOR of constants or
11192 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11194 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11195 unsigned &MaskValue) {
11197 unsigned NumElems = BuildVector->getNumOperands();
11199 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11200 // We don't handle the >2 lanes case right now.
11201 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11205 unsigned NumElemsInLane = NumElems / NumLanes;
11207 // Blend for v16i16 should be symmetric for the both lanes.
11208 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11209 SDValue EltCond = BuildVector->getOperand(i);
11210 SDValue SndLaneEltCond =
11211 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11213 int Lane1Cond = -1, Lane2Cond = -1;
11214 if (isa<ConstantSDNode>(EltCond))
11215 Lane1Cond = !isZero(EltCond);
11216 if (isa<ConstantSDNode>(SndLaneEltCond))
11217 Lane2Cond = !isZero(SndLaneEltCond);
11219 unsigned LaneMask = 0;
11220 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11221 // Lane1Cond != 0, means we want the first argument.
11222 // Lane1Cond == 0, means we want the second argument.
11223 // The encoding of this argument is 0 for the first argument, 1
11224 // for the second. Therefore, invert the condition.
11225 LaneMask = !Lane1Cond << i;
11226 else if (Lane1Cond < 0)
11227 LaneMask = !Lane2Cond << i;
11231 MaskValue |= LaneMask;
11233 MaskValue |= LaneMask << NumElemsInLane;
11238 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11239 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11240 const X86Subtarget *Subtarget,
11241 SelectionDAG &DAG) {
11242 SDValue Cond = Op.getOperand(0);
11243 SDValue LHS = Op.getOperand(1);
11244 SDValue RHS = Op.getOperand(2);
11246 MVT VT = Op.getSimpleValueType();
11248 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11250 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11252 // Only non-legal VSELECTs reach this lowering, convert those into generic
11253 // shuffles and re-use the shuffle lowering path for blends.
11254 SmallVector<int, 32> Mask;
11255 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11256 SDValue CondElt = CondBV->getOperand(i);
11258 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11260 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11263 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11264 // A vselect where all conditions and data are constants can be optimized into
11265 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11266 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11267 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11268 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11271 // Try to lower this to a blend-style vector shuffle. This can handle all
11272 // constant condition cases.
11273 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11276 // Variable blends are only legal from SSE4.1 onward.
11277 if (!Subtarget->hasSSE41())
11280 // Only some types will be legal on some subtargets. If we can emit a legal
11281 // VSELECT-matching blend, return Op, and but if we need to expand, return
11283 switch (Op.getSimpleValueType().SimpleTy) {
11285 // Most of the vector types have blends past SSE4.1.
11289 // The byte blends for AVX vectors were introduced only in AVX2.
11290 if (Subtarget->hasAVX2())
11297 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11298 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11301 // FIXME: We should custom lower this by fixing the condition and using i8
11307 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11308 MVT VT = Op.getSimpleValueType();
11311 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11314 if (VT.getSizeInBits() == 8) {
11315 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11316 Op.getOperand(0), Op.getOperand(1));
11317 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11318 DAG.getValueType(VT));
11319 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11322 if (VT.getSizeInBits() == 16) {
11323 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11324 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11326 return DAG.getNode(
11327 ISD::TRUNCATE, dl, MVT::i16,
11328 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11329 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11330 Op.getOperand(1)));
11331 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11332 Op.getOperand(0), Op.getOperand(1));
11333 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11334 DAG.getValueType(VT));
11335 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11338 if (VT == MVT::f32) {
11339 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11340 // the result back to FR32 register. It's only worth matching if the
11341 // result has a single use which is a store or a bitcast to i32. And in
11342 // the case of a store, it's not worth it if the index is a constant 0,
11343 // because a MOVSSmr can be used instead, which is smaller and faster.
11344 if (!Op.hasOneUse())
11346 SDNode *User = *Op.getNode()->use_begin();
11347 if ((User->getOpcode() != ISD::STORE ||
11348 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11349 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11350 (User->getOpcode() != ISD::BITCAST ||
11351 User->getValueType(0) != MVT::i32))
11353 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11354 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11356 return DAG.getBitcast(MVT::f32, Extract);
11359 if (VT == MVT::i32 || VT == MVT::i64) {
11360 // ExtractPS/pextrq works with constant index.
11361 if (isa<ConstantSDNode>(Op.getOperand(1)))
11367 /// Extract one bit from mask vector, like v16i1 or v8i1.
11368 /// AVX-512 feature.
11370 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11371 SDValue Vec = Op.getOperand(0);
11373 MVT VecVT = Vec.getSimpleValueType();
11374 SDValue Idx = Op.getOperand(1);
11375 MVT EltVT = Op.getSimpleValueType();
11377 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11378 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11379 "Unexpected vector type in ExtractBitFromMaskVector");
11381 // variable index can't be handled in mask registers,
11382 // extend vector to VR512
11383 if (!isa<ConstantSDNode>(Idx)) {
11384 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11385 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11386 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11387 ExtVT.getVectorElementType(), Ext, Idx);
11388 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11391 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11392 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11393 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11394 rc = getRegClassFor(MVT::v16i1);
11395 unsigned MaxSift = rc->getSize()*8 - 1;
11396 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11397 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11398 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11399 DAG.getConstant(MaxSift, dl, MVT::i8));
11400 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11401 DAG.getIntPtrConstant(0, dl));
11405 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11406 SelectionDAG &DAG) const {
11408 SDValue Vec = Op.getOperand(0);
11409 MVT VecVT = Vec.getSimpleValueType();
11410 SDValue Idx = Op.getOperand(1);
11412 if (Op.getSimpleValueType() == MVT::i1)
11413 return ExtractBitFromMaskVector(Op, DAG);
11415 if (!isa<ConstantSDNode>(Idx)) {
11416 if (VecVT.is512BitVector() ||
11417 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11418 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11421 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11422 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11423 MaskEltVT.getSizeInBits());
11425 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11426 auto PtrVT = getPointerTy(DAG.getDataLayout());
11427 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11428 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11429 DAG.getConstant(0, dl, PtrVT));
11430 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11431 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11432 DAG.getConstant(0, dl, PtrVT));
11437 // If this is a 256-bit vector result, first extract the 128-bit vector and
11438 // then extract the element from the 128-bit vector.
11439 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11441 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11442 // Get the 128-bit vector.
11443 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11444 MVT EltVT = VecVT.getVectorElementType();
11446 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11447 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11449 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11450 // this can be done with a mask.
11451 IdxVal &= ElemsPerChunk - 1;
11452 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11453 DAG.getConstant(IdxVal, dl, MVT::i32));
11456 assert(VecVT.is128BitVector() && "Unexpected vector length");
11458 if (Subtarget->hasSSE41())
11459 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11462 MVT VT = Op.getSimpleValueType();
11463 // TODO: handle v16i8.
11464 if (VT.getSizeInBits() == 16) {
11465 SDValue Vec = Op.getOperand(0);
11466 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11468 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11469 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11470 DAG.getBitcast(MVT::v4i32, Vec),
11471 Op.getOperand(1)));
11472 // Transform it so it match pextrw which produces a 32-bit result.
11473 MVT EltVT = MVT::i32;
11474 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11475 Op.getOperand(0), Op.getOperand(1));
11476 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11477 DAG.getValueType(VT));
11478 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11481 if (VT.getSizeInBits() == 32) {
11482 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11486 // SHUFPS the element to the lowest double word, then movss.
11487 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11488 MVT VVT = Op.getOperand(0).getSimpleValueType();
11489 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11490 DAG.getUNDEF(VVT), Mask);
11491 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11492 DAG.getIntPtrConstant(0, dl));
11495 if (VT.getSizeInBits() == 64) {
11496 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11497 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11498 // to match extract_elt for f64.
11499 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11503 // UNPCKHPD the element to the lowest double word, then movsd.
11504 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11505 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11506 int Mask[2] = { 1, -1 };
11507 MVT VVT = Op.getOperand(0).getSimpleValueType();
11508 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11509 DAG.getUNDEF(VVT), Mask);
11510 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11511 DAG.getIntPtrConstant(0, dl));
11517 /// Insert one bit to mask vector, like v16i1 or v8i1.
11518 /// AVX-512 feature.
11520 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11522 SDValue Vec = Op.getOperand(0);
11523 SDValue Elt = Op.getOperand(1);
11524 SDValue Idx = Op.getOperand(2);
11525 MVT VecVT = Vec.getSimpleValueType();
11527 if (!isa<ConstantSDNode>(Idx)) {
11528 // Non constant index. Extend source and destination,
11529 // insert element and then truncate the result.
11530 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11531 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11532 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11533 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11534 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11535 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11538 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11539 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11541 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11542 DAG.getConstant(IdxVal, dl, MVT::i8));
11543 if (Vec.getOpcode() == ISD::UNDEF)
11545 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11548 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11549 SelectionDAG &DAG) const {
11550 MVT VT = Op.getSimpleValueType();
11551 MVT EltVT = VT.getVectorElementType();
11553 if (EltVT == MVT::i1)
11554 return InsertBitToMaskVector(Op, DAG);
11557 SDValue N0 = Op.getOperand(0);
11558 SDValue N1 = Op.getOperand(1);
11559 SDValue N2 = Op.getOperand(2);
11560 if (!isa<ConstantSDNode>(N2))
11562 auto *N2C = cast<ConstantSDNode>(N2);
11563 unsigned IdxVal = N2C->getZExtValue();
11565 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11566 // into that, and then insert the subvector back into the result.
11567 if (VT.is256BitVector() || VT.is512BitVector()) {
11568 // With a 256-bit vector, we can insert into the zero element efficiently
11569 // using a blend if we have AVX or AVX2 and the right data type.
11570 if (VT.is256BitVector() && IdxVal == 0) {
11571 // TODO: It is worthwhile to cast integer to floating point and back
11572 // and incur a domain crossing penalty if that's what we'll end up
11573 // doing anyway after extracting to a 128-bit vector.
11574 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11575 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11576 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11577 N2 = DAG.getIntPtrConstant(1, dl);
11578 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11582 // Get the desired 128-bit vector chunk.
11583 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11585 // Insert the element into the desired chunk.
11586 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11587 assert(isPowerOf2_32(NumEltsIn128));
11588 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11589 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11591 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11592 DAG.getConstant(IdxIn128, dl, MVT::i32));
11594 // Insert the changed part back into the bigger vector
11595 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11597 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11599 if (Subtarget->hasSSE41()) {
11600 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11602 if (VT == MVT::v8i16) {
11603 Opc = X86ISD::PINSRW;
11605 assert(VT == MVT::v16i8);
11606 Opc = X86ISD::PINSRB;
11609 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11611 if (N1.getValueType() != MVT::i32)
11612 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11613 if (N2.getValueType() != MVT::i32)
11614 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11615 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11618 if (EltVT == MVT::f32) {
11619 // Bits [7:6] of the constant are the source select. This will always be
11620 // zero here. The DAG Combiner may combine an extract_elt index into
11621 // these bits. For example (insert (extract, 3), 2) could be matched by
11622 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11623 // Bits [5:4] of the constant are the destination select. This is the
11624 // value of the incoming immediate.
11625 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11626 // combine either bitwise AND or insert of float 0.0 to set these bits.
11628 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11629 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11630 // If this is an insertion of 32-bits into the low 32-bits of
11631 // a vector, we prefer to generate a blend with immediate rather
11632 // than an insertps. Blends are simpler operations in hardware and so
11633 // will always have equal or better performance than insertps.
11634 // But if optimizing for size and there's a load folding opportunity,
11635 // generate insertps because blendps does not have a 32-bit memory
11637 N2 = DAG.getIntPtrConstant(1, dl);
11638 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11639 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11641 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11642 // Create this as a scalar to vector..
11643 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11644 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11647 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11648 // PINSR* works with constant index.
11653 if (EltVT == MVT::i8)
11656 if (EltVT.getSizeInBits() == 16) {
11657 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11658 // as its second argument.
11659 if (N1.getValueType() != MVT::i32)
11660 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11661 if (N2.getValueType() != MVT::i32)
11662 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11663 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11668 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11670 MVT OpVT = Op.getSimpleValueType();
11672 // If this is a 256-bit vector result, first insert into a 128-bit
11673 // vector and then insert into the 256-bit vector.
11674 if (!OpVT.is128BitVector()) {
11675 // Insert into a 128-bit vector.
11676 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11677 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11678 OpVT.getVectorNumElements() / SizeFactor);
11680 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11682 // Insert the 128-bit vector.
11683 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11686 if (OpVT == MVT::v1i64 &&
11687 Op.getOperand(0).getValueType() == MVT::i64)
11688 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11690 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11691 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11692 return DAG.getBitcast(
11693 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11696 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11697 // a simple subregister reference or explicit instructions to grab
11698 // upper bits of a vector.
11699 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11700 SelectionDAG &DAG) {
11702 SDValue In = Op.getOperand(0);
11703 SDValue Idx = Op.getOperand(1);
11704 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11705 MVT ResVT = Op.getSimpleValueType();
11706 MVT InVT = In.getSimpleValueType();
11708 if (Subtarget->hasFp256()) {
11709 if (ResVT.is128BitVector() &&
11710 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11711 isa<ConstantSDNode>(Idx)) {
11712 return Extract128BitVector(In, IdxVal, DAG, dl);
11714 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11715 isa<ConstantSDNode>(Idx)) {
11716 return Extract256BitVector(In, IdxVal, DAG, dl);
11722 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11723 // simple superregister reference or explicit instructions to insert
11724 // the upper bits of a vector.
11725 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11726 SelectionDAG &DAG) {
11727 if (!Subtarget->hasAVX())
11731 SDValue Vec = Op.getOperand(0);
11732 SDValue SubVec = Op.getOperand(1);
11733 SDValue Idx = Op.getOperand(2);
11735 if (!isa<ConstantSDNode>(Idx))
11738 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11739 MVT OpVT = Op.getSimpleValueType();
11740 MVT SubVecVT = SubVec.getSimpleValueType();
11742 // Fold two 16-byte subvector loads into one 32-byte load:
11743 // (insert_subvector (insert_subvector undef, (load addr), 0),
11744 // (load addr + 16), Elts/2)
11746 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11747 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11748 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11749 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11750 if (Idx2 && Idx2->getZExtValue() == 0) {
11751 SDValue SubVec2 = Vec.getOperand(1);
11752 // If needed, look through a bitcast to get to the load.
11753 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11754 SubVec2 = SubVec2.getOperand(0);
11756 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11758 unsigned Alignment = FirstLd->getAlignment();
11759 unsigned AS = FirstLd->getAddressSpace();
11760 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11761 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11762 OpVT, AS, Alignment, &Fast) && Fast) {
11763 SDValue Ops[] = { SubVec2, SubVec };
11764 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11771 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11772 SubVecVT.is128BitVector())
11773 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11775 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11776 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11778 if (OpVT.getVectorElementType() == MVT::i1) {
11779 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11781 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11782 SDValue Undef = DAG.getUNDEF(OpVT);
11783 unsigned NumElems = OpVT.getVectorNumElements();
11784 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11786 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11787 // Zero upper bits of the Vec
11788 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11789 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11791 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11793 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11794 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11797 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11799 // Zero upper bits of the Vec2
11800 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11801 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11802 // Zero lower bits of the Vec
11803 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11804 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11805 // Merge them together
11806 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11812 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11813 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11814 // one of the above mentioned nodes. It has to be wrapped because otherwise
11815 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11816 // be used to form addressing mode. These wrapped nodes will be selected
11819 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11820 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11822 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11823 // global base reg.
11824 unsigned char OpFlag = 0;
11825 unsigned WrapperKind = X86ISD::Wrapper;
11826 CodeModel::Model M = DAG.getTarget().getCodeModel();
11828 if (Subtarget->isPICStyleRIPRel() &&
11829 (M == CodeModel::Small || M == CodeModel::Kernel))
11830 WrapperKind = X86ISD::WrapperRIP;
11831 else if (Subtarget->isPICStyleGOT())
11832 OpFlag = X86II::MO_GOTOFF;
11833 else if (Subtarget->isPICStyleStubPIC())
11834 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11836 auto PtrVT = getPointerTy(DAG.getDataLayout());
11837 SDValue Result = DAG.getTargetConstantPool(
11838 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11840 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11841 // With PIC, the address is actually $g + Offset.
11844 DAG.getNode(ISD::ADD, DL, PtrVT,
11845 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11851 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11852 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11854 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11855 // global base reg.
11856 unsigned char OpFlag = 0;
11857 unsigned WrapperKind = X86ISD::Wrapper;
11858 CodeModel::Model M = DAG.getTarget().getCodeModel();
11860 if (Subtarget->isPICStyleRIPRel() &&
11861 (M == CodeModel::Small || M == CodeModel::Kernel))
11862 WrapperKind = X86ISD::WrapperRIP;
11863 else if (Subtarget->isPICStyleGOT())
11864 OpFlag = X86II::MO_GOTOFF;
11865 else if (Subtarget->isPICStyleStubPIC())
11866 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11868 auto PtrVT = getPointerTy(DAG.getDataLayout());
11869 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11871 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11873 // With PIC, the address is actually $g + Offset.
11876 DAG.getNode(ISD::ADD, DL, PtrVT,
11877 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11883 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11884 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11886 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11887 // global base reg.
11888 unsigned char OpFlag = 0;
11889 unsigned WrapperKind = X86ISD::Wrapper;
11890 CodeModel::Model M = DAG.getTarget().getCodeModel();
11892 if (Subtarget->isPICStyleRIPRel() &&
11893 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11894 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11895 OpFlag = X86II::MO_GOTPCREL;
11896 WrapperKind = X86ISD::WrapperRIP;
11897 } else if (Subtarget->isPICStyleGOT()) {
11898 OpFlag = X86II::MO_GOT;
11899 } else if (Subtarget->isPICStyleStubPIC()) {
11900 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11901 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11902 OpFlag = X86II::MO_DARWIN_NONLAZY;
11905 auto PtrVT = getPointerTy(DAG.getDataLayout());
11906 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11909 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11911 // With PIC, the address is actually $g + Offset.
11912 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11913 !Subtarget->is64Bit()) {
11915 DAG.getNode(ISD::ADD, DL, PtrVT,
11916 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11919 // For symbols that require a load from a stub to get the address, emit the
11921 if (isGlobalStubReference(OpFlag))
11922 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11923 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11924 false, false, false, 0);
11930 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11931 // Create the TargetBlockAddressAddress node.
11932 unsigned char OpFlags =
11933 Subtarget->ClassifyBlockAddressReference();
11934 CodeModel::Model M = DAG.getTarget().getCodeModel();
11935 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11936 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11938 auto PtrVT = getPointerTy(DAG.getDataLayout());
11939 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11941 if (Subtarget->isPICStyleRIPRel() &&
11942 (M == CodeModel::Small || M == CodeModel::Kernel))
11943 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11945 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11947 // With PIC, the address is actually $g + Offset.
11948 if (isGlobalRelativeToPICBase(OpFlags)) {
11949 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11950 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11957 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11958 int64_t Offset, SelectionDAG &DAG) const {
11959 // Create the TargetGlobalAddress node, folding in the constant
11960 // offset if it is legal.
11961 unsigned char OpFlags =
11962 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11963 CodeModel::Model M = DAG.getTarget().getCodeModel();
11964 auto PtrVT = getPointerTy(DAG.getDataLayout());
11966 if (OpFlags == X86II::MO_NO_FLAG &&
11967 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11968 // A direct static reference to a global.
11969 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11972 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11975 if (Subtarget->isPICStyleRIPRel() &&
11976 (M == CodeModel::Small || M == CodeModel::Kernel))
11977 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11979 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11981 // With PIC, the address is actually $g + Offset.
11982 if (isGlobalRelativeToPICBase(OpFlags)) {
11983 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11984 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11987 // For globals that require a load from a stub to get the address, emit the
11989 if (isGlobalStubReference(OpFlags))
11990 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11991 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11992 false, false, false, 0);
11994 // If there was a non-zero offset that we didn't fold, create an explicit
11995 // addition for it.
11997 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11998 DAG.getConstant(Offset, dl, PtrVT));
12004 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12005 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12006 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12007 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12011 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12012 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12013 unsigned char OperandFlags, bool LocalDynamic = false) {
12014 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12015 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12017 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12018 GA->getValueType(0),
12022 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12026 SDValue Ops[] = { Chain, TGA, *InFlag };
12027 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12029 SDValue Ops[] = { Chain, TGA };
12030 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12033 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12034 MFI->setAdjustsStack(true);
12035 MFI->setHasCalls(true);
12037 SDValue Flag = Chain.getValue(1);
12038 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12041 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12043 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12046 SDLoc dl(GA); // ? function entry point might be better
12047 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12048 DAG.getNode(X86ISD::GlobalBaseReg,
12049 SDLoc(), PtrVT), InFlag);
12050 InFlag = Chain.getValue(1);
12052 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12055 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12057 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12059 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12060 X86::RAX, X86II::MO_TLSGD);
12063 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12069 // Get the start address of the TLS block for this module.
12070 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12071 .getInfo<X86MachineFunctionInfo>();
12072 MFI->incNumLocalDynamicTLSAccesses();
12076 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12077 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12080 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12081 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12082 InFlag = Chain.getValue(1);
12083 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12084 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12087 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12091 unsigned char OperandFlags = X86II::MO_DTPOFF;
12092 unsigned WrapperKind = X86ISD::Wrapper;
12093 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12094 GA->getValueType(0),
12095 GA->getOffset(), OperandFlags);
12096 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12098 // Add x@dtpoff with the base.
12099 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12102 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12103 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12104 const EVT PtrVT, TLSModel::Model model,
12105 bool is64Bit, bool isPIC) {
12108 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12109 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12110 is64Bit ? 257 : 256));
12112 SDValue ThreadPointer =
12113 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12114 MachinePointerInfo(Ptr), false, false, false, 0);
12116 unsigned char OperandFlags = 0;
12117 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12119 unsigned WrapperKind = X86ISD::Wrapper;
12120 if (model == TLSModel::LocalExec) {
12121 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12122 } else if (model == TLSModel::InitialExec) {
12124 OperandFlags = X86II::MO_GOTTPOFF;
12125 WrapperKind = X86ISD::WrapperRIP;
12127 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12130 llvm_unreachable("Unexpected model");
12133 // emit "addl x@ntpoff,%eax" (local exec)
12134 // or "addl x@indntpoff,%eax" (initial exec)
12135 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12137 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12138 GA->getOffset(), OperandFlags);
12139 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12141 if (model == TLSModel::InitialExec) {
12142 if (isPIC && !is64Bit) {
12143 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12144 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12148 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12149 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12150 false, false, false, 0);
12153 // The address of the thread local variable is the add of the thread
12154 // pointer with the offset of the variable.
12155 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12159 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12161 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12162 const GlobalValue *GV = GA->getGlobal();
12163 auto PtrVT = getPointerTy(DAG.getDataLayout());
12165 if (Subtarget->isTargetELF()) {
12166 if (DAG.getTarget().Options.EmulatedTLS)
12167 return LowerToTLSEmulatedModel(GA, DAG);
12168 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12170 case TLSModel::GeneralDynamic:
12171 if (Subtarget->is64Bit())
12172 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12173 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12174 case TLSModel::LocalDynamic:
12175 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12176 Subtarget->is64Bit());
12177 case TLSModel::InitialExec:
12178 case TLSModel::LocalExec:
12179 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12180 DAG.getTarget().getRelocationModel() ==
12183 llvm_unreachable("Unknown TLS model.");
12186 if (Subtarget->isTargetDarwin()) {
12187 // Darwin only has one model of TLS. Lower to that.
12188 unsigned char OpFlag = 0;
12189 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12190 X86ISD::WrapperRIP : X86ISD::Wrapper;
12192 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12193 // global base reg.
12194 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12195 !Subtarget->is64Bit();
12197 OpFlag = X86II::MO_TLVP_PIC_BASE;
12199 OpFlag = X86II::MO_TLVP;
12201 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12202 GA->getValueType(0),
12203 GA->getOffset(), OpFlag);
12204 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12206 // With PIC32, the address is actually $g + Offset.
12208 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12209 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12212 // Lowering the machine isd will make sure everything is in the right
12214 SDValue Chain = DAG.getEntryNode();
12215 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12216 SDValue Args[] = { Chain, Offset };
12217 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12219 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12220 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12221 MFI->setAdjustsStack(true);
12223 // And our return value (tls address) is in the standard call return value
12225 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12226 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12229 if (Subtarget->isTargetKnownWindowsMSVC() ||
12230 Subtarget->isTargetWindowsGNU()) {
12231 // Just use the implicit TLS architecture
12232 // Need to generate someting similar to:
12233 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12235 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12236 // mov rcx, qword [rdx+rcx*8]
12237 // mov eax, .tls$:tlsvar
12238 // [rax+rcx] contains the address
12239 // Windows 64bit: gs:0x58
12240 // Windows 32bit: fs:__tls_array
12243 SDValue Chain = DAG.getEntryNode();
12245 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12246 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12247 // use its literal value of 0x2C.
12248 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12249 ? Type::getInt8PtrTy(*DAG.getContext(),
12251 : Type::getInt32PtrTy(*DAG.getContext(),
12254 SDValue TlsArray = Subtarget->is64Bit()
12255 ? DAG.getIntPtrConstant(0x58, dl)
12256 : (Subtarget->isTargetWindowsGNU()
12257 ? DAG.getIntPtrConstant(0x2C, dl)
12258 : DAG.getExternalSymbol("_tls_array", PtrVT));
12260 SDValue ThreadPointer =
12261 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12265 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12266 res = ThreadPointer;
12268 // Load the _tls_index variable
12269 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12270 if (Subtarget->is64Bit())
12271 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12272 MachinePointerInfo(), MVT::i32, false, false,
12275 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12278 auto &DL = DAG.getDataLayout();
12280 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12281 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12283 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12286 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12289 // Get the offset of start of .tls section
12290 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12291 GA->getValueType(0),
12292 GA->getOffset(), X86II::MO_SECREL);
12293 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12295 // The address of the thread local variable is the add of the thread
12296 // pointer with the offset of the variable.
12297 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12300 llvm_unreachable("TLS not implemented for this target.");
12303 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12304 /// and take a 2 x i32 value to shift plus a shift amount.
12305 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12306 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12307 MVT VT = Op.getSimpleValueType();
12308 unsigned VTBits = VT.getSizeInBits();
12310 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12311 SDValue ShOpLo = Op.getOperand(0);
12312 SDValue ShOpHi = Op.getOperand(1);
12313 SDValue ShAmt = Op.getOperand(2);
12314 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12315 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12317 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12318 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12319 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12320 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12321 : DAG.getConstant(0, dl, VT);
12323 SDValue Tmp2, Tmp3;
12324 if (Op.getOpcode() == ISD::SHL_PARTS) {
12325 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12326 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12328 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12329 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12332 // If the shift amount is larger or equal than the width of a part we can't
12333 // rely on the results of shld/shrd. Insert a test and select the appropriate
12334 // values for large shift amounts.
12335 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12336 DAG.getConstant(VTBits, dl, MVT::i8));
12337 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12338 AndNode, DAG.getConstant(0, dl, MVT::i8));
12341 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12342 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12343 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12345 if (Op.getOpcode() == ISD::SHL_PARTS) {
12346 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12347 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12349 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12350 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12353 SDValue Ops[2] = { Lo, Hi };
12354 return DAG.getMergeValues(Ops, dl);
12357 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12358 SelectionDAG &DAG) const {
12359 SDValue Src = Op.getOperand(0);
12360 MVT SrcVT = Src.getSimpleValueType();
12361 MVT VT = Op.getSimpleValueType();
12364 if (SrcVT.isVector()) {
12365 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12366 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12367 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12368 DAG.getUNDEF(SrcVT)));
12370 if (SrcVT.getVectorElementType() == MVT::i1) {
12371 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12372 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12373 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12378 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12379 "Unknown SINT_TO_FP to lower!");
12381 // These are really Legal; return the operand so the caller accepts it as
12383 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12385 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12386 Subtarget->is64Bit()) {
12390 unsigned Size = SrcVT.getSizeInBits()/8;
12391 MachineFunction &MF = DAG.getMachineFunction();
12392 auto PtrVT = getPointerTy(MF.getDataLayout());
12393 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12394 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12395 SDValue Chain = DAG.getStore(
12396 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12397 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12399 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12402 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12404 SelectionDAG &DAG) const {
12408 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12410 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12412 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12414 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12416 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12417 MachineMemOperand *MMO;
12419 int SSFI = FI->getIndex();
12420 MMO = DAG.getMachineFunction().getMachineMemOperand(
12421 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12422 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12424 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12425 StackSlot = StackSlot.getOperand(1);
12427 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12428 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12430 Tys, Ops, SrcVT, MMO);
12433 Chain = Result.getValue(1);
12434 SDValue InFlag = Result.getValue(2);
12436 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12437 // shouldn't be necessary except that RFP cannot be live across
12438 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12439 MachineFunction &MF = DAG.getMachineFunction();
12440 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12441 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12442 auto PtrVT = getPointerTy(MF.getDataLayout());
12443 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12444 Tys = DAG.getVTList(MVT::Other);
12446 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12448 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12449 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12450 MachineMemOperand::MOStore, SSFISize, SSFISize);
12452 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12453 Ops, Op.getValueType(), MMO);
12454 Result = DAG.getLoad(
12455 Op.getValueType(), DL, Chain, StackSlot,
12456 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12457 false, false, false, 0);
12463 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12464 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12465 SelectionDAG &DAG) const {
12466 // This algorithm is not obvious. Here it is what we're trying to output:
12469 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12470 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12472 haddpd %xmm0, %xmm0
12474 pshufd $0x4e, %xmm0, %xmm1
12480 LLVMContext *Context = DAG.getContext();
12482 // Build some magic constants.
12483 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12484 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12485 auto PtrVT = getPointerTy(DAG.getDataLayout());
12486 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12488 SmallVector<Constant*,2> CV1;
12490 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12491 APInt(64, 0x4330000000000000ULL))));
12493 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12494 APInt(64, 0x4530000000000000ULL))));
12495 Constant *C1 = ConstantVector::get(CV1);
12496 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12498 // Load the 64-bit value into an XMM register.
12499 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12502 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12503 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12504 false, false, false, 16);
12506 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12509 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12510 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12511 false, false, false, 16);
12512 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12513 // TODO: Are there any fast-math-flags to propagate here?
12514 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12517 if (Subtarget->hasSSE3()) {
12518 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12519 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12521 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12522 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12524 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12525 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12528 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12529 DAG.getIntPtrConstant(0, dl));
12532 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12533 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12534 SelectionDAG &DAG) const {
12536 // FP constant to bias correct the final result.
12537 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12540 // Load the 32-bit value into an XMM register.
12541 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12544 // Zero out the upper parts of the register.
12545 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12547 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12548 DAG.getBitcast(MVT::v2f64, Load),
12549 DAG.getIntPtrConstant(0, dl));
12551 // Or the load with the bias.
12552 SDValue Or = DAG.getNode(
12553 ISD::OR, dl, MVT::v2i64,
12554 DAG.getBitcast(MVT::v2i64,
12555 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12556 DAG.getBitcast(MVT::v2i64,
12557 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12559 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12560 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12562 // Subtract the bias.
12563 // TODO: Are there any fast-math-flags to propagate here?
12564 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12566 // Handle final rounding.
12567 MVT DestVT = Op.getSimpleValueType();
12569 if (DestVT.bitsLT(MVT::f64))
12570 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12571 DAG.getIntPtrConstant(0, dl));
12572 if (DestVT.bitsGT(MVT::f64))
12573 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12575 // Handle final rounding.
12579 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12580 const X86Subtarget &Subtarget) {
12581 // The algorithm is the following:
12582 // #ifdef __SSE4_1__
12583 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12584 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12585 // (uint4) 0x53000000, 0xaa);
12587 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12588 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12590 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12591 // return (float4) lo + fhi;
12593 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12594 // reassociate the two FADDs, and if we do that, the algorithm fails
12595 // spectacularly (PR24512).
12596 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12597 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12598 // there's also the MachineCombiner reassociations happening on Machine IR.
12599 if (DAG.getTarget().Options.UnsafeFPMath)
12603 SDValue V = Op->getOperand(0);
12604 MVT VecIntVT = V.getSimpleValueType();
12605 bool Is128 = VecIntVT == MVT::v4i32;
12606 MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12607 // If we convert to something else than the supported type, e.g., to v4f64,
12609 if (VecFloatVT != Op->getSimpleValueType(0))
12612 unsigned NumElts = VecIntVT.getVectorNumElements();
12613 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12614 "Unsupported custom type");
12615 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12617 // In the #idef/#else code, we have in common:
12618 // - The vector of constants:
12624 // Create the splat vector for 0x4b000000.
12625 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12626 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12627 CstLow, CstLow, CstLow, CstLow};
12628 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12629 makeArrayRef(&CstLowArray[0], NumElts));
12630 // Create the splat vector for 0x53000000.
12631 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12632 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12633 CstHigh, CstHigh, CstHigh, CstHigh};
12634 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12635 makeArrayRef(&CstHighArray[0], NumElts));
12637 // Create the right shift.
12638 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12639 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12640 CstShift, CstShift, CstShift, CstShift};
12641 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12642 makeArrayRef(&CstShiftArray[0], NumElts));
12643 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12646 if (Subtarget.hasSSE41()) {
12647 MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12648 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12649 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12650 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12651 // Low will be bitcasted right away, so do not bother bitcasting back to its
12653 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12654 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12655 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12656 // (uint4) 0x53000000, 0xaa);
12657 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12658 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12659 // High will be bitcasted right away, so do not bother bitcasting back to
12660 // its original type.
12661 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12662 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12664 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12665 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12666 CstMask, CstMask, CstMask);
12667 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12668 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12669 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12671 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12672 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12675 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12676 SDValue CstFAdd = DAG.getConstantFP(
12677 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12678 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12679 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12680 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12681 makeArrayRef(&CstFAddArray[0], NumElts));
12683 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12684 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12685 // TODO: Are there any fast-math-flags to propagate here?
12687 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12688 // return (float4) lo + fhi;
12689 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12690 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12693 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12694 SelectionDAG &DAG) const {
12695 SDValue N0 = Op.getOperand(0);
12696 MVT SVT = N0.getSimpleValueType();
12699 switch (SVT.SimpleTy) {
12701 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12706 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12707 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12708 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12712 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12715 assert(Subtarget->hasAVX512());
12716 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12717 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12721 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12722 SelectionDAG &DAG) const {
12723 SDValue N0 = Op.getOperand(0);
12725 auto PtrVT = getPointerTy(DAG.getDataLayout());
12727 if (Op.getSimpleValueType().isVector())
12728 return lowerUINT_TO_FP_vec(Op, DAG);
12730 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12731 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12732 // the optimization here.
12733 if (DAG.SignBitIsZero(N0))
12734 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12736 MVT SrcVT = N0.getSimpleValueType();
12737 MVT DstVT = Op.getSimpleValueType();
12739 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12740 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12741 // Conversions from unsigned i32 to f32/f64 are legal,
12742 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12746 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12747 return LowerUINT_TO_FP_i64(Op, DAG);
12748 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12749 return LowerUINT_TO_FP_i32(Op, DAG);
12750 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12753 // Make a 64-bit buffer, and use it to build an FILD.
12754 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12755 if (SrcVT == MVT::i32) {
12756 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12757 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12758 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12759 StackSlot, MachinePointerInfo(),
12761 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12762 OffsetSlot, MachinePointerInfo(),
12764 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12768 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12769 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12770 StackSlot, MachinePointerInfo(),
12772 // For i64 source, we need to add the appropriate power of 2 if the input
12773 // was negative. This is the same as the optimization in
12774 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12775 // we must be careful to do the computation in x87 extended precision, not
12776 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12777 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12778 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12779 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12780 MachineMemOperand::MOLoad, 8, 8);
12782 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12783 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12784 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12787 APInt FF(32, 0x5F800000ULL);
12789 // Check whether the sign bit is set.
12790 SDValue SignSet = DAG.getSetCC(
12791 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12792 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12794 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12795 SDValue FudgePtr = DAG.getConstantPool(
12796 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12798 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12799 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12800 SDValue Four = DAG.getIntPtrConstant(4, dl);
12801 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12803 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12805 // Load the value out, extending it from f32 to f80.
12806 // FIXME: Avoid the extend by constructing the right constant pool?
12807 SDValue Fudge = DAG.getExtLoad(
12808 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12809 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12810 false, false, false, 4);
12811 // Extend everything to 80 bits to force it to be done on x87.
12812 // TODO: Are there any fast-math-flags to propagate here?
12813 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12814 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12815 DAG.getIntPtrConstant(0, dl));
12818 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12819 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12820 // just return an <SDValue(), SDValue()> pair.
12821 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12822 // to i16, i32 or i64, and we lower it to a legal sequence.
12823 // If lowered to the final integer result we return a <result, SDValue()> pair.
12824 // Otherwise we lower it to a sequence ending with a FIST, return a
12825 // <FIST, StackSlot> pair, and the caller is responsible for loading
12826 // the final integer result from StackSlot.
12827 std::pair<SDValue,SDValue>
12828 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12829 bool IsSigned, bool IsReplace) const {
12832 EVT DstTy = Op.getValueType();
12833 EVT TheVT = Op.getOperand(0).getValueType();
12834 auto PtrVT = getPointerTy(DAG.getDataLayout());
12836 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12837 // f16 must be promoted before using the lowering in this routine.
12838 // fp128 does not use this lowering.
12839 return std::make_pair(SDValue(), SDValue());
12842 // If using FIST to compute an unsigned i64, we'll need some fixup
12843 // to handle values above the maximum signed i64. A FIST is always
12844 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12845 bool UnsignedFixup = !IsSigned &&
12846 DstTy == MVT::i64 &&
12847 (!Subtarget->is64Bit() ||
12848 !isScalarFPTypeInSSEReg(TheVT));
12850 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12851 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12852 // The low 32 bits of the fist result will have the correct uint32 result.
12853 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12857 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12858 DstTy.getSimpleVT() >= MVT::i16 &&
12859 "Unknown FP_TO_INT to lower!");
12861 // These are really Legal.
12862 if (DstTy == MVT::i32 &&
12863 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12864 return std::make_pair(SDValue(), SDValue());
12865 if (Subtarget->is64Bit() &&
12866 DstTy == MVT::i64 &&
12867 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12868 return std::make_pair(SDValue(), SDValue());
12870 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12872 MachineFunction &MF = DAG.getMachineFunction();
12873 unsigned MemSize = DstTy.getSizeInBits()/8;
12874 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12875 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12878 switch (DstTy.getSimpleVT().SimpleTy) {
12879 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12880 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12881 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12882 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12885 SDValue Chain = DAG.getEntryNode();
12886 SDValue Value = Op.getOperand(0);
12887 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12889 if (UnsignedFixup) {
12891 // Conversion to unsigned i64 is implemented with a select,
12892 // depending on whether the source value fits in the range
12893 // of a signed i64. Let Thresh be the FP equivalent of
12894 // 0x8000000000000000ULL.
12896 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12897 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12898 // Fist-to-mem64 FistSrc
12899 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12900 // to XOR'ing the high 32 bits with Adjust.
12902 // Being a power of 2, Thresh is exactly representable in all FP formats.
12903 // For X87 we'd like to use the smallest FP type for this constant, but
12904 // for DAG type consistency we have to match the FP operand type.
12906 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12907 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12908 bool LosesInfo = false;
12909 if (TheVT == MVT::f64)
12910 // The rounding mode is irrelevant as the conversion should be exact.
12911 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12913 else if (TheVT == MVT::f80)
12914 Status = Thresh.convert(APFloat::x87DoubleExtended,
12915 APFloat::rmNearestTiesToEven, &LosesInfo);
12917 assert(Status == APFloat::opOK && !LosesInfo &&
12918 "FP conversion should have been exact");
12920 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12922 SDValue Cmp = DAG.getSetCC(DL,
12923 getSetCCResultType(DAG.getDataLayout(),
12924 *DAG.getContext(), TheVT),
12925 Value, ThreshVal, ISD::SETLT);
12926 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12927 DAG.getConstant(0, DL, MVT::i32),
12928 DAG.getConstant(0x80000000, DL, MVT::i32));
12929 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12930 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12931 *DAG.getContext(), TheVT),
12932 Value, ThreshVal, ISD::SETLT);
12933 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12936 // FIXME This causes a redundant load/store if the SSE-class value is already
12937 // in memory, such as if it is on the callstack.
12938 if (isScalarFPTypeInSSEReg(TheVT)) {
12939 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12940 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12941 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12943 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12945 Chain, StackSlot, DAG.getValueType(TheVT)
12948 MachineMemOperand *MMO =
12949 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12950 MachineMemOperand::MOLoad, MemSize, MemSize);
12951 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12952 Chain = Value.getValue(1);
12953 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12954 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12957 MachineMemOperand *MMO =
12958 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12959 MachineMemOperand::MOStore, MemSize, MemSize);
12961 if (UnsignedFixup) {
12963 // Insert the FIST, load its result as two i32's,
12964 // and XOR the high i32 with Adjust.
12966 SDValue FistOps[] = { Chain, Value, StackSlot };
12967 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12968 FistOps, DstTy, MMO);
12970 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12971 MachinePointerInfo(),
12972 false, false, false, 0);
12973 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12974 DAG.getConstant(4, DL, PtrVT));
12976 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12977 MachinePointerInfo(),
12978 false, false, false, 0);
12979 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12981 if (Subtarget->is64Bit()) {
12982 // Join High32 and Low32 into a 64-bit result.
12983 // (High32 << 32) | Low32
12984 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12985 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12986 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12987 DAG.getConstant(32, DL, MVT::i8));
12988 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12989 return std::make_pair(Result, SDValue());
12992 SDValue ResultOps[] = { Low32, High32 };
12994 SDValue pair = IsReplace
12995 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12996 : DAG.getMergeValues(ResultOps, DL);
12997 return std::make_pair(pair, SDValue());
12999 // Build the FP_TO_INT*_IN_MEM
13000 SDValue Ops[] = { Chain, Value, StackSlot };
13001 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13003 return std::make_pair(FIST, StackSlot);
13007 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13008 const X86Subtarget *Subtarget) {
13009 MVT VT = Op->getSimpleValueType(0);
13010 SDValue In = Op->getOperand(0);
13011 MVT InVT = In.getSimpleValueType();
13014 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13015 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13017 // Optimize vectors in AVX mode:
13020 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13021 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13022 // Concat upper and lower parts.
13025 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13026 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13027 // Concat upper and lower parts.
13030 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13031 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13032 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13035 if (Subtarget->hasInt256())
13036 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13038 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13039 SDValue Undef = DAG.getUNDEF(InVT);
13040 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13041 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13042 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13044 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13045 VT.getVectorNumElements()/2);
13047 OpLo = DAG.getBitcast(HVT, OpLo);
13048 OpHi = DAG.getBitcast(HVT, OpHi);
13050 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13053 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13054 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13055 MVT VT = Op->getSimpleValueType(0);
13056 SDValue In = Op->getOperand(0);
13057 MVT InVT = In.getSimpleValueType();
13059 unsigned int NumElts = VT.getVectorNumElements();
13060 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13063 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13064 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13066 assert(InVT.getVectorElementType() == MVT::i1);
13067 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13069 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13071 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13073 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13074 if (VT.is512BitVector())
13076 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13079 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13080 SelectionDAG &DAG) {
13081 if (Subtarget->hasFp256())
13082 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13088 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13089 SelectionDAG &DAG) {
13091 MVT VT = Op.getSimpleValueType();
13092 SDValue In = Op.getOperand(0);
13093 MVT SVT = In.getSimpleValueType();
13095 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13096 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13098 if (Subtarget->hasFp256())
13099 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13102 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13103 VT.getVectorNumElements() != SVT.getVectorNumElements());
13107 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13109 MVT VT = Op.getSimpleValueType();
13110 SDValue In = Op.getOperand(0);
13111 MVT InVT = In.getSimpleValueType();
13113 if (VT == MVT::i1) {
13114 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13115 "Invalid scalar TRUNCATE operation");
13116 if (InVT.getSizeInBits() >= 32)
13118 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13119 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13121 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13122 "Invalid TRUNCATE operation");
13124 // move vector to mask - truncate solution for SKX
13125 if (VT.getVectorElementType() == MVT::i1) {
13126 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13127 Subtarget->hasBWI())
13128 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13129 if ((InVT.is256BitVector() || InVT.is128BitVector())
13130 && InVT.getScalarSizeInBits() <= 16 &&
13131 Subtarget->hasBWI() && Subtarget->hasVLX())
13132 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13133 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13134 Subtarget->hasDQI())
13135 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13136 if ((InVT.is256BitVector() || InVT.is128BitVector())
13137 && InVT.getScalarSizeInBits() >= 32 &&
13138 Subtarget->hasDQI() && Subtarget->hasVLX())
13139 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13142 if (VT.getVectorElementType() == MVT::i1) {
13143 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13144 unsigned NumElts = InVT.getVectorNumElements();
13145 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13146 if (InVT.getSizeInBits() < 512) {
13147 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13148 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13153 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13154 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13155 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13158 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13159 if (Subtarget->hasAVX512()) {
13160 // word to byte only under BWI
13161 if (InVT == MVT::v16i16 && !Subtarget->hasBWI()) // v16i16 -> v16i8
13162 return DAG.getNode(X86ISD::VTRUNC, DL, VT,
13163 DAG.getNode(X86ISD::VSEXT, DL, MVT::v16i32, In));
13164 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13166 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13167 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13168 if (Subtarget->hasInt256()) {
13169 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13170 In = DAG.getBitcast(MVT::v8i32, In);
13171 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13173 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13174 DAG.getIntPtrConstant(0, DL));
13177 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13178 DAG.getIntPtrConstant(0, DL));
13179 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13180 DAG.getIntPtrConstant(2, DL));
13181 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13182 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13183 static const int ShufMask[] = {0, 2, 4, 6};
13184 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13187 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13188 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13189 if (Subtarget->hasInt256()) {
13190 In = DAG.getBitcast(MVT::v32i8, In);
13192 SmallVector<SDValue,32> pshufbMask;
13193 for (unsigned i = 0; i < 2; ++i) {
13194 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13195 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13196 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13197 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13198 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13199 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13200 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13201 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13202 for (unsigned j = 0; j < 8; ++j)
13203 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13205 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13206 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13207 In = DAG.getBitcast(MVT::v4i64, In);
13209 static const int ShufMask[] = {0, 2, -1, -1};
13210 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13212 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13213 DAG.getIntPtrConstant(0, DL));
13214 return DAG.getBitcast(VT, In);
13217 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13218 DAG.getIntPtrConstant(0, DL));
13220 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13221 DAG.getIntPtrConstant(4, DL));
13223 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13224 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13226 // The PSHUFB mask:
13227 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13228 -1, -1, -1, -1, -1, -1, -1, -1};
13230 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13231 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13232 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13234 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13235 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13237 // The MOVLHPS Mask:
13238 static const int ShufMask2[] = {0, 1, 4, 5};
13239 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13240 return DAG.getBitcast(MVT::v8i16, res);
13243 // Handle truncation of V256 to V128 using shuffles.
13244 if (!VT.is128BitVector() || !InVT.is256BitVector())
13247 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13249 unsigned NumElems = VT.getVectorNumElements();
13250 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13252 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13253 // Prepare truncation shuffle mask
13254 for (unsigned i = 0; i != NumElems; ++i)
13255 MaskVec[i] = i * 2;
13256 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13257 DAG.getUNDEF(NVT), &MaskVec[0]);
13258 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13259 DAG.getIntPtrConstant(0, DL));
13262 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13263 SelectionDAG &DAG) const {
13264 assert(!Op.getSimpleValueType().isVector());
13266 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13267 /*IsSigned=*/ true, /*IsReplace=*/ false);
13268 SDValue FIST = Vals.first, StackSlot = Vals.second;
13269 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13270 if (!FIST.getNode())
13273 if (StackSlot.getNode())
13274 // Load the result.
13275 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13276 FIST, StackSlot, MachinePointerInfo(),
13277 false, false, false, 0);
13279 // The node is the result.
13283 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13284 SelectionDAG &DAG) const {
13285 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13286 /*IsSigned=*/ false, /*IsReplace=*/ false);
13287 SDValue FIST = Vals.first, StackSlot = Vals.second;
13288 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13289 if (!FIST.getNode())
13292 if (StackSlot.getNode())
13293 // Load the result.
13294 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13295 FIST, StackSlot, MachinePointerInfo(),
13296 false, false, false, 0);
13298 // The node is the result.
13302 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13304 MVT VT = Op.getSimpleValueType();
13305 SDValue In = Op.getOperand(0);
13306 MVT SVT = In.getSimpleValueType();
13308 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13310 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13311 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13312 In, DAG.getUNDEF(SVT)));
13315 /// The only differences between FABS and FNEG are the mask and the logic op.
13316 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13317 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13318 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13319 "Wrong opcode for lowering FABS or FNEG.");
13321 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13323 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13324 // into an FNABS. We'll lower the FABS after that if it is still in use.
13326 for (SDNode *User : Op->uses())
13327 if (User->getOpcode() == ISD::FNEG)
13331 MVT VT = Op.getSimpleValueType();
13333 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13334 // decide if we should generate a 16-byte constant mask when we only need 4 or
13335 // 8 bytes for the scalar case.
13341 if (VT.isVector()) {
13343 EltVT = VT.getVectorElementType();
13344 NumElts = VT.getVectorNumElements();
13346 // There are no scalar bitwise logical SSE/AVX instructions, so we
13347 // generate a 16-byte vector constant and logic op even for the scalar case.
13348 // Using a 16-byte mask allows folding the load of the mask with
13349 // the logic op, so it can save (~4 bytes) on code size.
13350 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13352 NumElts = (VT == MVT::f64) ? 2 : 4;
13355 unsigned EltBits = EltVT.getSizeInBits();
13356 LLVMContext *Context = DAG.getContext();
13357 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13359 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13360 Constant *C = ConstantInt::get(*Context, MaskElt);
13361 C = ConstantVector::getSplat(NumElts, C);
13362 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13363 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13364 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13366 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13367 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13368 false, false, false, Alignment);
13370 SDValue Op0 = Op.getOperand(0);
13371 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13373 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13374 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13377 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13379 // For the scalar case extend to a 128-bit vector, perform the logic op,
13380 // and extract the scalar result back out.
13381 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13382 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13383 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13384 DAG.getIntPtrConstant(0, dl));
13387 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13388 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13389 LLVMContext *Context = DAG.getContext();
13390 SDValue Op0 = Op.getOperand(0);
13391 SDValue Op1 = Op.getOperand(1);
13393 MVT VT = Op.getSimpleValueType();
13394 MVT SrcVT = Op1.getSimpleValueType();
13396 // If second operand is smaller, extend it first.
