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/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/WinEHFuncInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
71 const X86Subtarget &STI)
72 : TargetLowering(TM), Subtarget(&STI) {
73 X86ScalarSSEf64 = Subtarget->hasSSE2();
74 X86ScalarSSEf32 = Subtarget->hasSSE1();
75 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
77 // Set up the TargetLowering object.
79 // X86 is weird. It always uses i8 for shift amounts and setcc results.
80 setBooleanContents(ZeroOrOneBooleanContent);
81 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
82 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
84 // For 64-bit, since we have so many registers, use the ILP scheduler.
85 // For 32-bit, use the register pressure specific scheduling.
86 // For Atom, always use ILP scheduling.
87 if (Subtarget->isAtom())
88 setSchedulingPreference(Sched::ILP);
89 else if (Subtarget->is64Bit())
90 setSchedulingPreference(Sched::ILP);
92 setSchedulingPreference(Sched::RegPressure);
93 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
94 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
96 // Bypass expensive divides on Atom when compiling with O2.
97 if (TM.getOptLevel() >= CodeGenOpt::Default) {
98 if (Subtarget->hasSlowDivide32())
99 addBypassSlowDiv(32, 8);
100 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
101 addBypassSlowDiv(64, 16);
104 if (Subtarget->isTargetKnownWindowsMSVC()) {
105 // Setup Windows compiler runtime calls.
106 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
107 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
108 setLibcallName(RTLIB::SREM_I64, "_allrem");
109 setLibcallName(RTLIB::UREM_I64, "_aullrem");
110 setLibcallName(RTLIB::MUL_I64, "_allmul");
111 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
112 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
118 if (Subtarget->isTargetDarwin()) {
119 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
120 setUseUnderscoreSetJmp(false);
121 setUseUnderscoreLongJmp(false);
122 } else if (Subtarget->isTargetWindowsGNU()) {
123 // MS runtime is weird: it exports _setjmp, but longjmp!
124 setUseUnderscoreSetJmp(true);
125 setUseUnderscoreLongJmp(false);
127 setUseUnderscoreSetJmp(true);
128 setUseUnderscoreLongJmp(true);
131 // Set up the register classes.
132 addRegisterClass(MVT::i8, &X86::GR8RegClass);
133 addRegisterClass(MVT::i16, &X86::GR16RegClass);
134 addRegisterClass(MVT::i32, &X86::GR32RegClass);
135 if (Subtarget->is64Bit())
136 addRegisterClass(MVT::i64, &X86::GR64RegClass);
138 for (MVT VT : MVT::integer_valuetypes())
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
141 // We don't accept any truncstore of integer registers.
142 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
143 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
145 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
146 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
147 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
149 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
151 // SETOEQ and SETUNE require checking two conditions.
152 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
153 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
155 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
159 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
161 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
162 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
165 if (Subtarget->is64Bit()) {
166 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
167 // f32/f64 are legal, f80 is custom.
168 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
172 } else if (!Subtarget->useSoftFloat()) {
173 // We have an algorithm for SSE2->double, and we turn this into a
174 // 64-bit FILD followed by conditional FADD for other targets.
175 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
176 // We have an algorithm for SSE2, and we turn this into a 64-bit
177 // FILD or VCVTUSI2SS/SD for other targets.
178 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
181 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
183 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
186 if (!Subtarget->useSoftFloat()) {
187 // SSE has no i16 to fp conversion, only i32
188 if (X86ScalarSSEf32) {
189 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
190 // f32 and f64 cases are Legal, f80 case is not
191 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
193 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
197 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
201 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
202 // are Legal, f80 is custom lowered.
203 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
204 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
206 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
208 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
209 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
211 if (X86ScalarSSEf32) {
212 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
213 // f32 and f64 cases are Legal, f80 case is not
214 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
216 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
217 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
220 // Handle FP_TO_UINT by promoting the destination to a larger signed
222 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
223 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
224 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
226 if (Subtarget->is64Bit()) {
227 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
228 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
229 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
233 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
235 } else if (!Subtarget->useSoftFloat()) {
236 // Since AVX is a superset of SSE3, only check for SSE here.
237 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
238 // Expand FP_TO_UINT into a select.
239 // FIXME: We would like to use a Custom expander here eventually to do
240 // the optimal thing for SSE vs. the default expansion in the legalizer.
241 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
243 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
244 // With SSE3 we can use fisttpll to convert to a signed i64; without
245 // SSE, we're stuck with a fistpll.
246 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
248 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
251 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
252 if (!X86ScalarSSEf64) {
253 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
254 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
257 // Without SSE, i64->f64 goes through memory.
258 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
262 // Scalar integer divide and remainder are lowered to use operations that
263 // produce two results, to match the available instructions. This exposes
264 // the two-result form to trivial CSE, which is able to combine x/y and x%y
265 // into a single instruction.
267 // Scalar integer multiply-high is also lowered to use two-result
268 // operations, to match the available instructions. However, plain multiply
269 // (low) operations are left as Legal, as there are single-result
270 // instructions for this in x86. Using the two-result multiply instructions
271 // when both high and low results are needed must be arranged by dagcombine.
272 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
273 setOperationAction(ISD::MULHS, VT, Expand);
274 setOperationAction(ISD::MULHU, VT, Expand);
275 setOperationAction(ISD::SDIV, VT, Expand);
276 setOperationAction(ISD::UDIV, VT, Expand);
277 setOperationAction(ISD::SREM, VT, Expand);
278 setOperationAction(ISD::UREM, VT, Expand);
280 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
281 setOperationAction(ISD::ADDC, VT, Custom);
282 setOperationAction(ISD::ADDE, VT, Custom);
283 setOperationAction(ISD::SUBC, VT, Custom);
284 setOperationAction(ISD::SUBE, VT, Custom);
287 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
288 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
289 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
290 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
291 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
292 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
293 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
294 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
295 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
296 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
297 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
301 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
302 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
303 if (Subtarget->is64Bit())
304 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
305 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
308 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
310 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
311 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
312 // is. We should promote the value to 64-bits to solve this.
313 // This is what the CRT headers do - `fmodf` is an inline header
314 // function casting to f64 and calling `fmod`.
315 setOperationAction(ISD::FREM , MVT::f32 , Promote);
317 setOperationAction(ISD::FREM , MVT::f32 , Expand);
320 setOperationAction(ISD::FREM , MVT::f64 , Expand);
321 setOperationAction(ISD::FREM , MVT::f80 , Expand);
322 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
324 // Promote the i8 variants and force them on up to i32 which has a shorter
326 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
327 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
328 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
329 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
330 if (Subtarget->hasBMI()) {
331 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
332 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
333 if (Subtarget->is64Bit())
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
336 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
337 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
338 if (Subtarget->is64Bit())
339 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
342 if (Subtarget->hasLZCNT()) {
343 // When promoting the i8 variants, force them to i32 for a shorter
345 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
346 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
348 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
351 if (Subtarget->is64Bit())
352 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
354 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
355 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
356 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
358 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
360 if (Subtarget->is64Bit()) {
361 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
362 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
366 // Special handling for half-precision floating point conversions.
367 // If we don't have F16C support, then lower half float conversions
368 // into library calls.
369 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
370 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
371 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
374 // There's never any support for operations beyond MVT::f32.
375 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
376 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
377 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
378 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
380 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
381 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
382 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
383 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
384 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
385 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
387 if (Subtarget->hasPOPCNT()) {
388 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
390 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
391 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
392 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
397 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
399 if (!Subtarget->hasMOVBE())
400 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
402 // These should be promoted to a larger select which is supported.
403 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
404 // X86 wants to expand cmov itself.
405 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
406 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
417 if (Subtarget->is64Bit()) {
418 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
421 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
422 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
423 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
424 // support continuation, user-level threading, and etc.. As a result, no
425 // other SjLj exception interfaces are implemented and please don't build
426 // your own exception handling based on them.
427 // LLVM/Clang supports zero-cost DWARF exception handling.
428 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
429 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
432 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
433 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
434 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
435 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
436 if (Subtarget->is64Bit())
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
438 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
439 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
440 if (Subtarget->is64Bit()) {
441 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
442 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
443 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
444 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
445 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
447 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
448 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
449 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
450 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
451 if (Subtarget->is64Bit()) {
452 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
453 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
454 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
457 if (Subtarget->hasSSE1())
458 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
460 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
462 // Expand certain atomics
463 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
464 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
465 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
466 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
469 if (Subtarget->hasCmpxchg16b()) {
470 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
473 // FIXME - use subtarget debug flags
474 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
475 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
476 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
479 if (Subtarget->isTarget64BitLP64()) {
480 setExceptionPointerRegister(X86::RAX);
481 setExceptionSelectorRegister(X86::RDX);
483 setExceptionPointerRegister(X86::EAX);
484 setExceptionSelectorRegister(X86::EDX);
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.getScalarType() == 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)) {
2883 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2884 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2885 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2886 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2887 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2888 MachinePointerInfo::getFixedStack(
2889 DAG.getMachineFunction(), UnwindHelpFI),
2890 /*isVolatile=*/true,
2891 /*isNonTemporal=*/false, /*Alignment=*/0);
2893 // Functions using Win32 EH are considered to have opaque SP adjustments
2894 // to force local variables to be addressed from the frame or base
2896 MFI->setHasOpaqueSPAdjustment(true);
2904 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2905 SDValue StackPtr, SDValue Arg,
2906 SDLoc dl, SelectionDAG &DAG,
2907 const CCValAssign &VA,
2908 ISD::ArgFlagsTy Flags) const {
2909 unsigned LocMemOffset = VA.getLocMemOffset();
2910 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2911 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2913 if (Flags.isByVal())
2914 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2916 return DAG.getStore(
2917 Chain, dl, Arg, PtrOff,
2918 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2922 /// Emit a load of return address if tail call
2923 /// optimization is performed and it is required.
2925 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2926 SDValue &OutRetAddr, SDValue Chain,
2927 bool IsTailCall, bool Is64Bit,
2928 int FPDiff, SDLoc dl) const {
2929 // Adjust the Return address stack slot.
2930 EVT VT = getPointerTy(DAG.getDataLayout());
2931 OutRetAddr = getReturnAddressFrameIndex(DAG);
2933 // Load the "old" Return address.
2934 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2935 false, false, false, 0);
2936 return SDValue(OutRetAddr.getNode(), 1);
2939 /// Emit a store of the return address if tail call
2940 /// optimization is performed and it is required (FPDiff!=0).
2941 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2942 SDValue Chain, SDValue RetAddrFrIdx,
2943 EVT PtrVT, unsigned SlotSize,
2944 int FPDiff, SDLoc dl) {
2945 // Store the return address to the appropriate stack slot.
2946 if (!FPDiff) return Chain;
2947 // Calculate the new stack slot for the return address.
2948 int NewReturnAddrFI =
2949 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2951 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2952 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2953 MachinePointerInfo::getFixedStack(
2954 DAG.getMachineFunction(), NewReturnAddrFI),
2959 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2960 /// operation of specified width.
2961 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
2963 unsigned NumElems = VT.getVectorNumElements();
2964 SmallVector<int, 8> Mask;
2965 Mask.push_back(NumElems);
2966 for (unsigned i = 1; i != NumElems; ++i)
2968 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2972 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2973 SmallVectorImpl<SDValue> &InVals) const {
2974 SelectionDAG &DAG = CLI.DAG;
2976 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2977 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2978 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2979 SDValue Chain = CLI.Chain;
2980 SDValue Callee = CLI.Callee;
2981 CallingConv::ID CallConv = CLI.CallConv;
2982 bool &isTailCall = CLI.IsTailCall;
2983 bool isVarArg = CLI.IsVarArg;
2985 MachineFunction &MF = DAG.getMachineFunction();
2986 bool Is64Bit = Subtarget->is64Bit();
2987 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2988 StructReturnType SR = callIsStructReturn(Outs);
2989 bool IsSibcall = false;
2990 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2991 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2993 if (Attr.getValueAsString() == "true")
2996 if (Subtarget->isPICStyleGOT() &&
2997 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2998 // If we are using a GOT, disable tail calls to external symbols with
2999 // default visibility. Tail calling such a symbol requires using a GOT
3000 // relocation, which forces early binding of the symbol. This breaks code
3001 // that require lazy function symbol resolution. Using musttail or
3002 // GuaranteedTailCallOpt will override this.
3003 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3004 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3005 G->getGlobal()->hasDefaultVisibility()))
3009 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3011 // Force this to be a tail call. The verifier rules are enough to ensure
3012 // that we can lower this successfully without moving the return address
3015 } else if (isTailCall) {
3016 // Check if it's really possible to do a tail call.
3017 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3018 isVarArg, SR != NotStructReturn,
3019 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3020 Outs, OutVals, Ins, DAG);
3022 // Sibcalls are automatically detected tailcalls which do not require
3024 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3031 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3032 "Var args not supported with calling convention fastcc, ghc or hipe");
3034 // Analyze operands of the call, assigning locations to each operand.