13397 if (SrcVT.bitsLT(VT)) {
13398 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13401 // And if it is bigger, shrink it first.
13402 if (SrcVT.bitsGT(VT)) {
13403 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13407 // At this point the operands and the result should have the same
13408 // type, and that won't be f80 since that is not custom lowered.
13410 const fltSemantics &Sem =
13411 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13412 const unsigned SizeInBits = VT.getSizeInBits();
13414 SmallVector<Constant *, 4> CV(
13415 VT == MVT::f64 ? 2 : 4,
13416 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13418 // First, clear all bits but the sign bit from the second operand (sign).
13419 CV[0] = ConstantFP::get(*Context,
13420 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13421 Constant *C = ConstantVector::get(CV);
13422 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13423 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13425 // Perform all logic operations as 16-byte vectors because there are no
13426 // scalar FP logic instructions in SSE. This allows load folding of the
13427 // constants into the logic instructions.
13428 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13430 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13431 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13432 false, false, false, 16);
13433 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13434 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13436 // Next, clear the sign bit from the first operand (magnitude).
13437 // If it's a constant, we can clear it here.
13438 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13439 APFloat APF = Op0CN->getValueAPF();
13440 // If the magnitude is a positive zero, the sign bit alone is enough.
13441 if (APF.isPosZero())
13442 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13443 DAG.getIntPtrConstant(0, dl));
13445 CV[0] = ConstantFP::get(*Context, APF);
13447 CV[0] = ConstantFP::get(
13449 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13451 C = ConstantVector::get(CV);
13452 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13454 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13455 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13456 false, false, false, 16);
13457 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13458 if (!isa<ConstantFPSDNode>(Op0)) {
13459 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13460 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13462 // OR the magnitude value with the sign bit.
13463 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13464 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13465 DAG.getIntPtrConstant(0, dl));
13468 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13469 SDValue N0 = Op.getOperand(0);
13471 MVT VT = Op.getSimpleValueType();
13473 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13474 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13475 DAG.getConstant(1, dl, VT));
13476 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13479 // Check whether an OR'd tree is PTEST-able.
13480 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13481 SelectionDAG &DAG) {
13482 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13484 if (!Subtarget->hasSSE41())
13487 if (!Op->hasOneUse())
13490 SDNode *N = Op.getNode();
13493 SmallVector<SDValue, 8> Opnds;
13494 DenseMap<SDValue, unsigned> VecInMap;
13495 SmallVector<SDValue, 8> VecIns;
13496 EVT VT = MVT::Other;
13498 // Recognize a special case where a vector is casted into wide integer to
13500 Opnds.push_back(N->getOperand(0));
13501 Opnds.push_back(N->getOperand(1));
13503 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13504 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13505 // BFS traverse all OR'd operands.
13506 if (I->getOpcode() == ISD::OR) {
13507 Opnds.push_back(I->getOperand(0));
13508 Opnds.push_back(I->getOperand(1));
13509 // Re-evaluate the number of nodes to be traversed.
13510 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13514 // Quit if a non-EXTRACT_VECTOR_ELT
13515 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13518 // Quit if without a constant index.
13519 SDValue Idx = I->getOperand(1);
13520 if (!isa<ConstantSDNode>(Idx))
13523 SDValue ExtractedFromVec = I->getOperand(0);
13524 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13525 if (M == VecInMap.end()) {
13526 VT = ExtractedFromVec.getValueType();
13527 // Quit if not 128/256-bit vector.
13528 if (!VT.is128BitVector() && !VT.is256BitVector())
13530 // Quit if not the same type.
13531 if (VecInMap.begin() != VecInMap.end() &&
13532 VT != VecInMap.begin()->first.getValueType())
13534 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13535 VecIns.push_back(ExtractedFromVec);
13537 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13540 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13541 "Not extracted from 128-/256-bit vector.");
13543 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13545 for (DenseMap<SDValue, unsigned>::const_iterator
13546 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13547 // Quit if not all elements are used.
13548 if (I->second != FullMask)
13552 MVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13554 // Cast all vectors into TestVT for PTEST.
13555 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13556 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13558 // If more than one full vectors are evaluated, OR them first before PTEST.
13559 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13560 // Each iteration will OR 2 nodes and append the result until there is only
13561 // 1 node left, i.e. the final OR'd value of all vectors.
13562 SDValue LHS = VecIns[Slot];
13563 SDValue RHS = VecIns[Slot + 1];
13564 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13567 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13568 VecIns.back(), VecIns.back());
13571 /// \brief return true if \c Op has a use that doesn't just read flags.
13572 static bool hasNonFlagsUse(SDValue Op) {
13573 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13575 SDNode *User = *UI;
13576 unsigned UOpNo = UI.getOperandNo();
13577 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13578 // Look pass truncate.
13579 UOpNo = User->use_begin().getOperandNo();
13580 User = *User->use_begin();
13583 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13584 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13590 /// Emit nodes that will be selected as "test Op0,Op0", or something
13592 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13593 SelectionDAG &DAG) const {
13594 if (Op.getValueType() == MVT::i1) {
13595 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13596 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13597 DAG.getConstant(0, dl, MVT::i8));
13599 // CF and OF aren't always set the way we want. Determine which
13600 // of these we need.
13601 bool NeedCF = false;
13602 bool NeedOF = false;
13605 case X86::COND_A: case X86::COND_AE:
13606 case X86::COND_B: case X86::COND_BE:
13609 case X86::COND_G: case X86::COND_GE:
13610 case X86::COND_L: case X86::COND_LE:
13611 case X86::COND_O: case X86::COND_NO: {
13612 // Check if we really need to set the
13613 // Overflow flag. If NoSignedWrap is present
13614 // that is not actually needed.
13615 switch (Op->getOpcode()) {
13620 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13621 if (BinNode->Flags.hasNoSignedWrap())
13631 // See if we can use the EFLAGS value from the operand instead of
13632 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13633 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13634 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13635 // Emit a CMP with 0, which is the TEST pattern.
13636 //if (Op.getValueType() == MVT::i1)
13637 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13638 // DAG.getConstant(0, MVT::i1));
13639 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13640 DAG.getConstant(0, dl, Op.getValueType()));
13642 unsigned Opcode = 0;
13643 unsigned NumOperands = 0;
13645 // Truncate operations may prevent the merge of the SETCC instruction
13646 // and the arithmetic instruction before it. Attempt to truncate the operands
13647 // of the arithmetic instruction and use a reduced bit-width instruction.
13648 bool NeedTruncation = false;
13649 SDValue ArithOp = Op;
13650 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13651 SDValue Arith = Op->getOperand(0);
13652 // Both the trunc and the arithmetic op need to have one user each.
13653 if (Arith->hasOneUse())
13654 switch (Arith.getOpcode()) {
13661 NeedTruncation = true;
13667 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13668 // which may be the result of a CAST. We use the variable 'Op', which is the
13669 // non-casted variable when we check for possible users.
13670 switch (ArithOp.getOpcode()) {
13672 // Due to an isel shortcoming, be conservative if this add is likely to be
13673 // selected as part of a load-modify-store instruction. When the root node
13674 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13675 // uses of other nodes in the match, such as the ADD in this case. This
13676 // leads to the ADD being left around and reselected, with the result being
13677 // two adds in the output. Alas, even if none our users are stores, that
13678 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13679 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13680 // climbing the DAG back to the root, and it doesn't seem to be worth the
13682 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13683 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13684 if (UI->getOpcode() != ISD::CopyToReg &&
13685 UI->getOpcode() != ISD::SETCC &&
13686 UI->getOpcode() != ISD::STORE)
13689 if (ConstantSDNode *C =
13690 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13691 // An add of one will be selected as an INC.
13692 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13693 Opcode = X86ISD::INC;
13698 // An add of negative one (subtract of one) will be selected as a DEC.
13699 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13700 Opcode = X86ISD::DEC;
13706 // Otherwise use a regular EFLAGS-setting add.
13707 Opcode = X86ISD::ADD;
13712 // If we have a constant logical shift that's only used in a comparison
13713 // against zero turn it into an equivalent AND. This allows turning it into
13714 // a TEST instruction later.
13715 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13716 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13717 EVT VT = Op.getValueType();
13718 unsigned BitWidth = VT.getSizeInBits();
13719 unsigned ShAmt = Op->getConstantOperandVal(1);
13720 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13722 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13723 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13724 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13725 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13727 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13728 DAG.getConstant(Mask, dl, VT));
13729 DAG.ReplaceAllUsesWith(Op, New);
13735 // If the primary and result isn't used, don't bother using X86ISD::AND,
13736 // because a TEST instruction will be better.
13737 if (!hasNonFlagsUse(Op))
13743 // Due to the ISEL shortcoming noted above, be conservative if this op is
13744 // likely to be selected as part of a load-modify-store instruction.
13745 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13746 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13747 if (UI->getOpcode() == ISD::STORE)
13750 // Otherwise use a regular EFLAGS-setting instruction.
13751 switch (ArithOp.getOpcode()) {
13752 default: llvm_unreachable("unexpected operator!");
13753 case ISD::SUB: Opcode = X86ISD::SUB; break;
13754 case ISD::XOR: Opcode = X86ISD::XOR; break;
13755 case ISD::AND: Opcode = X86ISD::AND; break;
13757 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13758 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13759 if (EFLAGS.getNode())
13762 Opcode = X86ISD::OR;
13776 return SDValue(Op.getNode(), 1);
13782 // If we found that truncation is beneficial, perform the truncation and
13784 if (NeedTruncation) {
13785 EVT VT = Op.getValueType();
13786 SDValue WideVal = Op->getOperand(0);
13787 EVT WideVT = WideVal.getValueType();
13788 unsigned ConvertedOp = 0;
13789 // Use a target machine opcode to prevent further DAGCombine
13790 // optimizations that may separate the arithmetic operations
13791 // from the setcc node.
13792 switch (WideVal.getOpcode()) {
13794 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13795 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13796 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13797 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13798 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13802 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13803 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13804 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13805 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13806 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13812 // Emit a CMP with 0, which is the TEST pattern.
13813 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13814 DAG.getConstant(0, dl, Op.getValueType()));
13816 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13817 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13819 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13820 DAG.ReplaceAllUsesWith(Op, New);
13821 return SDValue(New.getNode(), 1);
13824 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13826 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13827 SDLoc dl, SelectionDAG &DAG) const {
13828 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13829 if (C->getAPIntValue() == 0)
13830 return EmitTest(Op0, X86CC, dl, DAG);
13832 assert(Op0.getValueType() != MVT::i1 &&
13833 "Unexpected comparison operation for MVT::i1 operands");
13836 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13837 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13838 // Do the comparison at i32 if it's smaller, besides the Atom case.
13839 // This avoids subregister aliasing issues. Keep the smaller reference
13840 // if we're optimizing for size, however, as that'll allow better folding
13841 // of memory operations.
13842 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13843 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13844 !Subtarget->isAtom()) {
13845 unsigned ExtendOp =
13846 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13847 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13848 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13850 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13851 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13852 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13854 return SDValue(Sub.getNode(), 1);
13856 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13859 /// Convert a comparison if required by the subtarget.
13860 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13861 SelectionDAG &DAG) const {
13862 // If the subtarget does not support the FUCOMI instruction, floating-point
13863 // comparisons have to be converted.
13864 if (Subtarget->hasCMov() ||
13865 Cmp.getOpcode() != X86ISD::CMP ||
13866 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13867 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13870 // The instruction selector will select an FUCOM instruction instead of
13871 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13872 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13873 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13875 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13876 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13877 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13878 DAG.getConstant(8, dl, MVT::i8));
13879 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13880 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13883 /// The minimum architected relative accuracy is 2^-12. We need one
13884 /// Newton-Raphson step to have a good float result (24 bits of precision).
13885 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13886 DAGCombinerInfo &DCI,
13887 unsigned &RefinementSteps,
13888 bool &UseOneConstNR) const {
13889 EVT VT = Op.getValueType();
13890 const char *RecipOp;
13892 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13893 // TODO: Add support for AVX512 (v16f32).
13894 // It is likely not profitable to do this for f64 because a double-precision
13895 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13896 // instructions: convert to single, rsqrtss, convert back to double, refine
13897 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13898 // along with FMA, this could be a throughput win.
13899 if (VT == MVT::f32 && Subtarget->hasSSE1())
13901 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13902 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13903 RecipOp = "vec-sqrtf";
13907 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13908 if (!Recips.isEnabled(RecipOp))
13911 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13912 UseOneConstNR = false;
13913 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13916 /// The minimum architected relative accuracy is 2^-12. We need one
13917 /// Newton-Raphson step to have a good float result (24 bits of precision).
13918 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13919 DAGCombinerInfo &DCI,
13920 unsigned &RefinementSteps) const {
13921 EVT VT = Op.getValueType();
13922 const char *RecipOp;
13924 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13925 // TODO: Add support for AVX512 (v16f32).
13926 // It is likely not profitable to do this for f64 because a double-precision
13927 // reciprocal estimate with refinement on x86 prior to FMA requires
13928 // 15 instructions: convert to single, rcpss, convert back to double, refine
13929 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13930 // along with FMA, this could be a throughput win.
13931 if (VT == MVT::f32 && Subtarget->hasSSE1())
13933 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13934 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13935 RecipOp = "vec-divf";
13939 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13940 if (!Recips.isEnabled(RecipOp))
13943 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13944 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13947 /// If we have at least two divisions that use the same divisor, convert to
13948 /// multplication by a reciprocal. This may need to be adjusted for a given
13949 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13950 /// This is because we still need one division to calculate the reciprocal and
13951 /// then we need two multiplies by that reciprocal as replacements for the
13952 /// original divisions.
13953 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13957 static bool isAllOnes(SDValue V) {
13958 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13959 return C && C->isAllOnesValue();
13962 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13963 /// if it's possible.
13964 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13965 SDLoc dl, SelectionDAG &DAG) const {
13966 SDValue Op0 = And.getOperand(0);
13967 SDValue Op1 = And.getOperand(1);
13968 if (Op0.getOpcode() == ISD::TRUNCATE)
13969 Op0 = Op0.getOperand(0);
13970 if (Op1.getOpcode() == ISD::TRUNCATE)
13971 Op1 = Op1.getOperand(0);
13974 if (Op1.getOpcode() == ISD::SHL)
13975 std::swap(Op0, Op1);
13976 if (Op0.getOpcode() == ISD::SHL) {
13977 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13978 if (And00C->getZExtValue() == 1) {
13979 // If we looked past a truncate, check that it's only truncating away
13981 unsigned BitWidth = Op0.getValueSizeInBits();
13982 unsigned AndBitWidth = And.getValueSizeInBits();
13983 if (BitWidth > AndBitWidth) {
13985 DAG.computeKnownBits(Op0, Zeros, Ones);
13986 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13990 RHS = Op0.getOperand(1);
13992 } else if (Op1.getOpcode() == ISD::Constant) {
13993 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13994 uint64_t AndRHSVal = AndRHS->getZExtValue();
13995 SDValue AndLHS = Op0;
13997 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13998 LHS = AndLHS.getOperand(0);
13999 RHS = AndLHS.getOperand(1);
14002 // Use BT if the immediate can't be encoded in a TEST instruction.
14003 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14005 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14009 if (LHS.getNode()) {
14010 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14011 // instruction. Since the shift amount is in-range-or-undefined, we know
14012 // that doing a bittest on the i32 value is ok. We extend to i32 because
14013 // the encoding for the i16 version is larger than the i32 version.
14014 // Also promote i16 to i32 for performance / code size reason.
14015 if (LHS.getValueType() == MVT::i8 ||
14016 LHS.getValueType() == MVT::i16)
14017 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14019 // If the operand types disagree, extend the shift amount to match. Since
14020 // BT ignores high bits (like shifts) we can use anyextend.
14021 if (LHS.getValueType() != RHS.getValueType())
14022 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14024 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14025 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14026 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14027 DAG.getConstant(Cond, dl, MVT::i8), BT);
14033 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14035 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14040 // SSE Condition code mapping:
14049 switch (SetCCOpcode) {
14050 default: llvm_unreachable("Unexpected SETCC condition");
14052 case ISD::SETEQ: SSECC = 0; break;
14054 case ISD::SETGT: Swap = true; // Fallthrough
14056 case ISD::SETOLT: SSECC = 1; break;
14058 case ISD::SETGE: Swap = true; // Fallthrough
14060 case ISD::SETOLE: SSECC = 2; break;
14061 case ISD::SETUO: SSECC = 3; break;
14063 case ISD::SETNE: SSECC = 4; break;
14064 case ISD::SETULE: Swap = true; // Fallthrough
14065 case ISD::SETUGE: SSECC = 5; break;
14066 case ISD::SETULT: Swap = true; // Fallthrough
14067 case ISD::SETUGT: SSECC = 6; break;
14068 case ISD::SETO: SSECC = 7; break;
14070 case ISD::SETONE: SSECC = 8; break;
14073 std::swap(Op0, Op1);
14078 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14079 // ones, and then concatenate the result back.
14080 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14081 MVT VT = Op.getSimpleValueType();
14083 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14084 "Unsupported value type for operation");
14086 unsigned NumElems = VT.getVectorNumElements();
14088 SDValue CC = Op.getOperand(2);
14090 // Extract the LHS vectors
14091 SDValue LHS = Op.getOperand(0);
14092 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14093 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14095 // Extract the RHS vectors
14096 SDValue RHS = Op.getOperand(1);
14097 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14098 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14100 // Issue the operation on the smaller types and concatenate the result back
14101 MVT EltVT = VT.getVectorElementType();
14102 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14103 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14104 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14105 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14108 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14109 SDValue Op0 = Op.getOperand(0);
14110 SDValue Op1 = Op.getOperand(1);
14111 SDValue CC = Op.getOperand(2);
14112 MVT VT = Op.getSimpleValueType();
14115 assert(Op0.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14116 "Unexpected type for boolean compare operation");
14117 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14118 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14119 DAG.getConstant(-1, dl, VT));
14120 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14121 DAG.getConstant(-1, dl, VT));
14122 switch (SetCCOpcode) {
14123 default: llvm_unreachable("Unexpected SETCC condition");
14125 // (x == y) -> ~(x ^ y)
14126 return DAG.getNode(ISD::XOR, dl, VT,
14127 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14128 DAG.getConstant(-1, dl, VT));
14130 // (x != y) -> (x ^ y)
14131 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14134 // (x > y) -> (x & ~y)
14135 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14138 // (x < y) -> (~x & y)
14139 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14142 // (x <= y) -> (~x | y)
14143 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14146 // (x >=y) -> (x | ~y)
14147 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14151 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14152 const X86Subtarget *Subtarget) {
14153 SDValue Op0 = Op.getOperand(0);
14154 SDValue Op1 = Op.getOperand(1);
14155 SDValue CC = Op.getOperand(2);
14156 MVT VT = Op.getSimpleValueType();
14159 assert(Op0.getSimpleValueType().getVectorElementType().getSizeInBits() >= 8 &&
14160 Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
14161 "Cannot set masked compare for this operation");
14163 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14165 bool Unsigned = false;
14168 switch (SetCCOpcode) {
14169 default: llvm_unreachable("Unexpected SETCC condition");
14170 case ISD::SETNE: SSECC = 4; break;
14171 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14172 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14173 case ISD::SETLT: Swap = true; //fall-through
14174 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14175 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14176 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14177 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14178 case ISD::SETULE: Unsigned = true; //fall-through
14179 case ISD::SETLE: SSECC = 2; break;
14183 std::swap(Op0, Op1);
14185 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14186 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14187 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14188 DAG.getConstant(SSECC, dl, MVT::i8));
14191 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14192 /// operand \p Op1. If non-trivial (for example because it's not constant)
14193 /// return an empty value.
14194 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14196 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14200 MVT VT = Op1.getSimpleValueType();
14201 MVT EVT = VT.getVectorElementType();
14202 unsigned n = VT.getVectorNumElements();
14203 SmallVector<SDValue, 8> ULTOp1;
14205 for (unsigned i = 0; i < n; ++i) {
14206 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14207 if (!Elt || Elt->isOpaque() || Elt->getSimpleValueType(0) != EVT)
14210 // Avoid underflow.
14211 APInt Val = Elt->getAPIntValue();
14215 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14218 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14221 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14222 SelectionDAG &DAG) {
14223 SDValue Op0 = Op.getOperand(0);
14224 SDValue Op1 = Op.getOperand(1);
14225 SDValue CC = Op.getOperand(2);
14226 MVT VT = Op.getSimpleValueType();
14227 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14228 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14233 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14234 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14237 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14238 unsigned Opc = X86ISD::CMPP;
14239 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14240 assert(VT.getVectorNumElements() <= 16);
14241 Opc = X86ISD::CMPM;
14243 // In the two special cases we can't handle, emit two comparisons.
14246 unsigned CombineOpc;
14247 if (SetCCOpcode == ISD::SETUEQ) {
14248 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14250 assert(SetCCOpcode == ISD::SETONE);
14251 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14254 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14255 DAG.getConstant(CC0, dl, MVT::i8));
14256 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14257 DAG.getConstant(CC1, dl, MVT::i8));
14258 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14260 // Handle all other FP comparisons here.
14261 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14262 DAG.getConstant(SSECC, dl, MVT::i8));
14265 MVT VTOp0 = Op0.getSimpleValueType();
14266 assert(VTOp0 == Op1.getSimpleValueType() &&
14267 "Expected operands with same type!");
14268 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14269 "Invalid number of packed elements for source and destination!");
14271 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14272 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14273 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14274 // legalizer firstly checks if the first operand in input to the setcc has
14275 // a legal type. If so, then it promotes the return type to that same type.
14276 // Otherwise, the return type is promoted to the 'next legal type' which,
14277 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14279 // We reach this code only if the following two conditions are met:
14280 // 1. Both return type and operand type have been promoted to wider types
14281 // by the type legalizer.
14282 // 2. The original operand type has been promoted to a 256-bit vector.
14284 // Note that condition 2. only applies for AVX targets.
14285 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14286 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14289 // The non-AVX512 code below works under the assumption that source and
14290 // destination types are the same.
14291 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14292 "Value types for source and destination must be the same!");
14294 // Break 256-bit integer vector compare into smaller ones.
14295 if (VT.is256BitVector() && !Subtarget->hasInt256())
14296 return Lower256IntVSETCC(Op, DAG);
14298 MVT OpVT = Op1.getSimpleValueType();
14299 if (OpVT.getVectorElementType() == MVT::i1)
14300 return LowerBoolVSETCC_AVX512(Op, DAG);
14302 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14303 if (Subtarget->hasAVX512()) {
14304 if (Op1.getSimpleValueType().is512BitVector() ||
14305 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14306 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14307 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14309 // In AVX-512 architecture setcc returns mask with i1 elements,
14310 // But there is no compare instruction for i8 and i16 elements in KNL.
14311 // We are not talking about 512-bit operands in this case, these
14312 // types are illegal.
14314 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14315 OpVT.getVectorElementType().getSizeInBits() >= 8))
14316 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14317 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14320 // Lower using XOP integer comparisons.
14321 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14322 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14323 // Translate compare code to XOP PCOM compare mode.
14324 unsigned CmpMode = 0;
14325 switch (SetCCOpcode) {
14326 default: llvm_unreachable("Unexpected SETCC condition");
14328 case ISD::SETLT: CmpMode = 0x00; break;
14330 case ISD::SETLE: CmpMode = 0x01; break;
14332 case ISD::SETGT: CmpMode = 0x02; break;
14334 case ISD::SETGE: CmpMode = 0x03; break;
14335 case ISD::SETEQ: CmpMode = 0x04; break;
14336 case ISD::SETNE: CmpMode = 0x05; break;
14339 // Are we comparing unsigned or signed integers?
14340 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14341 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14343 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14344 DAG.getConstant(CmpMode, dl, MVT::i8));
14347 // We are handling one of the integer comparisons here. Since SSE only has
14348 // GT and EQ comparisons for integer, swapping operands and multiple
14349 // operations may be required for some comparisons.
14351 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14352 bool Subus = false;
14354 switch (SetCCOpcode) {
14355 default: llvm_unreachable("Unexpected SETCC condition");
14356 case ISD::SETNE: Invert = true;
14357 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14358 case ISD::SETLT: Swap = true;
14359 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14360 case ISD::SETGE: Swap = true;
14361 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14362 Invert = true; break;
14363 case ISD::SETULT: Swap = true;
14364 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14365 FlipSigns = true; break;
14366 case ISD::SETUGE: Swap = true;
14367 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14368 FlipSigns = true; Invert = true; break;
14371 // Special case: Use min/max operations for SETULE/SETUGE
14372 MVT VET = VT.getVectorElementType();
14374 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14375 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14378 switch (SetCCOpcode) {
14380 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14381 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14384 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14387 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14388 if (!MinMax && hasSubus) {
14389 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14391 // t = psubus Op0, Op1
14392 // pcmpeq t, <0..0>
14393 switch (SetCCOpcode) {
14395 case ISD::SETULT: {
14396 // If the comparison is against a constant we can turn this into a
14397 // setule. With psubus, setule does not require a swap. This is
14398 // beneficial because the constant in the register is no longer
14399 // destructed as the destination so it can be hoisted out of a loop.
14400 // Only do this pre-AVX since vpcmp* is no longer destructive.
14401 if (Subtarget->hasAVX())
14403 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14404 if (ULEOp1.getNode()) {
14406 Subus = true; Invert = false; Swap = false;
14410 // Psubus is better than flip-sign because it requires no inversion.
14411 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14412 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14416 Opc = X86ISD::SUBUS;
14422 std::swap(Op0, Op1);
14424 // Check that the operation in question is available (most are plain SSE2,
14425 // but PCMPGTQ and PCMPEQQ have different requirements).
14426 if (VT == MVT::v2i64) {
14427 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14428 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14430 // First cast everything to the right type.
14431 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14432 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14434 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14435 // bits of the inputs before performing those operations. The lower
14436 // compare is always unsigned.
14439 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14441 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14442 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14443 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14444 Sign, Zero, Sign, Zero);
14446 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14447 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14449 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14450 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14451 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14453 // Create masks for only the low parts/high parts of the 64 bit integers.
14454 static const int MaskHi[] = { 1, 1, 3, 3 };
14455 static const int MaskLo[] = { 0, 0, 2, 2 };
14456 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14457 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14458 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14460 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14461 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14464 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14466 return DAG.getBitcast(VT, Result);
14469 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14470 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14471 // pcmpeqd + pshufd + pand.
14472 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14474 // First cast everything to the right type.
14475 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14476 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14479 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14481 // Make sure the lower and upper halves are both all-ones.
14482 static const int Mask[] = { 1, 0, 3, 2 };
14483 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14484 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14487 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14489 return DAG.getBitcast(VT, Result);
14493 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14494 // bits of the inputs before performing those operations.
14496 MVT EltVT = VT.getVectorElementType();
14497 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14499 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14500 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14503 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14505 // If the logical-not of the result is required, perform that now.
14507 Result = DAG.getNOT(dl, Result, VT);
14510 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14513 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14514 getZeroVector(VT, Subtarget, DAG, dl));
14519 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14521 MVT VT = Op.getSimpleValueType();
14523 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14525 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14526 && "SetCC type must be 8-bit or 1-bit integer");
14527 SDValue Op0 = Op.getOperand(0);
14528 SDValue Op1 = Op.getOperand(1);
14530 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14532 // Optimize to BT if possible.
14533 // Lower (X & (1 << N)) == 0 to BT(X, N).
14534 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14535 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14536 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14537 Op1.getOpcode() == ISD::Constant &&
14538 cast<ConstantSDNode>(Op1)->isNullValue() &&
14539 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14540 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14541 if (NewSetCC.getNode()) {
14543 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14548 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14550 if (Op1.getOpcode() == ISD::Constant &&
14551 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14552 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14553 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14555 // If the input is a setcc, then reuse the input setcc or use a new one with
14556 // the inverted condition.
14557 if (Op0.getOpcode() == X86ISD::SETCC) {
14558 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14559 bool Invert = (CC == ISD::SETNE) ^
14560 cast<ConstantSDNode>(Op1)->isNullValue();
14564 CCode = X86::GetOppositeBranchCondition(CCode);
14565 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14566 DAG.getConstant(CCode, dl, MVT::i8),
14567 Op0.getOperand(1));
14569 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14573 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14574 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14575 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14577 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14578 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14581 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14582 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14583 if (X86CC == X86::COND_INVALID)
14586 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14587 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14588 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14589 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14591 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14595 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14596 static bool isX86LogicalCmp(SDValue Op) {
14597 unsigned Opc = Op.getNode()->getOpcode();
14598 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14599 Opc == X86ISD::SAHF)
14601 if (Op.getResNo() == 1 &&
14602 (Opc == X86ISD::ADD ||
14603 Opc == X86ISD::SUB ||
14604 Opc == X86ISD::ADC ||
14605 Opc == X86ISD::SBB ||
14606 Opc == X86ISD::SMUL ||
14607 Opc == X86ISD::UMUL ||
14608 Opc == X86ISD::INC ||
14609 Opc == X86ISD::DEC ||
14610 Opc == X86ISD::OR ||
14611 Opc == X86ISD::XOR ||
14612 Opc == X86ISD::AND))
14615 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14621 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14622 if (V.getOpcode() != ISD::TRUNCATE)
14625 SDValue VOp0 = V.getOperand(0);
14626 unsigned InBits = VOp0.getValueSizeInBits();
14627 unsigned Bits = V.getValueSizeInBits();
14628 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14631 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14632 bool addTest = true;
14633 SDValue Cond = Op.getOperand(0);
14634 SDValue Op1 = Op.getOperand(1);
14635 SDValue Op2 = Op.getOperand(2);
14637 MVT VT = Op1.getSimpleValueType();
14640 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14641 // are available or VBLENDV if AVX is available.
14642 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14643 if (Cond.getOpcode() == ISD::SETCC &&
14644 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14645 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14646 VT == Cond.getOperand(0).getSimpleValueType() && Cond->hasOneUse()) {
14647 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14648 int SSECC = translateX86FSETCC(
14649 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14652 if (Subtarget->hasAVX512()) {
14653 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14654 DAG.getConstant(SSECC, DL, MVT::i8));
14655 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14658 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14659 DAG.getConstant(SSECC, DL, MVT::i8));
14661 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14662 // of 3 logic instructions for size savings and potentially speed.
14663 // Unfortunately, there is no scalar form of VBLENDV.
14665 // If either operand is a constant, don't try this. We can expect to
14666 // optimize away at least one of the logic instructions later in that
14667 // case, so that sequence would be faster than a variable blend.
14669 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14670 // uses XMM0 as the selection register. That may need just as many
14671 // instructions as the AND/ANDN/OR sequence due to register moves, so
14674 if (Subtarget->hasAVX() &&
14675 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14677 // Convert to vectors, do a VSELECT, and convert back to scalar.
14678 // All of the conversions should be optimized away.
14680 MVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14681 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14682 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14683 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14685 MVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14686 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14688 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14690 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14691 VSel, DAG.getIntPtrConstant(0, DL));
14693 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14694 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14695 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14699 if (VT.isVector() && VT.getVectorElementType() == MVT::i1) {
14701 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14702 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14703 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14704 Op1Scalar = Op1.getOperand(0);
14706 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14707 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14708 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14709 Op2Scalar = Op2.getOperand(0);
14710 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14711 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14712 Op1Scalar.getValueType(),
14713 Cond, Op1Scalar, Op2Scalar);
14714 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14715 return DAG.getBitcast(VT, newSelect);
14716 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14717 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14718 DAG.getIntPtrConstant(0, DL));
14722 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14723 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14724 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14725 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14726 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14727 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14728 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14730 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14733 if (Cond.getOpcode() == ISD::SETCC) {
14734 SDValue NewCond = LowerSETCC(Cond, DAG);
14735 if (NewCond.getNode())
14739 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14740 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14741 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14742 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14743 if (Cond.getOpcode() == X86ISD::SETCC &&
14744 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14745 isZero(Cond.getOperand(1).getOperand(1))) {
14746 SDValue Cmp = Cond.getOperand(1);
14748 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14750 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14751 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14752 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14754 SDValue CmpOp0 = Cmp.getOperand(0);
14755 // Apply further optimizations for special cases
14756 // (select (x != 0), -1, 0) -> neg & sbb
14757 // (select (x == 0), 0, -1) -> neg & sbb
14758 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14759 if (YC->isNullValue() &&
14760 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14761 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14762 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14763 DAG.getConstant(0, DL,
14764 CmpOp0.getValueType()),
14766 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14767 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14768 SDValue(Neg.getNode(), 1));
14772 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14773 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14774 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14776 SDValue Res = // Res = 0 or -1.
14777 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14778 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14780 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14781 Res = DAG.getNOT(DL, Res, Res.getValueType());
14783 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14784 if (!N2C || !N2C->isNullValue())
14785 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14790 // Look past (and (setcc_carry (cmp ...)), 1).
14791 if (Cond.getOpcode() == ISD::AND &&
14792 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14793 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14794 if (C && C->getAPIntValue() == 1)
14795 Cond = Cond.getOperand(0);
14798 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14799 // setting operand in place of the X86ISD::SETCC.
14800 unsigned CondOpcode = Cond.getOpcode();
14801 if (CondOpcode == X86ISD::SETCC ||
14802 CondOpcode == X86ISD::SETCC_CARRY) {
14803 CC = Cond.getOperand(0);
14805 SDValue Cmp = Cond.getOperand(1);
14806 unsigned Opc = Cmp.getOpcode();
14807 MVT VT = Op.getSimpleValueType();
14809 bool IllegalFPCMov = false;
14810 if (VT.isFloatingPoint() && !VT.isVector() &&
14811 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14812 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14814 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14815 Opc == X86ISD::BT) { // FIXME
14819 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14820 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14821 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14822 Cond.getOperand(0).getValueType() != MVT::i8)) {
14823 SDValue LHS = Cond.getOperand(0);
14824 SDValue RHS = Cond.getOperand(1);
14825 unsigned X86Opcode;
14828 switch (CondOpcode) {
14829 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14830 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14831 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14832 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14833 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14834 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14835 default: llvm_unreachable("unexpected overflowing operator");
14837 if (CondOpcode == ISD::UMULO)
14838 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14841 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14843 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14845 if (CondOpcode == ISD::UMULO)
14846 Cond = X86Op.getValue(2);
14848 Cond = X86Op.getValue(1);
14850 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14855 // Look past the truncate if the high bits are known zero.
14856 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14857 Cond = Cond.getOperand(0);
14859 // We know the result of AND is compared against zero. Try to match
14861 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14862 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14863 if (NewSetCC.getNode()) {
14864 CC = NewSetCC.getOperand(0);
14865 Cond = NewSetCC.getOperand(1);
14872 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14873 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14876 // a < b ? -1 : 0 -> RES = ~setcc_carry
14877 // a < b ? 0 : -1 -> RES = setcc_carry
14878 // a >= b ? -1 : 0 -> RES = setcc_carry
14879 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14880 if (Cond.getOpcode() == X86ISD::SUB) {
14881 Cond = ConvertCmpIfNecessary(Cond, DAG);
14882 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14884 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14885 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14886 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14887 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14889 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14890 return DAG.getNOT(DL, Res, Res.getValueType());
14895 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14896 // widen the cmov and push the truncate through. This avoids introducing a new
14897 // branch during isel and doesn't add any extensions.
14898 if (Op.getValueType() == MVT::i8 &&
14899 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14900 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14901 if (T1.getValueType() == T2.getValueType() &&
14902 // Blacklist CopyFromReg to avoid partial register stalls.
14903 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14904 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14905 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14906 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14910 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14911 // condition is true.
14912 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14913 SDValue Ops[] = { Op2, Op1, CC, Cond };
14914 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14917 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14918 const X86Subtarget *Subtarget,
14919 SelectionDAG &DAG) {
14920 MVT VT = Op->getSimpleValueType(0);
14921 SDValue In = Op->getOperand(0);
14922 MVT InVT = In.getSimpleValueType();
14923 MVT VTElt = VT.getVectorElementType();
14924 MVT InVTElt = InVT.getVectorElementType();
14928 if ((InVTElt == MVT::i1) &&
14929 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14930 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14932 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14933 VTElt.getSizeInBits() <= 16)) ||
14935 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14936 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14938 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14939 VTElt.getSizeInBits() >= 32))))
14940 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14942 unsigned int NumElts = VT.getVectorNumElements();
14944 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14947 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14948 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14949 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14950 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14953 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14954 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14956 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14959 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14961 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14962 if (VT.is512BitVector())
14964 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14967 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14968 const X86Subtarget *Subtarget,
14969 SelectionDAG &DAG) {
14970 SDValue In = Op->getOperand(0);
14971 MVT VT = Op->getSimpleValueType(0);
14972 MVT InVT = In.getSimpleValueType();
14973 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14975 MVT InSVT = InVT.getVectorElementType();
14976 assert(VT.getVectorElementType().getSizeInBits() > InSVT.getSizeInBits());
14978 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14980 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14985 // SSE41 targets can use the pmovsx* instructions directly.
14986 if (Subtarget->hasSSE41())
14987 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14989 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14993 // As SRAI is only available on i16/i32 types, we expand only up to i32
14994 // and handle i64 separately.
14995 while (CurrVT != VT && CurrVT.getVectorElementType() != MVT::i32) {
14996 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14997 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14998 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14999 Curr = DAG.getBitcast(CurrVT, Curr);
15002 SDValue SignExt = Curr;
15003 if (CurrVT != InVT) {
15004 unsigned SignExtShift =
15005 CurrVT.getVectorElementType().getSizeInBits() - InSVT.getSizeInBits();
15006 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15007 DAG.getConstant(SignExtShift, dl, MVT::i8));
15013 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15014 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15015 DAG.getConstant(31, dl, MVT::i8));
15016 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15017 return DAG.getBitcast(VT, Ext);
15023 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15024 SelectionDAG &DAG) {
15025 MVT VT = Op->getSimpleValueType(0);
15026 SDValue In = Op->getOperand(0);
15027 MVT InVT = In.getSimpleValueType();
15030 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15031 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15033 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15034 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15035 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15038 if (Subtarget->hasInt256())
15039 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15041 // Optimize vectors in AVX mode
15042 // Sign extend v8i16 to v8i32 and
15045 // Divide input vector into two parts
15046 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15047 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15048 // concat the vectors to original VT
15050 unsigned NumElems = InVT.getVectorNumElements();
15051 SDValue Undef = DAG.getUNDEF(InVT);
15053 SmallVector<int,8> ShufMask1(NumElems, -1);
15054 for (unsigned i = 0; i != NumElems/2; ++i)
15057 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15059 SmallVector<int,8> ShufMask2(NumElems, -1);
15060 for (unsigned i = 0; i != NumElems/2; ++i)
15061 ShufMask2[i] = i + NumElems/2;
15063 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15065 MVT HalfVT = MVT::getVectorVT(VT.getVectorElementType(),
15066 VT.getVectorNumElements()/2);
15068 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15069 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15071 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15074 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15075 // may emit an illegal shuffle but the expansion is still better than scalar
15076 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15077 // we'll emit a shuffle and a arithmetic shift.
15078 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15079 // TODO: It is possible to support ZExt by zeroing the undef values during
15080 // the shuffle phase or after the shuffle.
15081 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15082 SelectionDAG &DAG) {
15083 MVT RegVT = Op.getSimpleValueType();
15084 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15085 assert(RegVT.isInteger() &&
15086 "We only custom lower integer vector sext loads.");
15088 // Nothing useful we can do without SSE2 shuffles.
15089 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15091 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15093 EVT MemVT = Ld->getMemoryVT();
15094 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15095 unsigned RegSz = RegVT.getSizeInBits();
15097 ISD::LoadExtType Ext = Ld->getExtensionType();
15099 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15100 && "Only anyext and sext are currently implemented.");
15101 assert(MemVT != RegVT && "Cannot extend to the same type");
15102 assert(MemVT.isVector() && "Must load a vector from memory");
15104 unsigned NumElems = RegVT.getVectorNumElements();
15105 unsigned MemSz = MemVT.getSizeInBits();
15106 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15108 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15109 // The only way in which we have a legal 256-bit vector result but not the
15110 // integer 256-bit operations needed to directly lower a sextload is if we
15111 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15112 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15113 // correctly legalized. We do this late to allow the canonical form of
15114 // sextload to persist throughout the rest of the DAG combiner -- it wants
15115 // to fold together any extensions it can, and so will fuse a sign_extend
15116 // of an sextload into a sextload targeting a wider value.
15118 if (MemSz == 128) {
15119 // Just switch this to a normal load.
15120 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15121 "it must be a legal 128-bit vector "
15123 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15124 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15125 Ld->isInvariant(), Ld->getAlignment());
15127 assert(MemSz < 128 &&
15128 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15129 // Do an sext load to a 128-bit vector type. We want to use the same
15130 // number of elements, but elements half as wide. This will end up being
15131 // recursively lowered by this routine, but will succeed as we definitely
15132 // have all the necessary features if we're using AVX1.
15134 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15135 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15137 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15138 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15139 Ld->isNonTemporal(), Ld->isInvariant(),
15140 Ld->getAlignment());
15143 // Replace chain users with the new chain.
15144 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15145 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15147 // Finally, do a normal sign-extend to the desired register.
15148 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15151 // All sizes must be a power of two.
15152 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15153 "Non-power-of-two elements are not custom lowered!");
15155 // Attempt to load the original value using scalar loads.
15156 // Find the largest scalar type that divides the total loaded size.
15157 MVT SclrLoadTy = MVT::i8;
15158 for (MVT Tp : MVT::integer_valuetypes()) {
15159 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15164 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15165 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15167 SclrLoadTy = MVT::f64;
15169 // Calculate the number of scalar loads that we need to perform
15170 // in order to load our vector from memory.
15171 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15173 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15174 "Can only lower sext loads with a single scalar load!");
15176 unsigned loadRegZize = RegSz;
15177 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15180 // Represent our vector as a sequence of elements which are the
15181 // largest scalar that we can load.
15182 EVT LoadUnitVecVT = EVT::getVectorVT(
15183 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15185 // Represent the data using the same element type that is stored in
15186 // memory. In practice, we ''widen'' MemVT.
15188 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15189 loadRegZize / MemVT.getScalarSizeInBits());
15191 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15192 "Invalid vector type");
15194 // We can't shuffle using an illegal type.
15195 assert(TLI.isTypeLegal(WideVecVT) &&
15196 "We only lower types that form legal widened vector types");
15198 SmallVector<SDValue, 8> Chains;
15199 SDValue Ptr = Ld->getBasePtr();
15200 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15201 TLI.getPointerTy(DAG.getDataLayout()));
15202 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15204 for (unsigned i = 0; i < NumLoads; ++i) {
15205 // Perform a single load.
15206 SDValue ScalarLoad =
15207 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15208 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15209 Ld->getAlignment());
15210 Chains.push_back(ScalarLoad.getValue(1));
15211 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15212 // another round of DAGCombining.
15214 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15216 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15217 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15219 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15222 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15224 // Bitcast the loaded value to a vector of the original element type, in
15225 // the size of the target vector type.
15226 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15227 unsigned SizeRatio = RegSz / MemSz;
15229 if (Ext == ISD::SEXTLOAD) {
15230 // If we have SSE4.1, we can directly emit a VSEXT node.
15231 if (Subtarget->hasSSE41()) {
15232 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15233 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15237 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15239 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15240 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15242 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15243 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15247 // Redistribute the loaded elements into the different locations.
15248 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15249 for (unsigned i = 0; i != NumElems; ++i)
15250 ShuffleVec[i * SizeRatio] = i;
15252 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15253 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15255 // Bitcast to the requested type.
15256 Shuff = DAG.getBitcast(RegVT, Shuff);
15257 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15261 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15262 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15263 // from the AND / OR.
15264 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15265 Opc = Op.getOpcode();
15266 if (Opc != ISD::OR && Opc != ISD::AND)
15268 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15269 Op.getOperand(0).hasOneUse() &&
15270 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15271 Op.getOperand(1).hasOneUse());
15274 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15275 // 1 and that the SETCC node has a single use.
15276 static bool isXor1OfSetCC(SDValue Op) {
15277 if (Op.getOpcode() != ISD::XOR)
15279 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15280 if (N1C && N1C->getAPIntValue() == 1) {
15281 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15282 Op.getOperand(0).hasOneUse();
15287 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15288 bool addTest = true;
15289 SDValue Chain = Op.getOperand(0);
15290 SDValue Cond = Op.getOperand(1);
15291 SDValue Dest = Op.getOperand(2);
15294 bool Inverted = false;
15296 if (Cond.getOpcode() == ISD::SETCC) {
15297 // Check for setcc([su]{add,sub,mul}o == 0).
15298 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15299 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15300 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15301 Cond.getOperand(0).getResNo() == 1 &&
15302 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15303 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15304 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15305 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15306 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15307 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15309 Cond = Cond.getOperand(0);
15311 SDValue NewCond = LowerSETCC(Cond, DAG);
15312 if (NewCond.getNode())
15317 // FIXME: LowerXALUO doesn't handle these!!
15318 else if (Cond.getOpcode() == X86ISD::ADD ||
15319 Cond.getOpcode() == X86ISD::SUB ||
15320 Cond.getOpcode() == X86ISD::SMUL ||
15321 Cond.getOpcode() == X86ISD::UMUL)
15322 Cond = LowerXALUO(Cond, DAG);
15325 // Look pass (and (setcc_carry (cmp ...)), 1).
15326 if (Cond.getOpcode() == ISD::AND &&
15327 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15328 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15329 if (C && C->getAPIntValue() == 1)
15330 Cond = Cond.getOperand(0);
15333 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15334 // setting operand in place of the X86ISD::SETCC.
15335 unsigned CondOpcode = Cond.getOpcode();
15336 if (CondOpcode == X86ISD::SETCC ||
15337 CondOpcode == X86ISD::SETCC_CARRY) {
15338 CC = Cond.getOperand(0);
15340 SDValue Cmp = Cond.getOperand(1);
15341 unsigned Opc = Cmp.getOpcode();
15342 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15343 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15347 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15351 // These can only come from an arithmetic instruction with overflow,
15352 // e.g. SADDO, UADDO.
15353 Cond = Cond.getNode()->getOperand(1);
15359 CondOpcode = Cond.getOpcode();
15360 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15361 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15362 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15363 Cond.getOperand(0).getValueType() != MVT::i8)) {
15364 SDValue LHS = Cond.getOperand(0);
15365 SDValue RHS = Cond.getOperand(1);
15366 unsigned X86Opcode;
15369 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15370 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15372 switch (CondOpcode) {
15373 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15375 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15377 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15380 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15381 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15385 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15388 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15389 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15390 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15391 default: llvm_unreachable("unexpected overflowing operator");
15394 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15395 if (CondOpcode == ISD::UMULO)
15396 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15399 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15401 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15403 if (CondOpcode == ISD::UMULO)
15404 Cond = X86Op.getValue(2);
15406 Cond = X86Op.getValue(1);
15408 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15412 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15413 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15414 if (CondOpc == ISD::OR) {
15415 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15416 // two branches instead of an explicit OR instruction with a
15418 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15419 isX86LogicalCmp(Cmp)) {
15420 CC = Cond.getOperand(0).getOperand(0);
15421 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15422 Chain, Dest, CC, Cmp);
15423 CC = Cond.getOperand(1).getOperand(0);
15427 } else { // ISD::AND
15428 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15429 // two branches instead of an explicit AND instruction with a
15430 // separate test. However, we only do this if this block doesn't
15431 // have a fall-through edge, because this requires an explicit
15432 // jmp when the condition is false.
15433 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15434 isX86LogicalCmp(Cmp) &&
15435 Op.getNode()->hasOneUse()) {
15436 X86::CondCode CCode =
15437 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15438 CCode = X86::GetOppositeBranchCondition(CCode);
15439 CC = DAG.getConstant(CCode, dl, MVT::i8);
15440 SDNode *User = *Op.getNode()->use_begin();
15441 // Look for an unconditional branch following this conditional branch.
15442 // We need this because we need to reverse the successors in order
15443 // to implement FCMP_OEQ.
15444 if (User->getOpcode() == ISD::BR) {
15445 SDValue FalseBB = User->getOperand(1);
15447 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15448 assert(NewBR == User);
15452 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15453 Chain, Dest, CC, Cmp);
15454 X86::CondCode CCode =
15455 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15456 CCode = X86::GetOppositeBranchCondition(CCode);
15457 CC = DAG.getConstant(CCode, dl, MVT::i8);
15463 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15464 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15465 // It should be transformed during dag combiner except when the condition
15466 // is set by a arithmetics with overflow node.
15467 X86::CondCode CCode =
15468 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15469 CCode = X86::GetOppositeBranchCondition(CCode);
15470 CC = DAG.getConstant(CCode, dl, MVT::i8);
15471 Cond = Cond.getOperand(0).getOperand(1);
15473 } else if (Cond.getOpcode() == ISD::SETCC &&
15474 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15475 // For FCMP_OEQ, we can emit
15476 // two branches instead of an explicit AND instruction with a
15477 // separate test. However, we only do this if this block doesn't
15478 // have a fall-through edge, because this requires an explicit
15479 // jmp when the condition is false.
15480 if (Op.getNode()->hasOneUse()) {
15481 SDNode *User = *Op.getNode()->use_begin();
15482 // Look for an unconditional branch following this conditional branch.
15483 // We need this because we need to reverse the successors in order
15484 // to implement FCMP_OEQ.
15485 if (User->getOpcode() == ISD::BR) {
15486 SDValue FalseBB = User->getOperand(1);
15488 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15489 assert(NewBR == User);
15493 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15494 Cond.getOperand(0), Cond.getOperand(1));
15495 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15496 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15497 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15498 Chain, Dest, CC, Cmp);
15499 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15504 } else if (Cond.getOpcode() == ISD::SETCC &&
15505 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15506 // For FCMP_UNE, we can emit
15507 // two branches instead of an explicit AND instruction with a
15508 // separate test. However, we only do this if this block doesn't
15509 // have a fall-through edge, because this requires an explicit
15510 // jmp when the condition is false.
15511 if (Op.getNode()->hasOneUse()) {
15512 SDNode *User = *Op.getNode()->use_begin();
15513 // Look for an unconditional branch following this conditional branch.
15514 // We need this because we need to reverse the successors in order
15515 // to implement FCMP_UNE.
15516 if (User->getOpcode() == ISD::BR) {
15517 SDValue FalseBB = User->getOperand(1);
15519 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15520 assert(NewBR == User);
15523 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15524 Cond.getOperand(0), Cond.getOperand(1));
15525 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15526 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15527 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15528 Chain, Dest, CC, Cmp);
15529 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15539 // Look pass the truncate if the high bits are known zero.
15540 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15541 Cond = Cond.getOperand(0);
15543 // We know the result of AND is compared against zero. Try to match
15545 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15546 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15547 if (NewSetCC.getNode()) {
15548 CC = NewSetCC.getOperand(0);
15549 Cond = NewSetCC.getOperand(1);
15556 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15557 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15558 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15560 Cond = ConvertCmpIfNecessary(Cond, DAG);
15561 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15562 Chain, Dest, CC, Cond);
15565 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15566 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15567 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15568 // that the guard pages used by the OS virtual memory manager are allocated in
15569 // correct sequence.
15571 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15572 SelectionDAG &DAG) const {
15573 MachineFunction &MF = DAG.getMachineFunction();
15574 bool SplitStack = MF.shouldSplitStack();
15575 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15580 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15581 SDNode* Node = Op.getNode();
15583 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15584 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15585 " not tell us which reg is the stack pointer!");
15586 EVT VT = Node->getValueType(0);
15587 SDValue Tmp1 = SDValue(Node, 0);
15588 SDValue Tmp2 = SDValue(Node, 1);
15589 SDValue Tmp3 = Node->getOperand(2);
15590 SDValue Chain = Tmp1.getOperand(0);
15592 // Chain the dynamic stack allocation so that it doesn't modify the stack
15593 // pointer when other instructions are using the stack.
15594 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15597 SDValue Size = Tmp2.getOperand(1);
15598 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15599 Chain = SP.getValue(1);
15600 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15601 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15602 unsigned StackAlign = TFI.getStackAlignment();
15603 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15604 if (Align > StackAlign)
15605 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15606 DAG.getConstant(-(uint64_t)Align, dl, VT));
15607 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15609 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15610 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15613 SDValue Ops[2] = { Tmp1, Tmp2 };
15614 return DAG.getMergeValues(Ops, dl);
15618 SDValue Chain = Op.getOperand(0);
15619 SDValue Size = Op.getOperand(1);
15620 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15621 EVT VT = Op.getNode()->getValueType(0);
15623 bool Is64Bit = Subtarget->is64Bit();
15624 MVT SPTy = getPointerTy(DAG.getDataLayout());
15627 MachineRegisterInfo &MRI = MF.getRegInfo();
15630 // The 64 bit implementation of segmented stacks needs to clobber both r10
15631 // r11. This makes it impossible to use it along with nested parameters.
15632 const Function *F = MF.getFunction();
15634 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15636 if (I->hasNestAttr())
15637 report_fatal_error("Cannot use segmented stacks with functions that "
15638 "have nested arguments.");
15641 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15642 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15643 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15644 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15645 DAG.getRegister(Vreg, SPTy));
15646 SDValue Ops1[2] = { Value, Chain };
15647 return DAG.getMergeValues(Ops1, dl);
15650 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15652 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15653 Flag = Chain.getValue(1);
15654 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15656 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15658 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15659 unsigned SPReg = RegInfo->getStackRegister();
15660 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15661 Chain = SP.getValue(1);
15664 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15665 DAG.getConstant(-(uint64_t)Align, dl, VT));
15666 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15669 SDValue Ops1[2] = { SP, Chain };
15670 return DAG.getMergeValues(Ops1, dl);
15674 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15675 MachineFunction &MF = DAG.getMachineFunction();
15676 auto PtrVT = getPointerTy(MF.getDataLayout());
15677 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15679 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15682 if (!Subtarget->is64Bit() ||
15683 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15684 // vastart just stores the address of the VarArgsFrameIndex slot into the
15685 // memory location argument.
15686 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15687 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15688 MachinePointerInfo(SV), false, false, 0);
15692 // gp_offset (0 - 6 * 8)
15693 // fp_offset (48 - 48 + 8 * 16)
15694 // overflow_arg_area (point to parameters coming in memory).
15696 SmallVector<SDValue, 8> MemOps;
15697 SDValue FIN = Op.getOperand(1);
15699 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15700 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15702 FIN, MachinePointerInfo(SV), false, false, 0);
15703 MemOps.push_back(Store);
15706 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15707 Store = DAG.getStore(Op.getOperand(0), DL,
15708 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15710 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15711 MemOps.push_back(Store);
15713 // Store ptr to overflow_arg_area
15714 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15715 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15716 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15717 MachinePointerInfo(SV, 8),
15719 MemOps.push_back(Store);
15721 // Store ptr to reg_save_area.
15722 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15723 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15724 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15725 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15726 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15727 MemOps.push_back(Store);
15728 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15731 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15732 assert(Subtarget->is64Bit() &&
15733 "LowerVAARG only handles 64-bit va_arg!");
15734 assert(Op.getNode()->getNumOperands() == 4);
15736 MachineFunction &MF = DAG.getMachineFunction();
15737 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15738 // The Win64 ABI uses char* instead of a structure.
15739 return DAG.expandVAArg(Op.getNode());
15741 SDValue Chain = Op.getOperand(0);
15742 SDValue SrcPtr = Op.getOperand(1);
15743 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15744 unsigned Align = Op.getConstantOperandVal(3);
15747 EVT ArgVT = Op.getNode()->getValueType(0);
15748 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15749 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15752 // Decide which area this value should be read from.
15753 // TODO: Implement the AMD64 ABI in its entirety. This simple
15754 // selection mechanism works only for the basic types.
15755 if (ArgVT == MVT::f80) {
15756 llvm_unreachable("va_arg for f80 not yet implemented");
15757 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15758 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15759 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15760 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15762 llvm_unreachable("Unhandled argument type in LowerVAARG");
15765 if (ArgMode == 2) {
15766 // Sanity Check: Make sure using fp_offset makes sense.
15767 assert(!Subtarget->useSoftFloat() &&
15768 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15769 Subtarget->hasSSE1());
15772 // Insert VAARG_64 node into the DAG
15773 // VAARG_64 returns two values: Variable Argument Address, Chain
15774 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15775 DAG.getConstant(ArgMode, dl, MVT::i8),
15776 DAG.getConstant(Align, dl, MVT::i32)};
15777 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15778 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15779 VTs, InstOps, MVT::i64,
15780 MachinePointerInfo(SV),
15782 /*Volatile=*/false,
15784 /*WriteMem=*/true);
15785 Chain = VAARG.getValue(1);
15787 // Load the next argument and return it
15788 return DAG.getLoad(ArgVT, dl,
15791 MachinePointerInfo(),
15792 false, false, false, 0);
15795 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15796 SelectionDAG &DAG) {
15797 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15798 // where a va_list is still an i8*.
15799 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15800 if (Subtarget->isCallingConvWin64(
15801 DAG.getMachineFunction().getFunction()->getCallingConv()))
15802 // Probably a Win64 va_copy.
15803 return DAG.expandVACopy(Op.getNode());
15805 SDValue Chain = Op.getOperand(0);
15806 SDValue DstPtr = Op.getOperand(1);
15807 SDValue SrcPtr = Op.getOperand(2);
15808 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15809 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15812 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15813 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15815 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15818 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15819 // amount is a constant. Takes immediate version of shift as input.
15820 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15821 SDValue SrcOp, uint64_t ShiftAmt,
15822 SelectionDAG &DAG) {
15823 MVT ElementType = VT.getVectorElementType();
15825 // Fold this packed shift into its first operand if ShiftAmt is 0.
15829 // Check for ShiftAmt >= element width
15830 if (ShiftAmt >= ElementType.getSizeInBits()) {
15831 if (Opc == X86ISD::VSRAI)
15832 ShiftAmt = ElementType.getSizeInBits() - 1;
15834 return DAG.getConstant(0, dl, VT);
15837 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15838 && "Unknown target vector shift-by-constant node");
15840 // Fold this packed vector shift into a build vector if SrcOp is a
15841 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15842 if (VT == SrcOp.getSimpleValueType() &&
15843 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15844 SmallVector<SDValue, 8> Elts;
15845 unsigned NumElts = SrcOp->getNumOperands();
15846 ConstantSDNode *ND;
15849 default: llvm_unreachable(nullptr);
15850 case X86ISD::VSHLI:
15851 for (unsigned i=0; i!=NumElts; ++i) {
15852 SDValue CurrentOp = SrcOp->getOperand(i);
15853 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15854 Elts.push_back(CurrentOp);
15857 ND = cast<ConstantSDNode>(CurrentOp);
15858 const APInt &C = ND->getAPIntValue();
15859 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15862 case X86ISD::VSRLI:
15863 for (unsigned i=0; i!=NumElts; ++i) {
15864 SDValue CurrentOp = SrcOp->getOperand(i);
15865 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15866 Elts.push_back(CurrentOp);
15869 ND = cast<ConstantSDNode>(CurrentOp);
15870 const APInt &C = ND->getAPIntValue();
15871 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15874 case X86ISD::VSRAI:
15875 for (unsigned i=0; i!=NumElts; ++i) {
15876 SDValue CurrentOp = SrcOp->getOperand(i);
15877 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15878 Elts.push_back(CurrentOp);
15881 ND = cast<ConstantSDNode>(CurrentOp);
15882 const APInt &C = ND->getAPIntValue();
15883 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15888 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15891 return DAG.getNode(Opc, dl, VT, SrcOp,
15892 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15895 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15896 // may or may not be a constant. Takes immediate version of shift as input.
15897 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15898 SDValue SrcOp, SDValue ShAmt,
15899 SelectionDAG &DAG) {
15900 MVT SVT = ShAmt.getSimpleValueType();
15901 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15903 // Catch shift-by-constant.
15904 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15905 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15906 CShAmt->getZExtValue(), DAG);
15908 // Change opcode to non-immediate version
15910 default: llvm_unreachable("Unknown target vector shift node");
15911 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15912 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15913 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15916 const X86Subtarget &Subtarget =
15917 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15918 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15919 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15920 // Let the shuffle legalizer expand this shift amount node.
15921 SDValue Op0 = ShAmt.getOperand(0);
15922 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15923 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15925 // Need to build a vector containing shift amount.
15926 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15927 SmallVector<SDValue, 4> ShOps;
15928 ShOps.push_back(ShAmt);
15929 if (SVT == MVT::i32) {
15930 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15931 ShOps.push_back(DAG.getUNDEF(SVT));
15933 ShOps.push_back(DAG.getUNDEF(SVT));
15935 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15936 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15939 // The return type has to be a 128-bit type with the same element
15940 // type as the input type.
15941 MVT EltVT = VT.getVectorElementType();
15942 MVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15944 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15945 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15948 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15949 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15950 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15951 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15952 SDValue PreservedSrc,
15953 const X86Subtarget *Subtarget,
15954 SelectionDAG &DAG) {
15955 MVT VT = Op.getSimpleValueType();
15956 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
15958 unsigned OpcodeSelect = ISD::VSELECT;
15961 if (isAllOnes(Mask))
15964 if (MaskVT.bitsGT(Mask.getSimpleValueType())) {
15965 MVT newMaskVT = MVT::getIntegerVT(MaskVT.getSizeInBits());
15966 VMask = DAG.getBitcast(MaskVT,
15967 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15969 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
15970 Mask.getSimpleValueType().getSizeInBits());
15971 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15972 // are extracted by EXTRACT_SUBVECTOR.
15973 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15974 DAG.getBitcast(BitcastVT, Mask),
15975 DAG.getIntPtrConstant(0, dl));
15978 switch (Op.getOpcode()) {
15980 case X86ISD::PCMPEQM:
15981 case X86ISD::PCMPGTM:
15983 case X86ISD::CMPMU:
15984 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15985 case X86ISD::VFPCLASS:
15986 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15987 case X86ISD::VTRUNC:
15988 case X86ISD::VTRUNCS:
15989 case X86ISD::VTRUNCUS:
15990 // We can't use ISD::VSELECT here because it is not always "Legal"
15991 // for the destination type. For example vpmovqb require only AVX512
15992 // and vselect that can operate on byte element type require BWI
15993 OpcodeSelect = X86ISD::SELECT;
15996 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15997 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15998 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16001 /// \brief Creates an SDNode for a predicated scalar operation.
16002 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16003 /// The mask is coming as MVT::i8 and it should be truncated
16004 /// to MVT::i1 while lowering masking intrinsics.
16005 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16006 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16007 /// for a scalar instruction.
16008 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16009 SDValue PreservedSrc,
16010 const X86Subtarget *Subtarget,
16011 SelectionDAG &DAG) {
16012 if (isAllOnes(Mask))
16015 MVT VT = Op.getSimpleValueType();
16017 // The mask should be of type MVT::i1
16018 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16020 if (Op.getOpcode() == X86ISD::FSETCC)
16021 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16022 if (Op.getOpcode() == X86ISD::VFPCLASS)
16023 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16025 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16026 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16027 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16030 static int getSEHRegistrationNodeSize(const Function *Fn) {
16031 if (!Fn->hasPersonalityFn())
16032 report_fatal_error(
16033 "querying registration node size for function without personality");
16034 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16035 // WinEHStatePass for the full struct definition.
16036 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16037 case EHPersonality::MSVC_X86SEH: return 24;
16038 case EHPersonality::MSVC_CXX: return 16;
16041 report_fatal_error("can only recover FP for MSVC EH personality functions");
16044 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
16045 /// function or when returning to a parent frame after catching an exception, we
16046 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16047 /// Here's the math:
16048 /// RegNodeBase = EntryEBP - RegNodeSize
16049 /// ParentFP = RegNodeBase - RegNodeFrameOffset
16050 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16051 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16052 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16053 SDValue EntryEBP) {
16054 MachineFunction &MF = DAG.getMachineFunction();
16057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16058 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16060 // It's possible that the parent function no longer has a personality function
16061 // if the exceptional code was optimized away, in which case we just return
16062 // the incoming EBP.
16063 if (!Fn->hasPersonalityFn())
16066 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16068 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16070 MCSymbol *OffsetSym =
16071 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16072 GlobalValue::getRealLinkageName(Fn->getName()));
16073 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16074 SDValue RegNodeFrameOffset =
16075 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16077 // RegNodeBase = EntryEBP - RegNodeSize
16078 // ParentFP = RegNodeBase - RegNodeFrameOffset
16079 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16080 DAG.getConstant(RegNodeSize, dl, PtrVT));
16081 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16084 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16085 SelectionDAG &DAG) {
16087 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16088 MVT VT = Op.getSimpleValueType();
16089 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16091 switch(IntrData->Type) {
16092 case INTR_TYPE_1OP:
16093 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16094 case INTR_TYPE_2OP:
16095 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16097 case INTR_TYPE_2OP_IMM8:
16098 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16099 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16100 case INTR_TYPE_3OP:
16101 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16102 Op.getOperand(2), Op.getOperand(3));
16103 case INTR_TYPE_4OP:
16104 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16105 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16106 case INTR_TYPE_1OP_MASK_RM: {
16107 SDValue Src = Op.getOperand(1);
16108 SDValue PassThru = Op.getOperand(2);
16109 SDValue Mask = Op.getOperand(3);
16110 SDValue RoundingMode;
16111 // We allways add rounding mode to the Node.
16112 // If the rounding mode is not specified, we add the
16113 // "current direction" mode.
16114 if (Op.getNumOperands() == 4)
16116 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16118 RoundingMode = Op.getOperand(4);
16119 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16120 if (IntrWithRoundingModeOpcode != 0)
16121 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16122 X86::STATIC_ROUNDING::CUR_DIRECTION)
16123 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16124 dl, Op.getValueType(), Src, RoundingMode),
16125 Mask, PassThru, Subtarget, DAG);
16126 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16128 Mask, PassThru, Subtarget, DAG);
16130 case INTR_TYPE_1OP_MASK: {
16131 SDValue Src = Op.getOperand(1);
16132 SDValue PassThru = Op.getOperand(2);
16133 SDValue Mask = Op.getOperand(3);
16134 // We add rounding mode to the Node when
16135 // - RM Opcode is specified and
16136 // - RM is not "current direction".
16137 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16138 if (IntrWithRoundingModeOpcode != 0) {
16139 SDValue Rnd = Op.getOperand(4);
16140 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16141 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16142 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16143 dl, Op.getValueType(),
16145 Mask, PassThru, Subtarget, DAG);
16148 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16149 Mask, PassThru, Subtarget, DAG);
16151 case INTR_TYPE_SCALAR_MASK: {
16152 SDValue Src1 = Op.getOperand(1);
16153 SDValue Src2 = Op.getOperand(2);
16154 SDValue passThru = Op.getOperand(3);
16155 SDValue Mask = Op.getOperand(4);
16156 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16157 Mask, passThru, Subtarget, DAG);
16159 case INTR_TYPE_SCALAR_MASK_RM: {
16160 SDValue Src1 = Op.getOperand(1);
16161 SDValue Src2 = Op.getOperand(2);
16162 SDValue Src0 = Op.getOperand(3);
16163 SDValue Mask = Op.getOperand(4);
16164 // There are 2 kinds of intrinsics in this group:
16165 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16166 // (2) With rounding mode and sae - 7 operands.
16167 if (Op.getNumOperands() == 6) {
16168 SDValue Sae = Op.getOperand(5);
16169 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16170 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16172 Mask, Src0, Subtarget, DAG);
16174 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16175 SDValue RoundingMode = Op.getOperand(5);
16176 SDValue Sae = Op.getOperand(6);
16177 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16178 RoundingMode, Sae),
16179 Mask, Src0, Subtarget, DAG);
16181 case INTR_TYPE_2OP_MASK:
16182 case INTR_TYPE_2OP_IMM8_MASK: {
16183 SDValue Src1 = Op.getOperand(1);
16184 SDValue Src2 = Op.getOperand(2);
16185 SDValue PassThru = Op.getOperand(3);
16186 SDValue Mask = Op.getOperand(4);
16188 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16189 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16191 // We specify 2 possible opcodes for intrinsics with rounding modes.
16192 // First, we check if the intrinsic may have non-default rounding mode,
16193 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16194 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16195 if (IntrWithRoundingModeOpcode != 0) {
16196 SDValue Rnd = Op.getOperand(5);
16197 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16198 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16199 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16200 dl, Op.getValueType(),
16202 Mask, PassThru, Subtarget, DAG);
16205 // TODO: Intrinsics should have fast-math-flags to propagate.
16206 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16207 Mask, PassThru, Subtarget, DAG);
16209 case INTR_TYPE_2OP_MASK_RM: {
16210 SDValue Src1 = Op.getOperand(1);
16211 SDValue Src2 = Op.getOperand(2);
16212 SDValue PassThru = Op.getOperand(3);
16213 SDValue Mask = Op.getOperand(4);
16214 // We specify 2 possible modes for intrinsics, with/without rounding
16216 // First, we check if the intrinsic have rounding mode (6 operands),
16217 // if not, we set rounding mode to "current".
16219 if (Op.getNumOperands() == 6)
16220 Rnd = Op.getOperand(5);
16222 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16223 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16225 Mask, PassThru, Subtarget, DAG);
16227 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16228 SDValue Src1 = Op.getOperand(1);
16229 SDValue Src2 = Op.getOperand(2);
16230 SDValue Src3 = Op.getOperand(3);
16231 SDValue PassThru = Op.getOperand(4);
16232 SDValue Mask = Op.getOperand(5);
16233 SDValue Sae = Op.getOperand(6);
16235 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16237 Mask, PassThru, Subtarget, DAG);
16239 case INTR_TYPE_3OP_MASK_RM: {
16240 SDValue Src1 = Op.getOperand(1);
16241 SDValue Src2 = Op.getOperand(2);
16242 SDValue Imm = Op.getOperand(3);
16243 SDValue PassThru = Op.getOperand(4);
16244 SDValue Mask = Op.getOperand(5);
16245 // We specify 2 possible modes for intrinsics, with/without rounding
16247 // First, we check if the intrinsic have rounding mode (7 operands),
16248 // if not, we set rounding mode to "current".
16250 if (Op.getNumOperands() == 7)
16251 Rnd = Op.getOperand(6);
16253 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16254 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16255 Src1, Src2, Imm, Rnd),
16256 Mask, PassThru, Subtarget, DAG);
16258 case INTR_TYPE_3OP_IMM8_MASK:
16259 case INTR_TYPE_3OP_MASK:
16260 case INSERT_SUBVEC: {
16261 SDValue Src1 = Op.getOperand(1);
16262 SDValue Src2 = Op.getOperand(2);
16263 SDValue Src3 = Op.getOperand(3);
16264 SDValue PassThru = Op.getOperand(4);
16265 SDValue Mask = Op.getOperand(5);
16267 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16268 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16269 else if (IntrData->Type == INSERT_SUBVEC) {
16270 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16271 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16272 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16273 Imm *= Src2.getSimpleValueType().getVectorNumElements();
16274 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16277 // We specify 2 possible opcodes for intrinsics with rounding modes.
16278 // First, we check if the intrinsic may have non-default rounding mode,
16279 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16280 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16281 if (IntrWithRoundingModeOpcode != 0) {
16282 SDValue Rnd = Op.getOperand(6);
16283 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16284 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16285 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16286 dl, Op.getValueType(),
16287 Src1, Src2, Src3, Rnd),
16288 Mask, PassThru, Subtarget, DAG);
16291 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16293 Mask, PassThru, Subtarget, DAG);
16295 case VPERM_3OP_MASKZ:
16296 case VPERM_3OP_MASK:
16299 case FMA_OP_MASK: {
16300 SDValue Src1 = Op.getOperand(1);
16301 SDValue Src2 = Op.getOperand(2);
16302 SDValue Src3 = Op.getOperand(3);
16303 SDValue Mask = Op.getOperand(4);
16304 MVT VT = Op.getSimpleValueType();
16305 SDValue PassThru = SDValue();
16307 // set PassThru element
16308 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16309 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16310 else if (IntrData->Type == FMA_OP_MASK3)
16315 // We specify 2 possible opcodes for intrinsics with rounding modes.
16316 // First, we check if the intrinsic may have non-default rounding mode,
16317 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16318 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16319 if (IntrWithRoundingModeOpcode != 0) {
16320 SDValue Rnd = Op.getOperand(5);
16321 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16322 X86::STATIC_ROUNDING::CUR_DIRECTION)
16323 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16324 dl, Op.getValueType(),
16325 Src1, Src2, Src3, Rnd),
16326 Mask, PassThru, Subtarget, DAG);
16328 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16329 dl, Op.getValueType(),
16331 Mask, PassThru, Subtarget, DAG);
16333 case TERLOG_OP_MASK:
16334 case TERLOG_OP_MASKZ: {
16335 SDValue Src1 = Op.getOperand(1);
16336 SDValue Src2 = Op.getOperand(2);
16337 SDValue Src3 = Op.getOperand(3);
16338 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16339 SDValue Mask = Op.getOperand(5);
16340 MVT VT = Op.getSimpleValueType();
16341 SDValue PassThru = Src1;
16342 // Set PassThru element.
16343 if (IntrData->Type == TERLOG_OP_MASKZ)
16344 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16346 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16347 Src1, Src2, Src3, Src4),
16348 Mask, PassThru, Subtarget, DAG);
16351 // FPclass intrinsics with mask
16352 SDValue Src1 = Op.getOperand(1);
16353 MVT VT = Src1.getSimpleValueType();
16354 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16355 SDValue Imm = Op.getOperand(2);
16356 SDValue Mask = Op.getOperand(3);
16357 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16358 Mask.getSimpleValueType().getSizeInBits());
16359 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16360 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16361 DAG.getTargetConstant(0, dl, MaskVT),
16363 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16364 DAG.getUNDEF(BitcastVT), FPclassMask,
16365 DAG.getIntPtrConstant(0, dl));
16366 return DAG.getBitcast(Op.getValueType(), Res);
16369 SDValue Src1 = Op.getOperand(1);
16370 SDValue Imm = Op.getOperand(2);
16371 SDValue Mask = Op.getOperand(3);
16372 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16373 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16374 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16375 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16378 case CMP_MASK_CC: {
16379 // Comparison intrinsics with masks.
16380 // Example of transformation:
16381 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16382 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16384 // (v8i1 (insert_subvector undef,
16385 // (v2i1 (and (PCMPEQM %a, %b),
16386 // (extract_subvector
16387 // (v8i1 (bitcast %mask)), 0))), 0))))
16388 MVT VT = Op.getOperand(1).getSimpleValueType();
16389 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16390 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16391 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16392 Mask.getSimpleValueType().getSizeInBits());
16394 if (IntrData->Type == CMP_MASK_CC) {
16395 SDValue CC = Op.getOperand(3);
16396 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16397 // We specify 2 possible opcodes for intrinsics with rounding modes.
16398 // First, we check if the intrinsic may have non-default rounding mode,
16399 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16400 if (IntrData->Opc1 != 0) {
16401 SDValue Rnd = Op.getOperand(5);
16402 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16403 X86::STATIC_ROUNDING::CUR_DIRECTION)
16404 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16405 Op.getOperand(2), CC, Rnd);
16407 //default rounding mode
16409 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16410 Op.getOperand(2), CC);
16413 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16414 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16417 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16418 DAG.getTargetConstant(0, dl,
16421 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16422 DAG.getUNDEF(BitcastVT), CmpMask,
16423 DAG.getIntPtrConstant(0, dl));
16424 return DAG.getBitcast(Op.getValueType(), Res);
16426 case CMP_MASK_SCALAR_CC: {
16427 SDValue Src1 = Op.getOperand(1);
16428 SDValue Src2 = Op.getOperand(2);
16429 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16430 SDValue Mask = Op.getOperand(4);
16433 if (IntrData->Opc1 != 0) {
16434 SDValue Rnd = Op.getOperand(5);
16435 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16436 X86::STATIC_ROUNDING::CUR_DIRECTION)
16437 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16439 //default rounding mode
16441 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16443 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16444 DAG.getTargetConstant(0, dl,
16448 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16449 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16450 DAG.getValueType(MVT::i1));
16452 case COMI: { // Comparison intrinsics
16453 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16454 SDValue LHS = Op.getOperand(1);
16455 SDValue RHS = Op.getOperand(2);
16456 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16457 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16458 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16459 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16460 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16461 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16464 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16465 Op.getOperand(1), Op.getOperand(2), DAG);
16467 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16468 Op.getSimpleValueType(),
16470 Op.getOperand(2), DAG),
16471 Op.getOperand(4), Op.getOperand(3), Subtarget,
16473 case COMPRESS_EXPAND_IN_REG: {
16474 SDValue Mask = Op.getOperand(3);
16475 SDValue DataToCompress = Op.getOperand(1);
16476 SDValue PassThru = Op.getOperand(2);
16477 if (isAllOnes(Mask)) // return data as is
16478 return Op.getOperand(1);
16480 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16482 Mask, PassThru, Subtarget, DAG);
16485 SDValue Mask = Op.getOperand(3);
16486 MVT VT = Op.getSimpleValueType();
16487 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16488 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16489 Mask.getSimpleValueType().getSizeInBits());
16491 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16492 DAG.getBitcast(BitcastVT, Mask),
16493 DAG.getIntPtrConstant(0, dl));
16494 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16503 default: return SDValue(); // Don't custom lower most intrinsics.
16505 case Intrinsic::x86_avx2_permd:
16506 case Intrinsic::x86_avx2_permps:
16507 // Operands intentionally swapped. Mask is last operand to intrinsic,
16508 // but second operand for node/instruction.
16509 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16510 Op.getOperand(2), Op.getOperand(1));
16512 // ptest and testp intrinsics. The intrinsic these come from are designed to
16513 // return an integer value, not just an instruction so lower it to the ptest
16514 // or testp pattern and a setcc for the result.
16515 case Intrinsic::x86_sse41_ptestz:
16516 case Intrinsic::x86_sse41_ptestc:
16517 case Intrinsic::x86_sse41_ptestnzc:
16518 case Intrinsic::x86_avx_ptestz_256:
16519 case Intrinsic::x86_avx_ptestc_256:
16520 case Intrinsic::x86_avx_ptestnzc_256:
16521 case Intrinsic::x86_avx_vtestz_ps:
16522 case Intrinsic::x86_avx_vtestc_ps:
16523 case Intrinsic::x86_avx_vtestnzc_ps:
16524 case Intrinsic::x86_avx_vtestz_pd:
16525 case Intrinsic::x86_avx_vtestc_pd:
16526 case Intrinsic::x86_avx_vtestnzc_pd:
16527 case Intrinsic::x86_avx_vtestz_ps_256:
16528 case Intrinsic::x86_avx_vtestc_ps_256:
16529 case Intrinsic::x86_avx_vtestnzc_ps_256:
16530 case Intrinsic::x86_avx_vtestz_pd_256:
16531 case Intrinsic::x86_avx_vtestc_pd_256:
16532 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16533 bool IsTestPacked = false;
16536 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16537 case Intrinsic::x86_avx_vtestz_ps:
16538 case Intrinsic::x86_avx_vtestz_pd:
16539 case Intrinsic::x86_avx_vtestz_ps_256:
16540 case Intrinsic::x86_avx_vtestz_pd_256:
16541 IsTestPacked = true; // Fallthrough
16542 case Intrinsic::x86_sse41_ptestz:
16543 case Intrinsic::x86_avx_ptestz_256:
16545 X86CC = X86::COND_E;
16547 case Intrinsic::x86_avx_vtestc_ps:
16548 case Intrinsic::x86_avx_vtestc_pd:
16549 case Intrinsic::x86_avx_vtestc_ps_256:
16550 case Intrinsic::x86_avx_vtestc_pd_256:
16551 IsTestPacked = true; // Fallthrough
16552 case Intrinsic::x86_sse41_ptestc:
16553 case Intrinsic::x86_avx_ptestc_256:
16555 X86CC = X86::COND_B;
16557 case Intrinsic::x86_avx_vtestnzc_ps:
16558 case Intrinsic::x86_avx_vtestnzc_pd:
16559 case Intrinsic::x86_avx_vtestnzc_ps_256:
16560 case Intrinsic::x86_avx_vtestnzc_pd_256:
16561 IsTestPacked = true; // Fallthrough
16562 case Intrinsic::x86_sse41_ptestnzc:
16563 case Intrinsic::x86_avx_ptestnzc_256:
16565 X86CC = X86::COND_A;
16569 SDValue LHS = Op.getOperand(1);
16570 SDValue RHS = Op.getOperand(2);
16571 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16572 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16573 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16574 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16575 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16577 case Intrinsic::x86_avx512_kortestz_w:
16578 case Intrinsic::x86_avx512_kortestc_w: {
16579 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16580 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16581 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16582 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16583 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16584 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16585 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16588 case Intrinsic::x86_sse42_pcmpistria128:
16589 case Intrinsic::x86_sse42_pcmpestria128:
16590 case Intrinsic::x86_sse42_pcmpistric128:
16591 case Intrinsic::x86_sse42_pcmpestric128:
16592 case Intrinsic::x86_sse42_pcmpistrio128:
16593 case Intrinsic::x86_sse42_pcmpestrio128:
16594 case Intrinsic::x86_sse42_pcmpistris128:
16595 case Intrinsic::x86_sse42_pcmpestris128:
16596 case Intrinsic::x86_sse42_pcmpistriz128:
16597 case Intrinsic::x86_sse42_pcmpestriz128: {
16601 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16602 case Intrinsic::x86_sse42_pcmpistria128:
16603 Opcode = X86ISD::PCMPISTRI;
16604 X86CC = X86::COND_A;
16606 case Intrinsic::x86_sse42_pcmpestria128:
16607 Opcode = X86ISD::PCMPESTRI;
16608 X86CC = X86::COND_A;
16610 case Intrinsic::x86_sse42_pcmpistric128:
16611 Opcode = X86ISD::PCMPISTRI;
16612 X86CC = X86::COND_B;
16614 case Intrinsic::x86_sse42_pcmpestric128:
16615 Opcode = X86ISD::PCMPESTRI;
16616 X86CC = X86::COND_B;
16618 case Intrinsic::x86_sse42_pcmpistrio128:
16619 Opcode = X86ISD::PCMPISTRI;
16620 X86CC = X86::COND_O;
16622 case Intrinsic::x86_sse42_pcmpestrio128:
16623 Opcode = X86ISD::PCMPESTRI;
16624 X86CC = X86::COND_O;
16626 case Intrinsic::x86_sse42_pcmpistris128:
16627 Opcode = X86ISD::PCMPISTRI;
16628 X86CC = X86::COND_S;
16630 case Intrinsic::x86_sse42_pcmpestris128:
16631 Opcode = X86ISD::PCMPESTRI;
16632 X86CC = X86::COND_S;
16634 case Intrinsic::x86_sse42_pcmpistriz128:
16635 Opcode = X86ISD::PCMPISTRI;
16636 X86CC = X86::COND_E;
16638 case Intrinsic::x86_sse42_pcmpestriz128:
16639 Opcode = X86ISD::PCMPESTRI;
16640 X86CC = X86::COND_E;
16643 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16644 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16645 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16646 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16647 DAG.getConstant(X86CC, dl, MVT::i8),
16648 SDValue(PCMP.getNode(), 1));
16649 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16652 case Intrinsic::x86_sse42_pcmpistri128:
16653 case Intrinsic::x86_sse42_pcmpestri128: {
16655 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16656 Opcode = X86ISD::PCMPISTRI;
16658 Opcode = X86ISD::PCMPESTRI;
16660 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16661 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16662 return DAG.getNode(Opcode, dl, VTs, NewOps);
16665 case Intrinsic::x86_seh_lsda: {
16666 // Compute the symbol for the LSDA. We know it'll get emitted later.
16667 MachineFunction &MF = DAG.getMachineFunction();
16668 SDValue Op1 = Op.getOperand(1);
16669 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16670 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16671 GlobalValue::getRealLinkageName(Fn->getName()));
16673 // Generate a simple absolute symbol reference. This intrinsic is only
16674 // supported on 32-bit Windows, which isn't PIC.
16675 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16676 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16679 case Intrinsic::x86_seh_recoverfp: {
16680 SDValue FnOp = Op.getOperand(1);
16681 SDValue IncomingFPOp = Op.getOperand(2);
16682 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16683 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16685 report_fatal_error(
16686 "llvm.x86.seh.recoverfp must take a function as the first argument");
16687 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16690 case Intrinsic::localaddress: {
16691 // Returns one of the stack, base, or frame pointer registers, depending on
16692 // which is used to reference local variables.
16693 MachineFunction &MF = DAG.getMachineFunction();
16694 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16696 if (RegInfo->hasBasePointer(MF))
16697 Reg = RegInfo->getBaseRegister();
16698 else // This function handles the SP or FP case.
16699 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16700 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16705 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16706 SDValue Src, SDValue Mask, SDValue Base,
16707 SDValue Index, SDValue ScaleOp, SDValue Chain,
16708 const X86Subtarget * Subtarget) {
16710 auto *C = cast<ConstantSDNode>(ScaleOp);
16711 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16712 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16713 Index.getSimpleValueType().getVectorNumElements());
16715 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16717 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16719 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16720 Mask.getSimpleValueType().getSizeInBits());
16722 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16723 // are extracted by EXTRACT_SUBVECTOR.
16724 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16725 DAG.getBitcast(BitcastVT, Mask),
16726 DAG.getIntPtrConstant(0, dl));
16728 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16729 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16730 SDValue Segment = DAG.getRegister(0, MVT::i32);
16731 if (Src.getOpcode() == ISD::UNDEF)
16732 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16733 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16734 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16735 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16736 return DAG.getMergeValues(RetOps, dl);
16739 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16740 SDValue Src, SDValue Mask, SDValue Base,
16741 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16743 auto *C = cast<ConstantSDNode>(ScaleOp);
16744 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16745 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16746 SDValue Segment = DAG.getRegister(0, MVT::i32);
16747 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16748 Index.getSimpleValueType().getVectorNumElements());
16750 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16752 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16754 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16755 Mask.getSimpleValueType().getSizeInBits());
16757 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16758 // are extracted by EXTRACT_SUBVECTOR.
16759 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16760 DAG.getBitcast(BitcastVT, Mask),
16761 DAG.getIntPtrConstant(0, dl));
16763 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16764 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16765 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16766 return SDValue(Res, 1);
16769 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16770 SDValue Mask, SDValue Base, SDValue Index,
16771 SDValue ScaleOp, SDValue Chain) {
16773 auto *C = cast<ConstantSDNode>(ScaleOp);
16774 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16775 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16776 SDValue Segment = DAG.getRegister(0, MVT::i32);
16778 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16780 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16782 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16784 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16785 //SDVTList VTs = DAG.getVTList(MVT::Other);
16786 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16787 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16788 return SDValue(Res, 0);
16791 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16792 // read performance monitor counters (x86_rdpmc).
16793 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16794 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16795 SmallVectorImpl<SDValue> &Results) {
16796 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16797 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16800 // The ECX register is used to select the index of the performance counter
16802 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16804 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16806 // Reads the content of a 64-bit performance counter and returns it in the
16807 // registers EDX:EAX.
16808 if (Subtarget->is64Bit()) {
16809 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16810 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16813 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16814 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16817 Chain = HI.getValue(1);
16819 if (Subtarget->is64Bit()) {
16820 // The EAX register is loaded with the low-order 32 bits. The EDX register
16821 // is loaded with the supported high-order bits of the counter.
16822 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16823 DAG.getConstant(32, DL, MVT::i8));
16824 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16825 Results.push_back(Chain);
16829 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16830 SDValue Ops[] = { LO, HI };
16831 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16832 Results.push_back(Pair);
16833 Results.push_back(Chain);
16836 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16837 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16838 // also used to custom lower READCYCLECOUNTER nodes.
16839 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16840 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16841 SmallVectorImpl<SDValue> &Results) {
16842 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16843 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16846 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16847 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16848 // and the EAX register is loaded with the low-order 32 bits.
16849 if (Subtarget->is64Bit()) {
16850 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16851 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16854 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16855 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16858 SDValue Chain = HI.getValue(1);
16860 if (Opcode == X86ISD::RDTSCP_DAG) {
16861 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16863 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16864 // the ECX register. Add 'ecx' explicitly to the chain.
16865 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16867 // Explicitly store the content of ECX at the location passed in input
16868 // to the 'rdtscp' intrinsic.
16869 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16870 MachinePointerInfo(), false, false, 0);
16873 if (Subtarget->is64Bit()) {
16874 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16875 // the EAX register is loaded with the low-order 32 bits.
16876 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16877 DAG.getConstant(32, DL, MVT::i8));
16878 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16879 Results.push_back(Chain);
16883 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16884 SDValue Ops[] = { LO, HI };
16885 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16886 Results.push_back(Pair);
16887 Results.push_back(Chain);
16890 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16891 SelectionDAG &DAG) {
16892 SmallVector<SDValue, 2> Results;
16894 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16896 return DAG.getMergeValues(Results, DL);
16899 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16900 SelectionDAG &DAG) {
16901 MachineFunction &MF = DAG.getMachineFunction();
16902 const Function *Fn = MF.getFunction();
16904 SDValue Chain = Op.getOperand(0);
16906 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16907 "using llvm.x86.seh.restoreframe requires a frame pointer");
16909 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16910 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16912 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16913 unsigned FrameReg =
16914 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16915 unsigned SPReg = RegInfo->getStackRegister();
16916 unsigned SlotSize = RegInfo->getSlotSize();
16918 // Get incoming EBP.
16919 SDValue IncomingEBP =
16920 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16922 // SP is saved in the first field of every registration node, so load
16923 // [EBP-RegNodeSize] into SP.
16924 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16925 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16926 DAG.getConstant(-RegNodeSize, dl, VT));
16928 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16929 false, VT.getScalarSizeInBits() / 8);
16930 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16932 if (!RegInfo->needsStackRealignment(MF)) {
16933 // Adjust EBP to point back to the original frame position.
16934 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16935 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16937 assert(RegInfo->hasBasePointer(MF) &&
16938 "functions with Win32 EH must use frame or base pointer register");
16940 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16941 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16942 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16944 // Reload the spilled EBP value, now that the stack and base pointers are
16946 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16947 X86FI->setHasSEHFramePtrSave(true);
16948 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16949 X86FI->setSEHFramePtrSaveIndex(FI);
16950 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16951 MachinePointerInfo(), false, false, false,
16952 VT.getScalarSizeInBits() / 8);
16953 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16959 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16960 /// return truncate Store/MaskedStore Node
16961 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16965 SDValue Mask = Op.getOperand(4);
16966 SDValue DataToTruncate = Op.getOperand(3);
16967 SDValue Addr = Op.getOperand(2);
16968 SDValue Chain = Op.getOperand(0);
16970 MVT VT = DataToTruncate.getSimpleValueType();
16971 MVT SVT = MVT::getVectorVT(ElementType, VT.getVectorNumElements());
16973 if (isAllOnes(Mask)) // return just a truncate store
16974 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16975 MachinePointerInfo(), SVT, false, false,
16976 SVT.getScalarSizeInBits()/8);
16978 MVT MaskVT = MVT::getVectorVT(MVT::i1, VT.getVectorNumElements());
16979 MVT BitcastVT = MVT::getVectorVT(MVT::i1,
16980 Mask.getSimpleValueType().getSizeInBits());
16981 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16982 // are extracted by EXTRACT_SUBVECTOR.
16983 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16984 DAG.getBitcast(BitcastVT, Mask),
16985 DAG.getIntPtrConstant(0, dl));
16987 MachineMemOperand *MMO = DAG.getMachineFunction().
16988 getMachineMemOperand(MachinePointerInfo(),
16989 MachineMemOperand::MOStore, SVT.getStoreSize(),
16990 SVT.getScalarSizeInBits()/8);
16992 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16993 VMask, SVT, MMO, true);
16996 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16997 SelectionDAG &DAG) {
16998 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17000 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17002 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
17003 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
17008 switch(IntrData->Type) {
17009 default: llvm_unreachable("Unknown Intrinsic Type");
17012 // Emit the node with the right value type.
17013 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17014 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17016 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17017 // Otherwise return the value from Rand, which is always 0, casted to i32.
17018 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17019 DAG.getConstant(1, dl, Op->getValueType(1)),
17020 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17021 SDValue(Result.getNode(), 1) };
17022 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17023 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17026 // Return { result, isValid, chain }.
17027 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17028 SDValue(Result.getNode(), 2));
17031 //gather(v1, mask, index, base, scale);
17032 SDValue Chain = Op.getOperand(0);
17033 SDValue Src = Op.getOperand(2);
17034 SDValue Base = Op.getOperand(3);
17035 SDValue Index = Op.getOperand(4);
17036 SDValue Mask = Op.getOperand(5);
17037 SDValue Scale = Op.getOperand(6);
17038 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17042 //scatter(base, mask, index, v1, scale);
17043 SDValue Chain = Op.getOperand(0);
17044 SDValue Base = Op.getOperand(2);
17045 SDValue Mask = Op.getOperand(3);
17046 SDValue Index = Op.getOperand(4);
17047 SDValue Src = Op.getOperand(5);
17048 SDValue Scale = Op.getOperand(6);
17049 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17053 SDValue Hint = Op.getOperand(6);
17054 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17055 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17056 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17057 SDValue Chain = Op.getOperand(0);
17058 SDValue Mask = Op.getOperand(2);
17059 SDValue Index = Op.getOperand(3);
17060 SDValue Base = Op.getOperand(4);
17061 SDValue Scale = Op.getOperand(5);
17062 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17064 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17066 SmallVector<SDValue, 2> Results;
17067 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17069 return DAG.getMergeValues(Results, dl);
17071 // Read Performance Monitoring Counters.
17073 SmallVector<SDValue, 2> Results;
17074 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17075 return DAG.getMergeValues(Results, dl);
17077 // XTEST intrinsics.
17079 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17080 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17081 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17082 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17084 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17085 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17086 Ret, SDValue(InTrans.getNode(), 1));
17090 SmallVector<SDValue, 2> Results;
17091 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17092 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17093 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17094 DAG.getConstant(-1, dl, MVT::i8));
17095 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17096 Op.getOperand(4), GenCF.getValue(1));
17097 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17098 Op.getOperand(5), MachinePointerInfo(),
17100 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17101 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17103 Results.push_back(SetCC);
17104 Results.push_back(Store);
17105 return DAG.getMergeValues(Results, dl);
17107 case COMPRESS_TO_MEM: {
17109 SDValue Mask = Op.getOperand(4);
17110 SDValue DataToCompress = Op.getOperand(3);
17111 SDValue Addr = Op.getOperand(2);
17112 SDValue Chain = Op.getOperand(0);
17114 MVT VT = DataToCompress.getSimpleValueType();
17115 if (isAllOnes(Mask)) // return just a store
17116 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17117 MachinePointerInfo(), false, false,
17118 VT.getScalarSizeInBits()/8);
17120 SDValue Compressed =
17121 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17122 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17123 return DAG.getStore(Chain, dl, Compressed, Addr,
17124 MachinePointerInfo(), false, false,
17125 VT.getScalarSizeInBits()/8);
17127 case TRUNCATE_TO_MEM_VI8:
17128 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17129 case TRUNCATE_TO_MEM_VI16:
17130 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17131 case TRUNCATE_TO_MEM_VI32:
17132 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17133 case EXPAND_FROM_MEM: {
17135 SDValue Mask = Op.getOperand(4);
17136 SDValue PassThru = Op.getOperand(3);
17137 SDValue Addr = Op.getOperand(2);
17138 SDValue Chain = Op.getOperand(0);
17139 MVT VT = Op.getSimpleValueType();
17141 if (isAllOnes(Mask)) // return just a load
17142 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17143 false, VT.getScalarSizeInBits()/8);
17145 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17146 false, false, false,
17147 VT.getScalarSizeInBits()/8);
17149 SDValue Results[] = {
17150 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17151 Mask, PassThru, Subtarget, DAG), Chain};
17152 return DAG.getMergeValues(Results, dl);
17157 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17158 SelectionDAG &DAG) const {
17159 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17160 MFI->setReturnAddressIsTaken(true);
17162 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17165 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17167 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17170 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17171 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17172 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17173 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17174 DAG.getNode(ISD::ADD, dl, PtrVT,
17175 FrameAddr, Offset),
17176 MachinePointerInfo(), false, false, false, 0);
17179 // Just load the return address.
17180 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17181 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17182 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17185 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17186 MachineFunction &MF = DAG.getMachineFunction();
17187 MachineFrameInfo *MFI = MF.getFrameInfo();
17188 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17189 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17190 EVT VT = Op.getValueType();
17192 MFI->setFrameAddressIsTaken(true);
17194 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17195 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17196 // is not possible to crawl up the stack without looking at the unwind codes
17198 int FrameAddrIndex = FuncInfo->getFAIndex();
17199 if (!FrameAddrIndex) {
17200 // Set up a frame object for the return address.
17201 unsigned SlotSize = RegInfo->getSlotSize();
17202 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17203 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17204 FuncInfo->setFAIndex(FrameAddrIndex);
17206 return DAG.getFrameIndex(FrameAddrIndex, VT);
17209 unsigned FrameReg =
17210 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17211 SDLoc dl(Op); // FIXME probably not meaningful
17212 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17213 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17214 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17215 "Invalid Frame Register!");
17216 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17218 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17219 MachinePointerInfo(),
17220 false, false, false, 0);
17224 // FIXME? Maybe this could be a TableGen attribute on some registers and
17225 // this table could be generated automatically from RegInfo.
17226 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17227 SelectionDAG &DAG) const {
17228 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17229 const MachineFunction &MF = DAG.getMachineFunction();
17231 unsigned Reg = StringSwitch<unsigned>(RegName)
17232 .Case("esp", X86::ESP)
17233 .Case("rsp", X86::RSP)
17234 .Case("ebp", X86::EBP)
17235 .Case("rbp", X86::RBP)
17238 if (Reg == X86::EBP || Reg == X86::RBP) {
17239 if (!TFI.hasFP(MF))
17240 report_fatal_error("register " + StringRef(RegName) +
17241 " is allocatable: function has no frame pointer");
17244 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17245 unsigned FrameReg =
17246 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17247 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17248 "Invalid Frame Register!");
17256 report_fatal_error("Invalid register name global variable");
17259 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17260 SelectionDAG &DAG) const {
17261 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17262 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17265 unsigned X86TargetLowering::getExceptionPointerRegister(
17266 const Constant *PersonalityFn) const {
17267 if (classifyEHPersonality(PersonalityFn) == EHPersonality::CoreCLR)
17268 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17270 return Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
17273 unsigned X86TargetLowering::getExceptionSelectorRegister(
17274 const Constant *PersonalityFn) const {
17275 // Funclet personalities don't use selectors (the runtime does the selection).
17276 assert(!isFuncletEHPersonality(classifyEHPersonality(PersonalityFn)));
17277 return Subtarget->isTarget64BitLP64() ? X86::RDX : X86::EDX;
17280 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17281 SDValue Chain = Op.getOperand(0);
17282 SDValue Offset = Op.getOperand(1);
17283 SDValue Handler = Op.getOperand(2);
17286 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17287 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17288 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17289 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17290 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17291 "Invalid Frame Register!");
17292 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17293 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17295 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17296 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17298 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17299 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17301 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17303 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17304 DAG.getRegister(StoreAddrReg, PtrVT));
17307 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17308 SelectionDAG &DAG) const {
17310 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17311 DAG.getVTList(MVT::i32, MVT::Other),
17312 Op.getOperand(0), Op.getOperand(1));
17315 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17316 SelectionDAG &DAG) const {
17318 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17319 Op.getOperand(0), Op.getOperand(1));
17322 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17323 return Op.getOperand(0);
17326 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17327 SelectionDAG &DAG) const {
17328 SDValue Root = Op.getOperand(0);
17329 SDValue Trmp = Op.getOperand(1); // trampoline
17330 SDValue FPtr = Op.getOperand(2); // nested function
17331 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17334 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17335 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17337 if (Subtarget->is64Bit()) {
17338 SDValue OutChains[6];
17340 // Large code-model.
17341 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17342 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17344 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17345 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17347 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17349 // Load the pointer to the nested function into R11.
17350 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17351 SDValue Addr = Trmp;
17352 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17353 Addr, MachinePointerInfo(TrmpAddr),
17356 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17357 DAG.getConstant(2, dl, MVT::i64));
17358 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17359 MachinePointerInfo(TrmpAddr, 2),
17362 // Load the 'nest' parameter value into R10.
17363 // R10 is specified in X86CallingConv.td
17364 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17365 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17366 DAG.getConstant(10, dl, MVT::i64));
17367 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17368 Addr, MachinePointerInfo(TrmpAddr, 10),
17371 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17372 DAG.getConstant(12, dl, MVT::i64));
17373 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17374 MachinePointerInfo(TrmpAddr, 12),
17377 // Jump to the nested function.