3035 SmallVector<CCValAssign, 16> ArgLocs;
3036 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3038 // Allocate shadow area for Win64
3040 CCInfo.AllocateStack(32, 8);
3042 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3044 // Get a count of how many bytes are to be pushed on the stack.
3045 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3047 // This is a sibcall. The memory operands are available in caller's
3048 // own caller's stack.
3050 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3051 canGuaranteeTCO(CallConv))
3052 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3055 if (isTailCall && !IsSibcall && !IsMustTail) {
3056 // Lower arguments at fp - stackoffset + fpdiff.
3057 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3059 FPDiff = NumBytesCallerPushed - NumBytes;
3061 // Set the delta of movement of the returnaddr stackslot.
3062 // But only set if delta is greater than previous delta.
3063 if (FPDiff < X86Info->getTCReturnAddrDelta())
3064 X86Info->setTCReturnAddrDelta(FPDiff);
3067 unsigned NumBytesToPush = NumBytes;
3068 unsigned NumBytesToPop = NumBytes;
3070 // If we have an inalloca argument, all stack space has already been allocated
3071 // for us and be right at the top of the stack. We don't support multiple
3072 // arguments passed in memory when using inalloca.
3073 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3075 if (!ArgLocs.back().isMemLoc())
3076 report_fatal_error("cannot use inalloca attribute on a register "
3078 if (ArgLocs.back().getLocMemOffset() != 0)
3079 report_fatal_error("any parameter with the inalloca attribute must be "
3080 "the only memory argument");
3084 Chain = DAG.getCALLSEQ_START(
3085 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3087 SDValue RetAddrFrIdx;
3088 // Load return address for tail calls.
3089 if (isTailCall && FPDiff)
3090 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3091 Is64Bit, FPDiff, dl);
3093 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3094 SmallVector<SDValue, 8> MemOpChains;
3097 // Walk the register/memloc assignments, inserting copies/loads. In the case
3098 // of tail call optimization arguments are handle later.
3099 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3100 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3101 // Skip inalloca arguments, they have already been written.
3102 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3103 if (Flags.isInAlloca())
3106 CCValAssign &VA = ArgLocs[i];
3107 EVT RegVT = VA.getLocVT();
3108 SDValue Arg = OutVals[i];
3109 bool isByVal = Flags.isByVal();
3111 // Promote the value if needed.
3112 switch (VA.getLocInfo()) {
3113 default: llvm_unreachable("Unknown loc info!");
3114 case CCValAssign::Full: break;
3115 case CCValAssign::SExt:
3116 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3118 case CCValAssign::ZExt:
3119 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3121 case CCValAssign::AExt:
3122 if (Arg.getValueType().isVector() &&
3123 Arg.getValueType().getScalarType() == MVT::i1)
3124 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3125 else if (RegVT.is128BitVector()) {
3126 // Special case: passing MMX values in XMM registers.
3127 Arg = DAG.getBitcast(MVT::i64, Arg);
3128 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3129 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3131 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3133 case CCValAssign::BCvt:
3134 Arg = DAG.getBitcast(RegVT, Arg);
3136 case CCValAssign::Indirect: {
3137 // Store the argument.
3138 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3139 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3140 Chain = DAG.getStore(
3141 Chain, dl, Arg, SpillSlot,
3142 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3149 if (VA.isRegLoc()) {
3150 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3151 if (isVarArg && IsWin64) {
3152 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3153 // shadow reg if callee is a varargs function.
3154 unsigned ShadowReg = 0;
3155 switch (VA.getLocReg()) {
3156 case X86::XMM0: ShadowReg = X86::RCX; break;
3157 case X86::XMM1: ShadowReg = X86::RDX; break;
3158 case X86::XMM2: ShadowReg = X86::R8; break;
3159 case X86::XMM3: ShadowReg = X86::R9; break;
3162 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3164 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3165 assert(VA.isMemLoc());
3166 if (!StackPtr.getNode())
3167 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3168 getPointerTy(DAG.getDataLayout()));
3169 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3170 dl, DAG, VA, Flags));
3174 if (!MemOpChains.empty())
3175 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3177 if (Subtarget->isPICStyleGOT()) {
3178 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3181 RegsToPass.push_back(std::make_pair(
3182 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3183 getPointerTy(DAG.getDataLayout()))));
3185 // If we are tail calling and generating PIC/GOT style code load the
3186 // address of the callee into ECX. The value in ecx is used as target of
3187 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3188 // for tail calls on PIC/GOT architectures. Normally we would just put the
3189 // address of GOT into ebx and then call target@PLT. But for tail calls
3190 // ebx would be restored (since ebx is callee saved) before jumping to the
3193 // Note: The actual moving to ECX is done further down.
3194 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3195 if (G && !G->getGlobal()->hasLocalLinkage() &&
3196 G->getGlobal()->hasDefaultVisibility())
3197 Callee = LowerGlobalAddress(Callee, DAG);
3198 else if (isa<ExternalSymbolSDNode>(Callee))
3199 Callee = LowerExternalSymbol(Callee, DAG);
3203 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3204 // From AMD64 ABI document:
3205 // For calls that may call functions that use varargs or stdargs
3206 // (prototype-less calls or calls to functions containing ellipsis (...) in
3207 // the declaration) %al is used as hidden argument to specify the number
3208 // of SSE registers used. The contents of %al do not need to match exactly
3209 // the number of registers, but must be an ubound on the number of SSE
3210 // registers used and is in the range 0 - 8 inclusive.
3212 // Count the number of XMM registers allocated.
3213 static const MCPhysReg XMMArgRegs[] = {
3214 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3215 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3217 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3218 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3219 && "SSE registers cannot be used when SSE is disabled");
3221 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3222 DAG.getConstant(NumXMMRegs, dl,
3226 if (isVarArg && IsMustTail) {
3227 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3228 for (const auto &F : Forwards) {
3229 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3230 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3234 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3235 // don't need this because the eligibility check rejects calls that require
3236 // shuffling arguments passed in memory.
3237 if (!IsSibcall && isTailCall) {
3238 // Force all the incoming stack arguments to be loaded from the stack
3239 // before any new outgoing arguments are stored to the stack, because the
3240 // outgoing stack slots may alias the incoming argument stack slots, and
3241 // the alias isn't otherwise explicit. This is slightly more conservative
3242 // than necessary, because it means that each store effectively depends
3243 // on every argument instead of just those arguments it would clobber.
3244 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3246 SmallVector<SDValue, 8> MemOpChains2;
3249 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3250 CCValAssign &VA = ArgLocs[i];
3253 assert(VA.isMemLoc());
3254 SDValue Arg = OutVals[i];
3255 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3256 // Skip inalloca arguments. They don't require any work.
3257 if (Flags.isInAlloca())
3259 // Create frame index.
3260 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3261 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3262 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3263 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3265 if (Flags.isByVal()) {
3266 // Copy relative to framepointer.
3267 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3268 if (!StackPtr.getNode())
3269 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3270 getPointerTy(DAG.getDataLayout()));
3271 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3274 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3278 // Store relative to framepointer.
3279 MemOpChains2.push_back(DAG.getStore(
3280 ArgChain, dl, Arg, FIN,
3281 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3286 if (!MemOpChains2.empty())
3287 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3289 // Store the return address to the appropriate stack slot.
3290 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3291 getPointerTy(DAG.getDataLayout()),
3292 RegInfo->getSlotSize(), FPDiff, dl);
3295 // Build a sequence of copy-to-reg nodes chained together with token chain
3296 // and flag operands which copy the outgoing args into registers.
3298 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3299 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3300 RegsToPass[i].second, InFlag);
3301 InFlag = Chain.getValue(1);
3304 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3305 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3306 // In the 64-bit large code model, we have to make all calls
3307 // through a register, since the call instruction's 32-bit
3308 // pc-relative offset may not be large enough to hold the whole
3310 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3311 // If the callee is a GlobalAddress node (quite common, every direct call
3312 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3314 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3316 // We should use extra load for direct calls to dllimported functions in
3318 const GlobalValue *GV = G->getGlobal();
3319 if (!GV->hasDLLImportStorageClass()) {
3320 unsigned char OpFlags = 0;
3321 bool ExtraLoad = false;
3322 unsigned WrapperKind = ISD::DELETED_NODE;
3324 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3325 // external symbols most go through the PLT in PIC mode. If the symbol
3326 // has hidden or protected visibility, or if it is static or local, then
3327 // we don't need to use the PLT - we can directly call it.
3328 if (Subtarget->isTargetELF() &&
3329 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3330 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3331 OpFlags = X86II::MO_PLT;
3332 } else if (Subtarget->isPICStyleStubAny() &&
3333 !GV->isStrongDefinitionForLinker() &&
3334 (!Subtarget->getTargetTriple().isMacOSX() ||
3335 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3336 // PC-relative references to external symbols should go through $stub,
3337 // unless we're building with the leopard linker or later, which
3338 // automatically synthesizes these stubs.
3339 OpFlags = X86II::MO_DARWIN_STUB;
3340 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3341 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3342 // If the function is marked as non-lazy, generate an indirect call
3343 // which loads from the GOT directly. This avoids runtime overhead
3344 // at the cost of eager binding (and one extra byte of encoding).
3345 OpFlags = X86II::MO_GOTPCREL;
3346 WrapperKind = X86ISD::WrapperRIP;
3350 Callee = DAG.getTargetGlobalAddress(
3351 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3353 // Add a wrapper if needed.
3354 if (WrapperKind != ISD::DELETED_NODE)
3355 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3356 getPointerTy(DAG.getDataLayout()), Callee);
3357 // Add extra indirection if needed.
3359 Callee = DAG.getLoad(
3360 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3361 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3364 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3365 unsigned char OpFlags = 0;
3367 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3368 // external symbols should go through the PLT.
3369 if (Subtarget->isTargetELF() &&
3370 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3371 OpFlags = X86II::MO_PLT;
3372 } else if (Subtarget->isPICStyleStubAny() &&
3373 (!Subtarget->getTargetTriple().isMacOSX() ||
3374 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3375 // PC-relative references to external symbols should go through $stub,
3376 // unless we're building with the leopard linker or later, which
3377 // automatically synthesizes these stubs.
3378 OpFlags = X86II::MO_DARWIN_STUB;
3381 Callee = DAG.getTargetExternalSymbol(
3382 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3383 } else if (Subtarget->isTarget64BitILP32() &&
3384 Callee->getValueType(0) == MVT::i32) {
3385 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3386 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3389 // Returns a chain & a flag for retval copy to use.
3390 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3391 SmallVector<SDValue, 8> Ops;
3393 if (!IsSibcall && isTailCall) {
3394 Chain = DAG.getCALLSEQ_END(Chain,
3395 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3396 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3397 InFlag = Chain.getValue(1);
3400 Ops.push_back(Chain);
3401 Ops.push_back(Callee);
3404 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3406 // Add argument registers to the end of the list so that they are known live
3408 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3409 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3410 RegsToPass[i].second.getValueType()));
3412 // Add a register mask operand representing the call-preserved registers.
3413 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3414 assert(Mask && "Missing call preserved mask for calling convention");
3416 // If this is an invoke in a 32-bit function using a funclet-based
3417 // personality, assume the function clobbers all registers. If an exception
3418 // is thrown, the runtime will not restore CSRs.
3419 // FIXME: Model this more precisely so that we can register allocate across
3420 // the normal edge and spill and fill across the exceptional edge.
3421 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3422 const Function *CallerFn = MF.getFunction();
3423 EHPersonality Pers =
3424 CallerFn->hasPersonalityFn()
3425 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3426 : EHPersonality::Unknown;
3427 if (isFuncletEHPersonality(Pers))
3428 Mask = RegInfo->getNoPreservedMask();
3431 Ops.push_back(DAG.getRegisterMask(Mask));
3433 if (InFlag.getNode())
3434 Ops.push_back(InFlag);
3438 //// If this is the first return lowered for this function, add the regs
3439 //// to the liveout set for the function.
3440 // This isn't right, although it's probably harmless on x86; liveouts
3441 // should be computed from returns not tail calls. Consider a void
3442 // function making a tail call to a function returning int.
3443 MF.getFrameInfo()->setHasTailCall();
3444 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3447 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3448 InFlag = Chain.getValue(1);
3450 // Create the CALLSEQ_END node.
3451 unsigned NumBytesForCalleeToPop;
3452 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3453 DAG.getTarget().Options.GuaranteedTailCallOpt))
3454 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3455 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3456 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3457 SR == StackStructReturn)
3458 // If this is a call to a struct-return function, the callee
3459 // pops the hidden struct pointer, so we have to push it back.
3460 // This is common for Darwin/X86, Linux & Mingw32 targets.
3461 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3462 NumBytesForCalleeToPop = 4;
3464 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3466 // Returns a flag for retval copy to use.