17378 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17379 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17380 DAG.getConstant(20, dl, MVT::i64));
17381 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17382 Addr, MachinePointerInfo(TrmpAddr, 20),
17385 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17386 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17387 DAG.getConstant(22, dl, MVT::i64));
17388 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17389 Addr, MachinePointerInfo(TrmpAddr, 22),
17392 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17394 const Function *Func =
17395 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17396 CallingConv::ID CC = Func->getCallingConv();
17401 llvm_unreachable("Unsupported calling convention");
17402 case CallingConv::C:
17403 case CallingConv::X86_StdCall: {
17404 // Pass 'nest' parameter in ECX.
17405 // Must be kept in sync with X86CallingConv.td
17406 NestReg = X86::ECX;
17408 // Check that ECX wasn't needed by an 'inreg' parameter.
17409 FunctionType *FTy = Func->getFunctionType();
17410 const AttributeSet &Attrs = Func->getAttributes();
17412 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17413 unsigned InRegCount = 0;
17416 for (FunctionType::param_iterator I = FTy->param_begin(),
17417 E = FTy->param_end(); I != E; ++I, ++Idx)
17418 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17419 auto &DL = DAG.getDataLayout();
17420 // FIXME: should only count parameters that are lowered to integers.
17421 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17424 if (InRegCount > 2) {
17425 report_fatal_error("Nest register in use - reduce number of inreg"
17431 case CallingConv::X86_FastCall:
17432 case CallingConv::X86_ThisCall:
17433 case CallingConv::Fast:
17434 // Pass 'nest' parameter in EAX.
17435 // Must be kept in sync with X86CallingConv.td
17436 NestReg = X86::EAX;
17440 SDValue OutChains[4];
17441 SDValue Addr, Disp;
17443 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17444 DAG.getConstant(10, dl, MVT::i32));
17445 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17447 // This is storing the opcode for MOV32ri.
17448 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17449 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17450 OutChains[0] = DAG.getStore(Root, dl,
17451 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17452 Trmp, MachinePointerInfo(TrmpAddr),
17455 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17456 DAG.getConstant(1, dl, MVT::i32));
17457 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17458 MachinePointerInfo(TrmpAddr, 1),
17461 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17462 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17463 DAG.getConstant(5, dl, MVT::i32));
17464 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17465 Addr, MachinePointerInfo(TrmpAddr, 5),
17468 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17469 DAG.getConstant(6, dl, MVT::i32));
17470 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17471 MachinePointerInfo(TrmpAddr, 6),
17474 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17478 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17479 SelectionDAG &DAG) const {
17481 The rounding mode is in bits 11:10 of FPSR, and has the following
17483 00 Round to nearest
17488 FLT_ROUNDS, on the other hand, expects the following:
17495 To perform the conversion, we do:
17496 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17499 MachineFunction &MF = DAG.getMachineFunction();
17500 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17501 unsigned StackAlignment = TFI.getStackAlignment();
17502 MVT VT = Op.getSimpleValueType();
17505 // Save FP Control Word to stack slot
17506 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17507 SDValue StackSlot =
17508 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17510 MachineMemOperand *MMO =
17511 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17512 MachineMemOperand::MOStore, 2, 2);
17514 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17515 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17516 DAG.getVTList(MVT::Other),
17517 Ops, MVT::i16, MMO);
17519 // Load FP Control Word from stack slot
17520 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17521 MachinePointerInfo(), false, false, false, 0);
17523 // Transform as necessary
17525 DAG.getNode(ISD::SRL, DL, MVT::i16,
17526 DAG.getNode(ISD::AND, DL, MVT::i16,
17527 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17528 DAG.getConstant(11, DL, MVT::i8));
17530 DAG.getNode(ISD::SRL, DL, MVT::i16,
17531 DAG.getNode(ISD::AND, DL, MVT::i16,
17532 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17533 DAG.getConstant(9, DL, MVT::i8));
17536 DAG.getNode(ISD::AND, DL, MVT::i16,
17537 DAG.getNode(ISD::ADD, DL, MVT::i16,
17538 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17539 DAG.getConstant(1, DL, MVT::i16)),
17540 DAG.getConstant(3, DL, MVT::i16));
17542 return DAG.getNode((VT.getSizeInBits() < 16 ?
17543 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17546 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17548 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17549 // to 512-bit vector.
17550 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17551 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17552 // split the vector, perform operation on it's Lo a Hi part and
17553 // concatenate the results.
17554 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17556 MVT VT = Op.getSimpleValueType();
17557 MVT EltVT = VT.getVectorElementType();
17558 unsigned NumElems = VT.getVectorNumElements();
17560 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17561 // Extend to 512 bit vector.
17562 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17563 "Unsupported value type for operation");
17565 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17566 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17567 DAG.getUNDEF(NewVT),
17569 DAG.getIntPtrConstant(0, dl));
17570 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17572 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17573 DAG.getIntPtrConstant(0, dl));
17576 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17577 "Unsupported element type");
17579 if (16 < NumElems) {
17580 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17582 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17583 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17585 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17586 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17588 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17591 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17593 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17594 "Unsupported value type for operation");
17596 // Use native supported vector instruction vplzcntd.
17597 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17598 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17599 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17600 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17602 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17605 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17606 SelectionDAG &DAG) {
17607 MVT VT = Op.getSimpleValueType();
17609 unsigned NumBits = VT.getSizeInBits();
17612 if (VT.isVector() && Subtarget->hasAVX512())
17613 return LowerVectorCTLZ_AVX512(Op, DAG);
17615 Op = Op.getOperand(0);
17616 if (VT == MVT::i8) {
17617 // Zero extend to i32 since there is not an i8 bsr.
17619 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17622 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17623 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17624 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17626 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17629 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17630 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17633 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17635 // Finally xor with NumBits-1.
17636 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17637 DAG.getConstant(NumBits - 1, dl, OpVT));
17640 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17644 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17645 SelectionDAG &DAG) {
17646 MVT VT = Op.getSimpleValueType();
17648 unsigned NumBits = VT.getSizeInBits();
17651 if (VT.isVector() && Subtarget->hasAVX512())
17652 return LowerVectorCTLZ_AVX512(Op, DAG);
17654 Op = Op.getOperand(0);
17655 if (VT == MVT::i8) {
17656 // Zero extend to i32 since there is not an i8 bsr.
17658 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17661 // Issue a bsr (scan bits in reverse).
17662 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17663 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17665 // And xor with NumBits-1.
17666 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17667 DAG.getConstant(NumBits - 1, dl, OpVT));
17670 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17674 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17675 MVT VT = Op.getSimpleValueType();
17676 unsigned NumBits = VT.getScalarSizeInBits();
17679 if (VT.isVector()) {
17680 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17682 SDValue N0 = Op.getOperand(0);
17683 SDValue Zero = DAG.getConstant(0, dl, VT);
17685 // lsb(x) = (x & -x)
17686 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17687 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17689 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17690 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17691 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17692 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17693 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17694 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17697 // cttz(x) = ctpop(lsb - 1)
17698 SDValue One = DAG.getConstant(1, dl, VT);
17699 return DAG.getNode(ISD::CTPOP, dl, VT,
17700 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17703 assert(Op.getOpcode() == ISD::CTTZ &&
17704 "Only scalar CTTZ requires custom lowering");
17706 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17707 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17708 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17710 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17713 DAG.getConstant(NumBits, dl, VT),
17714 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17717 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17720 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17721 // ones, and then concatenate the result back.
17722 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17723 MVT VT = Op.getSimpleValueType();
17725 assert(VT.is256BitVector() && VT.isInteger() &&
17726 "Unsupported value type for operation");
17728 unsigned NumElems = VT.getVectorNumElements();
17731 // Extract the LHS vectors
17732 SDValue LHS = Op.getOperand(0);
17733 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17734 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17736 // Extract the RHS vectors
17737 SDValue RHS = Op.getOperand(1);
17738 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17739 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17741 MVT EltVT = VT.getVectorElementType();
17742 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17744 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17745 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17746 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17749 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17750 if (Op.getValueType() == MVT::i1)
17751 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17752 Op.getOperand(0), Op.getOperand(1));
17753 assert(Op.getSimpleValueType().is256BitVector() &&
17754 Op.getSimpleValueType().isInteger() &&
17755 "Only handle AVX 256-bit vector integer operation");
17756 return Lower256IntArith(Op, DAG);
17759 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17760 if (Op.getValueType() == MVT::i1)
17761 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17762 Op.getOperand(0), Op.getOperand(1));
17763 assert(Op.getSimpleValueType().is256BitVector() &&
17764 Op.getSimpleValueType().isInteger() &&
17765 "Only handle AVX 256-bit vector integer operation");
17766 return Lower256IntArith(Op, DAG);
17769 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17770 assert(Op.getSimpleValueType().is256BitVector() &&
17771 Op.getSimpleValueType().isInteger() &&
17772 "Only handle AVX 256-bit vector integer operation");
17773 return Lower256IntArith(Op, DAG);
17776 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17777 SelectionDAG &DAG) {
17779 MVT VT = Op.getSimpleValueType();
17782 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17784 // Decompose 256-bit ops into smaller 128-bit ops.
17785 if (VT.is256BitVector() && !Subtarget->hasInt256())
17786 return Lower256IntArith(Op, DAG);
17788 SDValue A = Op.getOperand(0);
17789 SDValue B = Op.getOperand(1);
17791 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17792 // pairs, multiply and truncate.
17793 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17794 if (Subtarget->hasInt256()) {
17795 if (VT == MVT::v32i8) {
17796 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17797 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17798 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17799 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17800 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17801 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17802 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17803 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17804 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17805 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17808 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17809 return DAG.getNode(
17810 ISD::TRUNCATE, dl, VT,
17811 DAG.getNode(ISD::MUL, dl, ExVT,
17812 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17813 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17816 assert(VT == MVT::v16i8 &&
17817 "Pre-AVX2 support only supports v16i8 multiplication");
17818 MVT ExVT = MVT::v8i16;
17820 // Extract the lo parts and sign extend to i16
17822 if (Subtarget->hasSSE41()) {
17823 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17824 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17826 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17827 -1, 4, -1, 5, -1, 6, -1, 7};
17828 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17829 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17830 ALo = DAG.getBitcast(ExVT, ALo);
17831 BLo = DAG.getBitcast(ExVT, BLo);
17832 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17833 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17836 // Extract the hi parts and sign extend to i16
17838 if (Subtarget->hasSSE41()) {
17839 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17840 -1, -1, -1, -1, -1, -1, -1, -1};
17841 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17842 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17843 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17844 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17846 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17847 -1, 12, -1, 13, -1, 14, -1, 15};
17848 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17849 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17850 AHi = DAG.getBitcast(ExVT, AHi);
17851 BHi = DAG.getBitcast(ExVT, BHi);
17852 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17853 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17856 // Multiply, mask the lower 8bits of the lo/hi results and pack
17857 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17858 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17859 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17860 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17861 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17864 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17865 if (VT == MVT::v4i32) {
17866 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17867 "Should not custom lower when pmuldq is available!");
17869 // Extract the odd parts.
17870 static const int UnpackMask[] = { 1, -1, 3, -1 };
17871 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17872 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17874 // Multiply the even parts.
17875 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17876 // Now multiply odd parts.
17877 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17879 Evens = DAG.getBitcast(VT, Evens);
17880 Odds = DAG.getBitcast(VT, Odds);
17882 // Merge the two vectors back together with a shuffle. This expands into 2
17884 static const int ShufMask[] = { 0, 4, 2, 6 };
17885 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17888 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17889 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17891 // Ahi = psrlqi(a, 32);
17892 // Bhi = psrlqi(b, 32);
17894 // AloBlo = pmuludq(a, b);
17895 // AloBhi = pmuludq(a, Bhi);
17896 // AhiBlo = pmuludq(Ahi, b);
17898 // AloBhi = psllqi(AloBhi, 32);
17899 // AhiBlo = psllqi(AhiBlo, 32);
17900 // return AloBlo + AloBhi + AhiBlo;
17902 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17903 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17905 SDValue AhiBlo = Ahi;
17906 SDValue AloBhi = Bhi;
17907 // Bit cast to 32-bit vectors for MULUDQ
17908 MVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17909 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17910 A = DAG.getBitcast(MulVT, A);
17911 B = DAG.getBitcast(MulVT, B);
17912 Ahi = DAG.getBitcast(MulVT, Ahi);
17913 Bhi = DAG.getBitcast(MulVT, Bhi);
17915 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17916 // After shifting right const values the result may be all-zero.
17917 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17918 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17919 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17921 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17922 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17923 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17926 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17927 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17930 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17931 assert(Subtarget->isTargetWin64() && "Unexpected target");
17932 EVT VT = Op.getValueType();
17933 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17934 "Unexpected return type for lowering");
17938 switch (Op->getOpcode()) {
17939 default: llvm_unreachable("Unexpected request for libcall!");
17940 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17941 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17942 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17943 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17944 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17945 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17949 SDValue InChain = DAG.getEntryNode();
17951 TargetLowering::ArgListTy Args;
17952 TargetLowering::ArgListEntry Entry;
17953 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17954 EVT ArgVT = Op->getOperand(i).getValueType();
17955 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17956 "Unexpected argument type for lowering");
17957 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17958 Entry.Node = StackPtr;
17959 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17961 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17962 Entry.Ty = PointerType::get(ArgTy,0);
17963 Entry.isSExt = false;
17964 Entry.isZExt = false;
17965 Args.push_back(Entry);
17968 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17969 getPointerTy(DAG.getDataLayout()));
17971 TargetLowering::CallLoweringInfo CLI(DAG);
17972 CLI.setDebugLoc(dl).setChain(InChain)
17973 .setCallee(getLibcallCallingConv(LC),
17974 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17975 Callee, std::move(Args), 0)
17976 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17978 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17979 return DAG.getBitcast(VT, CallInfo.first);
17982 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17983 SelectionDAG &DAG) {
17984 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17985 MVT VT = Op0.getSimpleValueType();
17988 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17989 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17991 // PMULxD operations multiply each even value (starting at 0) of LHS with
17992 // the related value of RHS and produce a widen result.
17993 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17994 // => <2 x i64> <ae|cg>
17996 // In other word, to have all the results, we need to perform two PMULxD:
17997 // 1. one with the even values.
17998 // 2. one with the odd values.
17999 // To achieve #2, with need to place the odd values at an even position.
18001 // Place the odd value at an even position (basically, shift all values 1
18002 // step to the left):
18003 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18004 // <a|b|c|d> => <b|undef|d|undef>
18005 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18006 // <e|f|g|h> => <f|undef|h|undef>
18007 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18009 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18011 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18012 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18014 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18015 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18016 // => <2 x i64> <ae|cg>
18017 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18018 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18019 // => <2 x i64> <bf|dh>
18020 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18022 // Shuffle it back into the right order.
18023 SDValue Highs, Lows;
18024 if (VT == MVT::v8i32) {
18025 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18026 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18027 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18028 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18030 const int HighMask[] = {1, 5, 3, 7};
18031 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18032 const int LowMask[] = {0, 4, 2, 6};
18033 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18036 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18037 // unsigned multiply.
18038 if (IsSigned && !Subtarget->hasSSE41()) {
18039 SDValue ShAmt = DAG.getConstant(
18041 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18042 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18043 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18044 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18045 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18047 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18048 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18051 // The first result of MUL_LOHI is actually the low value, followed by the
18053 SDValue Ops[] = {Lows, Highs};
18054 return DAG.getMergeValues(Ops, dl);
18057 // Return true if the required (according to Opcode) shift-imm form is natively
18058 // supported by the Subtarget
18059 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18061 if (VT.getScalarSizeInBits() < 16)
18064 if (VT.is512BitVector() &&
18065 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18068 bool LShift = VT.is128BitVector() ||
18069 (VT.is256BitVector() && Subtarget->hasInt256());
18071 bool AShift = LShift && (Subtarget->hasVLX() ||
18072 (VT != MVT::v2i64 && VT != MVT::v4i64));
18073 return (Opcode == ISD::SRA) ? AShift : LShift;
18076 // The shift amount is a variable, but it is the same for all vector lanes.
18077 // These instructions are defined together with shift-immediate.
18079 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18081 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18084 // Return true if the required (according to Opcode) variable-shift form is
18085 // natively supported by the Subtarget
18086 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18089 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18092 // vXi16 supported only on AVX-512, BWI
18093 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18096 if (VT.is512BitVector() || Subtarget->hasVLX())
18099 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18100 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18101 return (Opcode == ISD::SRA) ? AShift : LShift;
18104 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18105 const X86Subtarget *Subtarget) {
18106 MVT VT = Op.getSimpleValueType();
18108 SDValue R = Op.getOperand(0);
18109 SDValue Amt = Op.getOperand(1);
18111 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18112 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18114 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18115 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18116 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18117 SDValue Ex = DAG.getBitcast(ExVT, R);
18119 if (ShiftAmt >= 32) {
18120 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18122 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18123 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18124 ShiftAmt - 32, DAG);
18125 if (VT == MVT::v2i64)
18126 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18127 if (VT == MVT::v4i64)
18128 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18129 {9, 1, 11, 3, 13, 5, 15, 7});
18131 // SRA upper i32, SHL whole i64 and select lower i32.
18132 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18135 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18136 Lower = DAG.getBitcast(ExVT, Lower);
18137 if (VT == MVT::v2i64)
18138 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18139 if (VT == MVT::v4i64)
18140 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18141 {8, 1, 10, 3, 12, 5, 14, 7});
18143 return DAG.getBitcast(VT, Ex);
18146 // Optimize shl/srl/sra with constant shift amount.
18147 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18148 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18149 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18151 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18152 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18154 // i64 SRA needs to be performed as partial shifts.
18155 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18156 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18157 return ArithmeticShiftRight64(ShiftAmt);
18159 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18160 unsigned NumElts = VT.getVectorNumElements();
18161 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18163 // Simple i8 add case
18164 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18165 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18167 // ashr(R, 7) === cmp_slt(R, 0)
18168 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18169 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18170 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18173 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18174 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18177 if (Op.getOpcode() == ISD::SHL) {
18178 // Make a large shift.
18179 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18181 SHL = DAG.getBitcast(VT, SHL);
18182 // Zero out the rightmost bits.
18183 SmallVector<SDValue, 32> V(
18184 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18185 return DAG.getNode(ISD::AND, dl, VT, SHL,
18186 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18188 if (Op.getOpcode() == ISD::SRL) {
18189 // Make a large shift.
18190 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18192 SRL = DAG.getBitcast(VT, SRL);
18193 // Zero out the leftmost bits.
18194 SmallVector<SDValue, 32> V(
18195 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18196 return DAG.getNode(ISD::AND, dl, VT, SRL,
18197 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18199 if (Op.getOpcode() == ISD::SRA) {
18200 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18201 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18202 SmallVector<SDValue, 32> V(NumElts,
18203 DAG.getConstant(128 >> ShiftAmt, dl,
18205 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18206 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18207 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18210 llvm_unreachable("Unknown shift opcode.");
18215 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18216 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18217 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18219 // Peek through any splat that was introduced for i64 shift vectorization.
18220 int SplatIndex = -1;
18221 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18222 if (SVN->isSplat()) {
18223 SplatIndex = SVN->getSplatIndex();
18224 Amt = Amt.getOperand(0);
18225 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18226 "Splat shuffle referencing second operand");
18229 if (Amt.getOpcode() != ISD::BITCAST ||
18230 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18233 Amt = Amt.getOperand(0);
18234 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18235 VT.getVectorNumElements();
18236 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18237 uint64_t ShiftAmt = 0;
18238 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18239 for (unsigned i = 0; i != Ratio; ++i) {
18240 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18244 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18247 // Check remaining shift amounts (if not a splat).
18248 if (SplatIndex < 0) {
18249 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18250 uint64_t ShAmt = 0;
18251 for (unsigned j = 0; j != Ratio; ++j) {
18252 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18256 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18258 if (ShAmt != ShiftAmt)
18263 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18264 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18266 if (Op.getOpcode() == ISD::SRA)
18267 return ArithmeticShiftRight64(ShiftAmt);
18273 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18274 const X86Subtarget* Subtarget) {
18275 MVT VT = Op.getSimpleValueType();
18277 SDValue R = Op.getOperand(0);
18278 SDValue Amt = Op.getOperand(1);
18280 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18281 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18283 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18284 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18286 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18288 MVT EltVT = VT.getVectorElementType();
18290 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18291 // Check if this build_vector node is doing a splat.
18292 // If so, then set BaseShAmt equal to the splat value.
18293 BaseShAmt = BV->getSplatValue();
18294 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18295 BaseShAmt = SDValue();
18297 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18298 Amt = Amt.getOperand(0);
18300 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18301 if (SVN && SVN->isSplat()) {
18302 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18303 SDValue InVec = Amt.getOperand(0);
18304 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18305 assert((SplatIdx < InVec.getSimpleValueType().getVectorNumElements()) &&
18306 "Unexpected shuffle index found!");
18307 BaseShAmt = InVec.getOperand(SplatIdx);
18308 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18309 if (ConstantSDNode *C =
18310 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18311 if (C->getZExtValue() == SplatIdx)
18312 BaseShAmt = InVec.getOperand(1);
18317 // Avoid introducing an extract element from a shuffle.
18318 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18319 DAG.getIntPtrConstant(SplatIdx, dl));
18323 if (BaseShAmt.getNode()) {
18324 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18325 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18326 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18327 else if (EltVT.bitsLT(MVT::i32))
18328 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18330 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18334 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18335 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18336 Amt.getOpcode() == ISD::BITCAST &&
18337 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18338 Amt = Amt.getOperand(0);
18339 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18340 VT.getVectorNumElements();
18341 std::vector<SDValue> Vals(Ratio);
18342 for (unsigned i = 0; i != Ratio; ++i)
18343 Vals[i] = Amt.getOperand(i);
18344 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18345 for (unsigned j = 0; j != Ratio; ++j)
18346 if (Vals[j] != Amt.getOperand(i + j))
18350 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18351 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18356 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18357 SelectionDAG &DAG) {
18358 MVT VT = Op.getSimpleValueType();
18360 SDValue R = Op.getOperand(0);
18361 SDValue Amt = Op.getOperand(1);
18363 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18364 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18366 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18369 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18372 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18375 // XOP has 128-bit variable logical/arithmetic shifts.
18376 // +ve/-ve Amt = shift left/right.
18377 if (Subtarget->hasXOP() &&
18378 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18379 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18380 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18381 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18382 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18384 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18385 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18386 if (Op.getOpcode() == ISD::SRA)
18387 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18390 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18391 // shifts per-lane and then shuffle the partial results back together.
18392 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18393 // Splat the shift amounts so the scalar shifts above will catch it.
18394 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18395 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18396 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18397 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18398 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18401 // i64 vector arithmetic shift can be emulated with the transform:
18402 // M = lshr(SIGN_BIT, Amt)
18403 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18404 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18405 Op.getOpcode() == ISD::SRA) {
18406 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18407 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18408 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18409 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18410 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18414 // If possible, lower this packed shift into a vector multiply instead of
18415 // expanding it into a sequence of scalar shifts.
18416 // Do this only if the vector shift count is a constant build_vector.
18417 if (Op.getOpcode() == ISD::SHL &&
18418 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18419 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18420 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18421 SmallVector<SDValue, 8> Elts;
18422 MVT SVT = VT.getVectorElementType();
18423 unsigned SVTBits = SVT.getSizeInBits();
18424 APInt One(SVTBits, 1);
18425 unsigned NumElems = VT.getVectorNumElements();
18427 for (unsigned i=0; i !=NumElems; ++i) {
18428 SDValue Op = Amt->getOperand(i);
18429 if (Op->getOpcode() == ISD::UNDEF) {
18430 Elts.push_back(Op);
18434 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18435 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18436 uint64_t ShAmt = C.getZExtValue();
18437 if (ShAmt >= SVTBits) {
18438 Elts.push_back(DAG.getUNDEF(SVT));
18441 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18443 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18444 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18447 // Lower SHL with variable shift amount.
18448 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18449 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18451 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18452 DAG.getConstant(0x3f800000U, dl, VT));
18453 Op = DAG.getBitcast(MVT::v4f32, Op);
18454 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18455 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18458 // If possible, lower this shift as a sequence of two shifts by
18459 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18461 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18463 // Could be rewritten as:
18464 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18466 // The advantage is that the two shifts from the example would be
18467 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18468 // the vector shift into four scalar shifts plus four pairs of vector
18470 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18471 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18472 unsigned TargetOpcode = X86ISD::MOVSS;
18473 bool CanBeSimplified;
18474 // The splat value for the first packed shift (the 'X' from the example).
18475 SDValue Amt1 = Amt->getOperand(0);
18476 // The splat value for the second packed shift (the 'Y' from the example).
18477 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18478 Amt->getOperand(2);
18480 // See if it is possible to replace this node with a sequence of
18481 // two shifts followed by a MOVSS/MOVSD
18482 if (VT == MVT::v4i32) {
18483 // Check if it is legal to use a MOVSS.
18484 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18485 Amt2 == Amt->getOperand(3);
18486 if (!CanBeSimplified) {
18487 // Otherwise, check if we can still simplify this node using a MOVSD.
18488 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18489 Amt->getOperand(2) == Amt->getOperand(3);
18490 TargetOpcode = X86ISD::MOVSD;
18491 Amt2 = Amt->getOperand(2);
18494 // Do similar checks for the case where the machine value type
18496 CanBeSimplified = Amt1 == Amt->getOperand(1);
18497 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18498 CanBeSimplified = Amt2 == Amt->getOperand(i);
18500 if (!CanBeSimplified) {
18501 TargetOpcode = X86ISD::MOVSD;
18502 CanBeSimplified = true;
18503 Amt2 = Amt->getOperand(4);
18504 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18505 CanBeSimplified = Amt1 == Amt->getOperand(i);
18506 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18507 CanBeSimplified = Amt2 == Amt->getOperand(j);
18511 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18512 isa<ConstantSDNode>(Amt2)) {
18513 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18514 MVT CastVT = MVT::v4i32;
18516 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18517 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18519 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18520 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18521 if (TargetOpcode == X86ISD::MOVSD)
18522 CastVT = MVT::v2i64;
18523 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18524 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18525 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18527 return DAG.getBitcast(VT, Result);
18531 // v4i32 Non Uniform Shifts.
18532 // If the shift amount is constant we can shift each lane using the SSE2
18533 // immediate shifts, else we need to zero-extend each lane to the lower i64
18534 // and shift using the SSE2 variable shifts.
18535 // The separate results can then be blended together.
18536 if (VT == MVT::v4i32) {
18537 unsigned Opc = Op.getOpcode();
18538 SDValue Amt0, Amt1, Amt2, Amt3;
18539 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18540 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18541 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18542 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18543 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18545 // ISD::SHL is handled above but we include it here for completeness.
18548 llvm_unreachable("Unknown target vector shift node");
18550 Opc = X86ISD::VSHL;
18553 Opc = X86ISD::VSRL;
18556 Opc = X86ISD::VSRA;
18559 // The SSE2 shifts use the lower i64 as the same shift amount for
18560 // all lanes and the upper i64 is ignored. These shuffle masks
18561 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18562 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18563 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18564 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18565 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18566 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18569 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18570 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18571 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18572 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18573 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18574 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18575 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18578 if (VT == MVT::v16i8 ||
18579 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18580 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18581 unsigned ShiftOpcode = Op->getOpcode();
18583 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18584 // On SSE41 targets we make use of the fact that VSELECT lowers
18585 // to PBLENDVB which selects bytes based just on the sign bit.
18586 if (Subtarget->hasSSE41()) {
18587 V0 = DAG.getBitcast(VT, V0);
18588 V1 = DAG.getBitcast(VT, V1);
18589 Sel = DAG.getBitcast(VT, Sel);
18590 return DAG.getBitcast(SelVT,
18591 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18593 // On pre-SSE41 targets we test for the sign bit by comparing to
18594 // zero - a negative value will set all bits of the lanes to true
18595 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18596 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18597 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18598 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18601 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18602 // We can safely do this using i16 shifts as we're only interested in
18603 // the 3 lower bits of each byte.
18604 Amt = DAG.getBitcast(ExtVT, Amt);
18605 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18606 Amt = DAG.getBitcast(VT, Amt);
18608 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18609 // r = VSELECT(r, shift(r, 4), a);
18611 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18612 R = SignBitSelect(VT, Amt, M, R);
18615 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18617 // r = VSELECT(r, shift(r, 2), a);
18618 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18619 R = SignBitSelect(VT, Amt, M, R);
18622 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18624 // return VSELECT(r, shift(r, 1), a);
18625 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18626 R = SignBitSelect(VT, Amt, M, R);
18630 if (Op->getOpcode() == ISD::SRA) {
18631 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18632 // so we can correctly sign extend. We don't care what happens to the
18634 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18635 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18636 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18637 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18638 ALo = DAG.getBitcast(ExtVT, ALo);
18639 AHi = DAG.getBitcast(ExtVT, AHi);
18640 RLo = DAG.getBitcast(ExtVT, RLo);
18641 RHi = DAG.getBitcast(ExtVT, RHi);
18643 // r = VSELECT(r, shift(r, 4), a);
18644 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18645 DAG.getConstant(4, dl, ExtVT));
18646 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18647 DAG.getConstant(4, dl, ExtVT));
18648 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18649 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18652 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18653 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18655 // r = VSELECT(r, shift(r, 2), a);
18656 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18657 DAG.getConstant(2, dl, ExtVT));
18658 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18659 DAG.getConstant(2, dl, ExtVT));
18660 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18661 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18664 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18665 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18667 // r = VSELECT(r, shift(r, 1), a);
18668 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18669 DAG.getConstant(1, dl, ExtVT));
18670 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18671 DAG.getConstant(1, dl, ExtVT));
18672 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18673 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18675 // Logical shift the result back to the lower byte, leaving a zero upper
18677 // meaning that we can safely pack with PACKUSWB.
18679 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18681 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18682 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18686 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18687 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18688 // solution better.
18689 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18690 MVT ExtVT = MVT::v8i32;
18692 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18693 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18694 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18695 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18696 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18699 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18700 MVT ExtVT = MVT::v8i32;
18701 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18702 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18703 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18704 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18705 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18706 ALo = DAG.getBitcast(ExtVT, ALo);
18707 AHi = DAG.getBitcast(ExtVT, AHi);
18708 RLo = DAG.getBitcast(ExtVT, RLo);
18709 RHi = DAG.getBitcast(ExtVT, RHi);
18710 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18711 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18712 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18713 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18714 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18717 if (VT == MVT::v8i16) {
18718 unsigned ShiftOpcode = Op->getOpcode();
18720 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18721 // On SSE41 targets we make use of the fact that VSELECT lowers
18722 // to PBLENDVB which selects bytes based just on the sign bit.
18723 if (Subtarget->hasSSE41()) {
18724 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18725 V0 = DAG.getBitcast(ExtVT, V0);
18726 V1 = DAG.getBitcast(ExtVT, V1);
18727 Sel = DAG.getBitcast(ExtVT, Sel);
18728 return DAG.getBitcast(
18729 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18731 // On pre-SSE41 targets we splat the sign bit - a negative value will
18732 // set all bits of the lanes to true and VSELECT uses that in
18733 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18735 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18736 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18739 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18740 if (Subtarget->hasSSE41()) {
18741 // On SSE41 targets we need to replicate the shift mask in both
18742 // bytes for PBLENDVB.
18745 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18746 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18748 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18751 // r = VSELECT(r, shift(r, 8), a);
18752 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18753 R = SignBitSelect(Amt, M, R);
18756 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18758 // r = VSELECT(r, shift(r, 4), a);
18759 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18760 R = SignBitSelect(Amt, M, R);
18763 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18765 // r = VSELECT(r, shift(r, 2), a);
18766 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18767 R = SignBitSelect(Amt, M, R);
18770 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18772 // return VSELECT(r, shift(r, 1), a);
18773 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18774 R = SignBitSelect(Amt, M, R);
18778 // Decompose 256-bit shifts into smaller 128-bit shifts.
18779 if (VT.is256BitVector()) {
18780 unsigned NumElems = VT.getVectorNumElements();
18781 MVT EltVT = VT.getVectorElementType();
18782 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18784 // Extract the two vectors
18785 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18786 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18788 // Recreate the shift amount vectors
18789 SDValue Amt1, Amt2;
18790 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18791 // Constant shift amount
18792 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18793 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18794 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18796 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18797 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18799 // Variable shift amount
18800 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18801 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18804 // Issue new vector shifts for the smaller types
18805 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18806 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18808 // Concatenate the result back
18809 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18815 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18816 SelectionDAG &DAG) {
18817 MVT VT = Op.getSimpleValueType();
18819 SDValue R = Op.getOperand(0);
18820 SDValue Amt = Op.getOperand(1);
18822 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18823 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18824 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18826 // XOP has 128-bit vector variable + immediate rotates.
18827 // +ve/-ve Amt = rotate left/right.
18829 // Split 256-bit integers.
18830 if (VT.is256BitVector())
18831 return Lower256IntArith(Op, DAG);
18833 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
18835 // Attempt to rotate by immediate.
18836 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18837 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18838 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18839 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18840 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18841 DAG.getConstant(RotateAmt, DL, MVT::i8));
18845 // Use general rotate by variable (per-element).
18846 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18849 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18850 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18851 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18852 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18853 // has only one use.
18854 SDNode *N = Op.getNode();
18855 SDValue LHS = N->getOperand(0);
18856 SDValue RHS = N->getOperand(1);
18857 unsigned BaseOp = 0;
18860 switch (Op.getOpcode()) {
18861 default: llvm_unreachable("Unknown ovf instruction!");
18863 // A subtract of one will be selected as a INC. Note that INC doesn't
18864 // set CF, so we can't do this for UADDO.
18865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18867 BaseOp = X86ISD::INC;
18868 Cond = X86::COND_O;
18871 BaseOp = X86ISD::ADD;
18872 Cond = X86::COND_O;
18875 BaseOp = X86ISD::ADD;
18876 Cond = X86::COND_B;
18879 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18880 // set CF, so we can't do this for USUBO.
18881 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18883 BaseOp = X86ISD::DEC;
18884 Cond = X86::COND_O;
18887 BaseOp = X86ISD::SUB;
18888 Cond = X86::COND_O;
18891 BaseOp = X86ISD::SUB;
18892 Cond = X86::COND_B;
18895 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18896 Cond = X86::COND_O;
18898 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18899 if (N->getValueType(0) == MVT::i8) {
18900 BaseOp = X86ISD::UMUL8;
18901 Cond = X86::COND_O;
18904 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18906 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18909 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18910 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18911 SDValue(Sum.getNode(), 2));
18913 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18917 // Also sets EFLAGS.
18918 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18919 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18922 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18923 DAG.getConstant(Cond, DL, MVT::i32),
18924 SDValue(Sum.getNode(), 1));
18926 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18929 /// Returns true if the operand type is exactly twice the native width, and
18930 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18931 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18932 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18933 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18934 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18937 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18938 else if (OpWidth == 128)
18939 return Subtarget->hasCmpxchg16b();
18944 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18945 return needsCmpXchgNb(SI->getValueOperand()->getType());
18948 // Note: this turns large loads into lock cmpxchg8b/16b.
18949 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18950 TargetLowering::AtomicExpansionKind
18951 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18952 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18953 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18954 : AtomicExpansionKind::None;
18957 TargetLowering::AtomicExpansionKind
18958 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18959 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18960 Type *MemType = AI->getType();
18962 // If the operand is too big, we must see if cmpxchg8/16b is available
18963 // and default to library calls otherwise.
18964 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18965 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18966 : AtomicExpansionKind::None;
18969 AtomicRMWInst::BinOp Op = AI->getOperation();
18972 llvm_unreachable("Unknown atomic operation");
18973 case AtomicRMWInst::Xchg:
18974 case AtomicRMWInst::Add:
18975 case AtomicRMWInst::Sub:
18976 // It's better to use xadd, xsub or xchg for these in all cases.
18977 return AtomicExpansionKind::None;
18978 case AtomicRMWInst::Or:
18979 case AtomicRMWInst::And:
18980 case AtomicRMWInst::Xor:
18981 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18982 // prefix to a normal instruction for these operations.
18983 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18984 : AtomicExpansionKind::None;
18985 case AtomicRMWInst::Nand:
18986 case AtomicRMWInst::Max:
18987 case AtomicRMWInst::Min:
18988 case AtomicRMWInst::UMax:
18989 case AtomicRMWInst::UMin:
18990 // These always require a non-trivial set of data operations on x86. We must
18991 // use a cmpxchg loop.
18992 return AtomicExpansionKind::CmpXChg;
18996 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18997 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18998 // no-sse2). There isn't any reason to disable it if the target processor
19000 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19004 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19005 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19006 Type *MemType = AI->getType();
19007 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19008 // there is no benefit in turning such RMWs into loads, and it is actually
19009 // harmful as it introduces a mfence.
19010 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19013 auto Builder = IRBuilder<>(AI);
19014 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19015 auto SynchScope = AI->getSynchScope();
19016 // We must restrict the ordering to avoid generating loads with Release or
19017 // ReleaseAcquire orderings.
19018 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19019 auto Ptr = AI->getPointerOperand();
19021 // Before the load we need a fence. Here is an example lifted from
19022 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19025 // x.store(1, relaxed);
19026 // r1 = y.fetch_add(0, release);
19028 // y.fetch_add(42, acquire);
19029 // r2 = x.load(relaxed);
19030 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19031 // lowered to just a load without a fence. A mfence flushes the store buffer,
19032 // making the optimization clearly correct.
19033 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19034 // otherwise, we might be able to be more aggressive on relaxed idempotent
19035 // rmw. In practice, they do not look useful, so we don't try to be
19036 // especially clever.
19037 if (SynchScope == SingleThread)
19038 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19039 // the IR level, so we must wrap it in an intrinsic.
19042 if (!hasMFENCE(*Subtarget))
19043 // FIXME: it might make sense to use a locked operation here but on a
19044 // different cache-line to prevent cache-line bouncing. In practice it
19045 // is probably a small win, and x86 processors without mfence are rare
19046 // enough that we do not bother.
19050 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19051 Builder.CreateCall(MFence, {});
19053 // Finally we can emit the atomic load.
19054 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19055 AI->getType()->getPrimitiveSizeInBits());
19056 Loaded->setAtomic(Order, SynchScope);
19057 AI->replaceAllUsesWith(Loaded);
19058 AI->eraseFromParent();
19062 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19063 SelectionDAG &DAG) {
19065 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19066 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19067 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19068 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19070 // The only fence that needs an instruction is a sequentially-consistent
19071 // cross-thread fence.
19072 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19073 if (hasMFENCE(*Subtarget))
19074 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19076 SDValue Chain = Op.getOperand(0);
19077 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19079 DAG.getRegister(X86::ESP, MVT::i32), // Base
19080 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19081 DAG.getRegister(0, MVT::i32), // Index
19082 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19083 DAG.getRegister(0, MVT::i32), // Segment.
19087 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19088 return SDValue(Res, 0);
19091 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19092 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19095 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19096 SelectionDAG &DAG) {
19097 MVT T = Op.getSimpleValueType();
19101 switch(T.SimpleTy) {
19102 default: llvm_unreachable("Invalid value type!");
19103 case MVT::i8: Reg = X86::AL; size = 1; break;
19104 case MVT::i16: Reg = X86::AX; size = 2; break;
19105 case MVT::i32: Reg = X86::EAX; size = 4; break;
19107 assert(Subtarget->is64Bit() && "Node not type legal!");
19108 Reg = X86::RAX; size = 8;
19111 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19112 Op.getOperand(2), SDValue());
19113 SDValue Ops[] = { cpIn.getValue(0),
19116 DAG.getTargetConstant(size, DL, MVT::i8),
19117 cpIn.getValue(1) };
19118 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19119 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19120 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19124 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19125 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19126 MVT::i32, cpOut.getValue(2));
19127 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19128 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19131 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19132 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19133 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19137 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19138 SelectionDAG &DAG) {
19139 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19140 MVT DstVT = Op.getSimpleValueType();
19142 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19143 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19144 if (DstVT != MVT::f64)
19145 // This conversion needs to be expanded.
19148 SDValue InVec = Op->getOperand(0);
19150 unsigned NumElts = SrcVT.getVectorNumElements();
19151 MVT SVT = SrcVT.getVectorElementType();
19153 // Widen the vector in input in the case of MVT::v2i32.
19154 // Example: from MVT::v2i32 to MVT::v4i32.
19155 SmallVector<SDValue, 16> Elts;
19156 for (unsigned i = 0, e = NumElts; i != e; ++i)
19157 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19158 DAG.getIntPtrConstant(i, dl)));
19160 // Explicitly mark the extra elements as Undef.
19161 Elts.append(NumElts, DAG.getUNDEF(SVT));
19163 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19164 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19165 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19166 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19167 DAG.getIntPtrConstant(0, dl));
19170 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19171 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19172 assert((DstVT == MVT::i64 ||
19173 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19174 "Unexpected custom BITCAST");
19175 // i64 <=> MMX conversions are Legal.
19176 if (SrcVT==MVT::i64 && DstVT.isVector())
19178 if (DstVT==MVT::i64 && SrcVT.isVector())
19180 // MMX <=> MMX conversions are Legal.
19181 if (SrcVT.isVector() && DstVT.isVector())
19183 // All other conversions need to be expanded.
19187 /// Compute the horizontal sum of bytes in V for the elements of VT.
19189 /// Requires V to be a byte vector and VT to be an integer vector type with
19190 /// wider elements than V's type. The width of the elements of VT determines
19191 /// how many bytes of V are summed horizontally to produce each element of the
19193 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19194 const X86Subtarget *Subtarget,
19195 SelectionDAG &DAG) {
19197 MVT ByteVecVT = V.getSimpleValueType();
19198 MVT EltVT = VT.getVectorElementType();
19199 int NumElts = VT.getVectorNumElements();
19200 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19201 "Expected value to have byte element type.");
19202 assert(EltVT != MVT::i8 &&
19203 "Horizontal byte sum only makes sense for wider elements!");
19204 unsigned VecSize = VT.getSizeInBits();
19205 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19207 // PSADBW instruction horizontally add all bytes and leave the result in i64
19208 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19209 if (EltVT == MVT::i64) {
19210 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19211 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19212 return DAG.getBitcast(VT, V);
19215 if (EltVT == MVT::i32) {
19216 // We unpack the low half and high half into i32s interleaved with zeros so
19217 // that we can use PSADBW to horizontally sum them. The most useful part of
19218 // this is that it lines up the results of two PSADBW instructions to be
19219 // two v2i64 vectors which concatenated are the 4 population counts. We can
19220 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19221 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19222 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19223 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19225 // Do the horizontal sums into two v2i64s.
19226 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19227 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19228 DAG.getBitcast(ByteVecVT, Low), Zeros);
19229 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19230 DAG.getBitcast(ByteVecVT, High), Zeros);
19232 // Merge them together.
19233 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19234 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19235 DAG.getBitcast(ShortVecVT, Low),
19236 DAG.getBitcast(ShortVecVT, High));
19238 return DAG.getBitcast(VT, V);
19241 // The only element type left is i16.
19242 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19244 // To obtain pop count for each i16 element starting from the pop count for
19245 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19246 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19247 // directly supported.
19248 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19249 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19250 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19251 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19252 DAG.getBitcast(ByteVecVT, V));
19253 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19256 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19257 const X86Subtarget *Subtarget,
19258 SelectionDAG &DAG) {
19259 MVT VT = Op.getSimpleValueType();
19260 MVT EltVT = VT.getVectorElementType();
19261 unsigned VecSize = VT.getSizeInBits();
19263 // Implement a lookup table in register by using an algorithm based on:
19264 // http://wm.ite.pl/articles/sse-popcount.html
19266 // The general idea is that every lower byte nibble in the input vector is an
19267 // index into a in-register pre-computed pop count table. We then split up the
19268 // input vector in two new ones: (1) a vector with only the shifted-right
19269 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19270 // masked out higher ones) for each byte. PSHUB is used separately with both
19271 // to index the in-register table. Next, both are added and the result is a
19272 // i8 vector where each element contains the pop count for input byte.
19274 // To obtain the pop count for elements != i8, we follow up with the same
19275 // approach and use additional tricks as described below.
19277 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19278 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19279 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19280 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19282 int NumByteElts = VecSize / 8;
19283 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19284 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19285 SmallVector<SDValue, 16> LUTVec;
19286 for (int i = 0; i < NumByteElts; ++i)
19287 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19288 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19289 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19290 DAG.getConstant(0x0F, DL, MVT::i8));
19291 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19294 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19295 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19296 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19299 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19301 // The input vector is used as the shuffle mask that index elements into the
19302 // LUT. After counting low and high nibbles, add the vector to obtain the
19303 // final pop count per i8 element.
19304 SDValue HighPopCnt =
19305 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19306 SDValue LowPopCnt =
19307 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19308 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19310 if (EltVT == MVT::i8)
19313 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19316 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19317 const X86Subtarget *Subtarget,
19318 SelectionDAG &DAG) {
19319 MVT VT = Op.getSimpleValueType();
19320 assert(VT.is128BitVector() &&
19321 "Only 128-bit vector bitmath lowering supported.");
19323 int VecSize = VT.getSizeInBits();
19324 MVT EltVT = VT.getVectorElementType();
19325 int Len = EltVT.getSizeInBits();
19327 // This is the vectorized version of the "best" algorithm from
19328 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19329 // with a minor tweak to use a series of adds + shifts instead of vector
19330 // multiplications. Implemented for all integer vector types. We only use
19331 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19332 // much faster, even faster than using native popcnt instructions.
19334 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19335 MVT VT = V.getSimpleValueType();
19336 SmallVector<SDValue, 32> Shifters(
19337 VT.getVectorNumElements(),
19338 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19339 return DAG.getNode(OpCode, DL, VT, V,
19340 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19342 auto GetMask = [&](SDValue V, APInt Mask) {
19343 MVT VT = V.getSimpleValueType();
19344 SmallVector<SDValue, 32> Masks(
19345 VT.getVectorNumElements(),
19346 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19347 return DAG.getNode(ISD::AND, DL, VT, V,
19348 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19351 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19352 // x86, so set the SRL type to have elements at least i16 wide. This is
19353 // correct because all of our SRLs are followed immediately by a mask anyways
19354 // that handles any bits that sneak into the high bits of the byte elements.
19355 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19359 // v = v - ((v >> 1) & 0x55555555...)
19361 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19362 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19363 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19365 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19366 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19367 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19368 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19369 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19371 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19372 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19373 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19374 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19376 // At this point, V contains the byte-wise population count, and we are
19377 // merely doing a horizontal sum if necessary to get the wider element
19379 if (EltVT == MVT::i8)
19382 return LowerHorizontalByteSum(
19383 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19387 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19388 SelectionDAG &DAG) {
19389 MVT VT = Op.getSimpleValueType();
19390 // FIXME: Need to add AVX-512 support here!
19391 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19392 "Unknown CTPOP type to handle");
19393 SDLoc DL(Op.getNode());
19394 SDValue Op0 = Op.getOperand(0);
19396 if (!Subtarget->hasSSSE3()) {
19397 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19398 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19399 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19402 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19403 unsigned NumElems = VT.getVectorNumElements();
19405 // Extract each 128-bit vector, compute pop count and concat the result.
19406 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19407 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19409 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19410 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19411 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19414 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19417 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19418 SelectionDAG &DAG) {
19419 assert(Op.getSimpleValueType().isVector() &&
19420 "We only do custom lowering for vector population count.");
19421 return LowerVectorCTPOP(Op, Subtarget, DAG);
19424 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19425 SDNode *Node = Op.getNode();
19427 EVT T = Node->getValueType(0);
19428 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19429 DAG.getConstant(0, dl, T), Node->getOperand(2));
19430 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19431 cast<AtomicSDNode>(Node)->getMemoryVT(),
19432 Node->getOperand(0),
19433 Node->getOperand(1), negOp,
19434 cast<AtomicSDNode>(Node)->getMemOperand(),
19435 cast<AtomicSDNode>(Node)->getOrdering(),
19436 cast<AtomicSDNode>(Node)->getSynchScope());
19439 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19440 SDNode *Node = Op.getNode();
19442 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19444 // Convert seq_cst store -> xchg
19445 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19446 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19447 // (The only way to get a 16-byte store is cmpxchg16b)
19448 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19449 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19450 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19451 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19452 cast<AtomicSDNode>(Node)->getMemoryVT(),
19453 Node->getOperand(0),
19454 Node->getOperand(1), Node->getOperand(2),
19455 cast<AtomicSDNode>(Node)->getMemOperand(),
19456 cast<AtomicSDNode>(Node)->getOrdering(),
19457 cast<AtomicSDNode>(Node)->getSynchScope());
19458 return Swap.getValue(1);
19460 // Other atomic stores have a simple pattern.
19464 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19465 MVT VT = Op.getNode()->getSimpleValueType(0);
19467 // Let legalize expand this if it isn't a legal type yet.
19468 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19471 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19474 bool ExtraOp = false;
19475 switch (Op.getOpcode()) {
19476 default: llvm_unreachable("Invalid code");
19477 case ISD::ADDC: Opc = X86ISD::ADD; break;
19478 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19479 case ISD::SUBC: Opc = X86ISD::SUB; break;
19480 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19484 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19486 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19487 Op.getOperand(1), Op.getOperand(2));
19490 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19491 SelectionDAG &DAG) {
19492 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19494 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19495 // which returns the values as { float, float } (in XMM0) or
19496 // { double, double } (which is returned in XMM0, XMM1).
19498 SDValue Arg = Op.getOperand(0);
19499 EVT ArgVT = Arg.getValueType();
19500 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19502 TargetLowering::ArgListTy Args;
19503 TargetLowering::ArgListEntry Entry;
19507 Entry.isSExt = false;
19508 Entry.isZExt = false;
19509 Args.push_back(Entry);
19511 bool isF64 = ArgVT == MVT::f64;
19512 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19513 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19514 // the results are returned via SRet in memory.
19515 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19516 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19518 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19520 Type *RetTy = isF64
19521 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19522 : (Type*)VectorType::get(ArgTy, 4);
19524 TargetLowering::CallLoweringInfo CLI(DAG);
19525 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19526 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19528 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19531 // Returned in xmm0 and xmm1.
19532 return CallResult.first;
19534 // Returned in bits 0:31 and 32:64 xmm0.