3468 Chain = DAG.getCALLSEQ_END(Chain,
3469 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3470 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3473 InFlag = Chain.getValue(1);
3476 // Handle result values, copying them out of physregs into vregs that we
3478 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3479 Ins, dl, DAG, InVals);
3482 //===----------------------------------------------------------------------===//
3483 // Fast Calling Convention (tail call) implementation
3484 //===----------------------------------------------------------------------===//
3486 // Like std call, callee cleans arguments, convention except that ECX is
3487 // reserved for storing the tail called function address. Only 2 registers are
3488 // free for argument passing (inreg). Tail call optimization is performed
3490 // * tailcallopt is enabled
3491 // * caller/callee are fastcc
3492 // On X86_64 architecture with GOT-style position independent code only local
3493 // (within module) calls are supported at the moment.
3494 // To keep the stack aligned according to platform abi the function
3495 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3496 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3497 // If a tail called function callee has more arguments than the caller the
3498 // caller needs to make sure that there is room to move the RETADDR to. This is
3499 // achieved by reserving an area the size of the argument delta right after the
3500 // original RETADDR, but before the saved framepointer or the spilled registers
3501 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3513 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3516 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3517 SelectionDAG& DAG) const {
3518 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3519 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3520 unsigned StackAlignment = TFI.getStackAlignment();
3521 uint64_t AlignMask = StackAlignment - 1;
3522 int64_t Offset = StackSize;
3523 unsigned SlotSize = RegInfo->getSlotSize();
3524 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3525 // Number smaller than 12 so just add the difference.
3526 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3528 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3529 Offset = ((~AlignMask) & Offset) + StackAlignment +
3530 (StackAlignment-SlotSize);
3535 /// Return true if the given stack call argument is already available in the
3536 /// same position (relatively) of the caller's incoming argument stack.
3538 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3539 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3540 const X86InstrInfo *TII) {
3541 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3543 if (Arg.getOpcode() == ISD::CopyFromReg) {
3544 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3545 if (!TargetRegisterInfo::isVirtualRegister(VR))
3547 MachineInstr *Def = MRI->getVRegDef(VR);
3550 if (!Flags.isByVal()) {
3551 if (!TII->isLoadFromStackSlot(Def, FI))
3554 unsigned Opcode = Def->getOpcode();
3555 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3556 Opcode == X86::LEA64_32r) &&
3557 Def->getOperand(1).isFI()) {
3558 FI = Def->getOperand(1).getIndex();
3559 Bytes = Flags.getByValSize();
3563 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3564 if (Flags.isByVal())
3565 // ByVal argument is passed in as a pointer but it's now being
3566 // dereferenced. e.g.
3567 // define @foo(%struct.X* %A) {
3568 // tail call @bar(%struct.X* byval %A)
3571 SDValue Ptr = Ld->getBasePtr();
3572 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3575 FI = FINode->getIndex();
3576 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3577 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3578 FI = FINode->getIndex();
3579 Bytes = Flags.getByValSize();
3583 assert(FI != INT_MAX);
3584 if (!MFI->isFixedObjectIndex(FI))
3586 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3589 /// Check whether the call is eligible for tail call optimization. Targets
3590 /// that want to do tail call optimization should implement this function.
3591 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3592 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3593 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3594 const SmallVectorImpl<ISD::OutputArg> &Outs,
3595 const SmallVectorImpl<SDValue> &OutVals,
3596 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3597 if (!mayTailCallThisCC(CalleeCC))
3600 // If -tailcallopt is specified, make fastcc functions tail-callable.
3601 MachineFunction &MF = DAG.getMachineFunction();
3602 const Function *CallerF = MF.getFunction();
3604 // If the function return type is x86_fp80 and the callee return type is not,
3605 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3606 // perform a tailcall optimization here.
3607 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3610 CallingConv::ID CallerCC = CallerF->getCallingConv();
3611 bool CCMatch = CallerCC == CalleeCC;
3612 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3613 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3615 // Win64 functions have extra shadow space for argument homing. Don't do the
3616 // sibcall if the caller and callee have mismatched expectations for this
3618 if (IsCalleeWin64 != IsCallerWin64)
3621 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3622 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3627 // Look for obvious safe cases to perform tail call optimization that do not
3628 // require ABI changes. This is what gcc calls sibcall.
3630 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3631 // emit a special epilogue.
3632 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3633 if (RegInfo->needsStackRealignment(MF))
3636 // Also avoid sibcall optimization if either caller or callee uses struct
3637 // return semantics.
3638 if (isCalleeStructRet || isCallerStructRet)
3641 // Do not sibcall optimize vararg calls unless all arguments are passed via
3643 if (isVarArg && !Outs.empty()) {
3644 // Optimizing for varargs on Win64 is unlikely to be safe without
3645 // additional testing.
3646 if (IsCalleeWin64 || IsCallerWin64)
3649 SmallVector<CCValAssign, 16> ArgLocs;
3650 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3653 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3654 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3655 if (!ArgLocs[i].isRegLoc())
3659 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3660 // stack. Therefore, if it's not used by the call it is not safe to optimize
3661 // this into a sibcall.
3662 bool Unused = false;
3663 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3670 SmallVector<CCValAssign, 16> RVLocs;
3671 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3673 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3674 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3675 CCValAssign &VA = RVLocs[i];
3676 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3681 // If the calling conventions do not match, then we'd better make sure the
3682 // results are returned in the same way as what the caller expects.
3684 SmallVector<CCValAssign, 16> RVLocs1;
3685 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3687 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3689 SmallVector<CCValAssign, 16> RVLocs2;
3690 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3692 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3694 if (RVLocs1.size() != RVLocs2.size())
3696 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3697 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3699 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3701 if (RVLocs1[i].isRegLoc()) {
3702 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3705 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3711 unsigned StackArgsSize = 0;
3713 // If the callee takes no arguments then go on to check the results of the
3715 if (!Outs.empty()) {
3716 // Check if stack adjustment is needed. For now, do not do this if any
3717 // argument is passed on the stack.
3718 SmallVector<CCValAssign, 16> ArgLocs;
3719 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3722 // Allocate shadow area for Win64
3724 CCInfo.AllocateStack(32, 8);
3726 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3727 StackArgsSize = CCInfo.getNextStackOffset();
3729 if (CCInfo.getNextStackOffset()) {
3730 // Check if the arguments are already laid out in the right way as
3731 // the caller's fixed stack objects.
3732 MachineFrameInfo *MFI = MF.getFrameInfo();
3733 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3734 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3735 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3736 CCValAssign &VA = ArgLocs[i];
3737 SDValue Arg = OutVals[i];
3738 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3739 if (VA.getLocInfo() == CCValAssign::Indirect)
3741 if (!VA.isRegLoc()) {
3742 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3749 // If the tailcall address may be in a register, then make sure it's
3750 // possible to register allocate for it. In 32-bit, the call address can
3751 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3752 // callee-saved registers are restored. These happen to be the same
3753 // registers used to pass 'inreg' arguments so watch out for those.
3754 if (!Subtarget->is64Bit() &&
3755 ((!isa<GlobalAddressSDNode>(Callee) &&
3756 !isa<ExternalSymbolSDNode>(Callee)) ||
3757 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3758 unsigned NumInRegs = 0;
3759 // In PIC we need an extra register to formulate the address computation
3761 unsigned MaxInRegs =
3762 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3764 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3765 CCValAssign &VA = ArgLocs[i];
3768 unsigned Reg = VA.getLocReg();
3771 case X86::EAX: case X86::EDX: case X86::ECX:
3772 if (++NumInRegs == MaxInRegs)
3780 bool CalleeWillPop =
3781 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3782 MF.getTarget().Options.GuaranteedTailCallOpt);
3784 if (unsigned BytesToPop =
3785 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3786 // If we have bytes to pop, the callee must pop them.
3787 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3788 if (!CalleePopMatches)
3790 } else if (CalleeWillPop && StackArgsSize > 0) {
3791 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3799 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3800 const TargetLibraryInfo *libInfo) const {
3801 return X86::createFastISel(funcInfo, libInfo);
3804 //===----------------------------------------------------------------------===//
3805 // Other Lowering Hooks
3806 //===----------------------------------------------------------------------===//
3808 static bool MayFoldLoad(SDValue Op) {
3809 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3812 static bool MayFoldIntoStore(SDValue Op) {
3813 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3816 static bool isTargetShuffle(unsigned Opcode) {
3818 default: return false;
3819 case X86ISD::BLENDI:
3820 case X86ISD::PSHUFB:
3821 case X86ISD::PSHUFD:
3822 case X86ISD::PSHUFHW:
3823 case X86ISD::PSHUFLW:
3825 case X86ISD::PALIGNR:
3826 case X86ISD::MOVLHPS:
3827 case X86ISD::MOVLHPD:
3828 case X86ISD::MOVHLPS:
3829 case X86ISD::MOVLPS:
3830 case X86ISD::MOVLPD:
3831 case X86ISD::MOVSHDUP:
3832 case X86ISD::MOVSLDUP:
3833 case X86ISD::MOVDDUP:
3836 case X86ISD::UNPCKL:
3837 case X86ISD::UNPCKH:
3838 case X86ISD::VPERMILPI:
3839 case X86ISD::VPERM2X128:
3840 case X86ISD::VPERMI:
3841 case X86ISD::VPERMV:
3842 case X86ISD::VPERMV3:
3847 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3848 SDValue V1, unsigned TargetMask,
3849 SelectionDAG &DAG) {
3851 default: llvm_unreachable("Unknown x86 shuffle node");
3852 case X86ISD::PSHUFD:
3853 case X86ISD::PSHUFHW:
3854 case X86ISD::PSHUFLW:
3855 case X86ISD::VPERMILPI:
3856 case X86ISD::VPERMI:
3857 return DAG.getNode(Opc, dl, VT, V1,
3858 DAG.getConstant(TargetMask, dl, MVT::i8));
3862 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, MVT VT,
3863 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3865 default: llvm_unreachable("Unknown x86 shuffle node");
3866 case X86ISD::MOVLHPS:
3867 case X86ISD::MOVLHPD:
3868 case X86ISD::MOVHLPS:
3869 case X86ISD::MOVLPS:
3870 case X86ISD::MOVLPD:
3873 case X86ISD::UNPCKL:
3874 case X86ISD::UNPCKH:
3875 return DAG.getNode(Opc, dl, VT, V1, V2);
3879 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3880 MachineFunction &MF = DAG.getMachineFunction();
3881 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3882 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3883 int ReturnAddrIndex = FuncInfo->getRAIndex();
3885 if (ReturnAddrIndex == 0) {
3886 // Set up a frame object for the return address.
3887 unsigned SlotSize = RegInfo->getSlotSize();
3888 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3891 FuncInfo->setRAIndex(ReturnAddrIndex);
3894 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3897 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3898 bool hasSymbolicDisplacement) {
3899 // Offset should fit into 32 bit immediate field.
3900 if (!isInt<32>(Offset))
3903 // If we don't have a symbolic displacement - we don't have any extra
3905 if (!hasSymbolicDisplacement)
3908 // FIXME: Some tweaks might be needed for medium code model.
3909 if (M != CodeModel::Small && M != CodeModel::Kernel)
3912 // For small code model we assume that latest object is 16MB before end of 31
3913 // bits boundary. We may also accept pretty large negative constants knowing
3914 // that all objects are in the positive half of address space.
3915 if (M == CodeModel::Small && Offset < 16*1024*1024)
3918 // For kernel code model we know that all object resist in the negative half
3919 // of 32bits address space. We may not accept negative offsets, since they may
3920 // be just off and we may accept pretty large positive ones.
3921 if (M == CodeModel::Kernel && Offset >= 0)
3927 /// Determines whether the callee is required to pop its own arguments.
3928 /// Callee pop is necessary to support tail calls.
3929 bool X86::isCalleePop(CallingConv::ID CallingConv,
3930 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3931 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3932 // can guarantee TCO.
3933 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3936 switch (CallingConv) {
3939 case CallingConv::X86_StdCall:
3940 case CallingConv::X86_FastCall:
3941 case CallingConv::X86_ThisCall:
3942 case CallingConv::X86_VectorCall:
3947 /// \brief Return true if the condition is an unsigned comparison operation.
3948 static bool isX86CCUnsigned(unsigned X86CC) {
3950 default: llvm_unreachable("Invalid integer condition!");
3951 case X86::COND_E: return true;
3952 case X86::COND_G: return false;
3953 case X86::COND_GE: return false;
3954 case X86::COND_L: return false;
3955 case X86::COND_LE: return false;
3956 case X86::COND_NE: return true;
3957 case X86::COND_B: return true;
3958 case X86::COND_A: return true;
3959 case X86::COND_BE: return true;
3960 case X86::COND_AE: return true;
3962 llvm_unreachable("covered switch fell through?!");
3965 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3966 /// condition code, returning the condition code and the LHS/RHS of the
3967 /// comparison to make.
3968 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3969 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3971 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3972 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3973 // X > -1 -> X == 0, jump !sign.
3974 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3975 return X86::COND_NS;
3977 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3978 // X < 0 -> X == 0, jump on sign.