19535 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19536 CallResult.first, DAG.getIntPtrConstant(0, dl));
19537 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19538 CallResult.first, DAG.getIntPtrConstant(1, dl));
19539 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19540 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19543 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19544 SelectionDAG &DAG) {
19545 assert(Subtarget->hasAVX512() &&
19546 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19548 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19549 MVT VT = N->getValue().getSimpleValueType();
19550 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19553 // X86 scatter kills mask register, so its type should be added to
19554 // the list of return values
19555 if (N->getNumValues() == 1) {
19556 SDValue Index = N->getIndex();
19557 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19558 !Index.getSimpleValueType().is512BitVector())
19559 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19561 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19562 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19563 N->getOperand(3), Index };
19565 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19566 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19567 return SDValue(NewScatter.getNode(), 0);
19572 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19573 SelectionDAG &DAG) {
19574 assert(Subtarget->hasAVX512() &&
19575 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19577 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19578 MVT VT = Op.getSimpleValueType();
19579 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19582 SDValue Index = N->getIndex();
19583 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19584 !Index.getSimpleValueType().is512BitVector()) {
19585 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19586 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19587 N->getOperand(3), Index };
19588 DAG.UpdateNodeOperands(N, Ops);
19593 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19594 SelectionDAG &DAG) const {
19595 // TODO: Eventually, the lowering of these nodes should be informed by or
19596 // deferred to the GC strategy for the function in which they appear. For
19597 // now, however, they must be lowered to something. Since they are logically
19598 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19599 // require special handling for these nodes), lower them as literal NOOPs for
19601 SmallVector<SDValue, 2> Ops;
19603 Ops.push_back(Op.getOperand(0));
19604 if (Op->getGluedNode())
19605 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19608 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19609 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19614 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19615 SelectionDAG &DAG) const {
19616 // TODO: Eventually, the lowering of these nodes should be informed by or
19617 // deferred to the GC strategy for the function in which they appear. For
19618 // now, however, they must be lowered to something. Since they are logically
19619 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19620 // require special handling for these nodes), lower them as literal NOOPs for
19622 SmallVector<SDValue, 2> Ops;
19624 Ops.push_back(Op.getOperand(0));
19625 if (Op->getGluedNode())
19626 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19629 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19630 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19635 /// LowerOperation - Provide custom lowering hooks for some operations.
19637 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19638 switch (Op.getOpcode()) {
19639 default: llvm_unreachable("Should not custom lower this!");
19640 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19641 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19642 return LowerCMP_SWAP(Op, Subtarget, DAG);
19643 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19644 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19645 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19646 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19647 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19648 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19649 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19650 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19651 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19652 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19653 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19654 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19655 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19656 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19657 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19658 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19659 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19660 case ISD::SHL_PARTS:
19661 case ISD::SRA_PARTS:
19662 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19663 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19664 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19665 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19666 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19667 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19668 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19669 case ISD::SIGN_EXTEND_VECTOR_INREG:
19670 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19671 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19672 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19673 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19674 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19676 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19677 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19678 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19679 case ISD::SETCC: return LowerSETCC(Op, DAG);
19680 case ISD::SELECT: return LowerSELECT(Op, DAG);
19681 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19682 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19683 case ISD::VASTART: return LowerVASTART(Op, DAG);
19684 case ISD::VAARG: return LowerVAARG(Op, DAG);
19685 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19686 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19687 case ISD::INTRINSIC_VOID:
19688 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19689 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19690 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19691 case ISD::FRAME_TO_ARGS_OFFSET:
19692 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19693 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19694 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19695 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19696 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19697 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19698 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19699 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19700 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19701 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19703 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19704 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19705 case ISD::UMUL_LOHI:
19706 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19707 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19710 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19716 case ISD::UMULO: return LowerXALUO(Op, DAG);
19717 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19718 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19722 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19723 case ISD::ADD: return LowerADD(Op, DAG);
19724 case ISD::SUB: return LowerSUB(Op, DAG);
19728 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19729 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19730 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19731 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19732 case ISD::GC_TRANSITION_START:
19733 return LowerGC_TRANSITION_START(Op, DAG);
19734 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19738 /// ReplaceNodeResults - Replace a node with an illegal result type
19739 /// with a new node built out of custom code.
19740 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19741 SmallVectorImpl<SDValue>&Results,
19742 SelectionDAG &DAG) const {
19744 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19745 switch (N->getOpcode()) {
19747 llvm_unreachable("Do not know how to custom type legalize this operation!");
19748 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19749 case X86ISD::FMINC:
19751 case X86ISD::FMAXC:
19752 case X86ISD::FMAX: {
19753 EVT VT = N->getValueType(0);
19754 assert(VT == MVT::v2f32 && "Unexpected type (!= v2f32) on FMIN/FMAX.");
19755 SDValue UNDEF = DAG.getUNDEF(VT);
19756 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19757 N->getOperand(0), UNDEF);
19758 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19759 N->getOperand(1), UNDEF);
19760 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19763 case ISD::SIGN_EXTEND_INREG:
19768 // We don't want to expand or promote these.
19775 case ISD::UDIVREM: {
19776 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19777 Results.push_back(V);
19780 case ISD::FP_TO_SINT:
19781 case ISD::FP_TO_UINT: {
19782 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19784 std::pair<SDValue,SDValue> Vals =
19785 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19786 SDValue FIST = Vals.first, StackSlot = Vals.second;
19787 if (FIST.getNode()) {
19788 EVT VT = N->getValueType(0);
19789 // Return a load from the stack slot.
19790 if (StackSlot.getNode())
19791 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19792 MachinePointerInfo(),
19793 false, false, false, 0));
19795 Results.push_back(FIST);
19799 case ISD::UINT_TO_FP: {
19800 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19801 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19802 N->getValueType(0) != MVT::v2f32)
19804 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19806 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19808 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19809 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19810 DAG.getBitcast(MVT::v2i64, VBias));
19811 Or = DAG.getBitcast(MVT::v2f64, Or);
19812 // TODO: Are there any fast-math-flags to propagate here?
19813 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19814 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19817 case ISD::FP_ROUND: {
19818 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19820 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19821 Results.push_back(V);
19824 case ISD::FP_EXTEND: {
19825 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19826 // No other ValueType for FP_EXTEND should reach this point.
19827 assert(N->getValueType(0) == MVT::v2f32 &&
19828 "Do not know how to legalize this Node");
19831 case ISD::INTRINSIC_W_CHAIN: {
19832 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19834 default : llvm_unreachable("Do not know how to custom type "
19835 "legalize this intrinsic operation!");
19836 case Intrinsic::x86_rdtsc:
19837 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19839 case Intrinsic::x86_rdtscp:
19840 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19842 case Intrinsic::x86_rdpmc:
19843 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19846 case ISD::READCYCLECOUNTER: {
19847 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19850 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19851 EVT T = N->getValueType(0);
19852 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19853 bool Regs64bit = T == MVT::i128;
19854 MVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19855 SDValue cpInL, cpInH;
19856 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19857 DAG.getConstant(0, dl, HalfT));
19858 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19859 DAG.getConstant(1, dl, HalfT));
19860 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19861 Regs64bit ? X86::RAX : X86::EAX,
19863 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19864 Regs64bit ? X86::RDX : X86::EDX,
19865 cpInH, cpInL.getValue(1));
19866 SDValue swapInL, swapInH;
19867 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19868 DAG.getConstant(0, dl, HalfT));
19869 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19870 DAG.getConstant(1, dl, HalfT));
19871 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19872 Regs64bit ? X86::RBX : X86::EBX,
19873 swapInL, cpInH.getValue(1));
19874 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19875 Regs64bit ? X86::RCX : X86::ECX,
19876 swapInH, swapInL.getValue(1));
19877 SDValue Ops[] = { swapInH.getValue(0),
19879 swapInH.getValue(1) };
19880 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19881 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19882 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19883 X86ISD::LCMPXCHG8_DAG;
19884 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19885 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19886 Regs64bit ? X86::RAX : X86::EAX,
19887 HalfT, Result.getValue(1));
19888 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19889 Regs64bit ? X86::RDX : X86::EDX,
19890 HalfT, cpOutL.getValue(2));
19891 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19893 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19894 MVT::i32, cpOutH.getValue(2));
19896 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19897 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19898 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19900 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19901 Results.push_back(Success);
19902 Results.push_back(EFLAGS.getValue(1));
19905 case ISD::ATOMIC_SWAP:
19906 case ISD::ATOMIC_LOAD_ADD:
19907 case ISD::ATOMIC_LOAD_SUB:
19908 case ISD::ATOMIC_LOAD_AND:
19909 case ISD::ATOMIC_LOAD_OR:
19910 case ISD::ATOMIC_LOAD_XOR:
19911 case ISD::ATOMIC_LOAD_NAND:
19912 case ISD::ATOMIC_LOAD_MIN:
19913 case ISD::ATOMIC_LOAD_MAX:
19914 case ISD::ATOMIC_LOAD_UMIN:
19915 case ISD::ATOMIC_LOAD_UMAX:
19916 case ISD::ATOMIC_LOAD: {
19917 // Delegate to generic TypeLegalization. Situations we can really handle
19918 // should have already been dealt with by AtomicExpandPass.cpp.
19921 case ISD::BITCAST: {
19922 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19923 EVT DstVT = N->getValueType(0);
19924 EVT SrcVT = N->getOperand(0)->getValueType(0);
19926 if (SrcVT != MVT::f64 ||
19927 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19930 unsigned NumElts = DstVT.getVectorNumElements();
19931 EVT SVT = DstVT.getVectorElementType();
19932 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19933 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19934 MVT::v2f64, N->getOperand(0));
19935 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19937 if (ExperimentalVectorWideningLegalization) {
19938 // If we are legalizing vectors by widening, we already have the desired
19939 // legal vector type, just return it.
19940 Results.push_back(ToVecInt);
19944 SmallVector<SDValue, 8> Elts;
19945 for (unsigned i = 0, e = NumElts; i != e; ++i)
19946 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19947 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19949 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19954 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19955 switch ((X86ISD::NodeType)Opcode) {
19956 case X86ISD::FIRST_NUMBER: break;
19957 case X86ISD::BSF: return "X86ISD::BSF";
19958 case X86ISD::BSR: return "X86ISD::BSR";
19959 case X86ISD::SHLD: return "X86ISD::SHLD";
19960 case X86ISD::SHRD: return "X86ISD::SHRD";
19961 case X86ISD::FAND: return "X86ISD::FAND";
19962 case X86ISD::FANDN: return "X86ISD::FANDN";
19963 case X86ISD::FOR: return "X86ISD::FOR";
19964 case X86ISD::FXOR: return "X86ISD::FXOR";
19965 case X86ISD::FILD: return "X86ISD::FILD";
19966 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19967 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19968 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19969 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19970 case X86ISD::FLD: return "X86ISD::FLD";
19971 case X86ISD::FST: return "X86ISD::FST";
19972 case X86ISD::CALL: return "X86ISD::CALL";
19973 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19974 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19975 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19976 case X86ISD::BT: return "X86ISD::BT";
19977 case X86ISD::CMP: return "X86ISD::CMP";
19978 case X86ISD::COMI: return "X86ISD::COMI";
19979 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19980 case X86ISD::CMPM: return "X86ISD::CMPM";
19981 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19982 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19983 case X86ISD::SETCC: return "X86ISD::SETCC";
19984 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19985 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19986 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19987 case X86ISD::CMOV: return "X86ISD::CMOV";
19988 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19989 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19990 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19991 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19992 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19993 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19994 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19995 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19996 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19997 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19998 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19999 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20000 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20001 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20002 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20003 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20004 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20005 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20006 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20007 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20008 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20009 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20010 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20011 case X86ISD::HADD: return "X86ISD::HADD";
20012 case X86ISD::HSUB: return "X86ISD::HSUB";
20013 case X86ISD::FHADD: return "X86ISD::FHADD";
20014 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20015 case X86ISD::ABS: return "X86ISD::ABS";
20016 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20017 case X86ISD::FMAX: return "X86ISD::FMAX";
20018 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20019 case X86ISD::FMIN: return "X86ISD::FMIN";
20020 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20021 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20022 case X86ISD::FMINC: return "X86ISD::FMINC";
20023 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20024 case X86ISD::FRCP: return "X86ISD::FRCP";
20025 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20026 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20027 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20028 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20029 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20030 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20031 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20032 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20033 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20034 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20035 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20036 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20037 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20038 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20039 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20040 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20041 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20042 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20043 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20044 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20045 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20046 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20047 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20048 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20049 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20050 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20051 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20052 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20053 case X86ISD::VSHL: return "X86ISD::VSHL";
20054 case X86ISD::VSRL: return "X86ISD::VSRL";
20055 case X86ISD::VSRA: return "X86ISD::VSRA";
20056 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20057 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20058 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20059 case X86ISD::CMPP: return "X86ISD::CMPP";
20060 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20061 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20062 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20063 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20064 case X86ISD::ADD: return "X86ISD::ADD";
20065 case X86ISD::SUB: return "X86ISD::SUB";
20066 case X86ISD::ADC: return "X86ISD::ADC";
20067 case X86ISD::SBB: return "X86ISD::SBB";
20068 case X86ISD::SMUL: return "X86ISD::SMUL";
20069 case X86ISD::UMUL: return "X86ISD::UMUL";
20070 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20071 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20072 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20073 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20074 case X86ISD::INC: return "X86ISD::INC";
20075 case X86ISD::DEC: return "X86ISD::DEC";
20076 case X86ISD::OR: return "X86ISD::OR";
20077 case X86ISD::XOR: return "X86ISD::XOR";
20078 case X86ISD::AND: return "X86ISD::AND";
20079 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20080 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20081 case X86ISD::PTEST: return "X86ISD::PTEST";
20082 case X86ISD::TESTP: return "X86ISD::TESTP";
20083 case X86ISD::TESTM: return "X86ISD::TESTM";
20084 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20085 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20086 case X86ISD::KTEST: return "X86ISD::KTEST";
20087 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20088 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20089 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20090 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20091 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20092 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20093 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20094 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20095 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20096 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20097 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20098 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20099 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20100 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20101 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20102 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20103 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20104 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20105 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20106 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20107 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20108 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20109 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20110 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20111 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20112 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20113 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20114 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20115 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20116 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20117 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20118 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20119 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20120 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20121 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20122 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20123 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20124 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20125 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20126 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20127 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20128 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20129 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20130 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20131 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20132 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20133 case X86ISD::SAHF: return "X86ISD::SAHF";
20134 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20135 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20136 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20137 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20138 case X86ISD::VPROT: return "X86ISD::VPROT";
20139 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20140 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20141 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20142 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20143 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20144 case X86ISD::FMADD: return "X86ISD::FMADD";
20145 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20146 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20147 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20148 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20149 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20150 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20151 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20152 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20153 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20154 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20155 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20156 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20157 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20158 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20159 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20160 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20161 case X86ISD::XTEST: return "X86ISD::XTEST";
20162 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20163 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20164 case X86ISD::SELECT: return "X86ISD::SELECT";
20165 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20166 case X86ISD::RCP28: return "X86ISD::RCP28";
20167 case X86ISD::EXP2: return "X86ISD::EXP2";
20168 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20169 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20170 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20171 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20172 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20173 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20174 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20175 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20176 case X86ISD::ADDS: return "X86ISD::ADDS";
20177 case X86ISD::SUBS: return "X86ISD::SUBS";
20178 case X86ISD::AVG: return "X86ISD::AVG";
20179 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20180 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20181 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20182 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20183 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20184 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20189 // isLegalAddressingMode - Return true if the addressing mode represented
20190 // by AM is legal for this target, for a load/store of the specified type.
20191 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20192 const AddrMode &AM, Type *Ty,
20193 unsigned AS) const {
20194 // X86 supports extremely general addressing modes.
20195 CodeModel::Model M = getTargetMachine().getCodeModel();
20196 Reloc::Model R = getTargetMachine().getRelocationModel();
20198 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20199 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20204 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20206 // If a reference to this global requires an extra load, we can't fold it.
20207 if (isGlobalStubReference(GVFlags))
20210 // If BaseGV requires a register for the PIC base, we cannot also have a
20211 // BaseReg specified.
20212 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20215 // If lower 4G is not available, then we must use rip-relative addressing.
20216 if ((M != CodeModel::Small || R != Reloc::Static) &&
20217 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20221 switch (AM.Scale) {
20227 // These scales always work.
20232 // These scales are formed with basereg+scalereg. Only accept if there is
20237 default: // Other stuff never works.
20244 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20245 unsigned Bits = Ty->getScalarSizeInBits();
20247 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20248 // particularly cheaper than those without.
20252 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20253 // variable shifts just as cheap as scalar ones.
20254 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20257 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20258 // fully general vector.
20262 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20263 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20265 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20266 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20267 return NumBits1 > NumBits2;
20270 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20271 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20274 if (!isTypeLegal(EVT::getEVT(Ty1)))
20277 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20279 // Assuming the caller doesn't have a zeroext or signext return parameter,
20280 // truncation all the way down to i1 is valid.
20284 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20285 return isInt<32>(Imm);
20288 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20289 // Can also use sub to handle negated immediates.
20290 return isInt<32>(Imm);
20293 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20294 if (!VT1.isInteger() || !VT2.isInteger())
20296 unsigned NumBits1 = VT1.getSizeInBits();
20297 unsigned NumBits2 = VT2.getSizeInBits();
20298 return NumBits1 > NumBits2;
20301 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20302 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20303 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20306 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20307 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20308 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20311 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20312 EVT VT1 = Val.getValueType();
20313 if (isZExtFree(VT1, VT2))
20316 if (Val.getOpcode() != ISD::LOAD)
20319 if (!VT1.isSimple() || !VT1.isInteger() ||
20320 !VT2.isSimple() || !VT2.isInteger())
20323 switch (VT1.getSimpleVT().SimpleTy) {
20328 // X86 has 8, 16, and 32-bit zero-extending loads.
20335 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20338 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20339 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20342 VT = VT.getScalarType();
20344 if (!VT.isSimple())
20347 switch (VT.getSimpleVT().SimpleTy) {
20358 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20359 // i16 instructions are longer (0x66 prefix) and potentially slower.
20360 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20363 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20364 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20365 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20366 /// are assumed to be legal.
20368 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20370 if (!VT.isSimple())
20373 // Not for i1 vectors
20374 if (VT.getSimpleVT().getScalarType() == MVT::i1)
20377 // Very little shuffling can be done for 64-bit vectors right now.
20378 if (VT.getSimpleVT().getSizeInBits() == 64)
20381 // We only care that the types being shuffled are legal. The lowering can
20382 // handle any possible shuffle mask that results.
20383 return isTypeLegal(VT.getSimpleVT());
20387 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20389 // Just delegate to the generic legality, clear masks aren't special.
20390 return isShuffleMaskLegal(Mask, VT);
20393 //===----------------------------------------------------------------------===//
20394 // X86 Scheduler Hooks
20395 //===----------------------------------------------------------------------===//
20397 /// Utility function to emit xbegin specifying the start of an RTM region.
20398 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20399 const TargetInstrInfo *TII) {
20400 DebugLoc DL = MI->getDebugLoc();
20402 const BasicBlock *BB = MBB->getBasicBlock();
20403 MachineFunction::iterator I = ++MBB->getIterator();
20405 // For the v = xbegin(), we generate
20416 MachineBasicBlock *thisMBB = MBB;
20417 MachineFunction *MF = MBB->getParent();
20418 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20419 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20420 MF->insert(I, mainMBB);
20421 MF->insert(I, sinkMBB);
20423 // Transfer the remainder of BB and its successor edges to sinkMBB.
20424 sinkMBB->splice(sinkMBB->begin(), MBB,
20425 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20426 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20430 // # fallthrough to mainMBB
20431 // # abortion to sinkMBB
20432 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20433 thisMBB->addSuccessor(mainMBB);
20434 thisMBB->addSuccessor(sinkMBB);
20438 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20439 mainMBB->addSuccessor(sinkMBB);
20442 // EAX is live into the sinkMBB
20443 sinkMBB->addLiveIn(X86::EAX);
20444 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20445 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20448 MI->eraseFromParent();
20452 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20453 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20454 // in the .td file.
20455 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20456 const TargetInstrInfo *TII) {
20458 switch (MI->getOpcode()) {
20459 default: llvm_unreachable("illegal opcode!");
20460 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20461 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20462 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20463 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20464 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20465 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20466 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20467 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20470 DebugLoc dl = MI->getDebugLoc();
20471 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20473 unsigned NumArgs = MI->getNumOperands();
20474 for (unsigned i = 1; i < NumArgs; ++i) {
20475 MachineOperand &Op = MI->getOperand(i);
20476 if (!(Op.isReg() && Op.isImplicit()))
20477 MIB.addOperand(Op);
20479 if (MI->hasOneMemOperand())
20480 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20482 BuildMI(*BB, MI, dl,
20483 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20484 .addReg(X86::XMM0);
20486 MI->eraseFromParent();
20490 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20491 // defs in an instruction pattern
20492 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20493 const TargetInstrInfo *TII) {
20495 switch (MI->getOpcode()) {
20496 default: llvm_unreachable("illegal opcode!");
20497 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20498 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20499 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20500 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20501 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20502 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20503 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20504 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20507 DebugLoc dl = MI->getDebugLoc();
20508 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20510 unsigned NumArgs = MI->getNumOperands(); // remove the results
20511 for (unsigned i = 1; i < NumArgs; ++i) {
20512 MachineOperand &Op = MI->getOperand(i);
20513 if (!(Op.isReg() && Op.isImplicit()))
20514 MIB.addOperand(Op);
20516 if (MI->hasOneMemOperand())
20517 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20519 BuildMI(*BB, MI, dl,
20520 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20523 MI->eraseFromParent();
20527 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20528 const X86Subtarget *Subtarget) {
20529 DebugLoc dl = MI->getDebugLoc();
20530 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20531 // Address into RAX/EAX, other two args into ECX, EDX.
20532 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20533 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20534 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20535 for (int i = 0; i < X86::AddrNumOperands; ++i)
20536 MIB.addOperand(MI->getOperand(i));
20538 unsigned ValOps = X86::AddrNumOperands;
20539 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20540 .addReg(MI->getOperand(ValOps).getReg());
20541 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20542 .addReg(MI->getOperand(ValOps+1).getReg());
20544 // The instruction doesn't actually take any operands though.
20545 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20547 MI->eraseFromParent(); // The pseudo is gone now.
20551 MachineBasicBlock *
20552 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20553 MachineBasicBlock *MBB) const {
20554 // Emit va_arg instruction on X86-64.
20556 // Operands to this pseudo-instruction:
20557 // 0 ) Output : destination address (reg)
20558 // 1-5) Input : va_list address (addr, i64mem)
20559 // 6 ) ArgSize : Size (in bytes) of vararg type
20560 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20561 // 8 ) Align : Alignment of type
20562 // 9 ) EFLAGS (implicit-def)
20564 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20565 static_assert(X86::AddrNumOperands == 5,
20566 "VAARG_64 assumes 5 address operands");
20568 unsigned DestReg = MI->getOperand(0).getReg();
20569 MachineOperand &Base = MI->getOperand(1);
20570 MachineOperand &Scale = MI->getOperand(2);
20571 MachineOperand &Index = MI->getOperand(3);
20572 MachineOperand &Disp = MI->getOperand(4);
20573 MachineOperand &Segment = MI->getOperand(5);
20574 unsigned ArgSize = MI->getOperand(6).getImm();
20575 unsigned ArgMode = MI->getOperand(7).getImm();
20576 unsigned Align = MI->getOperand(8).getImm();
20578 // Memory Reference
20579 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20580 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20581 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20583 // Machine Information
20584 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20585 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20586 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20587 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20588 DebugLoc DL = MI->getDebugLoc();
20590 // struct va_list {
20593 // i64 overflow_area (address)
20594 // i64 reg_save_area (address)
20596 // sizeof(va_list) = 24
20597 // alignment(va_list) = 8
20599 unsigned TotalNumIntRegs = 6;
20600 unsigned TotalNumXMMRegs = 8;
20601 bool UseGPOffset = (ArgMode == 1);
20602 bool UseFPOffset = (ArgMode == 2);
20603 unsigned MaxOffset = TotalNumIntRegs * 8 +
20604 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20606 /* Align ArgSize to a multiple of 8 */
20607 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20608 bool NeedsAlign = (Align > 8);
20610 MachineBasicBlock *thisMBB = MBB;
20611 MachineBasicBlock *overflowMBB;
20612 MachineBasicBlock *offsetMBB;
20613 MachineBasicBlock *endMBB;
20615 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20616 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20617 unsigned OffsetReg = 0;
20619 if (!UseGPOffset && !UseFPOffset) {
20620 // If we only pull from the overflow region, we don't create a branch.
20621 // We don't need to alter control flow.
20622 OffsetDestReg = 0; // unused
20623 OverflowDestReg = DestReg;
20625 offsetMBB = nullptr;
20626 overflowMBB = thisMBB;
20629 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20630 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20631 // If not, pull from overflow_area. (branch to overflowMBB)
20636 // offsetMBB overflowMBB
20641 // Registers for the PHI in endMBB
20642 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20643 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20645 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20646 MachineFunction *MF = MBB->getParent();
20647 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20648 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20649 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20651 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20653 // Insert the new basic blocks
20654 MF->insert(MBBIter, offsetMBB);
20655 MF->insert(MBBIter, overflowMBB);
20656 MF->insert(MBBIter, endMBB);
20658 // Transfer the remainder of MBB and its successor edges to endMBB.
20659 endMBB->splice(endMBB->begin(), thisMBB,
20660 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20661 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20663 // Make offsetMBB and overflowMBB successors of thisMBB
20664 thisMBB->addSuccessor(offsetMBB);
20665 thisMBB->addSuccessor(overflowMBB);
20667 // endMBB is a successor of both offsetMBB and overflowMBB
20668 offsetMBB->addSuccessor(endMBB);
20669 overflowMBB->addSuccessor(endMBB);
20671 // Load the offset value into a register
20672 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20673 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20677 .addDisp(Disp, UseFPOffset ? 4 : 0)
20678 .addOperand(Segment)
20679 .setMemRefs(MMOBegin, MMOEnd);
20681 // Check if there is enough room left to pull this argument.
20682 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20684 .addImm(MaxOffset + 8 - ArgSizeA8);
20686 // Branch to "overflowMBB" if offset >= max
20687 // Fall through to "offsetMBB" otherwise
20688 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20689 .addMBB(overflowMBB);
20692 // In offsetMBB, emit code to use the reg_save_area.
20694 assert(OffsetReg != 0);
20696 // Read the reg_save_area address.
20697 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20698 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20703 .addOperand(Segment)
20704 .setMemRefs(MMOBegin, MMOEnd);
20706 // Zero-extend the offset
20707 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20708 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20711 .addImm(X86::sub_32bit);
20713 // Add the offset to the reg_save_area to get the final address.
20714 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20715 .addReg(OffsetReg64)
20716 .addReg(RegSaveReg);
20718 // Compute the offset for the next argument
20719 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20720 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20722 .addImm(UseFPOffset ? 16 : 8);
20724 // Store it back into the va_list.
20725 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20729 .addDisp(Disp, UseFPOffset ? 4 : 0)
20730 .addOperand(Segment)
20731 .addReg(NextOffsetReg)
20732 .setMemRefs(MMOBegin, MMOEnd);
20735 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20740 // Emit code to use overflow area
20743 // Load the overflow_area address into a register.
20744 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20745 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20750 .addOperand(Segment)
20751 .setMemRefs(MMOBegin, MMOEnd);
20753 // If we need to align it, do so. Otherwise, just copy the address
20754 // to OverflowDestReg.
20756 // Align the overflow address
20757 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20758 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20760 // aligned_addr = (addr + (align-1)) & ~(align-1)
20761 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20762 .addReg(OverflowAddrReg)
20765 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20767 .addImm(~(uint64_t)(Align-1));
20769 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20770 .addReg(OverflowAddrReg);
20773 // Compute the next overflow address after this argument.
20774 // (the overflow address should be kept 8-byte aligned)
20775 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20776 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20777 .addReg(OverflowDestReg)
20778 .addImm(ArgSizeA8);
20780 // Store the new overflow address.
20781 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20786 .addOperand(Segment)
20787 .addReg(NextAddrReg)
20788 .setMemRefs(MMOBegin, MMOEnd);
20790 // If we branched, emit the PHI to the front of endMBB.
20792 BuildMI(*endMBB, endMBB->begin(), DL,
20793 TII->get(X86::PHI), DestReg)
20794 .addReg(OffsetDestReg).addMBB(offsetMBB)
20795 .addReg(OverflowDestReg).addMBB(overflowMBB);
20798 // Erase the pseudo instruction
20799 MI->eraseFromParent();
20804 MachineBasicBlock *
20805 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20807 MachineBasicBlock *MBB) const {
20808 // Emit code to save XMM registers to the stack. The ABI says that the
20809 // number of registers to save is given in %al, so it's theoretically
20810 // possible to do an indirect jump trick to avoid saving all of them,
20811 // however this code takes a simpler approach and just executes all
20812 // of the stores if %al is non-zero. It's less code, and it's probably
20813 // easier on the hardware branch predictor, and stores aren't all that
20814 // expensive anyway.
20816 // Create the new basic blocks. One block contains all the XMM stores,
20817 // and one block is the final destination regardless of whether any
20818 // stores were performed.
20819 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20820 MachineFunction *F = MBB->getParent();
20821 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20822 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20823 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20824 F->insert(MBBIter, XMMSaveMBB);
20825 F->insert(MBBIter, EndMBB);
20827 // Transfer the remainder of MBB and its successor edges to EndMBB.
20828 EndMBB->splice(EndMBB->begin(), MBB,
20829 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20830 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20832 // The original block will now fall through to the XMM save block.
20833 MBB->addSuccessor(XMMSaveMBB);
20834 // The XMMSaveMBB will fall through to the end block.
20835 XMMSaveMBB->addSuccessor(EndMBB);
20837 // Now add the instructions.
20838 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20839 DebugLoc DL = MI->getDebugLoc();
20841 unsigned CountReg = MI->getOperand(0).getReg();
20842 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20843 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20845 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20846 // If %al is 0, branch around the XMM save block.
20847 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20848 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20849 MBB->addSuccessor(EndMBB);
20852 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20853 // that was just emitted, but clearly shouldn't be "saved".
20854 assert((MI->getNumOperands() <= 3 ||
20855 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20856 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20857 && "Expected last argument to be EFLAGS");
20858 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20859 // In the XMM save block, save all the XMM argument registers.
20860 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20861 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20862 MachineMemOperand *MMO = F->getMachineMemOperand(
20863 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20864 MachineMemOperand::MOStore,
20865 /*Size=*/16, /*Align=*/16);
20866 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20867 .addFrameIndex(RegSaveFrameIndex)
20868 .addImm(/*Scale=*/1)
20869 .addReg(/*IndexReg=*/0)
20870 .addImm(/*Disp=*/Offset)
20871 .addReg(/*Segment=*/0)
20872 .addReg(MI->getOperand(i).getReg())
20873 .addMemOperand(MMO);
20876 MI->eraseFromParent(); // The pseudo instruction is gone now.
20881 // The EFLAGS operand of SelectItr might be missing a kill marker
20882 // because there were multiple uses of EFLAGS, and ISel didn't know
20883 // which to mark. Figure out whether SelectItr should have had a
20884 // kill marker, and set it if it should. Returns the correct kill
20886 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20887 MachineBasicBlock* BB,
20888 const TargetRegisterInfo* TRI) {
20889 // Scan forward through BB for a use/def of EFLAGS.
20890 MachineBasicBlock::iterator miI(std::next(SelectItr));
20891 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20892 const MachineInstr& mi = *miI;
20893 if (mi.readsRegister(X86::EFLAGS))
20895 if (mi.definesRegister(X86::EFLAGS))
20896 break; // Should have kill-flag - update below.
20899 // If we hit the end of the block, check whether EFLAGS is live into a
20901 if (miI == BB->end()) {
20902 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20903 sEnd = BB->succ_end();
20904 sItr != sEnd; ++sItr) {
20905 MachineBasicBlock* succ = *sItr;
20906 if (succ->isLiveIn(X86::EFLAGS))
20911 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20912 // out. SelectMI should have a kill flag on EFLAGS.
20913 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20917 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20918 // together with other CMOV pseudo-opcodes into a single basic-block with
20919 // conditional jump around it.
20920 static bool isCMOVPseudo(MachineInstr *MI) {
20921 switch (MI->getOpcode()) {
20922 case X86::CMOV_FR32:
20923 case X86::CMOV_FR64:
20924 case X86::CMOV_GR8:
20925 case X86::CMOV_GR16:
20926 case X86::CMOV_GR32:
20927 case X86::CMOV_RFP32:
20928 case X86::CMOV_RFP64:
20929 case X86::CMOV_RFP80:
20930 case X86::CMOV_V2F64:
20931 case X86::CMOV_V2I64:
20932 case X86::CMOV_V4F32:
20933 case X86::CMOV_V4F64:
20934 case X86::CMOV_V4I64:
20935 case X86::CMOV_V16F32:
20936 case X86::CMOV_V8F32:
20937 case X86::CMOV_V8F64:
20938 case X86::CMOV_V8I64:
20939 case X86::CMOV_V8I1:
20940 case X86::CMOV_V16I1:
20941 case X86::CMOV_V32I1:
20942 case X86::CMOV_V64I1:
20950 MachineBasicBlock *
20951 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20952 MachineBasicBlock *BB) const {
20953 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20954 DebugLoc DL = MI->getDebugLoc();
20956 // To "insert" a SELECT_CC instruction, we actually have to insert the
20957 // diamond control-flow pattern. The incoming instruction knows the
20958 // destination vreg to set, the condition code register to branch on, the
20959 // true/false values to select between, and a branch opcode to use.
20960 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20961 MachineFunction::iterator It = ++BB->getIterator();
20966 // cmpTY ccX, r1, r2
20968 // fallthrough --> copy0MBB
20969 MachineBasicBlock *thisMBB = BB;
20970 MachineFunction *F = BB->getParent();
20972 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20973 // as described above, by inserting a BB, and then making a PHI at the join
20974 // point to select the true and false operands of the CMOV in the PHI.
20976 // The code also handles two different cases of multiple CMOV opcodes
20980 // In this case, there are multiple CMOVs in a row, all which are based on
20981 // the same condition setting (or the exact opposite condition setting).
20982 // In this case we can lower all the CMOVs using a single inserted BB, and
20983 // then make a number of PHIs at the join point to model the CMOVs. The only
20984 // trickiness here, is that in a case like:
20986 // t2 = CMOV cond1 t1, f1
20987 // t3 = CMOV cond1 t2, f2
20989 // when rewriting this into PHIs, we have to perform some renaming on the
20990 // temps since you cannot have a PHI operand refer to a PHI result earlier
20991 // in the same block. The "simple" but wrong lowering would be:
20993 // t2 = PHI t1(BB1), f1(BB2)
20994 // t3 = PHI t2(BB1), f2(BB2)
20996 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20997 // renaming is to note that on the path through BB1, t2 is really just a
20998 // copy of t1, and do that renaming, properly generating:
21000 // t2 = PHI t1(BB1), f1(BB2)
21001 // t3 = PHI t1(BB1), f2(BB2)
21003 // Case 2, we lower cascaded CMOVs such as
21005 // (CMOV (CMOV F, T, cc1), T, cc2)
21007 // to two successives branches. For that, we look for another CMOV as the
21008 // following instruction.
21010 // Without this, we would add a PHI between the two jumps, which ends up
21011 // creating a few copies all around. For instance, for
21013 // (sitofp (zext (fcmp une)))
21015 // we would generate:
21017 // ucomiss %xmm1, %xmm0
21018 // movss <1.0f>, %xmm0
21019 // movaps %xmm0, %xmm1
21021 // xorps %xmm1, %xmm1
21024 // movaps %xmm1, %xmm0
21028 // because this custom-inserter would have generated:
21040 // A: X = ...; Y = ...
21042 // C: Z = PHI [X, A], [Y, B]
21044 // E: PHI [X, C], [Z, D]
21046 // If we lower both CMOVs in a single step, we can instead generate:
21058 // A: X = ...; Y = ...
21060 // E: PHI [X, A], [X, C], [Y, D]
21062 // Which, in our sitofp/fcmp example, gives us something like:
21064 // ucomiss %xmm1, %xmm0
21065 // movss <1.0f>, %xmm0
21068 // xorps %xmm0, %xmm0
21072 MachineInstr *CascadedCMOV = nullptr;
21073 MachineInstr *LastCMOV = MI;
21074 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21075 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21076 MachineBasicBlock::iterator NextMIIt =
21077 std::next(MachineBasicBlock::iterator(MI));
21079 // Check for case 1, where there are multiple CMOVs with the same condition
21080 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21081 // number of jumps the most.
21083 if (isCMOVPseudo(MI)) {
21084 // See if we have a string of CMOVS with the same condition.
21085 while (NextMIIt != BB->end() &&
21086 isCMOVPseudo(NextMIIt) &&
21087 (NextMIIt->getOperand(3).getImm() == CC ||
21088 NextMIIt->getOperand(3).getImm() == OppCC)) {
21089 LastCMOV = &*NextMIIt;
21094 // This checks for case 2, but only do this if we didn't already find
21095 // case 1, as indicated by LastCMOV == MI.
21096 if (LastCMOV == MI &&
21097 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21098 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21099 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21100 CascadedCMOV = &*NextMIIt;
21103 MachineBasicBlock *jcc1MBB = nullptr;
21105 // If we have a cascaded CMOV, we lower it to two successive branches to
21106 // the same block. EFLAGS is used by both, so mark it as live in the second.
21107 if (CascadedCMOV) {
21108 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21109 F->insert(It, jcc1MBB);
21110 jcc1MBB->addLiveIn(X86::EFLAGS);
21113 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21114 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21115 F->insert(It, copy0MBB);
21116 F->insert(It, sinkMBB);
21118 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21119 // live into the sink and copy blocks.
21120 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21122 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21123 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21124 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21125 copy0MBB->addLiveIn(X86::EFLAGS);
21126 sinkMBB->addLiveIn(X86::EFLAGS);
21129 // Transfer the remainder of BB and its successor edges to sinkMBB.
21130 sinkMBB->splice(sinkMBB->begin(), BB,
21131 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21132 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21134 // Add the true and fallthrough blocks as its successors.
21135 if (CascadedCMOV) {
21136 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21137 BB->addSuccessor(jcc1MBB);
21139 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21140 // jump to the sinkMBB.
21141 jcc1MBB->addSuccessor(copy0MBB);
21142 jcc1MBB->addSuccessor(sinkMBB);
21144 BB->addSuccessor(copy0MBB);
21147 // The true block target of the first (or only) branch is always sinkMBB.
21148 BB->addSuccessor(sinkMBB);
21150 // Create the conditional branch instruction.
21151 unsigned Opc = X86::GetCondBranchFromCond(CC);
21152 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21154 if (CascadedCMOV) {
21155 unsigned Opc2 = X86::GetCondBranchFromCond(
21156 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21157 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21161 // %FalseValue = ...
21162 // # fallthrough to sinkMBB
21163 copy0MBB->addSuccessor(sinkMBB);
21166 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21168 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21169 MachineBasicBlock::iterator MIItEnd =
21170 std::next(MachineBasicBlock::iterator(LastCMOV));
21171 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21172 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21173 MachineInstrBuilder MIB;
21175 // As we are creating the PHIs, we have to be careful if there is more than
21176 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21177 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21178 // That also means that PHI construction must work forward from earlier to
21179 // later, and that the code must maintain a mapping from earlier PHI's
21180 // destination registers, and the registers that went into the PHI.
21182 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21183 unsigned DestReg = MIIt->getOperand(0).getReg();
21184 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21185 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21187 // If this CMOV we are generating is the opposite condition from
21188 // the jump we generated, then we have to swap the operands for the
21189 // PHI that is going to be generated.
21190 if (MIIt->getOperand(3).getImm() == OppCC)
21191 std::swap(Op1Reg, Op2Reg);
21193 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21194 Op1Reg = RegRewriteTable[Op1Reg].first;
21196 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21197 Op2Reg = RegRewriteTable[Op2Reg].second;
21199 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21200 TII->get(X86::PHI), DestReg)
21201 .addReg(Op1Reg).addMBB(copy0MBB)
21202 .addReg(Op2Reg).addMBB(thisMBB);
21204 // Add this PHI to the rewrite table.
21205 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21208 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21209 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21210 if (CascadedCMOV) {
21211 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21212 // Copy the PHI result to the register defined by the second CMOV.
21213 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21214 DL, TII->get(TargetOpcode::COPY),
21215 CascadedCMOV->getOperand(0).getReg())
21216 .addReg(MI->getOperand(0).getReg());
21217 CascadedCMOV->eraseFromParent();
21220 // Now remove the CMOV(s).
21221 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21222 (MIIt++)->eraseFromParent();
21227 MachineBasicBlock *
21228 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21229 MachineBasicBlock *BB) const {
21230 // Combine the following atomic floating-point modification pattern:
21231 // a.store(reg OP a.load(acquire), release)
21232 // Transform them into:
21233 // OPss (%gpr), %xmm
21234 // movss %xmm, (%gpr)
21235 // Or sd equivalent for 64-bit operations.
21237 switch (MI->getOpcode()) {
21238 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21239 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21240 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21242 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21243 DebugLoc DL = MI->getDebugLoc();
21244 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21245 MachineOperand MSrc = MI->getOperand(0);
21246 unsigned VSrc = MI->getOperand(5).getReg();
21247 const MachineOperand &Disp = MI->getOperand(3);
21248 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21249 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21250 if (hasDisp && MSrc.isReg())
21251 MSrc.setIsKill(false);
21252 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21253 .addOperand(/*Base=*/MSrc)
21254 .addImm(/*Scale=*/1)
21255 .addReg(/*Index=*/0)
21256 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21258 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21259 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21261 .addOperand(/*Base=*/MSrc)
21262 .addImm(/*Scale=*/1)
21263 .addReg(/*Index=*/0)
21264 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21265 .addReg(/*Segment=*/0);
21266 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21267 MI->eraseFromParent(); // The pseudo instruction is gone now.
21271 MachineBasicBlock *
21272 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21273 MachineBasicBlock *BB) const {
21274 MachineFunction *MF = BB->getParent();
21275 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21276 DebugLoc DL = MI->getDebugLoc();
21277 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21279 assert(MF->shouldSplitStack());
21281 const bool Is64Bit = Subtarget->is64Bit();
21282 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21284 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21285 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21288 // ... [Till the alloca]
21289 // If stacklet is not large enough, jump to mallocMBB
21292 // Allocate by subtracting from RSP
21293 // Jump to continueMBB
21296 // Allocate by call to runtime
21300 // [rest of original BB]
21303 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21304 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21305 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21307 MachineRegisterInfo &MRI = MF->getRegInfo();
21308 const TargetRegisterClass *AddrRegClass =
21309 getRegClassFor(getPointerTy(MF->getDataLayout()));
21311 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21312 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21313 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21314 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21315 sizeVReg = MI->getOperand(1).getReg(),
21316 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21318 MachineFunction::iterator MBBIter = ++BB->getIterator();
21320 MF->insert(MBBIter, bumpMBB);
21321 MF->insert(MBBIter, mallocMBB);
21322 MF->insert(MBBIter, continueMBB);
21324 continueMBB->splice(continueMBB->begin(), BB,
21325 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21326 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21328 // Add code to the main basic block to check if the stack limit has been hit,
21329 // and if so, jump to mallocMBB otherwise to bumpMBB.
21330 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21331 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21332 .addReg(tmpSPVReg).addReg(sizeVReg);
21333 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21334 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21335 .addReg(SPLimitVReg);
21336 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21338 // bumpMBB simply decreases the stack pointer, since we know the current
21339 // stacklet has enough space.
21340 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21341 .addReg(SPLimitVReg);
21342 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21343 .addReg(SPLimitVReg);
21344 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21346 // Calls into a routine in libgcc to allocate more space from the heap.
21347 const uint32_t *RegMask =
21348 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21350 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21352 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21353 .addExternalSymbol("__morestack_allocate_stack_space")
21354 .addRegMask(RegMask)
21355 .addReg(X86::RDI, RegState::Implicit)
21356 .addReg(X86::RAX, RegState::ImplicitDefine);
21357 } else if (Is64Bit) {
21358 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21360 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21361 .addExternalSymbol("__morestack_allocate_stack_space")
21362 .addRegMask(RegMask)
21363 .addReg(X86::EDI, RegState::Implicit)
21364 .addReg(X86::EAX, RegState::ImplicitDefine);
21366 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21368 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21369 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21370 .addExternalSymbol("__morestack_allocate_stack_space")
21371 .addRegMask(RegMask)
21372 .addReg(X86::EAX, RegState::ImplicitDefine);
21376 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21379 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21380 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21381 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21383 // Set up the CFG correctly.
21384 BB->addSuccessor(bumpMBB);
21385 BB->addSuccessor(mallocMBB);
21386 mallocMBB->addSuccessor(continueMBB);
21387 bumpMBB->addSuccessor(continueMBB);
21389 // Take care of the PHI nodes.
21390 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21391 MI->getOperand(0).getReg())
21392 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21393 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21395 // Delete the original pseudo instruction.
21396 MI->eraseFromParent();
21399 return continueMBB;
21402 MachineBasicBlock *
21403 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21404 MachineBasicBlock *BB) const {
21405 assert(!Subtarget->isTargetMachO());
21406 DebugLoc DL = MI->getDebugLoc();
21407 MachineInstr *ResumeMI = Subtarget->getFrameLowering()->emitStackProbe(
21408 *BB->getParent(), *BB, MI, DL, false);
21409 MachineBasicBlock *ResumeBB = ResumeMI->getParent();
21410 MI->eraseFromParent(); // The pseudo instruction is gone now.
21414 MachineBasicBlock *
21415 X86TargetLowering::EmitLoweredCatchRet(MachineInstr *MI,
21416 MachineBasicBlock *BB) const {
21417 MachineFunction *MF = BB->getParent();
21418 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21419 MachineBasicBlock *TargetMBB = MI->getOperand(0).getMBB();
21420 DebugLoc DL = MI->getDebugLoc();
21422 assert(!isAsynchronousEHPersonality(
21423 classifyEHPersonality(MF->getFunction()->getPersonalityFn())) &&
21424 "SEH does not use catchret!");
21426 // Only 32-bit EH needs to worry about manually restoring stack pointers.
21427 if (!Subtarget->is32Bit())
21430 // C++ EH creates a new target block to hold the restore code, and wires up
21431 // the new block to the return destination with a normal JMP_4.
21432 MachineBasicBlock *RestoreMBB =
21433 MF->CreateMachineBasicBlock(BB->getBasicBlock());
21434 assert(BB->succ_size() == 1);
21435 MF->insert(std::next(BB->getIterator()), RestoreMBB);
21436 RestoreMBB->transferSuccessorsAndUpdatePHIs(BB);
21437 BB->addSuccessor(RestoreMBB);
21438 MI->getOperand(0).setMBB(RestoreMBB);
21440 auto RestoreMBBI = RestoreMBB->begin();
21441 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::EH_RESTORE));
21442 BuildMI(*RestoreMBB, RestoreMBBI, DL, TII.get(X86::JMP_4)).addMBB(TargetMBB);
21446 MachineBasicBlock *
21447 X86TargetLowering::EmitLoweredCatchPad(MachineInstr *MI,
21448 MachineBasicBlock *BB) const {
21449 MachineFunction *MF = BB->getParent();
21450 const Constant *PerFn = MF->getFunction()->getPersonalityFn();
21451 bool IsSEH = isAsynchronousEHPersonality(classifyEHPersonality(PerFn));
21452 // Only 32-bit SEH requires special handling for catchpad.
21453 if (IsSEH && Subtarget->is32Bit()) {
21454 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21455 DebugLoc DL = MI->getDebugLoc();
21456 BuildMI(*BB, MI, DL, TII.get(X86::EH_RESTORE));
21458 MI->eraseFromParent();
21462 MachineBasicBlock *
21463 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21464 MachineBasicBlock *BB) const {
21465 // This is pretty easy. We're taking the value that we received from
21466 // our load from the relocation, sticking it in either RDI (x86-64)
21467 // or EAX and doing an indirect call. The return value will then
21468 // be in the normal return register.
21469 MachineFunction *F = BB->getParent();
21470 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21471 DebugLoc DL = MI->getDebugLoc();
21473 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21474 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21476 // Get a register mask for the lowered call.
21477 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21478 // proper register mask.
21479 const uint32_t *RegMask =
21480 Subtarget->is64Bit() ?
21481 Subtarget->getRegisterInfo()->getDarwinTLSCallPreservedMask() :
21482 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21483 if (Subtarget->is64Bit()) {
21484 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21485 TII->get(X86::MOV64rm), X86::RDI)
21487 .addImm(0).addReg(0)
21488 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21489 MI->getOperand(3).getTargetFlags())
21491 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21492 addDirectMem(MIB, X86::RDI);
21493 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21494 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21495 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21496 TII->get(X86::MOV32rm), X86::EAX)
21498 .addImm(0).addReg(0)
21499 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21500 MI->getOperand(3).getTargetFlags())
21502 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21503 addDirectMem(MIB, X86::EAX);
21504 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21506 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21507 TII->get(X86::MOV32rm), X86::EAX)
21508 .addReg(TII->getGlobalBaseReg(F))
21509 .addImm(0).addReg(0)
21510 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21511 MI->getOperand(3).getTargetFlags())
21513 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21514 addDirectMem(MIB, X86::EAX);
21515 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21518 MI->eraseFromParent(); // The pseudo instruction is gone now.
21522 MachineBasicBlock *
21523 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21524 MachineBasicBlock *MBB) const {
21525 DebugLoc DL = MI->getDebugLoc();
21526 MachineFunction *MF = MBB->getParent();
21527 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21528 MachineRegisterInfo &MRI = MF->getRegInfo();
21530 const BasicBlock *BB = MBB->getBasicBlock();
21531 MachineFunction::iterator I = ++MBB->getIterator();
21533 // Memory Reference
21534 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21535 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21538 unsigned MemOpndSlot = 0;
21540 unsigned CurOp = 0;
21542 DstReg = MI->getOperand(CurOp++).getReg();
21543 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21544 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21545 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21546 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21548 MemOpndSlot = CurOp;
21550 MVT PVT = getPointerTy(MF->getDataLayout());
21551 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21552 "Invalid Pointer Size!");
21554 // For v = setjmp(buf), we generate
21557 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21558 // SjLjSetup restoreMBB
21564 // v = phi(main, restore)
21567 // if base pointer being used, load it from frame
21570 MachineBasicBlock *thisMBB = MBB;
21571 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21572 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21573 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21574 MF->insert(I, mainMBB);
21575 MF->insert(I, sinkMBB);
21576 MF->push_back(restoreMBB);
21577 restoreMBB->setHasAddressTaken();
21579 MachineInstrBuilder MIB;
21581 // Transfer the remainder of BB and its successor edges to sinkMBB.
21582 sinkMBB->splice(sinkMBB->begin(), MBB,
21583 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21584 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21587 unsigned PtrStoreOpc = 0;
21588 unsigned LabelReg = 0;
21589 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21590 Reloc::Model RM = MF->getTarget().getRelocationModel();
21591 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21592 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21594 // Prepare IP either in reg or imm.
21595 if (!UseImmLabel) {
21596 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21597 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21598 LabelReg = MRI.createVirtualRegister(PtrRC);
21599 if (Subtarget->is64Bit()) {
21600 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21604 .addMBB(restoreMBB)
21607 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21608 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21609 .addReg(XII->getGlobalBaseReg(MF))
21612 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21616 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21618 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21619 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21620 if (i == X86::AddrDisp)
21621 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21623 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21626 MIB.addReg(LabelReg);
21628 MIB.addMBB(restoreMBB);
21629 MIB.setMemRefs(MMOBegin, MMOEnd);
21631 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21632 .addMBB(restoreMBB);
21634 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21635 MIB.addRegMask(RegInfo->getNoPreservedMask());
21636 thisMBB->addSuccessor(mainMBB);
21637 thisMBB->addSuccessor(restoreMBB);
21641 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21642 mainMBB->addSuccessor(sinkMBB);
21645 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21646 TII->get(X86::PHI), DstReg)
21647 .addReg(mainDstReg).addMBB(mainMBB)
21648 .addReg(restoreDstReg).addMBB(restoreMBB);
21651 if (RegInfo->hasBasePointer(*MF)) {
21652 const bool Uses64BitFramePtr =
21653 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21654 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21655 X86FI->setRestoreBasePointer(MF);
21656 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21657 unsigned BasePtr = RegInfo->getBaseRegister();
21658 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21659 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21660 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21661 .setMIFlag(MachineInstr::FrameSetup);
21663 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21664 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21665 restoreMBB->addSuccessor(sinkMBB);
21667 MI->eraseFromParent();
21671 MachineBasicBlock *
21672 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21673 MachineBasicBlock *MBB) const {
21674 DebugLoc DL = MI->getDebugLoc();
21675 MachineFunction *MF = MBB->getParent();
21676 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21677 MachineRegisterInfo &MRI = MF->getRegInfo();
21679 // Memory Reference
21680 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21681 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21683 MVT PVT = getPointerTy(MF->getDataLayout());
21684 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21685 "Invalid Pointer Size!");
21687 const TargetRegisterClass *RC =
21688 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21689 unsigned Tmp = MRI.createVirtualRegister(RC);
21690 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21691 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21692 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21693 unsigned SP = RegInfo->getStackRegister();
21695 MachineInstrBuilder MIB;
21697 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21698 const int64_t SPOffset = 2 * PVT.getStoreSize();
21700 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21701 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21704 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21705 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21706 MIB.addOperand(MI->getOperand(i));
21707 MIB.setMemRefs(MMOBegin, MMOEnd);
21709 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21710 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21711 if (i == X86::AddrDisp)
21712 MIB.addDisp(MI->getOperand(i), LabelOffset);
21714 MIB.addOperand(MI->getOperand(i));
21716 MIB.setMemRefs(MMOBegin, MMOEnd);
21718 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21719 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21720 if (i == X86::AddrDisp)
21721 MIB.addDisp(MI->getOperand(i), SPOffset);
21723 MIB.addOperand(MI->getOperand(i));
21725 MIB.setMemRefs(MMOBegin, MMOEnd);
21727 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21729 MI->eraseFromParent();
21733 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21734 // accumulator loops. Writing back to the accumulator allows the coalescer
21735 // to remove extra copies in the loop.
21736 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21737 MachineBasicBlock *
21738 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21739 MachineBasicBlock *MBB) const {
21740 MachineOperand &AddendOp = MI->getOperand(3);
21742 // Bail out early if the addend isn't a register - we can't switch these.
21743 if (!AddendOp.isReg())
21746 MachineFunction &MF = *MBB->getParent();
21747 MachineRegisterInfo &MRI = MF.getRegInfo();
21749 // Check whether the addend is defined by a PHI:
21750 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21751 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21752 if (!AddendDef.isPHI())
21755 // Look for the following pattern:
21757 // %addend = phi [%entry, 0], [%loop, %result]
21759 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21763 // %addend = phi [%entry, 0], [%loop, %result]
21765 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21767 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21768 assert(AddendDef.getOperand(i).isReg());
21769 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21770 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21771 if (&PHISrcInst == MI) {
21772 // Found a matching instruction.
21773 unsigned NewFMAOpc = 0;
21774 switch (MI->getOpcode()) {
21775 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21776 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21777 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21778 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21779 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21780 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21781 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21782 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21783 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21784 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21785 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21786 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21787 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21788 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21789 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21790 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21791 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21792 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21793 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21794 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21796 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21797 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21798 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21799 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21800 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21801 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21802 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21803 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21804 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21805 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21806 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21807 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21808 default: llvm_unreachable("Unrecognized FMA variant.");
21811 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21812 MachineInstrBuilder MIB =
21813 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21814 .addOperand(MI->getOperand(0))
21815 .addOperand(MI->getOperand(3))
21816 .addOperand(MI->getOperand(2))
21817 .addOperand(MI->getOperand(1));
21818 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21819 MI->eraseFromParent();
21826 MachineBasicBlock *
21827 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21828 MachineBasicBlock *BB) const {
21829 switch (MI->getOpcode()) {
21830 default: llvm_unreachable("Unexpected instr type to insert");
21831 case X86::TAILJMPd64:
21832 case X86::TAILJMPr64:
21833 case X86::TAILJMPm64:
21834 case X86::TAILJMPd64_REX:
21835 case X86::TAILJMPr64_REX:
21836 case X86::TAILJMPm64_REX:
21837 llvm_unreachable("TAILJMP64 would not be touched here.");
21838 case X86::TCRETURNdi64:
21839 case X86::TCRETURNri64:
21840 case X86::TCRETURNmi64:
21842 case X86::WIN_ALLOCA:
21843 return EmitLoweredWinAlloca(MI, BB);
21844 case X86::CATCHRET:
21845 return EmitLoweredCatchRet(MI, BB);
21846 case X86::CATCHPAD:
21847 return EmitLoweredCatchPad(MI, BB);
21848 case X86::SEG_ALLOCA_32:
21849 case X86::SEG_ALLOCA_64:
21850 return EmitLoweredSegAlloca(MI, BB);
21851 case X86::TLSCall_32:
21852 case X86::TLSCall_64:
21853 return EmitLoweredTLSCall(MI, BB);
21854 case X86::CMOV_FR32:
21855 case X86::CMOV_FR64:
21856 case X86::CMOV_GR8:
21857 case X86::CMOV_GR16:
21858 case X86::CMOV_GR32:
21859 case X86::CMOV_RFP32:
21860 case X86::CMOV_RFP64:
21861 case X86::CMOV_RFP80:
21862 case X86::CMOV_V2F64:
21863 case X86::CMOV_V2I64:
21864 case X86::CMOV_V4F32:
21865 case X86::CMOV_V4F64:
21866 case X86::CMOV_V4I64:
21867 case X86::CMOV_V16F32:
21868 case X86::CMOV_V8F32:
21869 case X86::CMOV_V8F64:
21870 case X86::CMOV_V8I64:
21871 case X86::CMOV_V8I1:
21872 case X86::CMOV_V16I1:
21873 case X86::CMOV_V32I1:
21874 case X86::CMOV_V64I1:
21875 return EmitLoweredSelect(MI, BB);
21877 case X86::RELEASE_FADD32mr:
21878 case X86::RELEASE_FADD64mr:
21879 return EmitLoweredAtomicFP(MI, BB);
21881 case X86::FP32_TO_INT16_IN_MEM:
21882 case X86::FP32_TO_INT32_IN_MEM:
21883 case X86::FP32_TO_INT64_IN_MEM:
21884 case X86::FP64_TO_INT16_IN_MEM:
21885 case X86::FP64_TO_INT32_IN_MEM:
21886 case X86::FP64_TO_INT64_IN_MEM:
21887 case X86::FP80_TO_INT16_IN_MEM:
21888 case X86::FP80_TO_INT32_IN_MEM:
21889 case X86::FP80_TO_INT64_IN_MEM: {
21890 MachineFunction *F = BB->getParent();
21891 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21892 DebugLoc DL = MI->getDebugLoc();
21894 // Change the floating point control register to use "round towards zero"
21895 // mode when truncating to an integer value.
21896 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21897 addFrameReference(BuildMI(*BB, MI, DL,
21898 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21900 // Load the old value of the high byte of the control word...
21902 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21903 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21906 // Set the high part to be round to zero...
21907 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21910 // Reload the modified control word now...
21911 addFrameReference(BuildMI(*BB, MI, DL,
21912 TII->get(X86::FLDCW16m)), CWFrameIdx);
21914 // Restore the memory image of control word to original value
21915 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21918 // Get the X86 opcode to use.
21920 switch (MI->getOpcode()) {
21921 default: llvm_unreachable("illegal opcode!");
21922 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21923 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21924 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21925 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21926 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21927 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21928 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21929 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21930 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21934 MachineOperand &Op = MI->getOperand(0);
21936 AM.BaseType = X86AddressMode::RegBase;
21937 AM.Base.Reg = Op.getReg();
21939 AM.BaseType = X86AddressMode::FrameIndexBase;
21940 AM.Base.FrameIndex = Op.getIndex();
21942 Op = MI->getOperand(1);
21944 AM.Scale = Op.getImm();
21945 Op = MI->getOperand(2);
21947 AM.IndexReg = Op.getImm();
21948 Op = MI->getOperand(3);
21949 if (Op.isGlobal()) {
21950 AM.GV = Op.getGlobal();
21952 AM.Disp = Op.getImm();
21954 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21955 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21957 // Reload the original control word now.
21958 addFrameReference(BuildMI(*BB, MI, DL,
21959 TII->get(X86::FLDCW16m)), CWFrameIdx);
21961 MI->eraseFromParent(); // The pseudo instruction is gone now.
21964 // String/text processing lowering.
21965 case X86::PCMPISTRM128REG:
21966 case X86::VPCMPISTRM128REG:
21967 case X86::PCMPISTRM128MEM:
21968 case X86::VPCMPISTRM128MEM:
21969 case X86::PCMPESTRM128REG:
21970 case X86::VPCMPESTRM128REG:
21971 case X86::PCMPESTRM128MEM:
21972 case X86::VPCMPESTRM128MEM:
21973 assert(Subtarget->hasSSE42() &&
21974 "Target must have SSE4.2 or AVX features enabled");
21975 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21977 // String/text processing lowering.
21978 case X86::PCMPISTRIREG:
21979 case X86::VPCMPISTRIREG:
21980 case X86::PCMPISTRIMEM:
21981 case X86::VPCMPISTRIMEM:
21982 case X86::PCMPESTRIREG:
21983 case X86::VPCMPESTRIREG:
21984 case X86::PCMPESTRIMEM:
21985 case X86::VPCMPESTRIMEM:
21986 assert(Subtarget->hasSSE42() &&
21987 "Target must have SSE4.2 or AVX features enabled");
21988 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21990 // Thread synchronization.
21992 return EmitMonitor(MI, BB, Subtarget);
21996 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21998 case X86::VASTART_SAVE_XMM_REGS:
21999 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
22001 case X86::VAARG_64:
22002 return EmitVAARG64WithCustomInserter(MI, BB);
22004 case X86::EH_SjLj_SetJmp32:
22005 case X86::EH_SjLj_SetJmp64:
22006 return emitEHSjLjSetJmp(MI, BB);
22008 case X86::EH_SjLj_LongJmp32:
22009 case X86::EH_SjLj_LongJmp64:
22010 return emitEHSjLjLongJmp(MI, BB);
22012 case TargetOpcode::STATEPOINT:
22013 // As an implementation detail, STATEPOINT shares the STACKMAP format at
22014 // this point in the process. We diverge later.
22015 return emitPatchPoint(MI, BB);
22017 case TargetOpcode::STACKMAP:
22018 case TargetOpcode::PATCHPOINT:
22019 return emitPatchPoint(MI, BB);
22021 case X86::VFMADDPDr213r:
22022 case X86::VFMADDPSr213r:
22023 case X86::VFMADDSDr213r:
22024 case X86::VFMADDSSr213r:
22025 case X86::VFMSUBPDr213r:
22026 case X86::VFMSUBPSr213r:
22027 case X86::VFMSUBSDr213r:
22028 case X86::VFMSUBSSr213r:
22029 case X86::VFNMADDPDr213r:
22030 case X86::VFNMADDPSr213r:
22031 case X86::VFNMADDSDr213r:
22032 case X86::VFNMADDSSr213r:
22033 case X86::VFNMSUBPDr213r:
22034 case X86::VFNMSUBPSr213r:
22035 case X86::VFNMSUBSDr213r:
22036 case X86::VFNMSUBSSr213r:
22037 case X86::VFMADDSUBPDr213r:
22038 case X86::VFMADDSUBPSr213r:
22039 case X86::VFMSUBADDPDr213r:
22040 case X86::VFMSUBADDPSr213r:
22041 case X86::VFMADDPDr213rY:
22042 case X86::VFMADDPSr213rY:
22043 case X86::VFMSUBPDr213rY:
22044 case X86::VFMSUBPSr213rY:
22045 case X86::VFNMADDPDr213rY:
22046 case X86::VFNMADDPSr213rY:
22047 case X86::VFNMSUBPDr213rY:
22048 case X86::VFNMSUBPSr213rY:
22049 case X86::VFMADDSUBPDr213rY:
22050 case X86::VFMADDSUBPSr213rY:
22051 case X86::VFMSUBADDPDr213rY:
22052 case X86::VFMSUBADDPSr213rY:
22053 return emitFMA3Instr(MI, BB);
22057 //===----------------------------------------------------------------------===//
22058 // X86 Optimization Hooks
22059 //===----------------------------------------------------------------------===//
22061 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22064 const SelectionDAG &DAG,
22065 unsigned Depth) const {
22066 unsigned BitWidth = KnownZero.getBitWidth();
22067 unsigned Opc = Op.getOpcode();
22068 assert((Opc >= ISD::BUILTIN_OP_END ||
22069 Opc == ISD::INTRINSIC_WO_CHAIN ||
22070 Opc == ISD::INTRINSIC_W_CHAIN ||
22071 Opc == ISD::INTRINSIC_VOID) &&
22072 "Should use MaskedValueIsZero if you don't know whether Op"
22073 " is a target node!");
22075 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22089 // These nodes' second result is a boolean.
22090 if (Op.getResNo() == 0)
22093 case X86ISD::SETCC:
22094 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22096 case ISD::INTRINSIC_WO_CHAIN: {
22097 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22098 unsigned NumLoBits = 0;
22101 case Intrinsic::x86_sse_movmsk_ps:
22102 case Intrinsic::x86_avx_movmsk_ps_256:
22103 case Intrinsic::x86_sse2_movmsk_pd:
22104 case Intrinsic::x86_avx_movmsk_pd_256:
22105 case Intrinsic::x86_mmx_pmovmskb:
22106 case Intrinsic::x86_sse2_pmovmskb_128:
22107 case Intrinsic::x86_avx2_pmovmskb: {
22108 // High bits of movmskp{s|d}, pmovmskb are known zero.
22110 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22111 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22112 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22113 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22114 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22115 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22116 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22117 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22119 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22128 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22130 const SelectionDAG &,
22131 unsigned Depth) const {
22132 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22133 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22134 return Op.getValueType().getScalarSizeInBits();
22140 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22141 /// node is a GlobalAddress + offset.
22142 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22143 const GlobalValue* &GA,
22144 int64_t &Offset) const {
22145 if (N->getOpcode() == X86ISD::Wrapper) {
22146 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22147 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22148 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22152 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22155 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22156 /// same as extracting the high 128-bit part of 256-bit vector and then
22157 /// inserting the result into the low part of a new 256-bit vector
22158 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22159 EVT VT = SVOp->getValueType(0);
22160 unsigned NumElems = VT.getVectorNumElements();
22162 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22163 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22164 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22165 SVOp->getMaskElt(j) >= 0)
22171 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22172 /// same as extracting the low 128-bit part of 256-bit vector and then
22173 /// inserting the result into the high part of a new 256-bit vector
22174 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22175 EVT VT = SVOp->getValueType(0);
22176 unsigned NumElems = VT.getVectorNumElements();
22178 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22179 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22180 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22181 SVOp->getMaskElt(j) >= 0)
22187 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22188 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22189 TargetLowering::DAGCombinerInfo &DCI,
22190 const X86Subtarget* Subtarget) {
22192 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22193 SDValue V1 = SVOp->getOperand(0);
22194 SDValue V2 = SVOp->getOperand(1);
22195 EVT VT = SVOp->getValueType(0);
22196 unsigned NumElems = VT.getVectorNumElements();
22198 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22199 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22203 // V UNDEF BUILD_VECTOR UNDEF
22205 // CONCAT_VECTOR CONCAT_VECTOR
22208 // RESULT: V + zero extended
22210 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22211 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22212 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22215 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22218 // To match the shuffle mask, the first half of the mask should
22219 // be exactly the first vector, and all the rest a splat with the
22220 // first element of the second one.
22221 for (unsigned i = 0; i != NumElems/2; ++i)
22222 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22223 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22226 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22227 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22228 if (Ld->hasNUsesOfValue(1, 0)) {
22229 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22230 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22232 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22234 Ld->getPointerInfo(),
22235 Ld->getAlignment(),
22236 false/*isVolatile*/, true/*ReadMem*/,
22237 false/*WriteMem*/);
22239 // Make sure the newly-created LOAD is in the same position as Ld in
22240 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22241 // and update uses of Ld's output chain to use the TokenFactor.
22242 if (Ld->hasAnyUseOfValue(1)) {
22243 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22244 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22245 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22246 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22247 SDValue(ResNode.getNode(), 1));
22250 return DAG.getBitcast(VT, ResNode);
22254 // Emit a zeroed vector and insert the desired subvector on its
22256 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22257 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22258 return DCI.CombineTo(N, InsV);
22261 //===--------------------------------------------------------------------===//
22262 // Combine some shuffles into subvector extracts and inserts:
22265 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22266 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22267 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22268 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22269 return DCI.CombineTo(N, InsV);
22272 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22273 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22274 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22275 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22276 return DCI.CombineTo(N, InsV);
22282 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22285 /// This is the leaf of the recursive combinine below. When we have found some
22286 /// chain of single-use x86 shuffle instructions and accumulated the combined
22287 /// shuffle mask represented by them, this will try to pattern match that mask
22288 /// into either a single instruction if there is a special purpose instruction
22289 /// for this operation, or into a PSHUFB instruction which is a fully general
22290 /// instruction but should only be used to replace chains over a certain depth.
22291 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22292 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22293 TargetLowering::DAGCombinerInfo &DCI,
22294 const X86Subtarget *Subtarget) {
22295 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22297 // Find the operand that enters the chain. Note that multiple uses are OK
22298 // here, we're not going to remove the operand we find.
22299 SDValue Input = Op.getOperand(0);
22300 while (Input.getOpcode() == ISD::BITCAST)
22301 Input = Input.getOperand(0);
22303 MVT VT = Input.getSimpleValueType();
22304 MVT RootVT = Root.getSimpleValueType();
22307 if (Mask.size() == 1) {
22308 int Index = Mask[0];
22309 assert((Index >= 0 || Index == SM_SentinelUndef ||
22310 Index == SM_SentinelZero) &&
22311 "Invalid shuffle index found!");
22313 // We may end up with an accumulated mask of size 1 as a result of
22314 // widening of shuffle operands (see function canWidenShuffleElements).
22315 // If the only shuffle index is equal to SM_SentinelZero then propagate
22316 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22317 // mask, and therefore the entire chain of shuffles can be folded away.
22318 if (Index == SM_SentinelZero)
22319 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22321 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22326 // Use the float domain if the operand type is a floating point type.
22327 bool FloatDomain = VT.isFloatingPoint();
22329 // For floating point shuffles, we don't have free copies in the shuffle
22330 // instructions or the ability to load as part of the instruction, so
22331 // canonicalize their shuffles to UNPCK or MOV variants.
22333 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22334 // vectors because it can have a load folded into it that UNPCK cannot. This
22335 // doesn't preclude something switching to the shorter encoding post-RA.
22337 // FIXME: Should teach these routines about AVX vector widths.
22338 if (FloatDomain && VT.is128BitVector()) {
22339 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22340 bool Lo = Mask.equals({0, 0});
22343 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22344 // is no slower than UNPCKLPD but has the option to fold the input operand
22345 // into even an unaligned memory load.
22346 if (Lo && Subtarget->hasSSE3()) {
22347 Shuffle = X86ISD::MOVDDUP;
22348 ShuffleVT = MVT::v2f64;
22350 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22351 // than the UNPCK variants.
22352 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22353 ShuffleVT = MVT::v4f32;
22355 if (Depth == 1 && Root->getOpcode() == Shuffle)
22356 return false; // Nothing to do!
22357 Op = DAG.getBitcast(ShuffleVT, Input);
22358 DCI.AddToWorklist(Op.getNode());
22359 if (Shuffle == X86ISD::MOVDDUP)
22360 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22362 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22363 DCI.AddToWorklist(Op.getNode());
22364 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22368 if (Subtarget->hasSSE3() &&
22369 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22370 bool Lo = Mask.equals({0, 0, 2, 2});
22371 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22372 MVT ShuffleVT = MVT::v4f32;
22373 if (Depth == 1 && Root->getOpcode() == Shuffle)
22374 return false; // Nothing to do!
22375 Op = DAG.getBitcast(ShuffleVT, Input);
22376 DCI.AddToWorklist(Op.getNode());
22377 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22378 DCI.AddToWorklist(Op.getNode());
22379 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22383 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22384 bool Lo = Mask.equals({0, 0, 1, 1});
22385 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22386 MVT ShuffleVT = MVT::v4f32;
22387 if (Depth == 1 && Root->getOpcode() == Shuffle)
22388 return false; // Nothing to do!
22389 Op = DAG.getBitcast(ShuffleVT, Input);
22390 DCI.AddToWorklist(Op.getNode());
22391 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22392 DCI.AddToWorklist(Op.getNode());
22393 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22399 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22400 // variants as none of these have single-instruction variants that are
22401 // superior to the UNPCK formulation.
22402 if (!FloatDomain && VT.is128BitVector() &&
22403 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22404 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22405 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22407 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22408 bool Lo = Mask[0] == 0;
22409 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22410 if (Depth == 1 && Root->getOpcode() == Shuffle)
22411 return false; // Nothing to do!
22413 switch (Mask.size()) {
22415 ShuffleVT = MVT::v8i16;
22418 ShuffleVT = MVT::v16i8;
22421 llvm_unreachable("Impossible mask size!");
22423 Op = DAG.getBitcast(ShuffleVT, Input);
22424 DCI.AddToWorklist(Op.getNode());
22425 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22426 DCI.AddToWorklist(Op.getNode());
22427 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22432 // Don't try to re-form single instruction chains under any circumstances now
22433 // that we've done encoding canonicalization for them.
22437 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22438 // can replace them with a single PSHUFB instruction profitably. Intel's
22439 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22440 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22441 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22442 SmallVector<SDValue, 16> PSHUFBMask;
22443 int NumBytes = VT.getSizeInBits() / 8;
22444 int Ratio = NumBytes / Mask.size();
22445 for (int i = 0; i < NumBytes; ++i) {
22446 if (Mask[i / Ratio] == SM_SentinelUndef) {
22447 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22450 int M = Mask[i / Ratio] != SM_SentinelZero
22451 ? Ratio * Mask[i / Ratio] + i % Ratio
22453 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22455 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22456 Op = DAG.getBitcast(ByteVT, Input);
22457 DCI.AddToWorklist(Op.getNode());
22458 SDValue PSHUFBMaskOp =
22459 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22460 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22461 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22462 DCI.AddToWorklist(Op.getNode());
22463 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22468 // Failed to find any combines.
22472 /// \brief Fully generic combining of x86 shuffle instructions.
22474 /// This should be the last combine run over the x86 shuffle instructions. Once
22475 /// they have been fully optimized, this will recursively consider all chains
22476 /// of single-use shuffle instructions, build a generic model of the cumulative
22477 /// shuffle operation, and check for simpler instructions which implement this
22478 /// operation. We use this primarily for two purposes:
22480 /// 1) Collapse generic shuffles to specialized single instructions when
22481 /// equivalent. In most cases, this is just an encoding size win, but
22482 /// sometimes we will collapse multiple generic shuffles into a single
22483 /// special-purpose shuffle.
22484 /// 2) Look for sequences of shuffle instructions with 3 or more total
22485 /// instructions, and replace them with the slightly more expensive SSSE3
22486 /// PSHUFB instruction if available. We do this as the last combining step
22487 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22488 /// a suitable short sequence of other instructions. The PHUFB will either
22489 /// use a register or have to read from memory and so is slightly (but only
22490 /// slightly) more expensive than the other shuffle instructions.
22492 /// Because this is inherently a quadratic operation (for each shuffle in
22493 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22494 /// This should never be an issue in practice as the shuffle lowering doesn't
22495 /// produce sequences of more than 8 instructions.
22497 /// FIXME: We will currently miss some cases where the redundant shuffling
22498 /// would simplify under the threshold for PSHUFB formation because of
22499 /// combine-ordering. To fix this, we should do the redundant instruction
22500 /// combining in this recursive walk.
22501 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22502 ArrayRef<int> RootMask,
22503 int Depth, bool HasPSHUFB,
22505 TargetLowering::DAGCombinerInfo &DCI,
22506 const X86Subtarget *Subtarget) {
22507 // Bound the depth of our recursive combine because this is ultimately
22508 // quadratic in nature.
22512 // Directly rip through bitcasts to find the underlying operand.
22513 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22514 Op = Op.getOperand(0);
22516 MVT VT = Op.getSimpleValueType();
22517 if (!VT.isVector())
22518 return false; // Bail if we hit a non-vector.
22520 assert(Root.getSimpleValueType().isVector() &&
22521 "Shuffles operate on vector types!");
22522 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22523 "Can only combine shuffles of the same vector register size.");
22525 if (!isTargetShuffle(Op.getOpcode()))
22527 SmallVector<int, 16> OpMask;
22529 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22530 // We only can combine unary shuffles which we can decode the mask for.
22531 if (!HaveMask || !IsUnary)
22534 assert(VT.getVectorNumElements() == OpMask.size() &&
22535 "Different mask size from vector size!");
22536 assert(((RootMask.size() > OpMask.size() &&
22537 RootMask.size() % OpMask.size() == 0) ||
22538 (OpMask.size() > RootMask.size() &&
22539 OpMask.size() % RootMask.size() == 0) ||
22540 OpMask.size() == RootMask.size()) &&
22541 "The smaller number of elements must divide the larger.");
22542 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22543 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22544 assert(((RootRatio == 1 && OpRatio == 1) ||
22545 (RootRatio == 1) != (OpRatio == 1)) &&
22546 "Must not have a ratio for both incoming and op masks!");
22548 SmallVector<int, 16> Mask;
22549 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22551 // Merge this shuffle operation's mask into our accumulated mask. Note that
22552 // this shuffle's mask will be the first applied to the input, followed by the
22553 // root mask to get us all the way to the root value arrangement. The reason
22554 // for this order is that we are recursing up the operation chain.
22555 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22556 int RootIdx = i / RootRatio;
22557 if (RootMask[RootIdx] < 0) {
22558 // This is a zero or undef lane, we're done.
22559 Mask.push_back(RootMask[RootIdx]);
22563 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22564 int OpIdx = RootMaskedIdx / OpRatio;
22565 if (OpMask[OpIdx] < 0) {
22566 // The incoming lanes are zero or undef, it doesn't matter which ones we
22568 Mask.push_back(OpMask[OpIdx]);
22572 // Ok, we have non-zero lanes, map them through.
22573 Mask.push_back(OpMask[OpIdx] * OpRatio +
22574 RootMaskedIdx % OpRatio);
22577 // See if we can recurse into the operand to combine more things.
22578 switch (Op.getOpcode()) {
22579 case X86ISD::PSHUFB:
22581 case X86ISD::PSHUFD:
22582 case X86ISD::PSHUFHW:
22583 case X86ISD::PSHUFLW:
22584 if (Op.getOperand(0).hasOneUse() &&
22585 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22586 HasPSHUFB, DAG, DCI, Subtarget))
22590 case X86ISD::UNPCKL:
22591 case X86ISD::UNPCKH:
22592 assert(Op.getOperand(0) == Op.getOperand(1) &&
22593 "We only combine unary shuffles!");
22594 // We can't check for single use, we have to check that this shuffle is the
22596 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22597 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22598 HasPSHUFB, DAG, DCI, Subtarget))
22603 // Minor canonicalization of the accumulated shuffle mask to make it easier
22604 // to match below. All this does is detect masks with squential pairs of
22605 // elements, and shrink them to the half-width mask. It does this in a loop
22606 // so it will reduce the size of the mask to the minimal width mask which
22607 // performs an equivalent shuffle.
22608 SmallVector<int, 16> WidenedMask;
22609 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22610 Mask = std::move(WidenedMask);
22611 WidenedMask.clear();
22614 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22618 /// \brief Get the PSHUF-style mask from PSHUF node.
22620 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22621 /// PSHUF-style masks that can be reused with such instructions.
22622 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22623 MVT VT = N.getSimpleValueType();
22624 SmallVector<int, 4> Mask;
22626 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22630 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22631 // matter. Check that the upper masks are repeats and remove them.
22632 if (VT.getSizeInBits() > 128) {
22633 int LaneElts = 128 / VT.getScalarSizeInBits();
22635 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22636 for (int j = 0; j < LaneElts; ++j)
22637 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22638 "Mask doesn't repeat in high 128-bit lanes!");
22640 Mask.resize(LaneElts);
22643 switch (N.getOpcode()) {
22644 case X86ISD::PSHUFD:
22646 case X86ISD::PSHUFLW:
22649 case X86ISD::PSHUFHW:
22650 Mask.erase(Mask.begin(), Mask.begin() + 4);
22651 for (int &M : Mask)
22655 llvm_unreachable("No valid shuffle instruction found!");
22659 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22661 /// We walk up the chain and look for a combinable shuffle, skipping over
22662 /// shuffles that we could hoist this shuffle's transformation past without
22663 /// altering anything.
22665 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22667 TargetLowering::DAGCombinerInfo &DCI) {
22668 assert(N.getOpcode() == X86ISD::PSHUFD &&
22669 "Called with something other than an x86 128-bit half shuffle!");
22672 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22673 // of the shuffles in the chain so that we can form a fresh chain to replace
22675 SmallVector<SDValue, 8> Chain;
22676 SDValue V = N.getOperand(0);
22677 for (; V.hasOneUse(); V = V.getOperand(0)) {
22678 switch (V.getOpcode()) {
22680 return SDValue(); // Nothing combined!
22683 // Skip bitcasts as we always know the type for the target specific
22687 case X86ISD::PSHUFD:
22688 // Found another dword shuffle.
22691 case X86ISD::PSHUFLW:
22692 // Check that the low words (being shuffled) are the identity in the
22693 // dword shuffle, and the high words are self-contained.
22694 if (Mask[0] != 0 || Mask[1] != 1 ||
22695 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22698 Chain.push_back(V);
22701 case X86ISD::PSHUFHW:
22702 // Check that the high words (being shuffled) are the identity in the
22703 // dword shuffle, and the low words are self-contained.
22704 if (Mask[2] != 2 || Mask[3] != 3 ||
22705 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22708 Chain.push_back(V);
22711 case X86ISD::UNPCKL:
22712 case X86ISD::UNPCKH:
22713 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22714 // shuffle into a preceding word shuffle.
22715 if (V.getSimpleValueType().getVectorElementType() != MVT::i8 &&
22716 V.getSimpleValueType().getVectorElementType() != MVT::i16)
22719 // Search for a half-shuffle which we can combine with.
22720 unsigned CombineOp =
22721 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22722 if (V.getOperand(0) != V.getOperand(1) ||
22723 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22725 Chain.push_back(V);
22726 V = V.getOperand(0);
22728 switch (V.getOpcode()) {
22730 return SDValue(); // Nothing to combine.
22732 case X86ISD::PSHUFLW:
22733 case X86ISD::PSHUFHW:
22734 if (V.getOpcode() == CombineOp)
22737 Chain.push_back(V);
22741 V = V.getOperand(0);
22745 } while (V.hasOneUse());
22748 // Break out of the loop if we break out of the switch.
22752 if (!V.hasOneUse())
22753 // We fell out of the loop without finding a viable combining instruction.
22756 // Merge this node's mask and our incoming mask.
22757 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22758 for (int &M : Mask)
22760 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22761 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22763 // Rebuild the chain around this new shuffle.
22764 while (!Chain.empty()) {
22765 SDValue W = Chain.pop_back_val();
22767 if (V.getValueType() != W.getOperand(0).getValueType())
22768 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22770 switch (W.getOpcode()) {
22772 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22774 case X86ISD::UNPCKL:
22775 case X86ISD::UNPCKH:
22776 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22779 case X86ISD::PSHUFD:
22780 case X86ISD::PSHUFLW:
22781 case X86ISD::PSHUFHW:
22782 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22786 if (V.getValueType() != N.getValueType())
22787 V = DAG.getBitcast(N.getValueType(), V);
22789 // Return the new chain to replace N.
22793 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22796 /// We walk up the chain, skipping shuffles of the other half and looking
22797 /// through shuffles which switch halves trying to find a shuffle of the same
22798 /// pair of dwords.
22799 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22801 TargetLowering::DAGCombinerInfo &DCI) {
22803 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22804 "Called with something other than an x86 128-bit half shuffle!");
22806 unsigned CombineOpcode = N.getOpcode();
22808 // Walk up a single-use chain looking for a combinable shuffle.
22809 SDValue V = N.getOperand(0);
22810 for (; V.hasOneUse(); V = V.getOperand(0)) {
22811 switch (V.getOpcode()) {
22813 return false; // Nothing combined!
22816 // Skip bitcasts as we always know the type for the target specific
22820 case X86ISD::PSHUFLW:
22821 case X86ISD::PSHUFHW:
22822 if (V.getOpcode() == CombineOpcode)
22825 // Other-half shuffles are no-ops.
22828 // Break out of the loop if we break out of the switch.
22832 if (!V.hasOneUse())
22833 // We fell out of the loop without finding a viable combining instruction.
22836 // Combine away the bottom node as its shuffle will be accumulated into
22837 // a preceding shuffle.
22838 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22840 // Record the old value.
22843 // Merge this node's mask and our incoming mask (adjusted to account for all
22844 // the pshufd instructions encountered).
22845 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22846 for (int &M : Mask)
22848 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22849 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22851 // Check that the shuffles didn't cancel each other out. If not, we need to
22852 // combine to the new one.
22854 // Replace the combinable shuffle with the combined one, updating all users
22855 // so that we re-evaluate the chain here.
22856 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22861 /// \brief Try to combine x86 target specific shuffles.
22862 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22863 TargetLowering::DAGCombinerInfo &DCI,
22864 const X86Subtarget *Subtarget) {
22866 MVT VT = N.getSimpleValueType();
22867 SmallVector<int, 4> Mask;
22869 switch (N.getOpcode()) {
22870 case X86ISD::PSHUFD:
22871 case X86ISD::PSHUFLW:
22872 case X86ISD::PSHUFHW:
22873 Mask = getPSHUFShuffleMask(N);
22874 assert(Mask.size() == 4);
22880 // Nuke no-op shuffles that show up after combining.
22881 if (isNoopShuffleMask(Mask))
22882 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22884 // Look for simplifications involving one or two shuffle instructions.
22885 SDValue V = N.getOperand(0);
22886 switch (N.getOpcode()) {
22889 case X86ISD::PSHUFLW:
22890 case X86ISD::PSHUFHW:
22891 assert(VT.getVectorElementType() == MVT::i16 && "Bad word shuffle type!");
22893 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22894 return SDValue(); // We combined away this shuffle, so we're done.
22896 // See if this reduces to a PSHUFD which is no more expensive and can
22897 // combine with more operations. Note that it has to at least flip the
22898 // dwords as otherwise it would have been removed as a no-op.
22899 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22900 int DMask[] = {0, 1, 2, 3};
22901 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22902 DMask[DOffset + 0] = DOffset + 1;
22903 DMask[DOffset + 1] = DOffset + 0;
22904 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22905 V = DAG.getBitcast(DVT, V);
22906 DCI.AddToWorklist(V.getNode());
22907 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22908 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22909 DCI.AddToWorklist(V.getNode());
22910 return DAG.getBitcast(VT, V);
22913 // Look for shuffle patterns which can be implemented as a single unpack.
22914 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22915 // only works when we have a PSHUFD followed by two half-shuffles.
22916 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22917 (V.getOpcode() == X86ISD::PSHUFLW ||
22918 V.getOpcode() == X86ISD::PSHUFHW) &&
22919 V.getOpcode() != N.getOpcode() &&
22921 SDValue D = V.getOperand(0);
22922 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22923 D = D.getOperand(0);
22924 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22925 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22926 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22927 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22928 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22930 for (int i = 0; i < 4; ++i) {
22931 WordMask[i + NOffset] = Mask[i] + NOffset;
22932 WordMask[i + VOffset] = VMask[i] + VOffset;
22934 // Map the word mask through the DWord mask.
22936 for (int i = 0; i < 8; ++i)
22937 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22938 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22939 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22940 // We can replace all three shuffles with an unpack.
22941 V = DAG.getBitcast(VT, D.getOperand(0));
22942 DCI.AddToWorklist(V.getNode());
22943 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22952 case X86ISD::PSHUFD:
22953 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22962 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22964 /// We combine this directly on the abstract vector shuffle nodes so it is
22965 /// easier to generically match. We also insert dummy vector shuffle nodes for
22966 /// the operands which explicitly discard the lanes which are unused by this
22967 /// operation to try to flow through the rest of the combiner the fact that
22968 /// they're unused.
22969 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22971 EVT VT = N->getValueType(0);
22973 // We only handle target-independent shuffles.
22974 // FIXME: It would be easy and harmless to use the target shuffle mask
22975 // extraction tool to support more.
22976 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22979 auto *SVN = cast<ShuffleVectorSDNode>(N);
22980 ArrayRef<int> Mask = SVN->getMask();
22981 SDValue V1 = N->getOperand(0);
22982 SDValue V2 = N->getOperand(1);
22984 // We require the first shuffle operand to be the SUB node, and the second to
22985 // be the ADD node.
22986 // FIXME: We should support the commuted patterns.
22987 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22990 // If there are other uses of these operations we can't fold them.
22991 if (!V1->hasOneUse() || !V2->hasOneUse())
22994 // Ensure that both operations have the same operands. Note that we can
22995 // commute the FADD operands.
22996 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22997 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22998 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
23001 // We're looking for blends between FADD and FSUB nodes. We insist on these
23002 // nodes being lined up in a specific expected pattern.
23003 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
23004 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
23005 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
23008 // Only specific types are legal at this point, assert so we notice if and
23009 // when these change.
23010 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
23011 VT == MVT::v4f64) &&
23012 "Unknown vector type encountered!");
23014 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
23017 /// PerformShuffleCombine - Performs several different shuffle combines.
23018 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
23019 TargetLowering::DAGCombinerInfo &DCI,
23020 const X86Subtarget *Subtarget) {
23022 SDValue N0 = N->getOperand(0);
23023 SDValue N1 = N->getOperand(1);
23024 EVT VT = N->getValueType(0);
23026 // Don't create instructions with illegal types after legalize types has run.
23027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23028 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23031 // If we have legalized the vector types, look for blends of FADD and FSUB
23032 // nodes that we can fuse into an ADDSUB node.
23033 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23034 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23037 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23038 if (Subtarget->hasFp256() && VT.is256BitVector() &&
23039 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23040 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23042 // During Type Legalization, when promoting illegal vector types,
23043 // the backend might introduce new shuffle dag nodes and bitcasts.
23045 // This code performs the following transformation:
23046 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23047 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23049 // We do this only if both the bitcast and the BINOP dag nodes have
23050 // one use. Also, perform this transformation only if the new binary
23051 // operation is legal. This is to avoid introducing dag nodes that
23052 // potentially need to be further expanded (or custom lowered) into a
23053 // less optimal sequence of dag nodes.
23054 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23055 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23056 N0.getOpcode() == ISD::BITCAST) {
23057 SDValue BC0 = N0.getOperand(0);
23058 EVT SVT = BC0.getValueType();
23059 unsigned Opcode = BC0.getOpcode();
23060 unsigned NumElts = VT.getVectorNumElements();
23062 if (BC0.hasOneUse() && SVT.isVector() &&
23063 SVT.getVectorNumElements() * 2 == NumElts &&
23064 TLI.isOperationLegal(Opcode, VT)) {
23065 bool CanFold = false;
23077 unsigned SVTNumElts = SVT.getVectorNumElements();
23078 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23079 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23080 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23081 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23082 CanFold = SVOp->getMaskElt(i) < 0;
23085 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23086 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23087 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23088 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23093 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23094 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23095 // consecutive, non-overlapping, and in the right order.
23096 SmallVector<SDValue, 16> Elts;
23097 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23098 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23100 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23103 if (isTargetShuffle(N->getOpcode())) {
23105 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23106 if (Shuffle.getNode())
23109 // Try recursively combining arbitrary sequences of x86 shuffle
23110 // instructions into higher-order shuffles. We do this after combining
23111 // specific PSHUF instruction sequences into their minimal form so that we
23112 // can evaluate how many specialized shuffle instructions are involved in
23113 // a particular chain.
23114 SmallVector<int, 1> NonceMask; // Just a placeholder.
23115 NonceMask.push_back(0);
23116 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23117 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23119 return SDValue(); // This routine will use CombineTo to replace N.
23125 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23126 /// specific shuffle of a load can be folded into a single element load.
23127 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23128 /// shuffles have been custom lowered so we need to handle those here.
23129 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23130 TargetLowering::DAGCombinerInfo &DCI) {
23131 if (DCI.isBeforeLegalizeOps())
23134 SDValue InVec = N->getOperand(0);
23135 SDValue EltNo = N->getOperand(1);
23137 if (!isa<ConstantSDNode>(EltNo))
23140 EVT OriginalVT = InVec.getValueType();
23142 if (InVec.getOpcode() == ISD::BITCAST) {
23143 // Don't duplicate a load with other uses.
23144 if (!InVec.hasOneUse())
23146 EVT BCVT = InVec.getOperand(0).getValueType();
23147 if (!BCVT.isVector() ||
23148 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23150 InVec = InVec.getOperand(0);
23153 EVT CurrentVT = InVec.getValueType();
23155 if (!isTargetShuffle(InVec.getOpcode()))
23158 // Don't duplicate a load with other uses.
23159 if (!InVec.hasOneUse())
23162 SmallVector<int, 16> ShuffleMask;
23164 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23165 ShuffleMask, UnaryShuffle))
23168 // Select the input vector, guarding against out of range extract vector.
23169 unsigned NumElems = CurrentVT.getVectorNumElements();
23170 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23171 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23172 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23173 : InVec.getOperand(1);
23175 // If inputs to shuffle are the same for both ops, then allow 2 uses
23176 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23177 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23179 if (LdNode.getOpcode() == ISD::BITCAST) {
23180 // Don't duplicate a load with other uses.
23181 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23184 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23185 LdNode = LdNode.getOperand(0);
23188 if (!ISD::isNormalLoad(LdNode.getNode()))
23191 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23193 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23196 EVT EltVT = N->getValueType(0);
23197 // If there's a bitcast before the shuffle, check if the load type and
23198 // alignment is valid.
23199 unsigned Align = LN0->getAlignment();
23200 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23201 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23202 EltVT.getTypeForEVT(*DAG.getContext()));
23204 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23207 // All checks match so transform back to vector_shuffle so that DAG combiner
23208 // can finish the job
23211 // Create shuffle node taking into account the case that its a unary shuffle
23212 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23213 : InVec.getOperand(1);
23214 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23215 InVec.getOperand(0), Shuffle,
23217 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23218 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23222 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23223 const X86Subtarget *Subtarget) {
23224 SDValue N0 = N->getOperand(0);
23225 EVT VT = N->getValueType(0);
23227 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23228 // special and don't usually play with other vector types, it's better to
23229 // handle them early to be sure we emit efficient code by avoiding
23230 // store-load conversions.
23231 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23232 N0.getValueType() == MVT::v2i32 &&
23233 isa<ConstantSDNode>(N0.getOperand(1))) {
23234 SDValue N00 = N0->getOperand(0);
23235 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23236 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23239 // Convert a bitcasted integer logic operation that has one bitcasted
23240 // floating-point operand and one constant operand into a floating-point
23241 // logic operation. This may create a load of the constant, but that is
23242 // cheaper than materializing the constant in an integer register and
23243 // transferring it to an SSE register or transferring the SSE operand to
23244 // integer register and back.
23246 switch (N0.getOpcode()) {
23247 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23248 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23249 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23250 default: return SDValue();
23252 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23253 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23254 isa<ConstantSDNode>(N0.getOperand(1)) &&
23255 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23256 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23257 SDValue N000 = N0.getOperand(0).getOperand(0);
23258 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23259 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23265 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23266 /// generation and convert it from being a bunch of shuffles and extracts
23267 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23268 /// storing the value and loading scalars back, while for x64 we should
23269 /// use 64-bit extracts and shifts.
23270 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23271 TargetLowering::DAGCombinerInfo &DCI) {
23272 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23275 SDValue InputVector = N->getOperand(0);
23276 SDLoc dl(InputVector);
23277 // Detect mmx to i32 conversion through a v2i32 elt extract.
23278 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23279 N->getValueType(0) == MVT::i32 &&
23280 InputVector.getValueType() == MVT::v2i32) {
23282 // The bitcast source is a direct mmx result.
23283 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23284 if (MMXSrc.getValueType() == MVT::x86mmx)
23285 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23286 N->getValueType(0),
23287 InputVector.getNode()->getOperand(0));
23289 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23290 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23291 MMXSrc.getValueType() == MVT::i64) {
23292 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23293 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23294 MMXSrcOp.getValueType() == MVT::v1i64 &&
23295 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23296 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23297 N->getValueType(0), MMXSrcOp.getOperand(0));
23301 EVT VT = N->getValueType(0);
23303 if (VT == MVT::i1 && isa<ConstantSDNode>(N->getOperand(1)) &&
23304 InputVector.getOpcode() == ISD::BITCAST &&
23305 isa<ConstantSDNode>(InputVector.getOperand(0))) {
23306 uint64_t ExtractedElt =
23307 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23308 uint64_t InputValue =
23309 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23310 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23311 return DAG.getConstant(Res, dl, MVT::i1);
23313 // Only operate on vectors of 4 elements, where the alternative shuffling
23314 // gets to be more expensive.
23315 if (InputVector.getValueType() != MVT::v4i32)
23318 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23319 // single use which is a sign-extend or zero-extend, and all elements are
23321 SmallVector<SDNode *, 4> Uses;
23322 unsigned ExtractedElements = 0;
23323 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23324 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23325 if (UI.getUse().getResNo() != InputVector.getResNo())
23328 SDNode *Extract = *UI;
23329 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23332 if (Extract->getValueType(0) != MVT::i32)
23334 if (!Extract->hasOneUse())
23336 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23337 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23339 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23342 // Record which element was extracted.
23343 ExtractedElements |=
23344 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23346 Uses.push_back(Extract);
23349 // If not all the elements were used, this may not be worthwhile.
23350 if (ExtractedElements != 15)
23353 // Ok, we've now decided to do the transformation.
23354 // If 64-bit shifts are legal, use the extract-shift sequence,
23355 // otherwise bounce the vector off the cache.
23356 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23359 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23360 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23361 auto &DL = DAG.getDataLayout();
23362 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23363 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23364 DAG.getConstant(0, dl, VecIdxTy));
23365 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23366 DAG.getConstant(1, dl, VecIdxTy));
23368 SDValue ShAmt = DAG.getConstant(
23369 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23370 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23371 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23372 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23373 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23374 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23375 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23377 // Store the value to a temporary stack slot.
23378 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23379 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23380 MachinePointerInfo(), false, false, 0);
23382 EVT ElementType = InputVector.getValueType().getVectorElementType();
23383 unsigned EltSize = ElementType.getSizeInBits() / 8;
23385 // Replace each use (extract) with a load of the appropriate element.
23386 for (unsigned i = 0; i < 4; ++i) {
23387 uint64_t Offset = EltSize * i;
23388 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23389 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23391 SDValue ScalarAddr =
23392 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23394 // Load the scalar.
23395 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23396 ScalarAddr, MachinePointerInfo(),
23397 false, false, false, 0);
23402 // Replace the extracts
23403 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23404 UE = Uses.end(); UI != UE; ++UI) {
23405 SDNode *Extract = *UI;
23407 SDValue Idx = Extract->getOperand(1);
23408 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23409 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23412 // The replacement was made in place; don't return anything.
23417 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23418 const X86Subtarget *Subtarget) {
23420 SDValue Cond = N->getOperand(0);
23421 SDValue LHS = N->getOperand(1);
23422 SDValue RHS = N->getOperand(2);
23424 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23425 SDValue CondSrc = Cond->getOperand(0);
23426 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23427 Cond = CondSrc->getOperand(0);
23430 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23433 // A vselect where all conditions and data are constants can be optimized into
23434 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23435 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23436 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23439 unsigned MaskValue = 0;
23440 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23443 MVT VT = N->getSimpleValueType(0);
23444 unsigned NumElems = VT.getVectorNumElements();
23445 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23446 for (unsigned i = 0; i < NumElems; ++i) {
23447 // Be sure we emit undef where we can.
23448 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23449 ShuffleMask[i] = -1;
23451 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23454 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23455 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23457 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23460 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23462 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23463 TargetLowering::DAGCombinerInfo &DCI,
23464 const X86Subtarget *Subtarget) {
23466 SDValue Cond = N->getOperand(0);
23467 // Get the LHS/RHS of the select.
23468 SDValue LHS = N->getOperand(1);
23469 SDValue RHS = N->getOperand(2);
23470 EVT VT = LHS.getValueType();
23471 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23473 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23474 // instructions match the semantics of the common C idiom x<y?x:y but not
23475 // x<=y?x:y, because of how they handle negative zero (which can be
23476 // ignored in unsafe-math mode).