3981 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3983 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3984 return X86::COND_LE;
3988 switch (SetCCOpcode) {
3989 default: llvm_unreachable("Invalid integer condition!");
3990 case ISD::SETEQ: return X86::COND_E;
3991 case ISD::SETGT: return X86::COND_G;
3992 case ISD::SETGE: return X86::COND_GE;
3993 case ISD::SETLT: return X86::COND_L;
3994 case ISD::SETLE: return X86::COND_LE;
3995 case ISD::SETNE: return X86::COND_NE;
3996 case ISD::SETULT: return X86::COND_B;
3997 case ISD::SETUGT: return X86::COND_A;
3998 case ISD::SETULE: return X86::COND_BE;
3999 case ISD::SETUGE: return X86::COND_AE;
4003 // First determine if it is required or is profitable to flip the operands.
4005 // If LHS is a foldable load, but RHS is not, flip the condition.
4006 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4007 !ISD::isNON_EXTLoad(RHS.getNode())) {
4008 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4009 std::swap(LHS, RHS);
4012 switch (SetCCOpcode) {
4018 std::swap(LHS, RHS);
4022 // On a floating point condition, the flags are set as follows:
4024 // 0 | 0 | 0 | X > Y
4025 // 0 | 0 | 1 | X < Y
4026 // 1 | 0 | 0 | X == Y
4027 // 1 | 1 | 1 | unordered
4028 switch (SetCCOpcode) {
4029 default: llvm_unreachable("Condcode should be pre-legalized away");
4031 case ISD::SETEQ: return X86::COND_E;
4032 case ISD::SETOLT: // flipped
4034 case ISD::SETGT: return X86::COND_A;
4035 case ISD::SETOLE: // flipped
4037 case ISD::SETGE: return X86::COND_AE;
4038 case ISD::SETUGT: // flipped
4040 case ISD::SETLT: return X86::COND_B;
4041 case ISD::SETUGE: // flipped
4043 case ISD::SETLE: return X86::COND_BE;
4045 case ISD::SETNE: return X86::COND_NE;
4046 case ISD::SETUO: return X86::COND_P;
4047 case ISD::SETO: return X86::COND_NP;
4049 case ISD::SETUNE: return X86::COND_INVALID;
4053 /// Is there a floating point cmov for the specific X86 condition code?
4054 /// Current x86 isa includes the following FP cmov instructions:
4055 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4056 static bool hasFPCMov(unsigned X86CC) {
4072 /// Returns true if the target can instruction select the
4073 /// specified FP immediate natively. If false, the legalizer will
4074 /// materialize the FP immediate as a load from a constant pool.
4075 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4076 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4077 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4083 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4084 ISD::LoadExtType ExtTy,
4086 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4087 // relocation target a movq or addq instruction: don't let the load shrink.
4088 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4089 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4090 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4091 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4095 /// \brief Returns true if it is beneficial to convert a load of a constant
4096 /// to just the constant itself.
4097 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4099 assert(Ty->isIntegerTy());
4101 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4102 if (BitSize == 0 || BitSize > 64)
4107 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4108 unsigned Index) const {
4109 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4112 return (Index == 0 || Index == ResVT.getVectorNumElements());
4115 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4116 // Speculate cttz only if we can directly use TZCNT.
4117 return Subtarget->hasBMI();
4120 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4121 // Speculate ctlz only if we can directly use LZCNT.
4122 return Subtarget->hasLZCNT();
4125 /// Return true if every element in Mask, beginning
4126 /// from position Pos and ending in Pos+Size is undef.
4127 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4128 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4134 /// Return true if Val is undef or if its value falls within the
4135 /// specified range (L, H].
4136 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4137 return (Val < 0) || (Val >= Low && Val < Hi);
4140 /// Val is either less than zero (undef) or equal to the specified value.
4141 static bool isUndefOrEqual(int Val, int CmpVal) {
4142 return (Val < 0 || Val == CmpVal);
4145 /// Return true if every element in Mask, beginning
4146 /// from position Pos and ending in Pos+Size, falls within the specified
4147 /// sequential range (Low, Low+Size]. or is undef.
4148 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4149 unsigned Pos, unsigned Size, int Low) {
4150 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4151 if (!isUndefOrEqual(Mask[i], Low))
4156 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4157 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4158 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4159 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4160 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4163 // The index should be aligned on a vecWidth-bit boundary.
4165 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4167 MVT VT = N->getSimpleValueType(0);
4168 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4169 bool Result = (Index * ElSize) % vecWidth == 0;
4174 /// Return true if the specified INSERT_SUBVECTOR
4175 /// operand specifies a subvector insert that is suitable for input to
4176 /// insertion of 128 or 256-bit subvectors
4177 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4178 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4179 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4181 // The index should be aligned on a vecWidth-bit boundary.
4183 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4185 MVT VT = N->getSimpleValueType(0);
4186 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4187 bool Result = (Index * ElSize) % vecWidth == 0;
4192 bool X86::isVINSERT128Index(SDNode *N) {
4193 return isVINSERTIndex(N, 128);
4196 bool X86::isVINSERT256Index(SDNode *N) {
4197 return isVINSERTIndex(N, 256);
4200 bool X86::isVEXTRACT128Index(SDNode *N) {
4201 return isVEXTRACTIndex(N, 128);
4204 bool X86::isVEXTRACT256Index(SDNode *N) {
4205 return isVEXTRACTIndex(N, 256);
4208 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4209 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4210 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4211 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4214 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4216 MVT VecVT = N->getOperand(0).getSimpleValueType();
4217 MVT ElVT = VecVT.getVectorElementType();
4219 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4220 return Index / NumElemsPerChunk;
4223 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4224 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4225 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4226 llvm_unreachable("Illegal insert subvector for VINSERT");
4229 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4231 MVT VecVT = N->getSimpleValueType(0);
4232 MVT ElVT = VecVT.getVectorElementType();
4234 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4235 return Index / NumElemsPerChunk;
4238 /// Return the appropriate immediate to extract the specified
4239 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4240 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4241 return getExtractVEXTRACTImmediate(N, 128);
4244 /// Return the appropriate immediate to extract the specified
4245 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4246 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4247 return getExtractVEXTRACTImmediate(N, 256);
4250 /// Return the appropriate immediate to insert at the specified
4251 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4252 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4253 return getInsertVINSERTImmediate(N, 128);
4256 /// Return the appropriate immediate to insert at the specified
4257 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4258 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4259 return getInsertVINSERTImmediate(N, 256);
4262 /// Returns true if V is a constant integer zero.
4263 static bool isZero(SDValue V) {
4264 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4265 return C && C->isNullValue();
4268 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4269 bool X86::isZeroNode(SDValue Elt) {
4272 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4273 return CFP->getValueAPF().isPosZero();
4277 // Build a vector of constants
4278 // Use an UNDEF node if MaskElt == -1.
4279 // Spilt 64-bit constants in the 32-bit mode.
4280 static SDValue getConstVector(ArrayRef<int> Values, MVT VT,
4282 SDLoc dl, bool IsMask = false) {
4284 SmallVector<SDValue, 32> Ops;
4287 MVT ConstVecVT = VT;
4288 unsigned NumElts = VT.getVectorNumElements();
4289 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4290 if (!In64BitMode && VT.getScalarType() == MVT::i64) {
4291 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4295 MVT EltVT = ConstVecVT.getVectorElementType();
4296 for (unsigned i = 0; i < NumElts; ++i) {
4297 bool IsUndef = Values[i] < 0 && IsMask;
4298 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4299 DAG.getConstant(Values[i], dl, EltVT);
4300 Ops.push_back(OpNode);
4302 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4303 DAG.getConstant(0, dl, EltVT));
4305 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4307 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4311 /// Returns a vector of specified type with all zero elements.
4312 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4313 SelectionDAG &DAG, SDLoc dl) {
4314 assert(VT.isVector() && "Expected a vector type");
4316 // Always build SSE zero vectors as <4 x i32> bitcasted
4317 // to their dest type. This ensures they get CSE'd.
4319 if (VT.is128BitVector()) { // SSE
4320 if (Subtarget->hasSSE2()) { // SSE2
4321 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4322 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4324 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4325 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4327 } else if (VT.is256BitVector()) { // AVX
4328 if (Subtarget->hasInt256()) { // AVX2
4329 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4330 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4331 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4333 // 256-bit logic and arithmetic instructions in AVX are all
4334 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4335 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4336 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4337 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4339 } else if (VT.is512BitVector()) { // AVX-512
4340 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4341 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4342 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4343 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4344 } else if (VT.getScalarType() == MVT::i1) {
4346 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4347 && "Unexpected vector type");
4348 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4349 && "Unexpected vector type");
4350 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4351 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4352 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4354 llvm_unreachable("Unexpected vector type");
4356 return DAG.getBitcast(VT, Vec);
4359 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4360 SelectionDAG &DAG, SDLoc dl,
4361 unsigned vectorWidth) {
4362 assert((vectorWidth == 128 || vectorWidth == 256) &&
4363 "Unsupported vector width");
4364 EVT VT = Vec.getValueType();
4365 EVT ElVT = VT.getVectorElementType();
4366 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4367 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4368 VT.getVectorNumElements()/Factor);
4370 // Extract from UNDEF is UNDEF.
4371 if (Vec.getOpcode() == ISD::UNDEF)
4372 return DAG.getUNDEF(ResultVT);
4374 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4375 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4376 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4378 // This is the index of the first element of the vectorWidth-bit chunk
4379 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4380 IdxVal &= ~(ElemsPerChunk - 1);
4382 // If the input is a buildvector just emit a smaller one.
4383 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4384 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4385 makeArrayRef(Vec->op_begin() + IdxVal, ElemsPerChunk));
4387 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4388 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4391 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4392 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4393 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4394 /// instructions or a simple subregister reference. Idx is an index in the
4395 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4396 /// lowering EXTRACT_VECTOR_ELT operations easier.
4397 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4398 SelectionDAG &DAG, SDLoc dl) {
4399 assert((Vec.getValueType().is256BitVector() ||
4400 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4401 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4404 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4405 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4406 SelectionDAG &DAG, SDLoc dl) {
4407 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4408 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4411 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4412 unsigned IdxVal, SelectionDAG &DAG,
4413 SDLoc dl, unsigned vectorWidth) {
4414 assert((vectorWidth == 128 || vectorWidth == 256) &&
4415 "Unsupported vector width");
4416 // Inserting UNDEF is Result
4417 if (Vec.getOpcode() == ISD::UNDEF)
4419 EVT VT = Vec.getValueType();
4420 EVT ElVT = VT.getVectorElementType();
4421 EVT ResultVT = Result.getValueType();
4423 // Insert the relevant vectorWidth bits.
4424 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4425 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
4427 // This is the index of the first element of the vectorWidth-bit chunk
4428 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
4429 IdxVal &= ~(ElemsPerChunk - 1);
4431 SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, dl);
4432 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4435 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4436 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4437 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4438 /// simple superregister reference. Idx is an index in the 128 bits
4439 /// we want. It need not be aligned to a 128-bit boundary. That makes
4440 /// lowering INSERT_VECTOR_ELT operations easier.
4441 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4442 SelectionDAG &DAG, SDLoc dl) {
4443 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4445 // For insertion into the zero index (low half) of a 256-bit vector, it is
4446 // more efficient to generate a blend with immediate instead of an insert*128.
4447 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4448 // extend the subvector to the size of the result vector. Make sure that
4449 // we are not recursing on that node by checking for undef here.
4450 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4451 Result.getOpcode() != ISD::UNDEF) {
4452 EVT ResultVT = Result.getValueType();
4453 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4454 SDValue Undef = DAG.getUNDEF(ResultVT);
4455 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4458 // The blend instruction, and therefore its mask, depend on the data type.
4459 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4460 if (ScalarType.isFloatingPoint()) {
4461 // Choose either vblendps (float) or vblendpd (double).
4462 unsigned ScalarSize = ScalarType.getSizeInBits();
4463 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4464 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4465 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4466 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4469 const X86Subtarget &Subtarget =
4470 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4472 // AVX2 is needed for 256-bit integer blend support.
4473 // Integers must be cast to 32-bit because there is only vpblendd;
4474 // vpblendw can't be used for this because it has a handicapped mask.
4476 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4477 // is still more efficient than using the wrong domain vinsertf128 that
4478 // will be created by InsertSubVector().
4479 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4481 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4482 Vec256 = DAG.getBitcast(CastVT, Vec256);
4483 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4484 return DAG.getBitcast(ResultVT, Vec256);
4487 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4490 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4491 SelectionDAG &DAG, SDLoc dl) {
4492 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4493 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4496 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4497 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4498 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4499 /// large BUILD_VECTORS.
4500 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4501 unsigned NumElems, SelectionDAG &DAG,
4503 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4504 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4507 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4508 unsigned NumElems, SelectionDAG &DAG,
4510 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4511 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4514 /// Returns a vector of specified type with all bits set.
4515 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4516 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4517 /// Then bitcast to their original type, ensuring they get CSE'd.
4518 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4519 SelectionDAG &DAG, SDLoc dl) {
4520 assert(VT.isVector() && "Expected a vector type");
4522 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4524 if (VT.is512BitVector()) {
4525 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4526 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4527 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4528 } else if (VT.is256BitVector()) {
4529 if (Subtarget->hasInt256()) { // AVX2
4530 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4531 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4533 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4534 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4536 } else if (VT.is128BitVector()) {
4537 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4539 llvm_unreachable("Unexpected vector type");
4541 return DAG.getBitcast(VT, Vec);
4544 /// Returns a vector_shuffle node for an unpackl operation.
4545 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4547 unsigned NumElems = VT.getVectorNumElements();
4548 SmallVector<int, 8> Mask;
4549 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4551 Mask.push_back(i + NumElems);
4553 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4556 /// Returns a vector_shuffle node for an unpackh operation.
4557 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4559 unsigned NumElems = VT.getVectorNumElements();
4560 SmallVector<int, 8> Mask;
4561 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4562 Mask.push_back(i + Half);
4563 Mask.push_back(i + NumElems + Half);
4565 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4568 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4569 /// This produces a shuffle where the low element of V2 is swizzled into the
4570 /// zero/undef vector, landing at element Idx.
4571 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4572 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4574 const X86Subtarget *Subtarget,
4575 SelectionDAG &DAG) {
4576 MVT VT = V2.getSimpleValueType();
4578 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4579 unsigned NumElems = VT.getVectorNumElements();
4580 SmallVector<int, 16> MaskVec;
4581 for (unsigned i = 0; i != NumElems; ++i)
4582 // If this is the insertion idx, put the low elt of V2 here.
4583 MaskVec.push_back(i == Idx ? NumElems : i);
4584 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4587 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4588 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4589 /// uses one source. Note that this will set IsUnary for shuffles which use a
4590 /// single input multiple times, and in those cases it will
4591 /// adjust the mask to only have indices within that single input.
4592 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4593 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4594 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4595 unsigned NumElems = VT.getVectorNumElements();
4599 bool IsFakeUnary = false;
4600 switch(N->getOpcode()) {
4601 case X86ISD::BLENDI:
4602 ImmN = N->getOperand(N->getNumOperands()-1);
4603 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4606 ImmN = N->getOperand(N->getNumOperands()-1);
4607 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4608 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4610 case X86ISD::UNPCKH:
4611 DecodeUNPCKHMask(VT, Mask);
4612 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4614 case X86ISD::UNPCKL:
4615 DecodeUNPCKLMask(VT, Mask);
4616 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4618 case X86ISD::MOVHLPS:
4619 DecodeMOVHLPSMask(NumElems, Mask);
4620 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4622 case X86ISD::MOVLHPS:
4623 DecodeMOVLHPSMask(NumElems, Mask);
4624 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4626 case X86ISD::PALIGNR:
4627 ImmN = N->getOperand(N->getNumOperands()-1);
4628 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4630 case X86ISD::PSHUFD:
4631 case X86ISD::VPERMILPI:
4632 ImmN = N->getOperand(N->getNumOperands()-1);
4633 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4636 case X86ISD::PSHUFHW:
4637 ImmN = N->getOperand(N->getNumOperands()-1);
4638 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4641 case X86ISD::PSHUFLW:
4642 ImmN = N->getOperand(N->getNumOperands()-1);
4643 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4646 case X86ISD::PSHUFB: {
4648 SDValue MaskNode = N->getOperand(1);
4649 while (MaskNode->getOpcode() == ISD::BITCAST)
4650 MaskNode = MaskNode->getOperand(0);
4652 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4653 // If we have a build-vector, then things are easy.
4654 EVT VT = MaskNode.getValueType();
4655 assert(VT.isVector() &&
4656 "Can't produce a non-vector with a build_vector!");
4657 if (!VT.isInteger())
4660 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4662 SmallVector<uint64_t, 32> RawMask;
4663 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4664 SDValue Op = MaskNode->getOperand(i);
4665 if (Op->getOpcode() == ISD::UNDEF) {
4666 RawMask.push_back((uint64_t)SM_SentinelUndef);
4669 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4672 APInt MaskElement = CN->getAPIntValue();
4674 // We now have to decode the element which could be any integer size and
4675 // extract each byte of it.
4676 for (int j = 0; j < NumBytesPerElement; ++j) {
4677 // Note that this is x86 and so always little endian: the low byte is
4678 // the first byte of the mask.
4679 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4680 MaskElement = MaskElement.lshr(8);
4683 DecodePSHUFBMask(RawMask, Mask);
4687 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4691 SDValue Ptr = MaskLoad->getBasePtr();
4692 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4693 Ptr->getOpcode() == X86ISD::WrapperRIP)
4694 Ptr = Ptr->getOperand(0);
4696 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4697 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4700 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4701 DecodePSHUFBMask(C, Mask);
4709 case X86ISD::VPERMI:
4710 ImmN = N->getOperand(N->getNumOperands()-1);
4711 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4716 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4718 case X86ISD::VPERM2X128:
4719 ImmN = N->getOperand(N->getNumOperands()-1);
4720 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4721 if (Mask.empty()) return false;
4722 // Mask only contains negative index if an element is zero.
4723 if (std::any_of(Mask.begin(), Mask.end(),
4724 [](int M){ return M == SM_SentinelZero; }))
4727 case X86ISD::MOVSLDUP:
4728 DecodeMOVSLDUPMask(VT, Mask);
4731 case X86ISD::MOVSHDUP:
4732 DecodeMOVSHDUPMask(VT, Mask);
4735 case X86ISD::MOVDDUP:
4736 DecodeMOVDDUPMask(VT, Mask);
4739 case X86ISD::MOVLHPD:
4740 case X86ISD::MOVLPD:
4741 case X86ISD::MOVLPS:
4742 // Not yet implemented
4744 case X86ISD::VPERMV: {
4746 SDValue MaskNode = N->getOperand(0);
4747 while (MaskNode->getOpcode() == ISD::BITCAST)
4748 MaskNode = MaskNode->getOperand(0);
4750 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4751 SmallVector<uint64_t, 32> RawMask;
4752 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4753 // If we have a build-vector, then things are easy.
4754 assert(MaskNode.getValueType().isInteger() &&
4755 MaskNode.getValueType().getVectorNumElements() ==
4756 VT.getVectorNumElements());
4758 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4759 SDValue Op = MaskNode->getOperand(i);
4760 if (Op->getOpcode() == ISD::UNDEF)
4761 RawMask.push_back((uint64_t)SM_SentinelUndef);
4762 else if (isa<ConstantSDNode>(Op)) {
4763 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4764 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4768 DecodeVPERMVMask(RawMask, Mask);
4771 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4772 unsigned NumEltsInMask = MaskNode->getNumOperands();
4773 MaskNode = MaskNode->getOperand(0);
4774 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4776 APInt MaskEltValue = CN->getAPIntValue();
4777 for (unsigned i = 0; i < NumEltsInMask; ++i)
4778 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4779 DecodeVPERMVMask(RawMask, Mask);
4782 // It may be a scalar load
4785 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4789 SDValue Ptr = MaskLoad->getBasePtr();
4790 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4791 Ptr->getOpcode() == X86ISD::WrapperRIP)
4792 Ptr = Ptr->getOperand(0);
4794 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4795 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4798 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4800 DecodeVPERMVMask(C, VT, Mask);
4807 case X86ISD::VPERMV3: {
4809 SDValue MaskNode = N->getOperand(1);
4810 while (MaskNode->getOpcode() == ISD::BITCAST)
4811 MaskNode = MaskNode->getOperand(1);
4813 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4814 // If we have a build-vector, then things are easy.
4815 assert(MaskNode.getValueType().isInteger() &&
4816 MaskNode.getValueType().getVectorNumElements() ==
4817 VT.getVectorNumElements());
4819 SmallVector<uint64_t, 32> RawMask;
4820 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4822 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4823 SDValue Op = MaskNode->getOperand(i);
4824 if (Op->getOpcode() == ISD::UNDEF)
4825 RawMask.push_back((uint64_t)SM_SentinelUndef);
4827 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4830 APInt MaskElement = CN->getAPIntValue();
4831 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4834 DecodeVPERMV3Mask(RawMask, Mask);
4838 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4842 SDValue Ptr = MaskLoad->getBasePtr();
4843 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4844 Ptr->getOpcode() == X86ISD::WrapperRIP)
4845 Ptr = Ptr->getOperand(0);
4847 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4848 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4851 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4853 DecodeVPERMV3Mask(C, VT, Mask);
4860 default: llvm_unreachable("unknown target shuffle node");
4863 // If we have a fake unary shuffle, the shuffle mask is spread across two
4864 // inputs that are actually the same node. Re-map the mask to always point
4865 // into the first input.
4868 if (M >= (int)Mask.size())
4874 /// Returns the scalar element that will make up the ith
4875 /// element of the result of the vector shuffle.
4876 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4879 return SDValue(); // Limit search depth.
4881 SDValue V = SDValue(N, 0);
4882 EVT VT = V.getValueType();
4883 unsigned Opcode = V.getOpcode();
4885 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4886 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4887 int Elt = SV->getMaskElt(Index);
4890 return DAG.getUNDEF(VT.getVectorElementType());
4892 unsigned NumElems = VT.getVectorNumElements();
4893 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4894 : SV->getOperand(1);
4895 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4898 // Recurse into target specific vector shuffles to find scalars.
4899 if (isTargetShuffle(Opcode)) {
4900 MVT ShufVT = V.getSimpleValueType();
4901 unsigned NumElems = ShufVT.getVectorNumElements();
4902 SmallVector<int, 16> ShuffleMask;
4905 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4908 int Elt = ShuffleMask[Index];
4910 return DAG.getUNDEF(ShufVT.getVectorElementType());
4912 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4914 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4918 // Actual nodes that may contain scalar elements
4919 if (Opcode == ISD::BITCAST) {
4920 V = V.getOperand(0);
4921 EVT SrcVT = V.getValueType();
4922 unsigned NumElems = VT.getVectorNumElements();
4924 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4928 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4929 return (Index == 0) ? V.getOperand(0)
4930 : DAG.getUNDEF(VT.getVectorElementType());
4932 if (V.getOpcode() == ISD::BUILD_VECTOR)
4933 return V.getOperand(Index);
4938 /// Custom lower build_vector of v16i8.
4939 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4940 unsigned NumNonZero, unsigned NumZero,
4942 const X86Subtarget* Subtarget,
4943 const TargetLowering &TLI) {
4951 // SSE4.1 - use PINSRB to insert each byte directly.
4952 if (Subtarget->hasSSE41()) {
4953 for (unsigned i = 0; i < 16; ++i) {
4954 bool isNonZero = (NonZeros & (1 << i)) != 0;
4958 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4960 V = DAG.getUNDEF(MVT::v16i8);
4963 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4964 MVT::v16i8, V, Op.getOperand(i),
4965 DAG.getIntPtrConstant(i, dl));
4972 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4973 for (unsigned i = 0; i < 16; ++i) {
4974 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4975 if (ThisIsNonZero && First) {
4977 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4979 V = DAG.getUNDEF(MVT::v8i16);
4984 SDValue ThisElt, LastElt;
4985 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4986 if (LastIsNonZero) {
4987 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4988 MVT::i16, Op.getOperand(i-1));
4990 if (ThisIsNonZero) {
4991 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4992 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4993 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4995 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4999 if (ThisElt.getNode())
5000 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5001 DAG.getIntPtrConstant(i/2, dl));
5005 return DAG.getBitcast(MVT::v16i8, V);
5008 /// Custom lower build_vector of v8i16.
5009 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5010 unsigned NumNonZero, unsigned NumZero,
5012 const X86Subtarget* Subtarget,
5013 const TargetLowering &TLI) {
5020 for (unsigned i = 0; i < 8; ++i) {
5021 bool isNonZero = (NonZeros & (1 << i)) != 0;
5025 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5027 V = DAG.getUNDEF(MVT::v8i16);
5030 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5031 MVT::v8i16, V, Op.getOperand(i),
5032 DAG.getIntPtrConstant(i, dl));
5039 /// Custom lower build_vector of v4i32 or v4f32.
5040 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5041 const X86Subtarget *Subtarget,
5042 const TargetLowering &TLI) {
5043 // Find all zeroable elements.
5044 std::bitset<4> Zeroable;
5045 for (int i=0; i < 4; ++i) {
5046 SDValue Elt = Op->getOperand(i);
5047 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5049 assert(Zeroable.size() - Zeroable.count() > 1 &&
5050 "We expect at least two non-zero elements!");
5052 // We only know how to deal with build_vector nodes where elements are either
5053 // zeroable or extract_vector_elt with constant index.
5054 SDValue FirstNonZero;
5055 unsigned FirstNonZeroIdx;
5056 for (unsigned i=0; i < 4; ++i) {
5059 SDValue Elt = Op->getOperand(i);
5060 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5061 !isa<ConstantSDNode>(Elt.getOperand(1)))
5063 // Make sure that this node is extracting from a 128-bit vector.