23477 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23478 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23479 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23480 (Subtarget->hasSSE2() ||
23481 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23482 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23484 unsigned Opcode = 0;
23485 // Check for x CC y ? x : y.
23486 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23487 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23491 // Converting this to a min would handle NaNs incorrectly, and swapping
23492 // the operands would cause it to handle comparisons between positive
23493 // and negative zero incorrectly.
23494 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23495 if (!DAG.getTarget().Options.UnsafeFPMath &&
23496 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23498 std::swap(LHS, RHS);
23500 Opcode = X86ISD::FMIN;
23503 // Converting this to a min would handle comparisons between positive
23504 // and negative zero incorrectly.
23505 if (!DAG.getTarget().Options.UnsafeFPMath &&
23506 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23508 Opcode = X86ISD::FMIN;
23511 // Converting this to a min would handle both negative zeros and NaNs
23512 // incorrectly, but we can swap the operands to fix both.
23513 std::swap(LHS, RHS);
23517 Opcode = X86ISD::FMIN;
23521 // Converting this to a max would handle comparisons between positive
23522 // and negative zero incorrectly.
23523 if (!DAG.getTarget().Options.UnsafeFPMath &&
23524 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23526 Opcode = X86ISD::FMAX;
23529 // Converting this to a max would handle NaNs incorrectly, and swapping
23530 // the operands would cause it to handle comparisons between positive
23531 // and negative zero incorrectly.
23532 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23533 if (!DAG.getTarget().Options.UnsafeFPMath &&
23534 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23536 std::swap(LHS, RHS);
23538 Opcode = X86ISD::FMAX;
23541 // Converting this to a max would handle both negative zeros and NaNs
23542 // incorrectly, but we can swap the operands to fix both.
23543 std::swap(LHS, RHS);
23547 Opcode = X86ISD::FMAX;
23550 // Check for x CC y ? y : x -- a min/max with reversed arms.
23551 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23552 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23556 // Converting this to a min would handle comparisons between positive
23557 // and negative zero incorrectly, and swapping the operands would
23558 // cause it to handle NaNs incorrectly.
23559 if (!DAG.getTarget().Options.UnsafeFPMath &&
23560 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23561 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23563 std::swap(LHS, RHS);
23565 Opcode = X86ISD::FMIN;
23568 // Converting this to a min would handle NaNs incorrectly.
23569 if (!DAG.getTarget().Options.UnsafeFPMath &&
23570 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23572 Opcode = X86ISD::FMIN;
23575 // Converting this to a min would handle both negative zeros and NaNs
23576 // incorrectly, but we can swap the operands to fix both.
23577 std::swap(LHS, RHS);
23581 Opcode = X86ISD::FMIN;
23585 // Converting this to a max would handle NaNs incorrectly.
23586 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23588 Opcode = X86ISD::FMAX;
23591 // Converting this to a max would handle comparisons between positive
23592 // and negative zero incorrectly, and swapping the operands would
23593 // cause it to handle NaNs incorrectly.
23594 if (!DAG.getTarget().Options.UnsafeFPMath &&
23595 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23596 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23598 std::swap(LHS, RHS);
23600 Opcode = X86ISD::FMAX;
23603 // Converting this to a max would handle both negative zeros and NaNs
23604 // incorrectly, but we can swap the operands to fix both.
23605 std::swap(LHS, RHS);
23609 Opcode = X86ISD::FMAX;
23615 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23618 EVT CondVT = Cond.getValueType();
23619 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23620 CondVT.getVectorElementType() == MVT::i1) {
23621 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23622 // lowering on KNL. In this case we convert it to
23623 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23624 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23625 // Since SKX these selects have a proper lowering.
23626 EVT OpVT = LHS.getValueType();
23627 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23628 (OpVT.getVectorElementType() == MVT::i8 ||
23629 OpVT.getVectorElementType() == MVT::i16) &&
23630 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23631 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23632 DCI.AddToWorklist(Cond.getNode());
23633 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23636 // If this is a select between two integer constants, try to do some
23638 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23639 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23640 // Don't do this for crazy integer types.
23641 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23642 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23643 // so that TrueC (the true value) is larger than FalseC.
23644 bool NeedsCondInvert = false;
23646 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23647 // Efficiently invertible.
23648 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23649 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23650 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23651 NeedsCondInvert = true;
23652 std::swap(TrueC, FalseC);
23655 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23656 if (FalseC->getAPIntValue() == 0 &&
23657 TrueC->getAPIntValue().isPowerOf2()) {
23658 if (NeedsCondInvert) // Invert the condition if needed.
23659 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23660 DAG.getConstant(1, DL, Cond.getValueType()));
23662 // Zero extend the condition if needed.
23663 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23665 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23666 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23667 DAG.getConstant(ShAmt, DL, MVT::i8));
23670 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23671 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23672 if (NeedsCondInvert) // Invert the condition if needed.
23673 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23674 DAG.getConstant(1, DL, Cond.getValueType()));
23676 // Zero extend the condition if needed.
23677 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23678 FalseC->getValueType(0), Cond);
23679 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23680 SDValue(FalseC, 0));
23683 // Optimize cases that will turn into an LEA instruction. This requires
23684 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23685 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23686 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23687 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23689 bool isFastMultiplier = false;
23691 switch ((unsigned char)Diff) {
23693 case 1: // result = add base, cond
23694 case 2: // result = lea base( , cond*2)
23695 case 3: // result = lea base(cond, cond*2)
23696 case 4: // result = lea base( , cond*4)
23697 case 5: // result = lea base(cond, cond*4)
23698 case 8: // result = lea base( , cond*8)
23699 case 9: // result = lea base(cond, cond*8)
23700 isFastMultiplier = true;
23705 if (isFastMultiplier) {
23706 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23707 if (NeedsCondInvert) // Invert the condition if needed.
23708 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23709 DAG.getConstant(1, DL, Cond.getValueType()));
23711 // Zero extend the condition if needed.
23712 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23714 // Scale the condition by the difference.
23716 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23717 DAG.getConstant(Diff, DL,
23718 Cond.getValueType()));
23720 // Add the base if non-zero.
23721 if (FalseC->getAPIntValue() != 0)
23722 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23723 SDValue(FalseC, 0));
23730 // Canonicalize max and min:
23731 // (x > y) ? x : y -> (x >= y) ? x : y
23732 // (x < y) ? x : y -> (x <= y) ? x : y
23733 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23734 // the need for an extra compare
23735 // against zero. e.g.
23736 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23738 // testl %edi, %edi
23740 // cmovgl %edi, %eax
23744 // cmovsl %eax, %edi
23745 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23746 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23747 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23748 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23753 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23754 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23755 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23756 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23761 // Early exit check
23762 if (!TLI.isTypeLegal(VT))
23765 // Match VSELECTs into subs with unsigned saturation.
23766 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23767 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23768 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23769 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23770 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23772 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23773 // left side invert the predicate to simplify logic below.
23775 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23777 CC = ISD::getSetCCInverse(CC, true);
23778 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23782 if (Other.getNode() && Other->getNumOperands() == 2 &&
23783 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23784 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23785 SDValue CondRHS = Cond->getOperand(1);
23787 // Look for a general sub with unsigned saturation first.
23788 // x >= y ? x-y : 0 --> subus x, y
23789 // x > y ? x-y : 0 --> subus x, y
23790 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23791 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23792 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23794 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23795 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23796 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23797 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23798 // If the RHS is a constant we have to reverse the const
23799 // canonicalization.
23800 // x > C-1 ? x+-C : 0 --> subus x, C
23801 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23802 CondRHSConst->getAPIntValue() ==
23803 (-OpRHSConst->getAPIntValue() - 1))
23804 return DAG.getNode(
23805 X86ISD::SUBUS, DL, VT, OpLHS,
23806 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23808 // Another special case: If C was a sign bit, the sub has been
23809 // canonicalized into a xor.
23810 // FIXME: Would it be better to use computeKnownBits to determine
23811 // whether it's safe to decanonicalize the xor?
23812 // x s< 0 ? x^C : 0 --> subus x, C
23813 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23814 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23815 OpRHSConst->getAPIntValue().isSignBit())
23816 // Note that we have to rebuild the RHS constant here to ensure we
23817 // don't rely on particular values of undef lanes.
23818 return DAG.getNode(
23819 X86ISD::SUBUS, DL, VT, OpLHS,
23820 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23825 // Simplify vector selection if condition value type matches vselect
23827 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23828 assert(Cond.getValueType().isVector() &&
23829 "vector select expects a vector selector!");
23831 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23832 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23834 // Try invert the condition if true value is not all 1s and false value
23836 if (!TValIsAllOnes && !FValIsAllZeros &&
23837 // Check if the selector will be produced by CMPP*/PCMP*
23838 Cond.getOpcode() == ISD::SETCC &&
23839 // Check if SETCC has already been promoted
23840 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23842 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23843 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23845 if (TValIsAllZeros || FValIsAllOnes) {
23846 SDValue CC = Cond.getOperand(2);
23847 ISD::CondCode NewCC =
23848 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23849 Cond.getOperand(0).getValueType().isInteger());
23850 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23851 std::swap(LHS, RHS);
23852 TValIsAllOnes = FValIsAllOnes;
23853 FValIsAllZeros = TValIsAllZeros;
23857 if (TValIsAllOnes || FValIsAllZeros) {
23860 if (TValIsAllOnes && FValIsAllZeros)
23862 else if (TValIsAllOnes)
23864 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23865 else if (FValIsAllZeros)
23866 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23867 DAG.getBitcast(CondVT, LHS));
23869 return DAG.getBitcast(VT, Ret);
23873 // We should generate an X86ISD::BLENDI from a vselect if its argument
23874 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23875 // constants. This specific pattern gets generated when we split a
23876 // selector for a 512 bit vector in a machine without AVX512 (but with
23877 // 256-bit vectors), during legalization:
23879 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23881 // Iff we find this pattern and the build_vectors are built from
23882 // constants, we translate the vselect into a shuffle_vector that we
23883 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23884 if ((N->getOpcode() == ISD::VSELECT ||
23885 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23886 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23887 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23888 if (Shuffle.getNode())
23892 // If this is a *dynamic* select (non-constant condition) and we can match
23893 // this node with one of the variable blend instructions, restructure the
23894 // condition so that the blends can use the high bit of each element and use
23895 // SimplifyDemandedBits to simplify the condition operand.
23896 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23897 !DCI.isBeforeLegalize() &&
23898 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23899 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
23901 // Don't optimize vector selects that map to mask-registers.
23905 // We can only handle the cases where VSELECT is directly legal on the
23906 // subtarget. We custom lower VSELECT nodes with constant conditions and
23907 // this makes it hard to see whether a dynamic VSELECT will correctly
23908 // lower, so we both check the operation's status and explicitly handle the
23909 // cases where a *dynamic* blend will fail even though a constant-condition
23910 // blend could be custom lowered.
23911 // FIXME: We should find a better way to handle this class of problems.
23912 // Potentially, we should combine constant-condition vselect nodes
23913 // pre-legalization into shuffles and not mark as many types as custom
23915 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23917 // FIXME: We don't support i16-element blends currently. We could and
23918 // should support them by making *all* the bits in the condition be set
23919 // rather than just the high bit and using an i8-element blend.
23920 if (VT.getVectorElementType() == MVT::i16)
23922 // Dynamic blending was only available from SSE4.1 onward.
23923 if (VT.is128BitVector() && !Subtarget->hasSSE41())
23925 // Byte blends are only available in AVX2
23926 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
23929 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23930 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23932 APInt KnownZero, KnownOne;
23933 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23934 DCI.isBeforeLegalizeOps());
23935 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23936 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23938 // If we changed the computation somewhere in the DAG, this change
23939 // will affect all users of Cond.
23940 // Make sure it is fine and update all the nodes so that we do not
23941 // use the generic VSELECT anymore. Otherwise, we may perform
23942 // wrong optimizations as we messed up with the actual expectation
23943 // for the vector boolean values.
23944 if (Cond != TLO.Old) {
23945 // Check all uses of that condition operand to check whether it will be
23946 // consumed by non-BLEND instructions, which may depend on all bits are
23948 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23950 if (I->getOpcode() != ISD::VSELECT)
23951 // TODO: Add other opcodes eventually lowered into BLEND.
23954 // Update all the users of the condition, before committing the change,
23955 // so that the VSELECT optimizations that expect the correct vector
23956 // boolean value will not be triggered.
23957 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23959 DAG.ReplaceAllUsesOfValueWith(
23961 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23962 Cond, I->getOperand(1), I->getOperand(2)));
23963 DCI.CommitTargetLoweringOpt(TLO);
23966 // At this point, only Cond is changed. Change the condition
23967 // just for N to keep the opportunity to optimize all other
23968 // users their own way.
23969 DAG.ReplaceAllUsesOfValueWith(
23971 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23972 TLO.New, N->getOperand(1), N->getOperand(2)));
23980 // Check whether a boolean test is testing a boolean value generated by
23981 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23984 // Simplify the following patterns:
23985 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23986 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23987 // to (Op EFLAGS Cond)
23989 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23990 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23991 // to (Op EFLAGS !Cond)
23993 // where Op could be BRCOND or CMOV.
23995 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23996 // Quit if not CMP and SUB with its value result used.
23997 if (Cmp.getOpcode() != X86ISD::CMP &&
23998 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
24001 // Quit if not used as a boolean value.
24002 if (CC != X86::COND_E && CC != X86::COND_NE)
24005 // Check CMP operands. One of them should be 0 or 1 and the other should be
24006 // an SetCC or extended from it.
24007 SDValue Op1 = Cmp.getOperand(0);
24008 SDValue Op2 = Cmp.getOperand(1);
24011 const ConstantSDNode* C = nullptr;
24012 bool needOppositeCond = (CC == X86::COND_E);
24013 bool checkAgainstTrue = false; // Is it a comparison against 1?
24015 if ((C = dyn_cast<ConstantSDNode>(Op1)))
24017 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
24019 else // Quit if all operands are not constants.
24022 if (C->getZExtValue() == 1) {
24023 needOppositeCond = !needOppositeCond;
24024 checkAgainstTrue = true;
24025 } else if (C->getZExtValue() != 0)
24026 // Quit if the constant is neither 0 or 1.
24029 bool truncatedToBoolWithAnd = false;
24030 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
24031 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
24032 SetCC.getOpcode() == ISD::TRUNCATE ||
24033 SetCC.getOpcode() == ISD::AND) {
24034 if (SetCC.getOpcode() == ISD::AND) {
24036 ConstantSDNode *CS;
24037 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
24038 CS->getZExtValue() == 1)
24040 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
24041 CS->getZExtValue() == 1)
24045 SetCC = SetCC.getOperand(OpIdx);
24046 truncatedToBoolWithAnd = true;
24048 SetCC = SetCC.getOperand(0);
24051 switch (SetCC.getOpcode()) {
24052 case X86ISD::SETCC_CARRY:
24053 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24054 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24055 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24056 // truncated to i1 using 'and'.
24057 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24059 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24060 "Invalid use of SETCC_CARRY!");
24062 case X86ISD::SETCC:
24063 // Set the condition code or opposite one if necessary.
24064 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24065 if (needOppositeCond)
24066 CC = X86::GetOppositeBranchCondition(CC);
24067 return SetCC.getOperand(1);
24068 case X86ISD::CMOV: {
24069 // Check whether false/true value has canonical one, i.e. 0 or 1.
24070 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24071 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24072 // Quit if true value is not a constant.
24075 // Quit if false value is not a constant.
24077 SDValue Op = SetCC.getOperand(0);
24078 // Skip 'zext' or 'trunc' node.
24079 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24080 Op.getOpcode() == ISD::TRUNCATE)
24081 Op = Op.getOperand(0);
24082 // A special case for rdrand/rdseed, where 0 is set if false cond is
24084 if ((Op.getOpcode() != X86ISD::RDRAND &&
24085 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24088 // Quit if false value is not the constant 0 or 1.
24089 bool FValIsFalse = true;
24090 if (FVal && FVal->getZExtValue() != 0) {
24091 if (FVal->getZExtValue() != 1)
24093 // If FVal is 1, opposite cond is needed.
24094 needOppositeCond = !needOppositeCond;
24095 FValIsFalse = false;
24097 // Quit if TVal is not the constant opposite of FVal.
24098 if (FValIsFalse && TVal->getZExtValue() != 1)
24100 if (!FValIsFalse && TVal->getZExtValue() != 0)
24102 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24103 if (needOppositeCond)
24104 CC = X86::GetOppositeBranchCondition(CC);
24105 return SetCC.getOperand(3);
24112 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24114 /// (X86or (X86setcc) (X86setcc))
24115 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24116 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24117 X86::CondCode &CC1, SDValue &Flags,
24119 if (Cond->getOpcode() == X86ISD::CMP) {
24120 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24121 if (!CondOp1C || !CondOp1C->isNullValue())
24124 Cond = Cond->getOperand(0);
24129 SDValue SetCC0, SetCC1;
24130 switch (Cond->getOpcode()) {
24131 default: return false;
24138 SetCC0 = Cond->getOperand(0);
24139 SetCC1 = Cond->getOperand(1);
24143 // Make sure we have SETCC nodes, using the same flags value.
24144 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24145 SetCC1.getOpcode() != X86ISD::SETCC ||
24146 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24149 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24150 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24151 Flags = SetCC0->getOperand(1);
24155 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24156 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24157 TargetLowering::DAGCombinerInfo &DCI,
24158 const X86Subtarget *Subtarget) {
24161 // If the flag operand isn't dead, don't touch this CMOV.
24162 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24165 SDValue FalseOp = N->getOperand(0);
24166 SDValue TrueOp = N->getOperand(1);
24167 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24168 SDValue Cond = N->getOperand(3);
24170 if (CC == X86::COND_E || CC == X86::COND_NE) {
24171 switch (Cond.getOpcode()) {
24175 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24176 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24177 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24183 Flags = checkBoolTestSetCCCombine(Cond, CC);
24184 if (Flags.getNode() &&
24185 // Extra check as FCMOV only supports a subset of X86 cond.
24186 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24187 SDValue Ops[] = { FalseOp, TrueOp,
24188 DAG.getConstant(CC, DL, MVT::i8), Flags };
24189 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24192 // If this is a select between two integer constants, try to do some
24193 // optimizations. Note that the operands are ordered the opposite of SELECT
24195 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24196 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24197 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24198 // larger than FalseC (the false value).
24199 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24200 CC = X86::GetOppositeBranchCondition(CC);
24201 std::swap(TrueC, FalseC);
24202 std::swap(TrueOp, FalseOp);
24205 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24206 // This is efficient for any integer data type (including i8/i16) and
24208 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24209 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24210 DAG.getConstant(CC, DL, MVT::i8), Cond);
24212 // Zero extend the condition if needed.
24213 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24215 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24216 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24217 DAG.getConstant(ShAmt, DL, MVT::i8));
24218 if (N->getNumValues() == 2) // Dead flag value?
24219 return DCI.CombineTo(N, Cond, SDValue());
24223 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24224 // for any integer data type, including i8/i16.
24225 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24226 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24227 DAG.getConstant(CC, DL, MVT::i8), Cond);
24229 // Zero extend the condition if needed.
24230 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24231 FalseC->getValueType(0), Cond);
24232 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24233 SDValue(FalseC, 0));
24235 if (N->getNumValues() == 2) // Dead flag value?
24236 return DCI.CombineTo(N, Cond, SDValue());
24240 // Optimize cases that will turn into an LEA instruction. This requires
24241 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24242 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24243 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24244 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24246 bool isFastMultiplier = false;
24248 switch ((unsigned char)Diff) {
24250 case 1: // result = add base, cond
24251 case 2: // result = lea base( , cond*2)
24252 case 3: // result = lea base(cond, cond*2)
24253 case 4: // result = lea base( , cond*4)
24254 case 5: // result = lea base(cond, cond*4)
24255 case 8: // result = lea base( , cond*8)
24256 case 9: // result = lea base(cond, cond*8)
24257 isFastMultiplier = true;
24262 if (isFastMultiplier) {
24263 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24264 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24265 DAG.getConstant(CC, DL, MVT::i8), Cond);
24266 // Zero extend the condition if needed.
24267 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24269 // Scale the condition by the difference.
24271 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24272 DAG.getConstant(Diff, DL, Cond.getValueType()));
24274 // Add the base if non-zero.
24275 if (FalseC->getAPIntValue() != 0)
24276 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24277 SDValue(FalseC, 0));
24278 if (N->getNumValues() == 2) // Dead flag value?
24279 return DCI.CombineTo(N, Cond, SDValue());
24286 // Handle these cases:
24287 // (select (x != c), e, c) -> select (x != c), e, x),
24288 // (select (x == c), c, e) -> select (x == c), x, e)
24289 // where the c is an integer constant, and the "select" is the combination
24290 // of CMOV and CMP.
24292 // The rationale for this change is that the conditional-move from a constant
24293 // needs two instructions, however, conditional-move from a register needs
24294 // only one instruction.
24296 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24297 // some instruction-combining opportunities. This opt needs to be
24298 // postponed as late as possible.
24300 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24301 // the DCI.xxxx conditions are provided to postpone the optimization as
24302 // late as possible.
24304 ConstantSDNode *CmpAgainst = nullptr;
24305 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24306 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24307 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24309 if (CC == X86::COND_NE &&
24310 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24311 CC = X86::GetOppositeBranchCondition(CC);
24312 std::swap(TrueOp, FalseOp);
24315 if (CC == X86::COND_E &&
24316 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24317 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24318 DAG.getConstant(CC, DL, MVT::i8), Cond };
24319 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24324 // Fold and/or of setcc's to double CMOV:
24325 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24326 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24328 // This combine lets us generate:
24329 // cmovcc1 (jcc1 if we don't have CMOV)
24335 // cmovne (jne if we don't have CMOV)
24336 // When we can't use the CMOV instruction, it might increase branch
24338 // When we can use CMOV, or when there is no mispredict, this improves
24339 // throughput and reduces register pressure.
24341 if (CC == X86::COND_NE) {
24343 X86::CondCode CC0, CC1;
24345 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24347 std::swap(FalseOp, TrueOp);
24348 CC0 = X86::GetOppositeBranchCondition(CC0);
24349 CC1 = X86::GetOppositeBranchCondition(CC1);
24352 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24354 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24355 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24356 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24357 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24365 /// PerformMulCombine - Optimize a single multiply with constant into two
24366 /// in order to implement it with two cheaper instructions, e.g.
24367 /// LEA + SHL, LEA + LEA.
24368 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24369 TargetLowering::DAGCombinerInfo &DCI) {
24370 // An imul is usually smaller than the alternative sequence.
24371 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24374 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24377 EVT VT = N->getValueType(0);
24378 if (VT != MVT::i64 && VT != MVT::i32)
24381 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24384 uint64_t MulAmt = C->getZExtValue();
24385 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24388 uint64_t MulAmt1 = 0;
24389 uint64_t MulAmt2 = 0;
24390 if ((MulAmt % 9) == 0) {
24392 MulAmt2 = MulAmt / 9;
24393 } else if ((MulAmt % 5) == 0) {
24395 MulAmt2 = MulAmt / 5;
24396 } else if ((MulAmt % 3) == 0) {
24398 MulAmt2 = MulAmt / 3;
24401 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24404 if (isPowerOf2_64(MulAmt2) &&
24405 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24406 // If second multiplifer is pow2, issue it first. We want the multiply by
24407 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24409 std::swap(MulAmt1, MulAmt2);
24412 if (isPowerOf2_64(MulAmt1))
24413 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24414 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24416 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24417 DAG.getConstant(MulAmt1, DL, VT));
24419 if (isPowerOf2_64(MulAmt2))
24420 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24421 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24423 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24424 DAG.getConstant(MulAmt2, DL, VT));
24426 // Do not add new nodes to DAG combiner worklist.
24427 DCI.CombineTo(N, NewMul, false);
24432 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24433 SDValue N0 = N->getOperand(0);
24434 SDValue N1 = N->getOperand(1);
24435 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24436 EVT VT = N0.getValueType();
24438 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24439 // since the result of setcc_c is all zero's or all ones.
24440 if (VT.isInteger() && !VT.isVector() &&
24441 N1C && N0.getOpcode() == ISD::AND &&
24442 N0.getOperand(1).getOpcode() == ISD::Constant) {
24443 SDValue N00 = N0.getOperand(0);
24444 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24445 APInt ShAmt = N1C->getAPIntValue();
24446 Mask = Mask.shl(ShAmt);
24447 bool MaskOK = false;
24448 // We can handle cases concerning bit-widening nodes containing setcc_c if
24449 // we carefully interrogate the mask to make sure we are semantics
24451 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24452 // of the underlying setcc_c operation if the setcc_c was zero extended.
24453 // Consider the following example:
24454 // zext(setcc_c) -> i32 0x0000FFFF
24455 // c1 -> i32 0x0000FFFF
24456 // c2 -> i32 0x00000001
24457 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24458 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24459 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24461 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24462 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24464 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24465 N00.getOpcode() == ISD::ANY_EXTEND) &&
24466 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24467 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24469 if (MaskOK && Mask != 0) {
24471 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24475 // Hardware support for vector shifts is sparse which makes us scalarize the
24476 // vector operations in many cases. Also, on sandybridge ADD is faster than
24478 // (shl V, 1) -> add V,V
24479 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24480 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24481 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24482 // We shift all of the values by one. In many cases we do not have
24483 // hardware support for this operation. This is better expressed as an ADD
24485 if (N1SplatC->getAPIntValue() == 1)
24486 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24492 /// \brief Returns a vector of 0s if the node in input is a vector logical
24493 /// shift by a constant amount which is known to be bigger than or equal
24494 /// to the vector element size in bits.
24495 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24496 const X86Subtarget *Subtarget) {
24497 EVT VT = N->getValueType(0);
24499 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24500 (!Subtarget->hasInt256() ||
24501 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24504 SDValue Amt = N->getOperand(1);
24506 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24507 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24508 APInt ShiftAmt = AmtSplat->getAPIntValue();
24509 unsigned MaxAmount =
24510 VT.getSimpleVT().getVectorElementType().getSizeInBits();
24512 // SSE2/AVX2 logical shifts always return a vector of 0s
24513 // if the shift amount is bigger than or equal to
24514 // the element size. The constant shift amount will be
24515 // encoded as a 8-bit immediate.
24516 if (ShiftAmt.trunc(8).uge(MaxAmount))
24517 return getZeroVector(VT, Subtarget, DAG, DL);
24523 /// PerformShiftCombine - Combine shifts.
24524 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24525 TargetLowering::DAGCombinerInfo &DCI,
24526 const X86Subtarget *Subtarget) {
24527 if (N->getOpcode() == ISD::SHL)
24528 if (SDValue V = PerformSHLCombine(N, DAG))
24531 // Try to fold this logical shift into a zero vector.
24532 if (N->getOpcode() != ISD::SRA)
24533 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24539 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24540 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24541 // and friends. Likewise for OR -> CMPNEQSS.
24542 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24543 TargetLowering::DAGCombinerInfo &DCI,
24544 const X86Subtarget *Subtarget) {
24547 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24548 // we're requiring SSE2 for both.
24549 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24550 SDValue N0 = N->getOperand(0);
24551 SDValue N1 = N->getOperand(1);
24552 SDValue CMP0 = N0->getOperand(1);
24553 SDValue CMP1 = N1->getOperand(1);
24556 // The SETCCs should both refer to the same CMP.
24557 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24560 SDValue CMP00 = CMP0->getOperand(0);
24561 SDValue CMP01 = CMP0->getOperand(1);
24562 EVT VT = CMP00.getValueType();
24564 if (VT == MVT::f32 || VT == MVT::f64) {
24565 bool ExpectingFlags = false;
24566 // Check for any users that want flags:
24567 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24568 !ExpectingFlags && UI != UE; ++UI)
24569 switch (UI->getOpcode()) {
24574 ExpectingFlags = true;
24576 case ISD::CopyToReg:
24577 case ISD::SIGN_EXTEND:
24578 case ISD::ZERO_EXTEND:
24579 case ISD::ANY_EXTEND:
24583 if (!ExpectingFlags) {
24584 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24585 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24587 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24588 X86::CondCode tmp = cc0;
24593 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24594 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24595 // FIXME: need symbolic constants for these magic numbers.
24596 // See X86ATTInstPrinter.cpp:printSSECC().
24597 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24598 if (Subtarget->hasAVX512()) {
24599 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24601 DAG.getConstant(x86cc, DL, MVT::i8));
24602 if (N->getValueType(0) != MVT::i1)
24603 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24607 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24608 CMP00.getValueType(), CMP00, CMP01,
24609 DAG.getConstant(x86cc, DL,
24612 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24613 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24615 if (is64BitFP && !Subtarget->is64Bit()) {
24616 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24617 // 64-bit integer, since that's not a legal type. Since
24618 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24619 // bits, but can do this little dance to extract the lowest 32 bits
24620 // and work with those going forward.
24621 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24623 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24624 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24625 Vector32, DAG.getIntPtrConstant(0, DL));
24629 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24630 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24631 DAG.getConstant(1, DL, IntVT));
24632 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24634 return OneBitOfTruth;
24642 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24643 /// so it can be folded inside ANDNP.
24644 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24645 EVT VT = N->getValueType(0);
24647 // Match direct AllOnes for 128 and 256-bit vectors
24648 if (ISD::isBuildVectorAllOnes(N))
24651 // Look through a bit convert.
24652 if (N->getOpcode() == ISD::BITCAST)
24653 N = N->getOperand(0).getNode();
24655 // Sometimes the operand may come from a insert_subvector building a 256-bit
24657 if (VT.is256BitVector() &&
24658 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24659 SDValue V1 = N->getOperand(0);
24660 SDValue V2 = N->getOperand(1);
24662 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24663 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24664 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24665 ISD::isBuildVectorAllOnes(V2.getNode()))
24672 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24673 // register. In most cases we actually compare or select YMM-sized registers
24674 // and mixing the two types creates horrible code. This method optimizes
24675 // some of the transition sequences.
24676 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24677 TargetLowering::DAGCombinerInfo &DCI,
24678 const X86Subtarget *Subtarget) {
24679 EVT VT = N->getValueType(0);
24680 if (!VT.is256BitVector())
24683 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24684 N->getOpcode() == ISD::ZERO_EXTEND ||
24685 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24687 SDValue Narrow = N->getOperand(0);
24688 EVT NarrowVT = Narrow->getValueType(0);
24689 if (!NarrowVT.is128BitVector())
24692 if (Narrow->getOpcode() != ISD::XOR &&
24693 Narrow->getOpcode() != ISD::AND &&
24694 Narrow->getOpcode() != ISD::OR)
24697 SDValue N0 = Narrow->getOperand(0);
24698 SDValue N1 = Narrow->getOperand(1);
24701 // The Left side has to be a trunc.
24702 if (N0.getOpcode() != ISD::TRUNCATE)
24705 // The type of the truncated inputs.
24706 EVT WideVT = N0->getOperand(0)->getValueType(0);
24710 // The right side has to be a 'trunc' or a constant vector.
24711 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24712 ConstantSDNode *RHSConstSplat = nullptr;
24713 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24714 RHSConstSplat = RHSBV->getConstantSplatNode();
24715 if (!RHSTrunc && !RHSConstSplat)
24718 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24720 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24723 // Set N0 and N1 to hold the inputs to the new wide operation.
24724 N0 = N0->getOperand(0);
24725 if (RHSConstSplat) {
24726 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getVectorElementType(),
24727 SDValue(RHSConstSplat, 0));
24728 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24729 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24730 } else if (RHSTrunc) {
24731 N1 = N1->getOperand(0);
24734 // Generate the wide operation.
24735 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24736 unsigned Opcode = N->getOpcode();
24738 case ISD::ANY_EXTEND:
24740 case ISD::ZERO_EXTEND: {
24741 unsigned InBits = NarrowVT.getScalarSizeInBits();
24742 APInt Mask = APInt::getAllOnesValue(InBits);
24743 Mask = Mask.zext(VT.getScalarSizeInBits());
24744 return DAG.getNode(ISD::AND, DL, VT,
24745 Op, DAG.getConstant(Mask, DL, VT));
24747 case ISD::SIGN_EXTEND:
24748 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24749 Op, DAG.getValueType(NarrowVT));
24751 llvm_unreachable("Unexpected opcode");
24755 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24756 TargetLowering::DAGCombinerInfo &DCI,
24757 const X86Subtarget *Subtarget) {
24758 SDValue N0 = N->getOperand(0);
24759 SDValue N1 = N->getOperand(1);
24762 // A vector zext_in_reg may be represented as a shuffle,
24763 // feeding into a bitcast (this represents anyext) feeding into
24764 // an and with a mask.
24765 // We'd like to try to combine that into a shuffle with zero
24766 // plus a bitcast, removing the and.
24767 if (N0.getOpcode() != ISD::BITCAST ||
24768 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24771 // The other side of the AND should be a splat of 2^C, where C
24772 // is the number of bits in the source type.
24773 if (N1.getOpcode() == ISD::BITCAST)
24774 N1 = N1.getOperand(0);
24775 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24777 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24779 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24780 EVT SrcType = Shuffle->getValueType(0);
24782 // We expect a single-source shuffle
24783 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24786 unsigned SrcSize = SrcType.getScalarSizeInBits();
24788 APInt SplatValue, SplatUndef;
24789 unsigned SplatBitSize;
24791 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24792 SplatBitSize, HasAnyUndefs))
24795 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24796 // Make sure the splat matches the mask we expect
24797 if (SplatBitSize > ResSize ||
24798 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24801 // Make sure the input and output size make sense
24802 if (SrcSize >= ResSize || ResSize % SrcSize)
24805 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24806 // The number of u's between each two values depends on the ratio between
24807 // the source and dest type.
24808 unsigned ZextRatio = ResSize / SrcSize;
24809 bool IsZext = true;
24810 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24811 if (i % ZextRatio) {
24812 if (Shuffle->getMaskElt(i) > 0) {
24818 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24819 // Expected element number
24829 // Ok, perform the transformation - replace the shuffle with
24830 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24831 // (instead of undef) where the k elements come from the zero vector.
24832 SmallVector<int, 8> Mask;
24833 unsigned NumElems = SrcType.getVectorNumElements();
24834 for (unsigned i = 0; i < NumElems; ++i)
24836 Mask.push_back(NumElems);
24838 Mask.push_back(i / ZextRatio);
24840 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24841 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24842 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24845 /// If both input operands of a logic op are being cast from floating point
24846 /// types, try to convert this into a floating point logic node to avoid
24847 /// unnecessary moves from SSE to integer registers.
24848 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24849 const X86Subtarget *Subtarget) {
24850 unsigned FPOpcode = ISD::DELETED_NODE;
24851 if (N->getOpcode() == ISD::AND)
24852 FPOpcode = X86ISD::FAND;
24853 else if (N->getOpcode() == ISD::OR)
24854 FPOpcode = X86ISD::FOR;
24855 else if (N->getOpcode() == ISD::XOR)
24856 FPOpcode = X86ISD::FXOR;
24858 assert(FPOpcode != ISD::DELETED_NODE &&
24859 "Unexpected input node for FP logic conversion");
24861 EVT VT = N->getValueType(0);
24862 SDValue N0 = N->getOperand(0);
24863 SDValue N1 = N->getOperand(1);
24865 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24866 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24867 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24868 SDValue N00 = N0.getOperand(0);
24869 SDValue N10 = N1.getOperand(0);
24870 EVT N00Type = N00.getValueType();
24871 EVT N10Type = N10.getValueType();
24872 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24873 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24874 return DAG.getBitcast(VT, FPLogic);
24880 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24881 TargetLowering::DAGCombinerInfo &DCI,
24882 const X86Subtarget *Subtarget) {
24883 if (DCI.isBeforeLegalizeOps())
24886 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24889 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24892 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24895 EVT VT = N->getValueType(0);
24896 SDValue N0 = N->getOperand(0);
24897 SDValue N1 = N->getOperand(1);
24900 // Create BEXTR instructions
24901 // BEXTR is ((X >> imm) & (2**size-1))
24902 if (VT == MVT::i32 || VT == MVT::i64) {
24903 // Check for BEXTR.
24904 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24905 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24906 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24907 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24908 if (MaskNode && ShiftNode) {
24909 uint64_t Mask = MaskNode->getZExtValue();
24910 uint64_t Shift = ShiftNode->getZExtValue();
24911 if (isMask_64(Mask)) {
24912 uint64_t MaskSize = countPopulation(Mask);
24913 if (Shift + MaskSize <= VT.getSizeInBits())
24914 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24915 DAG.getConstant(Shift | (MaskSize << 8), DL,
24924 // Want to form ANDNP nodes:
24925 // 1) In the hopes of then easily combining them with OR and AND nodes
24926 // to form PBLEND/PSIGN.
24927 // 2) To match ANDN packed intrinsics
24928 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24931 // Check LHS for vnot
24932 if (N0.getOpcode() == ISD::XOR &&
24933 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24934 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24935 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24937 // Check RHS for vnot
24938 if (N1.getOpcode() == ISD::XOR &&
24939 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24940 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24941 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24946 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24947 TargetLowering::DAGCombinerInfo &DCI,
24948 const X86Subtarget *Subtarget) {
24949 if (DCI.isBeforeLegalizeOps())
24952 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24955 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24958 SDValue N0 = N->getOperand(0);
24959 SDValue N1 = N->getOperand(1);
24960 EVT VT = N->getValueType(0);
24962 // look for psign/blend
24963 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24964 if (!Subtarget->hasSSSE3() ||
24965 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24968 // Canonicalize pandn to RHS
24969 if (N0.getOpcode() == X86ISD::ANDNP)
24971 // or (and (m, y), (pandn m, x))
24972 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24973 SDValue Mask = N1.getOperand(0);
24974 SDValue X = N1.getOperand(1);
24976 if (N0.getOperand(0) == Mask)
24977 Y = N0.getOperand(1);
24978 if (N0.getOperand(1) == Mask)
24979 Y = N0.getOperand(0);
24981 // Check to see if the mask appeared in both the AND and ANDNP and
24985 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24986 // Look through mask bitcast.
24987 if (Mask.getOpcode() == ISD::BITCAST)
24988 Mask = Mask.getOperand(0);
24989 if (X.getOpcode() == ISD::BITCAST)
24990 X = X.getOperand(0);
24991 if (Y.getOpcode() == ISD::BITCAST)
24992 Y = Y.getOperand(0);
24994 EVT MaskVT = Mask.getValueType();
24996 // Validate that the Mask operand is a vector sra node.
24997 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24998 // there is no psrai.b
24999 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
25000 unsigned SraAmt = ~0;
25001 if (Mask.getOpcode() == ISD::SRA) {
25002 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
25003 if (auto *AmtConst = AmtBV->getConstantSplatNode())
25004 SraAmt = AmtConst->getZExtValue();
25005 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
25006 SDValue SraC = Mask.getOperand(1);
25007 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
25009 if ((SraAmt + 1) != EltBits)
25014 // Now we know we at least have a plendvb with the mask val. See if
25015 // we can form a psignb/w/d.
25016 // psign = x.type == y.type == mask.type && y = sub(0, x);
25017 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
25018 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
25019 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
25020 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
25021 "Unsupported VT for PSIGN");
25022 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
25023 return DAG.getBitcast(VT, Mask);
25025 // PBLENDVB only available on SSE 4.1
25026 if (!Subtarget->hasSSE41())
25029 MVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
25031 X = DAG.getBitcast(BlendVT, X);
25032 Y = DAG.getBitcast(BlendVT, Y);
25033 Mask = DAG.getBitcast(BlendVT, Mask);
25034 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
25035 return DAG.getBitcast(VT, Mask);
25039 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
25042 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
25043 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
25045 // SHLD/SHRD instructions have lower register pressure, but on some
25046 // platforms they have higher latency than the equivalent
25047 // series of shifts/or that would otherwise be generated.
25048 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25049 // have higher latencies and we are not optimizing for size.
25050 if (!OptForSize && Subtarget->isSHLDSlow())
25053 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25055 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25057 if (!N0.hasOneUse() || !N1.hasOneUse())
25060 SDValue ShAmt0 = N0.getOperand(1);
25061 if (ShAmt0.getValueType() != MVT::i8)
25063 SDValue ShAmt1 = N1.getOperand(1);
25064 if (ShAmt1.getValueType() != MVT::i8)
25066 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25067 ShAmt0 = ShAmt0.getOperand(0);
25068 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25069 ShAmt1 = ShAmt1.getOperand(0);
25072 unsigned Opc = X86ISD::SHLD;
25073 SDValue Op0 = N0.getOperand(0);
25074 SDValue Op1 = N1.getOperand(0);
25075 if (ShAmt0.getOpcode() == ISD::SUB) {
25076 Opc = X86ISD::SHRD;
25077 std::swap(Op0, Op1);
25078 std::swap(ShAmt0, ShAmt1);
25081 unsigned Bits = VT.getSizeInBits();
25082 if (ShAmt1.getOpcode() == ISD::SUB) {
25083 SDValue Sum = ShAmt1.getOperand(0);
25084 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25085 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25086 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25087 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25088 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25089 return DAG.getNode(Opc, DL, VT,
25091 DAG.getNode(ISD::TRUNCATE, DL,
25094 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25095 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25097 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25098 return DAG.getNode(Opc, DL, VT,
25099 N0.getOperand(0), N1.getOperand(0),
25100 DAG.getNode(ISD::TRUNCATE, DL,
25107 // Generate NEG and CMOV for integer abs.
25108 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25109 EVT VT = N->getValueType(0);
25111 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25112 // 8-bit integer abs to NEG and CMOV.
25113 if (VT.isInteger() && VT.getSizeInBits() == 8)
25116 SDValue N0 = N->getOperand(0);
25117 SDValue N1 = N->getOperand(1);
25120 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25121 // and change it to SUB and CMOV.
25122 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25123 N0.getOpcode() == ISD::ADD &&
25124 N0.getOperand(1) == N1 &&
25125 N1.getOpcode() == ISD::SRA &&
25126 N1.getOperand(0) == N0.getOperand(0))
25127 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25128 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25129 // Generate SUB & CMOV.
25130 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25131 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25133 SDValue Ops[] = { N0.getOperand(0), Neg,
25134 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25135 SDValue(Neg.getNode(), 1) };
25136 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25141 // Try to turn tests against the signbit in the form of:
25142 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25145 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25146 // This is only worth doing if the output type is i8.
25147 if (N->getValueType(0) != MVT::i8)
25150 SDValue N0 = N->getOperand(0);
25151 SDValue N1 = N->getOperand(1);
25153 // We should be performing an xor against a truncated shift.
25154 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25157 // Make sure we are performing an xor against one.
25158 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25161 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25162 SDValue Shift = N0.getOperand(0);
25163 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25166 // Make sure we are truncating from one of i16, i32 or i64.
25167 EVT ShiftTy = Shift.getValueType();
25168 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25171 // Make sure the shift amount extracts the sign bit.
25172 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25173 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25176 // Create a greater-than comparison against -1.
25177 // N.B. Using SETGE against 0 works but we want a canonical looking
25178 // comparison, using SETGT matches up with what TranslateX86CC.
25180 SDValue ShiftOp = Shift.getOperand(0);
25181 EVT ShiftOpTy = ShiftOp.getValueType();
25182 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25183 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25187 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25188 TargetLowering::DAGCombinerInfo &DCI,
25189 const X86Subtarget *Subtarget) {
25190 if (DCI.isBeforeLegalizeOps())
25193 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25196 if (Subtarget->hasCMov())
25197 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25200 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25206 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25207 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25208 TargetLowering::DAGCombinerInfo &DCI,
25209 const X86Subtarget *Subtarget) {
25210 LoadSDNode *Ld = cast<LoadSDNode>(N);
25211 EVT RegVT = Ld->getValueType(0);
25212 EVT MemVT = Ld->getMemoryVT();
25214 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25216 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25217 // into two 16-byte operations.
25218 ISD::LoadExtType Ext = Ld->getExtensionType();
25220 unsigned AddressSpace = Ld->getAddressSpace();
25221 unsigned Alignment = Ld->getAlignment();
25222 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25223 Ext == ISD::NON_EXTLOAD &&
25224 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25225 AddressSpace, Alignment, &Fast) && !Fast) {
25226 unsigned NumElems = RegVT.getVectorNumElements();
25230 SDValue Ptr = Ld->getBasePtr();
25231 SDValue Increment =
25232 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25234 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25236 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25237 Ld->getPointerInfo(), Ld->isVolatile(),
25238 Ld->isNonTemporal(), Ld->isInvariant(),
25240 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25241 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25242 Ld->getPointerInfo(), Ld->isVolatile(),
25243 Ld->isNonTemporal(), Ld->isInvariant(),
25244 std::min(16U, Alignment));
25245 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25247 Load2.getValue(1));
25249 SDValue NewVec = DAG.getUNDEF(RegVT);
25250 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25251 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25252 return DCI.CombineTo(N, NewVec, TF, true);
25258 /// PerformMLOADCombine - Resolve extending loads
25259 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25260 TargetLowering::DAGCombinerInfo &DCI,
25261 const X86Subtarget *Subtarget) {
25262 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25263 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25266 EVT VT = Mld->getValueType(0);
25267 unsigned NumElems = VT.getVectorNumElements();
25268 EVT LdVT = Mld->getMemoryVT();
25271 assert(LdVT != VT && "Cannot extend to the same type");
25272 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25273 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25274 // From, To sizes and ElemCount must be pow of two
25275 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25276 "Unexpected size for extending masked load");
25278 unsigned SizeRatio = ToSz / FromSz;
25279 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25281 // Create a type on which we perform the shuffle
25282 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25283 LdVT.getScalarType(), NumElems*SizeRatio);
25284 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25286 // Convert Src0 value
25287 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25288 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25289 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25290 for (unsigned i = 0; i != NumElems; ++i)
25291 ShuffleVec[i] = i * SizeRatio;
25293 // Can't shuffle using an illegal type.
25294 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25295 "WideVecVT should be legal");
25296 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25297 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25299 // Prepare the new mask
25301 SDValue Mask = Mld->getMask();
25302 if (Mask.getValueType() == VT) {
25303 // Mask and original value have the same type
25304 NewMask = DAG.getBitcast(WideVecVT, Mask);
25305 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25306 for (unsigned i = 0; i != NumElems; ++i)
25307 ShuffleVec[i] = i * SizeRatio;
25308 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25309 ShuffleVec[i] = NumElems*SizeRatio;
25310 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25311 DAG.getConstant(0, dl, WideVecVT),
25315 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25316 unsigned WidenNumElts = NumElems*SizeRatio;
25317 unsigned MaskNumElts = VT.getVectorNumElements();
25318 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25321 unsigned NumConcat = WidenNumElts / MaskNumElts;
25322 SmallVector<SDValue, 16> Ops(NumConcat);
25323 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25325 for (unsigned i = 1; i != NumConcat; ++i)
25328 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25331 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25332 Mld->getBasePtr(), NewMask, WideSrc0,
25333 Mld->getMemoryVT(), Mld->getMemOperand(),
25335 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25336 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25338 /// PerformMSTORECombine - Resolve truncating stores
25339 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25340 const X86Subtarget *Subtarget) {
25341 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25342 if (!Mst->isTruncatingStore())
25345 EVT VT = Mst->getValue().getValueType();
25346 unsigned NumElems = VT.getVectorNumElements();
25347 EVT StVT = Mst->getMemoryVT();
25350 assert(StVT != VT && "Cannot truncate to the same type");
25351 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25352 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25354 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25356 // The truncating store is legal in some cases. For example
25357 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25358 // are designated for truncate store.
25359 // In this case we don't need any further transformations.
25360 if (TLI.isTruncStoreLegal(VT, StVT))
25363 // From, To sizes and ElemCount must be pow of two
25364 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25365 "Unexpected size for truncating masked store");
25366 // We are going to use the original vector elt for storing.
25367 // Accumulated smaller vector elements must be a multiple of the store size.
25368 assert (((NumElems * FromSz) % ToSz) == 0 &&
25369 "Unexpected ratio for truncating masked store");
25371 unsigned SizeRatio = FromSz / ToSz;
25372 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25374 // Create a type on which we perform the shuffle
25375 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25376 StVT.getScalarType(), NumElems*SizeRatio);
25378 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25380 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25381 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25382 for (unsigned i = 0; i != NumElems; ++i)
25383 ShuffleVec[i] = i * SizeRatio;
25385 // Can't shuffle using an illegal type.
25386 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25387 "WideVecVT should be legal");
25389 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25390 DAG.getUNDEF(WideVecVT),
25394 SDValue Mask = Mst->getMask();
25395 if (Mask.getValueType() == VT) {
25396 // Mask and original value have the same type
25397 NewMask = DAG.getBitcast(WideVecVT, Mask);
25398 for (unsigned i = 0; i != NumElems; ++i)
25399 ShuffleVec[i] = i * SizeRatio;
25400 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25401 ShuffleVec[i] = NumElems*SizeRatio;
25402 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25403 DAG.getConstant(0, dl, WideVecVT),
25407 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25408 unsigned WidenNumElts = NumElems*SizeRatio;
25409 unsigned MaskNumElts = VT.getVectorNumElements();
25410 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25413 unsigned NumConcat = WidenNumElts / MaskNumElts;
25414 SmallVector<SDValue, 16> Ops(NumConcat);
25415 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25417 for (unsigned i = 1; i != NumConcat; ++i)
25420 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25423 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25424 NewMask, StVT, Mst->getMemOperand(), false);
25426 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25427 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25428 const X86Subtarget *Subtarget) {
25429 StoreSDNode *St = cast<StoreSDNode>(N);
25430 EVT VT = St->getValue().getValueType();
25431 EVT StVT = St->getMemoryVT();
25433 SDValue StoredVal = St->getOperand(1);
25434 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25436 // If we are saving a concatenation of two XMM registers and 32-byte stores
25437 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25439 unsigned AddressSpace = St->getAddressSpace();
25440 unsigned Alignment = St->getAlignment();
25441 if (VT.is256BitVector() && StVT == VT &&
25442 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25443 AddressSpace, Alignment, &Fast) && !Fast) {
25444 unsigned NumElems = VT.getVectorNumElements();
25448 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25449 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25452 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25453 SDValue Ptr0 = St->getBasePtr();
25454 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25456 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25457 St->getPointerInfo(), St->isVolatile(),
25458 St->isNonTemporal(), Alignment);
25459 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25460 St->getPointerInfo(), St->isVolatile(),
25461 St->isNonTemporal(),
25462 std::min(16U, Alignment));
25463 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25466 // Optimize trunc store (of multiple scalars) to shuffle and store.