5064 MVT VT = Elt.getOperand(0).getSimpleValueType();
5065 if (!VT.is128BitVector())
5067 if (!FirstNonZero.getNode()) {
5069 FirstNonZeroIdx = i;
5073 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5074 SDValue V1 = FirstNonZero.getOperand(0);
5075 MVT VT = V1.getSimpleValueType();
5077 // See if this build_vector can be lowered as a blend with zero.
5079 unsigned EltMaskIdx, EltIdx;
5081 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5082 if (Zeroable[EltIdx]) {
5083 // The zero vector will be on the right hand side.
5084 Mask[EltIdx] = EltIdx+4;
5088 Elt = Op->getOperand(EltIdx);
5089 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5090 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5091 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5093 Mask[EltIdx] = EltIdx;
5097 // Let the shuffle legalizer deal with blend operations.
5098 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5099 if (V1.getSimpleValueType() != VT)
5100 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5101 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5104 // See if we can lower this build_vector to a INSERTPS.
5105 if (!Subtarget->hasSSE41())
5108 SDValue V2 = Elt.getOperand(0);
5109 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5112 bool CanFold = true;
5113 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5117 SDValue Current = Op->getOperand(i);
5118 SDValue SrcVector = Current->getOperand(0);
5121 CanFold = SrcVector == V1 &&
5122 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5128 assert(V1.getNode() && "Expected at least two non-zero elements!");
5129 if (V1.getSimpleValueType() != MVT::v4f32)
5130 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5131 if (V2.getSimpleValueType() != MVT::v4f32)
5132 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5134 // Ok, we can emit an INSERTPS instruction.
5135 unsigned ZMask = Zeroable.to_ulong();
5137 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5138 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5140 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5141 DAG.getIntPtrConstant(InsertPSMask, DL));
5142 return DAG.getBitcast(VT, Result);
5145 /// Return a vector logical shift node.
5146 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5147 unsigned NumBits, SelectionDAG &DAG,
5148 const TargetLowering &TLI, SDLoc dl) {
5149 assert(VT.is128BitVector() && "Unknown type for VShift");
5150 MVT ShVT = MVT::v2i64;
5151 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5152 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5153 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5154 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5155 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5156 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5160 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5162 // Check if the scalar load can be widened into a vector load. And if
5163 // the address is "base + cst" see if the cst can be "absorbed" into
5164 // the shuffle mask.
5165 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5166 SDValue Ptr = LD->getBasePtr();
5167 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5169 EVT PVT = LD->getValueType(0);
5170 if (PVT != MVT::i32 && PVT != MVT::f32)
5175 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5176 FI = FINode->getIndex();
5178 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5179 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5180 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5181 Offset = Ptr.getConstantOperandVal(1);
5182 Ptr = Ptr.getOperand(0);
5187 // FIXME: 256-bit vector instructions don't require a strict alignment,
5188 // improve this code to support it better.
5189 unsigned RequiredAlign = VT.getSizeInBits()/8;
5190 SDValue Chain = LD->getChain();
5191 // Make sure the stack object alignment is at least 16 or 32.
5192 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5193 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5194 if (MFI->isFixedObjectIndex(FI)) {
5195 // Can't change the alignment. FIXME: It's possible to compute
5196 // the exact stack offset and reference FI + adjust offset instead.
5197 // If someone *really* cares about this. That's the way to implement it.
5200 MFI->setObjectAlignment(FI, RequiredAlign);
5204 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5205 // Ptr + (Offset & ~15).
5208 if ((Offset % RequiredAlign) & 3)
5210 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5213 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5214 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5217 int EltNo = (Offset - StartOffset) >> 2;
5218 unsigned NumElems = VT.getVectorNumElements();
5220 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5221 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5222 LD->getPointerInfo().getWithOffset(StartOffset),
5223 false, false, false, 0);
5225 SmallVector<int, 8> Mask(NumElems, EltNo);
5227 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5233 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5234 /// elements can be replaced by a single large load which has the same value as
5235 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5237 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5239 /// FIXME: we'd also like to handle the case where the last elements are zero
5240 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5241 /// There's even a handy isZeroNode for that purpose.
5242 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5243 SDLoc &DL, SelectionDAG &DAG,
5244 bool isAfterLegalize) {
5245 unsigned NumElems = Elts.size();
5247 LoadSDNode *LDBase = nullptr;
5248 unsigned LastLoadedElt = -1U;
5250 // For each element in the initializer, see if we've found a load or an undef.
5251 // If we don't find an initial load element, or later load elements are
5252 // non-consecutive, bail out.
5253 for (unsigned i = 0; i < NumElems; ++i) {
5254 SDValue Elt = Elts[i];
5255 // Look through a bitcast.
5256 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5257 Elt = Elt.getOperand(0);
5258 if (!Elt.getNode() ||
5259 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5262 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5264 LDBase = cast<LoadSDNode>(Elt.getNode());
5268 if (Elt.getOpcode() == ISD::UNDEF)
5271 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5272 EVT LdVT = Elt.getValueType();
5273 // Each loaded element must be the correct fractional portion of the
5274 // requested vector load.
5275 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5277 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5282 // If we have found an entire vector of loads and undefs, then return a large
5283 // load of the entire vector width starting at the base pointer. If we found
5284 // consecutive loads for the low half, generate a vzext_load node.
5285 if (LastLoadedElt == NumElems - 1) {
5286 assert(LDBase && "Did not find base load for merging consecutive loads");
5287 EVT EltVT = LDBase->getValueType(0);
5288 // Ensure that the input vector size for the merged loads matches the
5289 // cumulative size of the input elements.
5290 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5293 if (isAfterLegalize &&
5294 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5297 SDValue NewLd = SDValue();
5299 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5300 LDBase->getPointerInfo(), LDBase->isVolatile(),
5301 LDBase->isNonTemporal(), LDBase->isInvariant(),
5302 LDBase->getAlignment());
5304 if (LDBase->hasAnyUseOfValue(1)) {
5305 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5307 SDValue(NewLd.getNode(), 1));
5308 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5309 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5310 SDValue(NewLd.getNode(), 1));
5316 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5317 //of a v4i32 / v4f32. It's probably worth generalizing.
5318 EVT EltVT = VT.getVectorElementType();
5319 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5320 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5321 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5322 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5324 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5325 LDBase->getPointerInfo(),
5326 LDBase->getAlignment(),
5327 false/*isVolatile*/, true/*ReadMem*/,
5330 // Make sure the newly-created LOAD is in the same position as LDBase in
5331 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5332 // update uses of LDBase's output chain to use the TokenFactor.
5333 if (LDBase->hasAnyUseOfValue(1)) {
5334 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5335 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5336 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5337 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5338 SDValue(ResNode.getNode(), 1));
5341 return DAG.getBitcast(VT, ResNode);
5346 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5347 /// to generate a splat value for the following cases:
5348 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5349 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5350 /// a scalar load, or a constant.
5351 /// The VBROADCAST node is returned when a pattern is found,
5352 /// or SDValue() otherwise.
5353 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5354 SelectionDAG &DAG) {
5355 // VBROADCAST requires AVX.
5356 // TODO: Splats could be generated for non-AVX CPUs using SSE
5357 // instructions, but there's less potential gain for only 128-bit vectors.
5358 if (!Subtarget->hasAVX())
5361 MVT VT = Op.getSimpleValueType();
5364 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5365 "Unsupported vector type for broadcast.");
5370 switch (Op.getOpcode()) {
5372 // Unknown pattern found.
5375 case ISD::BUILD_VECTOR: {
5376 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5377 BitVector UndefElements;
5378 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5380 // We need a splat of a single value to use broadcast, and it doesn't
5381 // make any sense if the value is only in one element of the vector.
5382 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5386 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5387 Ld.getOpcode() == ISD::ConstantFP);
5389 // Make sure that all of the users of a non-constant load are from the
5390 // BUILD_VECTOR node.
5391 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5396 case ISD::VECTOR_SHUFFLE: {
5397 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5399 // Shuffles must have a splat mask where the first element is
5401 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5404 SDValue Sc = Op.getOperand(0);
5405 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5406 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5408 if (!Subtarget->hasInt256())
5411 // Use the register form of the broadcast instruction available on AVX2.
5412 if (VT.getSizeInBits() >= 256)
5413 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5414 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5417 Ld = Sc.getOperand(0);
5418 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5419 Ld.getOpcode() == ISD::ConstantFP);
5421 // The scalar_to_vector node and the suspected
5422 // load node must have exactly one user.
5423 // Constants may have multiple users.
5425 // AVX-512 has register version of the broadcast
5426 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5427 Ld.getValueType().getSizeInBits() >= 32;
5428 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5435 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5436 bool IsGE256 = (VT.getSizeInBits() >= 256);
5438 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5439 // instruction to save 8 or more bytes of constant pool data.
5440 // TODO: If multiple splats are generated to load the same constant,
5441 // it may be detrimental to overall size. There needs to be a way to detect
5442 // that condition to know if this is truly a size win.
5443 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5445 // Handle broadcasting a single constant scalar from the constant pool
5447 // On Sandybridge (no AVX2), it is still better to load a constant vector
5448 // from the constant pool and not to broadcast it from a scalar.
5449 // But override that restriction when optimizing for size.
5450 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5451 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5452 EVT CVT = Ld.getValueType();
5453 assert(!CVT.isVector() && "Must not broadcast a vector type");
5455 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5456 // For size optimization, also splat v2f64 and v2i64, and for size opt
5457 // with AVX2, also splat i8 and i16.
5458 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5459 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5460 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5461 const Constant *C = nullptr;
5462 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5463 C = CI->getConstantIntValue();
5464 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5465 C = CF->getConstantFPValue();
5467 assert(C && "Invalid constant type");
5469 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5471 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5472 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5474 CVT, dl, DAG.getEntryNode(), CP,
5475 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5476 false, false, Alignment);
5478 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5482 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5484 // Handle AVX2 in-register broadcasts.
5485 if (!IsLoad && Subtarget->hasInt256() &&
5486 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5487 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5489 // The scalar source must be a normal load.
5493 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5494 (Subtarget->hasVLX() && ScalarSize == 64))
5495 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5497 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5498 // double since there is no vbroadcastsd xmm
5499 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5500 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5501 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5504 // Unsupported broadcast.
5508 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5509 /// underlying vector and index.
5511 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5513 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5515 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5516 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5519 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5521 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5523 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5524 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5527 // In this case the vector is the extract_subvector expression and the index
5528 // is 2, as specified by the shuffle.
5529 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5530 SDValue ShuffleVec = SVOp->getOperand(0);
5531 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5532 assert(ShuffleVecVT.getVectorElementType() ==
5533 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5535 int ShuffleIdx = SVOp->getMaskElt(Idx);
5536 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5537 ExtractedFromVec = ShuffleVec;
5543 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5544 MVT VT = Op.getSimpleValueType();
5546 // Skip if insert_vec_elt is not supported.
5547 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5548 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5552 unsigned NumElems = Op.getNumOperands();
5556 SmallVector<unsigned, 4> InsertIndices;
5557 SmallVector<int, 8> Mask(NumElems, -1);
5559 for (unsigned i = 0; i != NumElems; ++i) {
5560 unsigned Opc = Op.getOperand(i).getOpcode();
5562 if (Opc == ISD::UNDEF)
5565 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5566 // Quit if more than 1 elements need inserting.
5567 if (InsertIndices.size() > 1)
5570 InsertIndices.push_back(i);
5574 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5575 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5576 // Quit if non-constant index.
5577 if (!isa<ConstantSDNode>(ExtIdx))
5579 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5581 // Quit if extracted from vector of different type.
5582 if (ExtractedFromVec.getValueType() != VT)
5585 if (!VecIn1.getNode())
5586 VecIn1 = ExtractedFromVec;
5587 else if (VecIn1 != ExtractedFromVec) {
5588 if (!VecIn2.getNode())
5589 VecIn2 = ExtractedFromVec;
5590 else if (VecIn2 != ExtractedFromVec)
5591 // Quit if more than 2 vectors to shuffle
5595 if (ExtractedFromVec == VecIn1)
5597 else if (ExtractedFromVec == VecIn2)
5598 Mask[i] = Idx + NumElems;
5601 if (!VecIn1.getNode())
5604 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5605 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5606 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5607 unsigned Idx = InsertIndices[i];
5608 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5609 DAG.getIntPtrConstant(Idx, DL));
5615 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5616 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5617 Op.getScalarValueSizeInBits() == 1 &&
5618 "Can not convert non-constant vector");
5619 uint64_t Immediate = 0;
5620 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5621 SDValue In = Op.getOperand(idx);
5622 if (In.getOpcode() != ISD::UNDEF)
5623 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5627 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5628 return DAG.getConstant(Immediate, dl, VT);
5630 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5632 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5634 MVT VT = Op.getSimpleValueType();
5635 assert((VT.getVectorElementType() == MVT::i1) &&
5636 "Unexpected type in LowerBUILD_VECTORvXi1!");
5639 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5640 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5641 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5642 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5645 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5646 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5647 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5648 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5651 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5652 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5653 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5654 return DAG.getBitcast(VT, Imm);
5655 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5656 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5657 DAG.getIntPtrConstant(0, dl));
5660 // Vector has one or more non-const elements
5661 uint64_t Immediate = 0;
5662 SmallVector<unsigned, 16> NonConstIdx;
5663 bool IsSplat = true;
5664 bool HasConstElts = false;
5666 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5667 SDValue In = Op.getOperand(idx);
5668 if (In.getOpcode() == ISD::UNDEF)
5670 if (!isa<ConstantSDNode>(In))
5671 NonConstIdx.push_back(idx);
5673 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5674 HasConstElts = true;
5678 else if (In != Op.getOperand(SplatIdx))
5682 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5684 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5685 DAG.getConstant(1, dl, VT),
5686 DAG.getConstant(0, dl, VT));
5688 // insert elements one by one
5692 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5693 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5695 else if (HasConstElts)
5696 Imm = DAG.getConstant(0, dl, VT);
5698 Imm = DAG.getUNDEF(VT);
5699 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5700 DstVec = DAG.getBitcast(VT, Imm);
5702 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5703 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5704 DAG.getIntPtrConstant(0, dl));
5707 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5708 unsigned InsertIdx = NonConstIdx[i];
5709 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5710 Op.getOperand(InsertIdx),
5711 DAG.getIntPtrConstant(InsertIdx, dl));
5716 /// \brief Return true if \p N implements a horizontal binop and return the
5717 /// operands for the horizontal binop into V0 and V1.
5719 /// This is a helper function of LowerToHorizontalOp().
5720 /// This function checks that the build_vector \p N in input implements a
5721 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5722 /// operation to match.
5723 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5724 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5725 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5728 /// This function only analyzes elements of \p N whose indices are
5729 /// in range [BaseIdx, LastIdx).
5730 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5732 unsigned BaseIdx, unsigned LastIdx,
5733 SDValue &V0, SDValue &V1) {
5734 EVT VT = N->getValueType(0);
5736 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5737 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5738 "Invalid Vector in input!");
5740 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5741 bool CanFold = true;
5742 unsigned ExpectedVExtractIdx = BaseIdx;
5743 unsigned NumElts = LastIdx - BaseIdx;
5744 V0 = DAG.getUNDEF(VT);
5745 V1 = DAG.getUNDEF(VT);
5747 // Check if N implements a horizontal binop.
5748 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5749 SDValue Op = N->getOperand(i + BaseIdx);
5752 if (Op->getOpcode() == ISD::UNDEF) {
5753 // Update the expected vector extract index.
5754 if (i * 2 == NumElts)
5755 ExpectedVExtractIdx = BaseIdx;
5756 ExpectedVExtractIdx += 2;
5760 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5765 SDValue Op0 = Op.getOperand(0);
5766 SDValue Op1 = Op.getOperand(1);
5768 // Try to match the following pattern:
5769 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5770 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5771 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5772 Op0.getOperand(0) == Op1.getOperand(0) &&
5773 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5774 isa<ConstantSDNode>(Op1.getOperand(1)));
5778 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5779 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5781 if (i * 2 < NumElts) {
5782 if (V0.getOpcode() == ISD::UNDEF) {
5783 V0 = Op0.getOperand(0);
5784 if (V0.getValueType() != VT)
5788 if (V1.getOpcode() == ISD::UNDEF) {
5789 V1 = Op0.getOperand(0);
5790 if (V1.getValueType() != VT)
5793 if (i * 2 == NumElts)
5794 ExpectedVExtractIdx = BaseIdx;
5797 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5798 if (I0 == ExpectedVExtractIdx)
5799 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5800 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5801 // Try to match the following dag sequence:
5802 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5803 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5807 ExpectedVExtractIdx += 2;
5813 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5814 /// a concat_vector.
5816 /// This is a helper function of LowerToHorizontalOp().
5817 /// This function expects two 256-bit vectors called V0 and V1.
5818 /// At first, each vector is split into two separate 128-bit vectors.
5819 /// Then, the resulting 128-bit vectors are used to implement two
5820 /// horizontal binary operations.
5822 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5824 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5825 /// the two new horizontal binop.
5826 /// When Mode is set, the first horizontal binop dag node would take as input
5827 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5828 /// horizontal binop dag node would take as input the lower 128-bit of V1
5829 /// and the upper 128-bit of V1.
5831 /// HADD V0_LO, V0_HI
5832 /// HADD V1_LO, V1_HI
5834 /// Otherwise, the first horizontal binop dag node takes as input the lower
5835 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5836 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5838 /// HADD V0_LO, V1_LO
5839 /// HADD V0_HI, V1_HI
5841 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5842 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5843 /// the upper 128-bits of the result.
5844 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5845 SDLoc DL, SelectionDAG &DAG,
5846 unsigned X86Opcode, bool Mode,
5847 bool isUndefLO, bool isUndefHI) {
5848 EVT VT = V0.getValueType();
5849 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5850 "Invalid nodes in input!");
5852 unsigned NumElts = VT.getVectorNumElements();
5853 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5854 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5855 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5856 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5857 EVT NewVT = V0_LO.getValueType();
5859 SDValue LO = DAG.getUNDEF(NewVT);
5860 SDValue HI = DAG.getUNDEF(NewVT);
5863 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5864 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5865 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5866 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5867 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5869 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5870 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5871 V1_LO->getOpcode() != ISD::UNDEF))
5872 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5874 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5875 V1_HI->getOpcode() != ISD::UNDEF))
5876 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5879 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5882 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5884 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5885 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5886 EVT VT = BV->getValueType(0);
5887 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5888 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5892 unsigned NumElts = VT.getVectorNumElements();
5893 SDValue InVec0 = DAG.getUNDEF(VT);
5894 SDValue InVec1 = DAG.getUNDEF(VT);
5896 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5897 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5899 // Odd-numbered elements in the input build vector are obtained from
5900 // adding two integer/float elements.
5901 // Even-numbered elements in the input build vector are obtained from
5902 // subtracting two integer/float elements.
5903 unsigned ExpectedOpcode = ISD::FSUB;
5904 unsigned NextExpectedOpcode = ISD::FADD;
5905 bool AddFound = false;
5906 bool SubFound = false;
5908 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5909 SDValue Op = BV->getOperand(i);
5911 // Skip 'undef' values.
5912 unsigned Opcode = Op.getOpcode();
5913 if (Opcode == ISD::UNDEF) {
5914 std::swap(ExpectedOpcode, NextExpectedOpcode);
5918 // Early exit if we found an unexpected opcode.
5919 if (Opcode != ExpectedOpcode)
5922 SDValue Op0 = Op.getOperand(0);
5923 SDValue Op1 = Op.getOperand(1);
5925 // Try to match the following pattern:
5926 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5927 // Early exit if we cannot match that sequence.
5928 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5929 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5930 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5931 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5932 Op0.getOperand(1) != Op1.getOperand(1))
5935 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5939 // We found a valid add/sub node. Update the information accordingly.
5945 // Update InVec0 and InVec1.
5946 if (InVec0.getOpcode() == ISD::UNDEF) {
5947 InVec0 = Op0.getOperand(0);
5948 if (InVec0.getValueType() != VT)
5951 if (InVec1.getOpcode() == ISD::UNDEF) {
5952 InVec1 = Op1.getOperand(0);
5953 if (InVec1.getValueType() != VT)
5957 // Make sure that operands in input to each add/sub node always
5958 // come from a same pair of vectors.
5959 if (InVec0 != Op0.getOperand(0)) {
5960 if (ExpectedOpcode == ISD::FSUB)
5963 // FADD is commutable. Try to commute the operands
5964 // and then test again.
5965 std::swap(Op0, Op1);
5966 if (InVec0 != Op0.getOperand(0))
5970 if (InVec1 != Op1.getOperand(0))
5973 // Update the pair of expected opcodes.
5974 std::swap(ExpectedOpcode, NextExpectedOpcode);
5977 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5978 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5979 InVec1.getOpcode() != ISD::UNDEF)
5980 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5985 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5986 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5987 const X86Subtarget *Subtarget,
5988 SelectionDAG &DAG) {
5989 EVT VT = BV->getValueType(0);
5990 unsigned NumElts = VT.getVectorNumElements();
5991 unsigned NumUndefsLO = 0;
5992 unsigned NumUndefsHI = 0;
5993 unsigned Half = NumElts/2;
5995 // Count the number of UNDEF operands in the build_vector in input.
5996 for (unsigned i = 0, e = Half; i != e; ++i)
5997 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6000 for (unsigned i = Half, e = NumElts; i != e; ++i)
6001 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6004 // Early exit if this is either a build_vector of all UNDEFs or all the
6005 // operands but one are UNDEF.
6006 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6010 SDValue InVec0, InVec1;
6011 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6012 // Try to match an SSE3 float HADD/HSUB.
6013 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6014 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6016 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6017 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6018 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6019 // Try to match an SSSE3 integer HADD/HSUB.
6020 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6021 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6023 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6024 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6027 if (!Subtarget->hasAVX())
6030 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6031 // Try to match an AVX horizontal add/sub of packed single/double
6032 // precision floating point values from 256-bit vectors.
6033 SDValue InVec2, InVec3;
6034 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6035 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6036 ((InVec0.getOpcode() == ISD::UNDEF ||
6037 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6038 ((InVec1.getOpcode() == ISD::UNDEF ||
6039 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6040 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6042 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6043 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6044 ((InVec0.getOpcode() == ISD::UNDEF ||
6045 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6046 ((InVec1.getOpcode() == ISD::UNDEF ||
6047 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6048 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6049 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6050 // Try to match an AVX2 horizontal add/sub of signed integers.
6051 SDValue InVec2, InVec3;
6053 bool CanFold = true;
6055 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6056 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6057 ((InVec0.getOpcode() == ISD::UNDEF ||
6058 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6059 ((InVec1.getOpcode() == ISD::UNDEF ||
6060 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6061 X86Opcode = X86ISD::HADD;
6062 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6063 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6064 ((InVec0.getOpcode() == ISD::UNDEF ||
6065 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6066 ((InVec1.getOpcode() == ISD::UNDEF ||
6067 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6068 X86Opcode = X86ISD::HSUB;
6073 // Fold this build_vector into a single horizontal add/sub.
6074 // Do this only if the target has AVX2.
6075 if (Subtarget->hasAVX2())
6076 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6078 // Do not try to expand this build_vector into a pair of horizontal
6079 // add/sub if we can emit a pair of scalar add/sub.
6080 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6083 // Convert this build_vector into a pair of horizontal binop followed by
6085 bool isUndefLO = NumUndefsLO == Half;
6086 bool isUndefHI = NumUndefsHI == Half;
6087 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6088 isUndefLO, isUndefHI);
6092 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6093 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6095 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6096 X86Opcode = X86ISD::HADD;
6097 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6098 X86Opcode = X86ISD::HSUB;
6099 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6100 X86Opcode = X86ISD::FHADD;
6101 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6102 X86Opcode = X86ISD::FHSUB;
6106 // Don't try to expand this build_vector into a pair of horizontal add/sub
6107 // if we can simply emit a pair of scalar add/sub.
6108 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6111 // Convert this build_vector into two horizontal add/sub followed by
6113 bool isUndefLO = NumUndefsLO == Half;
6114 bool isUndefHI = NumUndefsHI == Half;
6115 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6116 isUndefLO, isUndefHI);
6123 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6126 MVT VT = Op.getSimpleValueType();
6127 MVT ExtVT = VT.getVectorElementType();
6128 unsigned NumElems = Op.getNumOperands();
6130 // Generate vectors for predicate vectors.
6131 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6132 return LowerBUILD_VECTORvXi1(Op, DAG);
6134 // Vectors containing all zeros can be matched by pxor and xorps later
6135 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6136 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6137 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6138 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6141 return getZeroVector(VT, Subtarget, DAG, dl);
6144 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6145 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6146 // vpcmpeqd on 256-bit vectors.
6147 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6148 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6151 if (!VT.is512BitVector())
6152 return getOnesVector(VT, Subtarget, DAG, dl);
6155 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6156 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6158 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6159 return HorizontalOp;
6160 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6163 unsigned EVTBits = ExtVT.getSizeInBits();
6165 unsigned NumZero = 0;
6166 unsigned NumNonZero = 0;
6167 unsigned NonZeros = 0;
6168 bool IsAllConstants = true;
6169 SmallSet<SDValue, 8> Values;
6170 for (unsigned i = 0; i < NumElems; ++i) {
6171 SDValue Elt = Op.getOperand(i);
6172 if (Elt.getOpcode() == ISD::UNDEF)
6175 if (Elt.getOpcode() != ISD::Constant &&
6176 Elt.getOpcode() != ISD::ConstantFP)
6177 IsAllConstants = false;
6178 if (X86::isZeroNode(Elt))
6181 NonZeros |= (1 << i);
6186 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6187 if (NumNonZero == 0)
6188 return DAG.getUNDEF(VT);
6190 // Special case for single non-zero, non-undef, element.