25467 // First, pack all of the elements in one place. Next, store to memory
25468 // in fewer chunks.
25469 if (St->isTruncatingStore() && VT.isVector()) {
25470 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25471 unsigned NumElems = VT.getVectorNumElements();
25472 assert(StVT != VT && "Cannot truncate to the same type");
25473 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25474 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25476 // The truncating store is legal in some cases. For example
25477 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25478 // are designated for truncate store.
25479 // In this case we don't need any further transformations.
25480 if (TLI.isTruncStoreLegal(VT, StVT))
25483 // From, To sizes and ElemCount must be pow of two
25484 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25485 // We are going to use the original vector elt for storing.
25486 // Accumulated smaller vector elements must be a multiple of the store size.
25487 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25489 unsigned SizeRatio = FromSz / ToSz;
25491 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25493 // Create a type on which we perform the shuffle
25494 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25495 StVT.getScalarType(), NumElems*SizeRatio);
25497 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25499 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25500 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25501 for (unsigned i = 0; i != NumElems; ++i)
25502 ShuffleVec[i] = i * SizeRatio;
25504 // Can't shuffle using an illegal type.
25505 if (!TLI.isTypeLegal(WideVecVT))
25508 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25509 DAG.getUNDEF(WideVecVT),
25511 // At this point all of the data is stored at the bottom of the
25512 // register. We now need to save it to mem.
25514 // Find the largest store unit
25515 MVT StoreType = MVT::i8;
25516 for (MVT Tp : MVT::integer_valuetypes()) {
25517 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25521 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25522 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25523 (64 <= NumElems * ToSz))
25524 StoreType = MVT::f64;
25526 // Bitcast the original vector into a vector of store-size units
25527 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25528 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25529 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25530 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25531 SmallVector<SDValue, 8> Chains;
25532 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25533 TLI.getPointerTy(DAG.getDataLayout()));
25534 SDValue Ptr = St->getBasePtr();
25536 // Perform one or more big stores into memory.
25537 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25538 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25539 StoreType, ShuffWide,
25540 DAG.getIntPtrConstant(i, dl));
25541 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25542 St->getPointerInfo(), St->isVolatile(),
25543 St->isNonTemporal(), St->getAlignment());
25544 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25545 Chains.push_back(Ch);
25548 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25551 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25552 // the FP state in cases where an emms may be missing.
25553 // A preferable solution to the general problem is to figure out the right
25554 // places to insert EMMS. This qualifies as a quick hack.
25556 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25557 if (VT.getSizeInBits() != 64)
25560 const Function *F = DAG.getMachineFunction().getFunction();
25561 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25563 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25564 if ((VT.isVector() ||
25565 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25566 isa<LoadSDNode>(St->getValue()) &&
25567 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25568 St->getChain().hasOneUse() && !St->isVolatile()) {
25569 SDNode* LdVal = St->getValue().getNode();
25570 LoadSDNode *Ld = nullptr;
25571 int TokenFactorIndex = -1;
25572 SmallVector<SDValue, 8> Ops;
25573 SDNode* ChainVal = St->getChain().getNode();
25574 // Must be a store of a load. We currently handle two cases: the load
25575 // is a direct child, and it's under an intervening TokenFactor. It is
25576 // possible to dig deeper under nested TokenFactors.
25577 if (ChainVal == LdVal)
25578 Ld = cast<LoadSDNode>(St->getChain());
25579 else if (St->getValue().hasOneUse() &&
25580 ChainVal->getOpcode() == ISD::TokenFactor) {
25581 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25582 if (ChainVal->getOperand(i).getNode() == LdVal) {
25583 TokenFactorIndex = i;
25584 Ld = cast<LoadSDNode>(St->getValue());
25586 Ops.push_back(ChainVal->getOperand(i));
25590 if (!Ld || !ISD::isNormalLoad(Ld))
25593 // If this is not the MMX case, i.e. we are just turning i64 load/store
25594 // into f64 load/store, avoid the transformation if there are multiple
25595 // uses of the loaded value.
25596 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25601 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25602 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25604 if (Subtarget->is64Bit() || F64IsLegal) {
25605 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25606 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25607 Ld->getPointerInfo(), Ld->isVolatile(),
25608 Ld->isNonTemporal(), Ld->isInvariant(),
25609 Ld->getAlignment());
25610 SDValue NewChain = NewLd.getValue(1);
25611 if (TokenFactorIndex != -1) {
25612 Ops.push_back(NewChain);
25613 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25615 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25616 St->getPointerInfo(),
25617 St->isVolatile(), St->isNonTemporal(),
25618 St->getAlignment());
25621 // Otherwise, lower to two pairs of 32-bit loads / stores.
25622 SDValue LoAddr = Ld->getBasePtr();
25623 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25624 DAG.getConstant(4, LdDL, MVT::i32));
25626 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25627 Ld->getPointerInfo(),
25628 Ld->isVolatile(), Ld->isNonTemporal(),
25629 Ld->isInvariant(), Ld->getAlignment());
25630 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25631 Ld->getPointerInfo().getWithOffset(4),
25632 Ld->isVolatile(), Ld->isNonTemporal(),
25634 MinAlign(Ld->getAlignment(), 4));
25636 SDValue NewChain = LoLd.getValue(1);
25637 if (TokenFactorIndex != -1) {
25638 Ops.push_back(LoLd);
25639 Ops.push_back(HiLd);
25640 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25643 LoAddr = St->getBasePtr();
25644 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25645 DAG.getConstant(4, StDL, MVT::i32));
25647 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25648 St->getPointerInfo(),
25649 St->isVolatile(), St->isNonTemporal(),
25650 St->getAlignment());
25651 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25652 St->getPointerInfo().getWithOffset(4),
25654 St->isNonTemporal(),
25655 MinAlign(St->getAlignment(), 4));
25656 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25659 // This is similar to the above case, but here we handle a scalar 64-bit
25660 // integer store that is extracted from a vector on a 32-bit target.
25661 // If we have SSE2, then we can treat it like a floating-point double
25662 // to get past legalization. The execution dependencies fixup pass will
25663 // choose the optimal machine instruction for the store if this really is
25664 // an integer or v2f32 rather than an f64.
25665 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25666 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25667 SDValue OldExtract = St->getOperand(1);
25668 SDValue ExtOp0 = OldExtract.getOperand(0);
25669 unsigned VecSize = ExtOp0.getValueSizeInBits();
25670 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25671 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25672 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25673 BitCast, OldExtract.getOperand(1));
25674 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25675 St->getPointerInfo(), St->isVolatile(),
25676 St->isNonTemporal(), St->getAlignment());
25682 /// Return 'true' if this vector operation is "horizontal"
25683 /// and return the operands for the horizontal operation in LHS and RHS. A
25684 /// horizontal operation performs the binary operation on successive elements
25685 /// of its first operand, then on successive elements of its second operand,
25686 /// returning the resulting values in a vector. For example, if
25687 /// A = < float a0, float a1, float a2, float a3 >
25689 /// B = < float b0, float b1, float b2, float b3 >
25690 /// then the result of doing a horizontal operation on A and B is
25691 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25692 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25693 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25694 /// set to A, RHS to B, and the routine returns 'true'.
25695 /// Note that the binary operation should have the property that if one of the
25696 /// operands is UNDEF then the result is UNDEF.
25697 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25698 // Look for the following pattern: if
25699 // A = < float a0, float a1, float a2, float a3 >
25700 // B = < float b0, float b1, float b2, float b3 >
25702 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25703 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25704 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25705 // which is A horizontal-op B.
25707 // At least one of the operands should be a vector shuffle.
25708 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25709 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25712 MVT VT = LHS.getSimpleValueType();
25714 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25715 "Unsupported vector type for horizontal add/sub");
25717 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25718 // operate independently on 128-bit lanes.
25719 unsigned NumElts = VT.getVectorNumElements();
25720 unsigned NumLanes = VT.getSizeInBits()/128;
25721 unsigned NumLaneElts = NumElts / NumLanes;
25722 assert((NumLaneElts % 2 == 0) &&
25723 "Vector type should have an even number of elements in each lane");
25724 unsigned HalfLaneElts = NumLaneElts/2;
25726 // View LHS in the form
25727 // LHS = VECTOR_SHUFFLE A, B, LMask
25728 // If LHS is not a shuffle then pretend it is the shuffle
25729 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25730 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25733 SmallVector<int, 16> LMask(NumElts);
25734 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25735 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25736 A = LHS.getOperand(0);
25737 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25738 B = LHS.getOperand(1);
25739 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25740 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25742 if (LHS.getOpcode() != ISD::UNDEF)
25744 for (unsigned i = 0; i != NumElts; ++i)
25748 // Likewise, view RHS in the form
25749 // RHS = VECTOR_SHUFFLE C, D, RMask
25751 SmallVector<int, 16> RMask(NumElts);
25752 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25753 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25754 C = RHS.getOperand(0);
25755 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25756 D = RHS.getOperand(1);
25757 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25758 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25760 if (RHS.getOpcode() != ISD::UNDEF)
25762 for (unsigned i = 0; i != NumElts; ++i)
25766 // Check that the shuffles are both shuffling the same vectors.
25767 if (!(A == C && B == D) && !(A == D && B == C))
25770 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25771 if (!A.getNode() && !B.getNode())
25774 // If A and B occur in reverse order in RHS, then "swap" them (which means
25775 // rewriting the mask).
25777 ShuffleVectorSDNode::commuteMask(RMask);
25779 // At this point LHS and RHS are equivalent to
25780 // LHS = VECTOR_SHUFFLE A, B, LMask
25781 // RHS = VECTOR_SHUFFLE A, B, RMask
25782 // Check that the masks correspond to performing a horizontal operation.
25783 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25784 for (unsigned i = 0; i != NumLaneElts; ++i) {
25785 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25787 // Ignore any UNDEF components.
25788 if (LIdx < 0 || RIdx < 0 ||
25789 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25790 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25793 // Check that successive elements are being operated on. If not, this is
25794 // not a horizontal operation.
25795 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25796 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25797 if (!(LIdx == Index && RIdx == Index + 1) &&
25798 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25803 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25804 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25808 /// Do target-specific dag combines on floating point adds.
25809 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25810 const X86Subtarget *Subtarget) {
25811 EVT VT = N->getValueType(0);
25812 SDValue LHS = N->getOperand(0);
25813 SDValue RHS = N->getOperand(1);
25815 // Try to synthesize horizontal adds from adds of shuffles.
25816 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25817 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25818 isHorizontalBinOp(LHS, RHS, true))
25819 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25823 /// Do target-specific dag combines on floating point subs.
25824 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25825 const X86Subtarget *Subtarget) {
25826 EVT VT = N->getValueType(0);
25827 SDValue LHS = N->getOperand(0);
25828 SDValue RHS = N->getOperand(1);
25830 // Try to synthesize horizontal subs from subs of shuffles.
25831 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25832 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25833 isHorizontalBinOp(LHS, RHS, false))
25834 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25838 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25839 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25840 const X86Subtarget *Subtarget) {
25841 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25843 // F[X]OR(0.0, x) -> x
25844 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25845 if (C->getValueAPF().isPosZero())
25846 return N->getOperand(1);
25848 // F[X]OR(x, 0.0) -> x
25849 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25850 if (C->getValueAPF().isPosZero())
25851 return N->getOperand(0);
25853 EVT VT = N->getValueType(0);
25854 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25856 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25857 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25859 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25860 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25861 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25862 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25863 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25868 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25869 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25870 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25872 // Only perform optimizations if UnsafeMath is used.
25873 if (!DAG.getTarget().Options.UnsafeFPMath)
25876 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25877 // into FMINC and FMAXC, which are Commutative operations.
25878 unsigned NewOp = 0;
25879 switch (N->getOpcode()) {
25880 default: llvm_unreachable("unknown opcode");
25881 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25882 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25885 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25886 N->getOperand(0), N->getOperand(1));
25889 /// Do target-specific dag combines on X86ISD::FAND nodes.
25890 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25891 // FAND(0.0, x) -> 0.0
25892 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25893 if (C->getValueAPF().isPosZero())
25894 return N->getOperand(0);
25896 // FAND(x, 0.0) -> 0.0
25897 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25898 if (C->getValueAPF().isPosZero())
25899 return N->getOperand(1);
25904 /// Do target-specific dag combines on X86ISD::FANDN nodes
25905 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25906 // FANDN(0.0, x) -> x
25907 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25908 if (C->getValueAPF().isPosZero())
25909 return N->getOperand(1);
25911 // FANDN(x, 0.0) -> 0.0
25912 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25913 if (C->getValueAPF().isPosZero())
25914 return N->getOperand(1);
25919 static SDValue PerformBTCombine(SDNode *N,
25921 TargetLowering::DAGCombinerInfo &DCI) {
25922 // BT ignores high bits in the bit index operand.
25923 SDValue Op1 = N->getOperand(1);
25924 if (Op1.hasOneUse()) {
25925 unsigned BitWidth = Op1.getValueSizeInBits();
25926 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25927 APInt KnownZero, KnownOne;
25928 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25929 !DCI.isBeforeLegalizeOps());
25930 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25931 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25932 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25933 DCI.CommitTargetLoweringOpt(TLO);
25938 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25939 SDValue Op = N->getOperand(0);
25940 if (Op.getOpcode() == ISD::BITCAST)
25941 Op = Op.getOperand(0);
25942 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25943 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25944 VT.getVectorElementType().getSizeInBits() ==
25945 OpVT.getVectorElementType().getSizeInBits()) {
25946 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25951 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25952 const X86Subtarget *Subtarget) {
25953 EVT VT = N->getValueType(0);
25954 if (!VT.isVector())
25957 SDValue N0 = N->getOperand(0);
25958 SDValue N1 = N->getOperand(1);
25959 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25962 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25963 // both SSE and AVX2 since there is no sign-extended shift right
25964 // operation on a vector with 64-bit elements.
25965 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25966 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25967 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25968 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25969 SDValue N00 = N0.getOperand(0);
25971 // EXTLOAD has a better solution on AVX2,
25972 // it may be replaced with X86ISD::VSEXT node.
25973 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25974 if (!ISD::isNormalLoad(N00.getNode()))
25977 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25978 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25980 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25986 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25987 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25988 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25989 /// eliminate extend, add, and shift instructions.
25990 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25991 const X86Subtarget *Subtarget) {
25992 // TODO: This should be valid for other integer types.
25993 EVT VT = Sext->getValueType(0);
25994 if (VT != MVT::i64)
25997 // We need an 'add nsw' feeding into the 'sext'.
25998 SDValue Add = Sext->getOperand(0);
25999 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
26002 // Having a constant operand to the 'add' ensures that we are not increasing
26003 // the instruction count because the constant is extended for free below.
26004 // A constant operand can also become the displacement field of an LEA.
26005 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
26009 // Don't make the 'add' bigger if there's no hope of combining it with some
26010 // other 'add' or 'shl' instruction.
26011 // TODO: It may be profitable to generate simpler LEA instructions in place
26012 // of single 'add' instructions, but the cost model for selecting an LEA
26013 // currently has a high threshold.
26014 bool HasLEAPotential = false;
26015 for (auto *User : Sext->uses()) {
26016 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
26017 HasLEAPotential = true;
26021 if (!HasLEAPotential)
26024 // Everything looks good, so pull the 'sext' ahead of the 'add'.
26025 int64_t AddConstant = AddOp1->getSExtValue();
26026 SDValue AddOp0 = Add.getOperand(0);
26027 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
26028 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
26030 // The wider add is guaranteed to not wrap because both operands are
26033 Flags.setNoSignedWrap(true);
26034 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
26037 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
26038 TargetLowering::DAGCombinerInfo &DCI,
26039 const X86Subtarget *Subtarget) {
26040 SDValue N0 = N->getOperand(0);
26041 EVT VT = N->getValueType(0);
26042 EVT SVT = VT.getScalarType();
26043 EVT InVT = N0.getValueType();
26044 EVT InSVT = InVT.getScalarType();
26047 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26048 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26049 // This exposes the sext to the sdivrem lowering, so that it directly extends
26050 // from AH (which we otherwise need to do contortions to access).
26051 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26052 InVT == MVT::i8 && VT == MVT::i32) {
26053 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26054 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26055 N0.getOperand(0), N0.getOperand(1));
26056 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26057 return R.getValue(1);
26060 if (!DCI.isBeforeLegalizeOps()) {
26061 if (InVT == MVT::i1) {
26062 SDValue Zero = DAG.getConstant(0, DL, VT);
26064 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26065 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26070 if (VT.isVector() && Subtarget->hasSSE2()) {
26071 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26072 EVT InVT = N.getValueType();
26073 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26074 Size / InVT.getScalarSizeInBits());
26075 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26076 DAG.getUNDEF(InVT));
26078 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26081 // If target-size is less than 128-bits, extend to a type that would extend
26082 // to 128 bits, extend that and extract the original target vector.
26083 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26084 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26085 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26086 unsigned Scale = 128 / VT.getSizeInBits();
26088 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26089 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26090 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26091 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26092 DAG.getIntPtrConstant(0, DL));
26095 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26096 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26097 if (VT.getSizeInBits() == 128 &&
26098 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26099 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26100 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26101 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26104 // On pre-AVX2 targets, split into 128-bit nodes of
26105 // ISD::SIGN_EXTEND_VECTOR_INREG.
26106 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26107 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26108 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26109 unsigned NumVecs = VT.getSizeInBits() / 128;
26110 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26111 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26112 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26114 SmallVector<SDValue, 8> Opnds;
26115 for (unsigned i = 0, Offset = 0; i != NumVecs;
26116 ++i, Offset += NumSubElts) {
26117 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26118 DAG.getIntPtrConstant(Offset, DL));
26119 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26120 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26121 Opnds.push_back(SrcVec);
26123 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26127 if (Subtarget->hasAVX() && VT.is256BitVector())
26128 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26131 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26137 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26138 const X86Subtarget* Subtarget) {
26140 EVT VT = N->getValueType(0);
26142 // Let legalize expand this if it isn't a legal type yet.
26143 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26146 EVT ScalarVT = VT.getScalarType();
26147 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26148 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26149 !Subtarget->hasAVX512()))
26152 SDValue A = N->getOperand(0);
26153 SDValue B = N->getOperand(1);
26154 SDValue C = N->getOperand(2);
26156 bool NegA = (A.getOpcode() == ISD::FNEG);
26157 bool NegB = (B.getOpcode() == ISD::FNEG);
26158 bool NegC = (C.getOpcode() == ISD::FNEG);
26160 // Negative multiplication when NegA xor NegB
26161 bool NegMul = (NegA != NegB);
26163 A = A.getOperand(0);
26165 B = B.getOperand(0);
26167 C = C.getOperand(0);
26171 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26173 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26175 return DAG.getNode(Opcode, dl, VT, A, B, C);
26178 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26179 TargetLowering::DAGCombinerInfo &DCI,
26180 const X86Subtarget *Subtarget) {
26181 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26182 // (and (i32 x86isd::setcc_carry), 1)
26183 // This eliminates the zext. This transformation is necessary because
26184 // ISD::SETCC is always legalized to i8.
26186 SDValue N0 = N->getOperand(0);
26187 EVT VT = N->getValueType(0);
26189 if (N0.getOpcode() == ISD::AND &&
26191 N0.getOperand(0).hasOneUse()) {
26192 SDValue N00 = N0.getOperand(0);
26193 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26194 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26195 if (!C || C->getZExtValue() != 1)
26197 return DAG.getNode(ISD::AND, dl, VT,
26198 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26199 N00.getOperand(0), N00.getOperand(1)),
26200 DAG.getConstant(1, dl, VT));
26204 if (N0.getOpcode() == ISD::TRUNCATE &&
26206 N0.getOperand(0).hasOneUse()) {
26207 SDValue N00 = N0.getOperand(0);
26208 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26209 return DAG.getNode(ISD::AND, dl, VT,
26210 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26211 N00.getOperand(0), N00.getOperand(1)),
26212 DAG.getConstant(1, dl, VT));
26216 if (VT.is256BitVector())
26217 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26220 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26221 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26222 // This exposes the zext to the udivrem lowering, so that it directly extends
26223 // from AH (which we otherwise need to do contortions to access).
26224 if (N0.getOpcode() == ISD::UDIVREM &&
26225 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26226 (VT == MVT::i32 || VT == MVT::i64)) {
26227 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26228 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26229 N0.getOperand(0), N0.getOperand(1));
26230 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26231 return R.getValue(1);
26237 // Optimize x == -y --> x+y == 0
26238 // x != -y --> x+y != 0
26239 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26240 const X86Subtarget* Subtarget) {
26241 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26242 SDValue LHS = N->getOperand(0);
26243 SDValue RHS = N->getOperand(1);
26244 EVT VT = N->getValueType(0);
26247 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26248 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26249 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26250 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26251 LHS.getOperand(1));
26252 return DAG.getSetCC(DL, N->getValueType(0), addV,
26253 DAG.getConstant(0, DL, addV.getValueType()), CC);
26255 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26256 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26257 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26258 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26259 RHS.getOperand(1));
26260 return DAG.getSetCC(DL, N->getValueType(0), addV,
26261 DAG.getConstant(0, DL, addV.getValueType()), CC);
26264 if (VT.getScalarType() == MVT::i1 &&
26265 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26267 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26268 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26269 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26271 if (!IsSEXT0 || !IsVZero1) {
26272 // Swap the operands and update the condition code.
26273 std::swap(LHS, RHS);
26274 CC = ISD::getSetCCSwappedOperands(CC);
26276 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26277 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26278 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26281 if (IsSEXT0 && IsVZero1) {
26282 assert(VT == LHS.getOperand(0).getValueType() &&
26283 "Uexpected operand type");
26284 if (CC == ISD::SETGT)
26285 return DAG.getConstant(0, DL, VT);
26286 if (CC == ISD::SETLE)
26287 return DAG.getConstant(1, DL, VT);
26288 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26289 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26291 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26292 "Unexpected condition code!");
26293 return LHS.getOperand(0);
26300 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26301 SDValue V0 = N->getOperand(0);
26302 SDValue V1 = N->getOperand(1);
26304 EVT VT = N->getValueType(0);
26306 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26307 // operands and changing the mask to 1. This saves us a bunch of
26308 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26309 // x86InstrInfo knows how to commute this back after instruction selection
26310 // if it would help register allocation.
26312 // TODO: If optimizing for size or a processor that doesn't suffer from
26313 // partial register update stalls, this should be transformed into a MOVSD
26314 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26316 if (VT == MVT::v2f64)
26317 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26318 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26319 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26320 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26326 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26327 // as "sbb reg,reg", since it can be extended without zext and produces
26328 // an all-ones bit which is more useful than 0/1 in some cases.
26329 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26332 return DAG.getNode(ISD::AND, DL, VT,
26333 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26334 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26336 DAG.getConstant(1, DL, VT));
26337 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26338 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26339 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26340 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26344 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26345 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26346 TargetLowering::DAGCombinerInfo &DCI,
26347 const X86Subtarget *Subtarget) {
26349 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26350 SDValue EFLAGS = N->getOperand(1);
26352 if (CC == X86::COND_A) {
26353 // Try to convert COND_A into COND_B in an attempt to facilitate
26354 // materializing "setb reg".
26356 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26357 // cannot take an immediate as its first operand.
26359 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26360 EFLAGS.getValueType().isInteger() &&
26361 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26362 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26363 EFLAGS.getNode()->getVTList(),
26364 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26365 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26366 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26370 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26371 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26373 if (CC == X86::COND_B)
26374 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26376 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26377 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26378 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26384 // Optimize branch condition evaluation.
26386 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26387 TargetLowering::DAGCombinerInfo &DCI,
26388 const X86Subtarget *Subtarget) {
26390 SDValue Chain = N->getOperand(0);
26391 SDValue Dest = N->getOperand(1);
26392 SDValue EFLAGS = N->getOperand(3);
26393 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26395 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26396 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26397 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26404 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26405 SelectionDAG &DAG) {
26406 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26407 // optimize away operation when it's from a constant.
26409 // The general transformation is:
26410 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26411 // AND(VECTOR_CMP(x,y), constant2)
26412 // constant2 = UNARYOP(constant)
26414 // Early exit if this isn't a vector operation, the operand of the
26415 // unary operation isn't a bitwise AND, or if the sizes of the operations
26416 // aren't the same.
26417 EVT VT = N->getValueType(0);
26418 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26419 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26420 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26423 // Now check that the other operand of the AND is a constant. We could
26424 // make the transformation for non-constant splats as well, but it's unclear
26425 // that would be a benefit as it would not eliminate any operations, just
26426 // perform one more step in scalar code before moving to the vector unit.
26427 if (BuildVectorSDNode *BV =
26428 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26429 // Bail out if the vector isn't a constant.
26430 if (!BV->isConstant())
26433 // Everything checks out. Build up the new and improved node.
26435 EVT IntVT = BV->getValueType(0);
26436 // Create a new constant of the appropriate type for the transformed
26438 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26439 // The AND node needs bitcasts to/from an integer vector type around it.
26440 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26441 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26442 N->getOperand(0)->getOperand(0), MaskConst);
26443 SDValue Res = DAG.getBitcast(VT, NewAnd);
26450 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26451 const X86Subtarget *Subtarget) {
26452 SDValue Op0 = N->getOperand(0);
26453 EVT VT = N->getValueType(0);
26454 EVT InVT = Op0.getValueType();
26455 EVT InSVT = InVT.getScalarType();
26456 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26458 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26459 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26460 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26462 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26463 InVT.getVectorNumElements());
26464 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26466 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26467 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26469 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26475 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26476 const X86Subtarget *Subtarget) {
26477 // First try to optimize away the conversion entirely when it's
26478 // conditionally from a constant. Vectors only.
26479 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26482 // Now move on to more general possibilities.
26483 SDValue Op0 = N->getOperand(0);
26484 EVT VT = N->getValueType(0);
26485 EVT InVT = Op0.getValueType();
26486 EVT InSVT = InVT.getScalarType();
26488 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26489 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26490 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26492 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26493 InVT.getVectorNumElements());
26494 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26495 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26498 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26499 // a 32-bit target where SSE doesn't support i64->FP operations.
26500 if (!Subtarget->useSoftFloat() && Op0.getOpcode() == ISD::LOAD) {
26501 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26502 EVT LdVT = Ld->getValueType(0);
26504 // This transformation is not supported if the result type is f16
26505 if (VT == MVT::f16)
26508 if (!Ld->isVolatile() && !VT.isVector() &&
26509 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26510 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26511 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26512 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26513 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26520 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26521 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26522 X86TargetLowering::DAGCombinerInfo &DCI) {
26523 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26524 // the result is either zero or one (depending on the input carry bit).
26525 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26526 if (X86::isZeroNode(N->getOperand(0)) &&
26527 X86::isZeroNode(N->getOperand(1)) &&
26528 // We don't have a good way to replace an EFLAGS use, so only do this when
26530 SDValue(N, 1).use_empty()) {
26532 EVT VT = N->getValueType(0);
26533 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26534 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26535 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26536 DAG.getConstant(X86::COND_B, DL,
26539 DAG.getConstant(1, DL, VT));
26540 return DCI.CombineTo(N, Res1, CarryOut);
26546 // fold (add Y, (sete X, 0)) -> adc 0, Y
26547 // (add Y, (setne X, 0)) -> sbb -1, Y
26548 // (sub (sete X, 0), Y) -> sbb 0, Y
26549 // (sub (setne X, 0), Y) -> adc -1, Y
26550 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26553 // Look through ZExts.
26554 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26555 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26558 SDValue SetCC = Ext.getOperand(0);
26559 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26562 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26563 if (CC != X86::COND_E && CC != X86::COND_NE)
26566 SDValue Cmp = SetCC.getOperand(1);
26567 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26568 !X86::isZeroNode(Cmp.getOperand(1)) ||
26569 !Cmp.getOperand(0).getValueType().isInteger())
26572 SDValue CmpOp0 = Cmp.getOperand(0);
26573 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26574 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26576 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26577 if (CC == X86::COND_NE)
26578 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26579 DL, OtherVal.getValueType(), OtherVal,
26580 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26582 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26583 DL, OtherVal.getValueType(), OtherVal,
26584 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26587 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26588 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26589 const X86Subtarget *Subtarget) {
26590 EVT VT = N->getValueType(0);
26591 SDValue Op0 = N->getOperand(0);
26592 SDValue Op1 = N->getOperand(1);
26594 // Try to synthesize horizontal adds from adds of shuffles.
26595 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26596 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26597 isHorizontalBinOp(Op0, Op1, true))
26598 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26600 return OptimizeConditionalInDecrement(N, DAG);
26603 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26604 const X86Subtarget *Subtarget) {
26605 SDValue Op0 = N->getOperand(0);
26606 SDValue Op1 = N->getOperand(1);
26608 // X86 can't encode an immediate LHS of a sub. See if we can push the
26609 // negation into a preceding instruction.
26610 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26611 // If the RHS of the sub is a XOR with one use and a constant, invert the
26612 // immediate. Then add one to the LHS of the sub so we can turn
26613 // X-Y -> X+~Y+1, saving one register.
26614 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26615 isa<ConstantSDNode>(Op1.getOperand(1))) {
26616 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26617 EVT VT = Op0.getValueType();
26618 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26620 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26621 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26622 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26626 // Try to synthesize horizontal adds from adds of shuffles.
26627 EVT VT = N->getValueType(0);
26628 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26629 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26630 isHorizontalBinOp(Op0, Op1, true))
26631 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26633 return OptimizeConditionalInDecrement(N, DAG);
26636 /// performVZEXTCombine - Performs build vector combines
26637 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26638 TargetLowering::DAGCombinerInfo &DCI,
26639 const X86Subtarget *Subtarget) {
26641 MVT VT = N->getSimpleValueType(0);
26642 SDValue Op = N->getOperand(0);
26643 MVT OpVT = Op.getSimpleValueType();
26644 MVT OpEltVT = OpVT.getVectorElementType();
26645 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26647 // (vzext (bitcast (vzext (x)) -> (vzext x)
26649 while (V.getOpcode() == ISD::BITCAST)
26650 V = V.getOperand(0);
26652 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26653 MVT InnerVT = V.getSimpleValueType();
26654 MVT InnerEltVT = InnerVT.getVectorElementType();
26656 // If the element sizes match exactly, we can just do one larger vzext. This
26657 // is always an exact type match as vzext operates on integer types.
26658 if (OpEltVT == InnerEltVT) {
26659 assert(OpVT == InnerVT && "Types must match for vzext!");
26660 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26663 // The only other way we can combine them is if only a single element of the
26664 // inner vzext is used in the input to the outer vzext.
26665 if (InnerEltVT.getSizeInBits() < InputBits)
26668 // In this case, the inner vzext is completely dead because we're going to
26669 // only look at bits inside of the low element. Just do the outer vzext on
26670 // a bitcast of the input to the inner.
26671 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26674 // Check if we can bypass extracting and re-inserting an element of an input
26675 // vector. Essentially:
26676 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26677 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26678 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26679 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26680 SDValue ExtractedV = V.getOperand(0);
26681 SDValue OrigV = ExtractedV.getOperand(0);
26682 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26683 if (ExtractIdx->getZExtValue() == 0) {
26684 MVT OrigVT = OrigV.getSimpleValueType();
26685 // Extract a subvector if necessary...
26686 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26687 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26688 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26689 OrigVT.getVectorNumElements() / Ratio);
26690 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26691 DAG.getIntPtrConstant(0, DL));
26693 Op = DAG.getBitcast(OpVT, OrigV);
26694 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26701 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26702 DAGCombinerInfo &DCI) const {
26703 SelectionDAG &DAG = DCI.DAG;
26704 switch (N->getOpcode()) {
26706 case ISD::EXTRACT_VECTOR_ELT:
26707 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26710 case X86ISD::SHRUNKBLEND:
26711 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26712 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26713 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26714 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26715 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26716 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26717 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26720 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26721 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26722 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26723 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26724 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26725 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26726 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26727 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26728 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26729 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26730 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26731 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26733 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26735 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26736 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26737 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26738 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26739 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26740 case ISD::ANY_EXTEND:
26741 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26742 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26743 case ISD::SIGN_EXTEND_INREG:
26744 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26745 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26746 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26747 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26748 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26749 case X86ISD::SHUFP: // Handle all target specific shuffles
26750 case X86ISD::PALIGNR:
26751 case X86ISD::UNPCKH:
26752 case X86ISD::UNPCKL:
26753 case X86ISD::MOVHLPS:
26754 case X86ISD::MOVLHPS:
26755 case X86ISD::PSHUFB:
26756 case X86ISD::PSHUFD:
26757 case X86ISD::PSHUFHW:
26758 case X86ISD::PSHUFLW:
26759 case X86ISD::MOVSS:
26760 case X86ISD::MOVSD:
26761 case X86ISD::VPERMILPI:
26762 case X86ISD::VPERM2X128:
26763 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26764 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26765 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26771 /// isTypeDesirableForOp - Return true if the target has native support for
26772 /// the specified value type and it is 'desirable' to use the type for the
26773 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26774 /// instruction encodings are longer and some i16 instructions are slow.
26775 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26776 if (!isTypeLegal(VT))
26778 if (VT != MVT::i16)
26785 case ISD::SIGN_EXTEND:
26786 case ISD::ZERO_EXTEND:
26787 case ISD::ANY_EXTEND:
26800 /// IsDesirableToPromoteOp - This method query the target whether it is
26801 /// beneficial for dag combiner to promote the specified node. If true, it
26802 /// should return the desired promotion type by reference.
26803 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26804 EVT VT = Op.getValueType();
26805 if (VT != MVT::i16)
26808 bool Promote = false;
26809 bool Commute = false;
26810 switch (Op.getOpcode()) {
26813 LoadSDNode *LD = cast<LoadSDNode>(Op);
26814 // If the non-extending load has a single use and it's not live out, then it
26815 // might be folded.
26816 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26817 Op.hasOneUse()*/) {
26818 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26819 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26820 // The only case where we'd want to promote LOAD (rather then it being
26821 // promoted as an operand is when it's only use is liveout.
26822 if (UI->getOpcode() != ISD::CopyToReg)
26829 case ISD::SIGN_EXTEND:
26830 case ISD::ZERO_EXTEND:
26831 case ISD::ANY_EXTEND:
26836 SDValue N0 = Op.getOperand(0);
26837 // Look out for (store (shl (load), x)).
26838 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26851 SDValue N0 = Op.getOperand(0);
26852 SDValue N1 = Op.getOperand(1);
26853 if (!Commute && MayFoldLoad(N1))
26855 // Avoid disabling potential load folding opportunities.
26856 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26858 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26868 //===----------------------------------------------------------------------===//
26869 // X86 Inline Assembly Support
26870 //===----------------------------------------------------------------------===//
26872 // Helper to match a string separated by whitespace.
26873 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26874 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26876 for (StringRef Piece : Pieces) {
26877 if (!S.startswith(Piece)) // Check if the piece matches.
26880 S = S.substr(Piece.size());
26881 StringRef::size_type Pos = S.find_first_not_of(" \t");
26882 if (Pos == 0) // We matched a prefix.
26891 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26893 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26894 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26895 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26896 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26898 if (AsmPieces.size() == 3)
26900 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26907 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26908 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26910 std::string AsmStr = IA->getAsmString();
26912 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26913 if (!Ty || Ty->getBitWidth() % 16 != 0)
26916 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26917 SmallVector<StringRef, 4> AsmPieces;
26918 SplitString(AsmStr, AsmPieces, ";\n");
26920 switch (AsmPieces.size()) {
26921 default: return false;
26923 // FIXME: this should verify that we are targeting a 486 or better. If not,
26924 // we will turn this bswap into something that will be lowered to logical
26925 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26926 // lower so don't worry about this.
26928 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26929 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26930 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26931 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26932 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26933 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26934 // No need to check constraints, nothing other than the equivalent of
26935 // "=r,0" would be valid here.
26936 return IntrinsicLowering::LowerToByteSwap(CI);
26939 // rorw $$8, ${0:w} --> llvm.bswap.i16
26940 if (CI->getType()->isIntegerTy(16) &&
26941 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26942 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26943 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26945 StringRef ConstraintsStr = IA->getConstraintString();
26946 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26947 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26948 if (clobbersFlagRegisters(AsmPieces))
26949 return IntrinsicLowering::LowerToByteSwap(CI);
26953 if (CI->getType()->isIntegerTy(32) &&
26954 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26955 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26956 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26957 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26959 StringRef ConstraintsStr = IA->getConstraintString();
26960 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26961 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26962 if (clobbersFlagRegisters(AsmPieces))
26963 return IntrinsicLowering::LowerToByteSwap(CI);
26966 if (CI->getType()->isIntegerTy(64)) {
26967 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26968 if (Constraints.size() >= 2 &&
26969 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26970 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26971 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26972 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26973 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26974 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26975 return IntrinsicLowering::LowerToByteSwap(CI);
26983 /// getConstraintType - Given a constraint letter, return the type of
26984 /// constraint it is for this target.
26985 X86TargetLowering::ConstraintType
26986 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26987 if (Constraint.size() == 1) {
26988 switch (Constraint[0]) {
26999 return C_RegisterClass;
27023 return TargetLowering::getConstraintType(Constraint);
27026 /// Examine constraint type and operand type and determine a weight value.
27027 /// This object must already have been set up with the operand type
27028 /// and the current alternative constraint selected.
27029 TargetLowering::ConstraintWeight
27030 X86TargetLowering::getSingleConstraintMatchWeight(
27031 AsmOperandInfo &info, const char *constraint) const {
27032 ConstraintWeight weight = CW_Invalid;
27033 Value *CallOperandVal = info.CallOperandVal;
27034 // If we don't have a value, we can't do a match,
27035 // but allow it at the lowest weight.
27036 if (!CallOperandVal)
27038 Type *type = CallOperandVal->getType();
27039 // Look at the constraint type.
27040 switch (*constraint) {
27042 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27053 if (CallOperandVal->getType()->isIntegerTy())
27054 weight = CW_SpecificReg;
27059 if (type->isFloatingPointTy())
27060 weight = CW_SpecificReg;
27063 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27064 weight = CW_SpecificReg;
27068 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27069 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27070 weight = CW_Register;
27073 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27074 if (C->getZExtValue() <= 31)
27075 weight = CW_Constant;
27079 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27080 if (C->getZExtValue() <= 63)
27081 weight = CW_Constant;
27085 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27086 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27087 weight = CW_Constant;
27091 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27092 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27093 weight = CW_Constant;
27097 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27098 if (C->getZExtValue() <= 3)
27099 weight = CW_Constant;
27103 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27104 if (C->getZExtValue() <= 0xff)
27105 weight = CW_Constant;
27110 if (isa<ConstantFP>(CallOperandVal)) {
27111 weight = CW_Constant;
27115 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27116 if ((C->getSExtValue() >= -0x80000000LL) &&
27117 (C->getSExtValue() <= 0x7fffffffLL))
27118 weight = CW_Constant;
27122 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27123 if (C->getZExtValue() <= 0xffffffff)
27124 weight = CW_Constant;
27131 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27132 /// with another that has more specific requirements based on the type of the
27133 /// corresponding operand.
27134 const char *X86TargetLowering::
27135 LowerXConstraint(EVT ConstraintVT) const {
27136 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27137 // 'f' like normal targets.
27138 if (ConstraintVT.isFloatingPoint()) {
27139 if (Subtarget->hasSSE2())
27141 if (Subtarget->hasSSE1())
27145 return TargetLowering::LowerXConstraint(ConstraintVT);
27148 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27149 /// vector. If it is invalid, don't add anything to Ops.
27150 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27151 std::string &Constraint,
27152 std::vector<SDValue>&Ops,
27153 SelectionDAG &DAG) const {
27156 // Only support length 1 constraints for now.
27157 if (Constraint.length() > 1) return;
27159 char ConstraintLetter = Constraint[0];
27160 switch (ConstraintLetter) {
27163 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27164 if (C->getZExtValue() <= 31) {
27165 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27166 Op.getValueType());
27172 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27173 if (C->getZExtValue() <= 63) {
27174 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27175 Op.getValueType());
27181 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27182 if (isInt<8>(C->getSExtValue())) {
27183 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27184 Op.getValueType());
27190 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27191 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27192 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27193 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27194 Op.getValueType());
27200 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27201 if (C->getZExtValue() <= 3) {
27202 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27203 Op.getValueType());
27209 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27210 if (C->getZExtValue() <= 255) {
27211 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27212 Op.getValueType());
27218 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27219 if (C->getZExtValue() <= 127) {
27220 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27221 Op.getValueType());
27227 // 32-bit signed value
27228 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27229 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27230 C->getSExtValue())) {
27231 // Widen to 64 bits here to get it sign extended.
27232 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27235 // FIXME gcc accepts some relocatable values here too, but only in certain
27236 // memory models; it's complicated.
27241 // 32-bit unsigned value
27242 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27243 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27244 C->getZExtValue())) {
27245 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27246 Op.getValueType());
27250 // FIXME gcc accepts some relocatable values here too, but only in certain
27251 // memory models; it's complicated.
27255 // Literal immediates are always ok.
27256 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27257 // Widen to 64 bits here to get it sign extended.
27258 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27262 // In any sort of PIC mode addresses need to be computed at runtime by
27263 // adding in a register or some sort of table lookup. These can't
27264 // be used as immediates.
27265 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27268 // If we are in non-pic codegen mode, we allow the address of a global (with
27269 // an optional displacement) to be used with 'i'.
27270 GlobalAddressSDNode *GA = nullptr;
27271 int64_t Offset = 0;
27273 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27275 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27276 Offset += GA->getOffset();
27278 } else if (Op.getOpcode() == ISD::ADD) {
27279 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27280 Offset += C->getZExtValue();
27281 Op = Op.getOperand(0);
27284 } else if (Op.getOpcode() == ISD::SUB) {
27285 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27286 Offset += -C->getZExtValue();
27287 Op = Op.getOperand(0);
27292 // Otherwise, this isn't something we can handle, reject it.
27296 const GlobalValue *GV = GA->getGlobal();
27297 // If we require an extra load to get this address, as in PIC mode, we
27298 // can't accept it.
27299 if (isGlobalStubReference(
27300 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27303 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27304 GA->getValueType(0), Offset);
27309 if (Result.getNode()) {
27310 Ops.push_back(Result);
27313 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27316 std::pair<unsigned, const TargetRegisterClass *>
27317 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27318 StringRef Constraint,
27320 // First, see if this is a constraint that directly corresponds to an LLVM
27322 if (Constraint.size() == 1) {
27323 // GCC Constraint Letters
27324 switch (Constraint[0]) {
27326 // TODO: Slight differences here in allocation order and leaving
27327 // RIP in the class. Do they matter any more here than they do
27328 // in the normal allocation?
27329 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27330 if (Subtarget->is64Bit()) {
27331 if (VT == MVT::i32 || VT == MVT::f32)
27332 return std::make_pair(0U, &X86::GR32RegClass);
27333 if (VT == MVT::i16)
27334 return std::make_pair(0U, &X86::GR16RegClass);
27335 if (VT == MVT::i8 || VT == MVT::i1)
27336 return std::make_pair(0U, &X86::GR8RegClass);
27337 if (VT == MVT::i64 || VT == MVT::f64)
27338 return std::make_pair(0U, &X86::GR64RegClass);
27341 // 32-bit fallthrough
27342 case 'Q': // Q_REGS
27343 if (VT == MVT::i32 || VT == MVT::f32)
27344 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27345 if (VT == MVT::i16)
27346 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27347 if (VT == MVT::i8 || VT == MVT::i1)
27348 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27349 if (VT == MVT::i64)
27350 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27352 case 'r': // GENERAL_REGS
27353 case 'l': // INDEX_REGS
27354 if (VT == MVT::i8 || VT == MVT::i1)
27355 return std::make_pair(0U, &X86::GR8RegClass);
27356 if (VT == MVT::i16)
27357 return std::make_pair(0U, &X86::GR16RegClass);
27358 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27359 return std::make_pair(0U, &X86::GR32RegClass);
27360 return std::make_pair(0U, &X86::GR64RegClass);
27361 case 'R': // LEGACY_REGS
27362 if (VT == MVT::i8 || VT == MVT::i1)
27363 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27364 if (VT == MVT::i16)
27365 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27366 if (VT == MVT::i32 || !Subtarget->is64Bit())
27367 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27368 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27369 case 'f': // FP Stack registers.
27370 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27371 // value to the correct fpstack register class.
27372 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27373 return std::make_pair(0U, &X86::RFP32RegClass);
27374 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27375 return std::make_pair(0U, &X86::RFP64RegClass);
27376 return std::make_pair(0U, &X86::RFP80RegClass);
27377 case 'y': // MMX_REGS if MMX allowed.
27378 if (!Subtarget->hasMMX()) break;
27379 return std::make_pair(0U, &X86::VR64RegClass);
27380 case 'Y': // SSE_REGS if SSE2 allowed
27381 if (!Subtarget->hasSSE2()) break;
27383 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27384 if (!Subtarget->hasSSE1()) break;
27386 switch (VT.SimpleTy) {
27388 // Scalar SSE types.
27391 return std::make_pair(0U, &X86::FR32RegClass);
27394 return std::make_pair(0U, &X86::FR64RegClass);
27402 return std::make_pair(0U, &X86::VR128RegClass);
27410 return std::make_pair(0U, &X86::VR256RegClass);
27415 return std::make_pair(0U, &X86::VR512RegClass);
27421 // Use the default implementation in TargetLowering to convert the register
27422 // constraint into a member of a register class.
27423 std::pair<unsigned, const TargetRegisterClass*> Res;
27424 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27426 // Not found as a standard register?
27428 // Map st(0) -> st(7) -> ST0
27429 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27430 tolower(Constraint[1]) == 's' &&
27431 tolower(Constraint[2]) == 't' &&
27432 Constraint[3] == '(' &&
27433 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27434 Constraint[5] == ')' &&
27435 Constraint[6] == '}') {
27437 Res.first = X86::FP0+Constraint[4]-'0';
27438 Res.second = &X86::RFP80RegClass;
27442 // GCC allows "st(0)" to be called just plain "st".
27443 if (StringRef("{st}").equals_lower(Constraint)) {
27444 Res.first = X86::FP0;
27445 Res.second = &X86::RFP80RegClass;
27450 if (StringRef("{flags}").equals_lower(Constraint)) {
27451 Res.first = X86::EFLAGS;
27452 Res.second = &X86::CCRRegClass;
27456 // 'A' means EAX + EDX.
27457 if (Constraint == "A") {
27458 Res.first = X86::EAX;
27459 Res.second = &X86::GR32_ADRegClass;
27465 // Otherwise, check to see if this is a register class of the wrong value
27466 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27467 // turn into {ax},{dx}.
27468 // MVT::Other is used to specify clobber names.
27469 if (Res.second->hasType(VT) || VT == MVT::Other)
27470 return Res; // Correct type already, nothing to do.
27472 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27473 // return "eax". This should even work for things like getting 64bit integer
27474 // registers when given an f64 type.
27475 const TargetRegisterClass *Class = Res.second;
27476 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27477 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27478 unsigned Size = VT.getSizeInBits();
27479 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27480 : Size == 16 ? MVT::i16
27481 : Size == 32 ? MVT::i32
27482 : Size == 64 ? MVT::i64
27484 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27486 Res.first = DestReg;
27487 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27488 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27489 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27490 : &X86::GR64RegClass;
27491 assert(Res.second->contains(Res.first) && "Register in register class");
27493 // No register found/type mismatch.
27495 Res.second = nullptr;
27497 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27498 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27499 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27500 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27501 Class == &X86::VR512RegClass) {
27502 // Handle references to XMM physical registers that got mapped into the
27503 // wrong class. This can happen with constraints like {xmm0} where the
27504 // target independent register mapper will just pick the first match it can
27505 // find, ignoring the required type.
27507 if (VT == MVT::f32 || VT == MVT::i32)
27508 Res.second = &X86::FR32RegClass;
27509 else if (VT == MVT::f64 || VT == MVT::i64)
27510 Res.second = &X86::FR64RegClass;
27511 else if (X86::VR128RegClass.hasType(VT))
27512 Res.second = &X86::VR128RegClass;
27513 else if (X86::VR256RegClass.hasType(VT))
27514 Res.second = &X86::VR256RegClass;
27515 else if (X86::VR512RegClass.hasType(VT))
27516 Res.second = &X86::VR512RegClass;
27518 // Type mismatch and not a clobber: Return an error;
27520 Res.second = nullptr;
27527 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27528 const AddrMode &AM, Type *Ty,
27529 unsigned AS) const {
27530 // Scaling factors are not free at all.
27531 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27532 // will take 2 allocations in the out of order engine instead of 1
27533 // for plain addressing mode, i.e. inst (reg1).
27535 // vaddps (%rsi,%drx), %ymm0, %ymm1
27536 // Requires two allocations (one for the load, one for the computation)
27538 // vaddps (%rsi), %ymm0, %ymm1
27539 // Requires just 1 allocation, i.e., freeing allocations for other operations
27540 // and having less micro operations to execute.
27542 // For some X86 architectures, this is even worse because for instance for
27543 // stores, the complex addressing mode forces the instruction to use the
27544 // "load" ports instead of the dedicated "store" port.
27545 // E.g., on Haswell:
27546 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27547 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27548 if (isLegalAddressingMode(DL, AM, Ty, AS))
27549 // Scale represents reg2 * scale, thus account for 1
27550 // as soon as we use a second register.
27551 return AM.Scale != 0;
27555 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27556 // Integer division on x86 is expensive. However, when aggressively optimizing
27557 // for code size, we prefer to use a div instruction, as it is usually smaller
27558 // than the alternative sequence.
27559 // The exception to this is vector division. Since x86 doesn't have vector
27560 // integer division, leaving the division as-is is a loss even in terms of
27561 // size, because it will have to be scalarized, while the alternative code
27562 // sequence can be performed in vector form.
27563 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27564 Attribute::MinSize);
27565 return OptSize && !VT.isVector();
27568 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27569 TargetLowering::ArgListTy& Args) const {
27570 // The MCU psABI requires some arguments to be passed in-register.
27571 // For regular calls, the inreg arguments are marked by the front-end.
27572 // However, for compiler generated library calls, we have to patch this
27574 if (!Subtarget->isTargetMCU() || !Args.size())
27577 unsigned FreeRegs = 3;
27578 for (auto &Arg : Args) {
27579 // For library functions, we do not expect any fancy types.
27580 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27581 unsigned SizeInRegs = (Size + 31) / 32;
27582 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27585 Arg.isInReg = true;
27586 FreeRegs -= SizeInRegs;