6191 if (NumNonZero == 1) {
6192 unsigned Idx = countTrailingZeros(NonZeros);
6193 SDValue Item = Op.getOperand(Idx);
6195 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6196 // the value are obviously zero, truncate the value to i32 and do the
6197 // insertion that way. Only do this if the value is non-constant or if the
6198 // value is a constant being inserted into element 0. It is cheaper to do
6199 // a constant pool load than it is to do a movd + shuffle.
6200 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6201 (!IsAllConstants || Idx == 0)) {
6202 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6204 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6205 EVT VecVT = MVT::v4i32;
6207 // Truncate the value (which may itself be a constant) to i32, and
6208 // convert it to a vector with movd (S2V+shuffle to zero extend).
6209 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6210 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6211 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6212 Item, Idx * 2, true, Subtarget, DAG));
6216 // If we have a constant or non-constant insertion into the low element of
6217 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6218 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6219 // depending on what the source datatype is.
6222 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6224 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6225 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6226 if (VT.is512BitVector()) {
6227 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6228 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6229 Item, DAG.getIntPtrConstant(0, dl));
6231 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6232 "Expected an SSE value type!");
6233 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6234 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6235 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6238 // We can't directly insert an i8 or i16 into a vector, so zero extend
6240 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6241 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6242 if (VT.is256BitVector()) {
6243 if (Subtarget->hasAVX()) {
6244 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6245 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6247 // Without AVX, we need to extend to a 128-bit vector and then
6248 // insert into the 256-bit vector.
6249 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6250 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6251 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6254 assert(VT.is128BitVector() && "Expected an SSE value type!");
6255 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6256 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6258 return DAG.getBitcast(VT, Item);
6262 // Is it a vector logical left shift?
6263 if (NumElems == 2 && Idx == 1 &&
6264 X86::isZeroNode(Op.getOperand(0)) &&
6265 !X86::isZeroNode(Op.getOperand(1))) {
6266 unsigned NumBits = VT.getSizeInBits();
6267 return getVShift(true, VT,
6268 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6269 VT, Op.getOperand(1)),
6270 NumBits/2, DAG, *this, dl);
6273 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6276 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6277 // is a non-constant being inserted into an element other than the low one,
6278 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6279 // movd/movss) to move this into the low element, then shuffle it into
6281 if (EVTBits == 32) {
6282 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6283 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6287 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6288 if (Values.size() == 1) {
6289 if (EVTBits == 32) {
6290 // Instead of a shuffle like this:
6291 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6292 // Check if it's possible to issue this instead.
6293 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6294 unsigned Idx = countTrailingZeros(NonZeros);
6295 SDValue Item = Op.getOperand(Idx);
6296 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6297 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6302 // A vector full of immediates; various special cases are already
6303 // handled, so this is best done with a single constant-pool load.
6307 // For AVX-length vectors, see if we can use a vector load to get all of the
6308 // elements, otherwise build the individual 128-bit pieces and use
6309 // shuffles to put them in place.
6310 if (VT.is256BitVector() || VT.is512BitVector()) {
6311 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6313 // Check for a build vector of consecutive loads.
6314 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6317 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6319 // Build both the lower and upper subvector.
6320 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6321 makeArrayRef(&V[0], NumElems/2));
6322 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6323 makeArrayRef(&V[NumElems / 2], NumElems/2));
6325 // Recreate the wider vector with the lower and upper part.
6326 if (VT.is256BitVector())
6327 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6328 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6331 // Let legalizer expand 2-wide build_vectors.
6332 if (EVTBits == 64) {
6333 if (NumNonZero == 1) {
6334 // One half is zero or undef.
6335 unsigned Idx = countTrailingZeros(NonZeros);
6336 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6337 Op.getOperand(Idx));
6338 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6343 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6344 if (EVTBits == 8 && NumElems == 16)
6345 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6349 if (EVTBits == 16 && NumElems == 8)
6350 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6354 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6355 if (EVTBits == 32 && NumElems == 4)
6356 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6359 // If element VT is == 32 bits, turn it into a number of shuffles.
6360 SmallVector<SDValue, 8> V(NumElems);
6361 if (NumElems == 4 && NumZero > 0) {
6362 for (unsigned i = 0; i < 4; ++i) {
6363 bool isZero = !(NonZeros & (1 << i));
6365 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6367 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6370 for (unsigned i = 0; i < 2; ++i) {
6371 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6374 V[i] = V[i*2]; // Must be a zero vector.
6377 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6380 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6383 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6388 bool Reverse1 = (NonZeros & 0x3) == 2;
6389 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6393 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6394 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6396 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6399 if (Values.size() > 1 && VT.is128BitVector()) {
6400 // Check for a build vector of consecutive loads.
6401 for (unsigned i = 0; i < NumElems; ++i)
6402 V[i] = Op.getOperand(i);
6404 // Check for elements which are consecutive loads.
6405 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6408 // Check for a build vector from mostly shuffle plus few inserting.
6409 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6412 // For SSE 4.1, use insertps to put the high elements into the low element.
6413 if (Subtarget->hasSSE41()) {
6415 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6416 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6418 Result = DAG.getUNDEF(VT);
6420 for (unsigned i = 1; i < NumElems; ++i) {
6421 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6422 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6423 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6428 // Otherwise, expand into a number of unpckl*, start by extending each of
6429 // our (non-undef) elements to the full vector width with the element in the
6430 // bottom slot of the vector (which generates no code for SSE).
6431 for (unsigned i = 0; i < NumElems; ++i) {
6432 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6433 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6435 V[i] = DAG.getUNDEF(VT);
6438 // Next, we iteratively mix elements, e.g. for v4f32:
6439 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6440 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6441 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6442 unsigned EltStride = NumElems >> 1;
6443 while (EltStride != 0) {
6444 for (unsigned i = 0; i < EltStride; ++i) {
6445 // If V[i+EltStride] is undef and this is the first round of mixing,
6446 // then it is safe to just drop this shuffle: V[i] is already in the
6447 // right place, the one element (since it's the first round) being
6448 // inserted as undef can be dropped. This isn't safe for successive
6449 // rounds because they will permute elements within both vectors.
6450 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6451 EltStride == NumElems/2)
6454 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6463 // 256-bit AVX can use the vinsertf128 instruction
6464 // to create 256-bit vectors from two other 128-bit ones.
6465 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6467 MVT ResVT = Op.getSimpleValueType();
6469 assert((ResVT.is256BitVector() ||
6470 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6472 SDValue V1 = Op.getOperand(0);
6473 SDValue V2 = Op.getOperand(1);
6474 unsigned NumElems = ResVT.getVectorNumElements();
6475 if (ResVT.is256BitVector())
6476 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6478 if (Op.getNumOperands() == 4) {
6479 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6480 ResVT.getVectorNumElements()/2);
6481 SDValue V3 = Op.getOperand(2);
6482 SDValue V4 = Op.getOperand(3);
6483 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6484 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6486 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6489 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6490 const X86Subtarget *Subtarget,
6491 SelectionDAG & DAG) {
6493 MVT ResVT = Op.getSimpleValueType();
6494 unsigned NumOfOperands = Op.getNumOperands();
6496 assert(isPowerOf2_32(NumOfOperands) &&
6497 "Unexpected number of operands in CONCAT_VECTORS");
6499 if (NumOfOperands > 2) {
6500 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6501 ResVT.getVectorNumElements()/2);
6502 SmallVector<SDValue, 2> Ops;
6503 for (unsigned i = 0; i < NumOfOperands/2; i++)
6504 Ops.push_back(Op.getOperand(i));
6505 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6507 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6508 Ops.push_back(Op.getOperand(i));
6509 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6510 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6513 SDValue V1 = Op.getOperand(0);
6514 SDValue V2 = Op.getOperand(1);
6515 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6516 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6518 if (IsZeroV1 && IsZeroV2)
6519 return getZeroVector(ResVT, Subtarget, DAG, dl);
6521 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6522 SDValue Undef = DAG.getUNDEF(ResVT);
6523 unsigned NumElems = ResVT.getVectorNumElements();
6524 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6526 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6527 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6531 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6532 // Zero the upper bits of V1
6533 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6534 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6537 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6540 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6541 const X86Subtarget *Subtarget,
6542 SelectionDAG &DAG) {
6543 MVT VT = Op.getSimpleValueType();
6544 if (VT.getVectorElementType() == MVT::i1)
6545 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6547 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6548 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6549 Op.getNumOperands() == 4)));
6551 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6552 // from two other 128-bit ones.
6554 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6555 return LowerAVXCONCAT_VECTORS(Op, DAG);
6558 //===----------------------------------------------------------------------===//
6559 // Vector shuffle lowering
6561 // This is an experimental code path for lowering vector shuffles on x86. It is
6562 // designed to handle arbitrary vector shuffles and blends, gracefully
6563 // degrading performance as necessary. It works hard to recognize idiomatic
6564 // shuffles and lower them to optimal instruction patterns without leaving
6565 // a framework that allows reasonably efficient handling of all vector shuffle
6567 //===----------------------------------------------------------------------===//
6569 /// \brief Tiny helper function to identify a no-op mask.
6571 /// This is a somewhat boring predicate function. It checks whether the mask
6572 /// array input, which is assumed to be a single-input shuffle mask of the kind
6573 /// used by the X86 shuffle instructions (not a fully general
6574 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6575 /// in-place shuffle are 'no-op's.
6576 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6577 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6578 if (Mask[i] != -1 && Mask[i] != i)
6583 /// \brief Helper function to classify a mask as a single-input mask.
6585 /// This isn't a generic single-input test because in the vector shuffle
6586 /// lowering we canonicalize single inputs to be the first input operand. This
6587 /// means we can more quickly test for a single input by only checking whether
6588 /// an input from the second operand exists. We also assume that the size of
6589 /// mask corresponds to the size of the input vectors which isn't true in the
6590 /// fully general case.
6591 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6593 if (M >= (int)Mask.size())
6598 /// \brief Test whether there are elements crossing 128-bit lanes in this
6601 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6602 /// and we routinely test for these.
6603 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6604 int LaneSize = 128 / VT.getScalarSizeInBits();
6605 int Size = Mask.size();
6606 for (int i = 0; i < Size; ++i)
6607 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6612 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6614 /// This checks a shuffle mask to see if it is performing the same
6615 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6616 /// that it is also not lane-crossing. It may however involve a blend from the
6617 /// same lane of a second vector.
6619 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6620 /// non-trivial to compute in the face of undef lanes. The representation is
6621 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6622 /// entries from both V1 and V2 inputs to the wider mask.
6624 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6625 SmallVectorImpl<int> &RepeatedMask) {
6626 int LaneSize = 128 / VT.getScalarSizeInBits();
6627 RepeatedMask.resize(LaneSize, -1);
6628 int Size = Mask.size();
6629 for (int i = 0; i < Size; ++i) {
6632 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6633 // This entry crosses lanes, so there is no way to model this shuffle.
6636 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6637 if (RepeatedMask[i % LaneSize] == -1)
6638 // This is the first non-undef entry in this slot of a 128-bit lane.
6639 RepeatedMask[i % LaneSize] =
6640 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6641 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6642 // Found a mismatch with the repeated mask.
6648 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6651 /// This is a fast way to test a shuffle mask against a fixed pattern:
6653 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6655 /// It returns true if the mask is exactly as wide as the argument list, and
6656 /// each element of the mask is either -1 (signifying undef) or the value given
6657 /// in the argument.
6658 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6659 ArrayRef<int> ExpectedMask) {
6660 if (Mask.size() != ExpectedMask.size())
6663 int Size = Mask.size();
6665 // If the values are build vectors, we can look through them to find
6666 // equivalent inputs that make the shuffles equivalent.
6667 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6668 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6670 for (int i = 0; i < Size; ++i)
6671 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6672 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6673 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6674 if (!MaskBV || !ExpectedBV ||
6675 MaskBV->getOperand(Mask[i] % Size) !=
6676 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6683 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6685 /// This helper function produces an 8-bit shuffle immediate corresponding to
6686 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6687 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6690 /// NB: We rely heavily on "undef" masks preserving the input lane.
6691 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6692 SelectionDAG &DAG) {
6693 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6694 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6695 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6696 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6697 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6700 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6701 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6702 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6703 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6704 return DAG.getConstant(Imm, DL, MVT::i8);
6707 /// \brief Compute whether each element of a shuffle is zeroable.
6709 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6710 /// Either it is an undef element in the shuffle mask, the element of the input
6711 /// referenced is undef, or the element of the input referenced is known to be
6712 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6713 /// as many lanes with this technique as possible to simplify the remaining
6715 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6716 SDValue V1, SDValue V2) {
6717 SmallBitVector Zeroable(Mask.size(), false);
6719 while (V1.getOpcode() == ISD::BITCAST)
6720 V1 = V1->getOperand(0);
6721 while (V2.getOpcode() == ISD::BITCAST)
6722 V2 = V2->getOperand(0);
6724 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6725 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6727 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6729 // Handle the easy cases.
6730 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6735 // If this is an index into a build_vector node (which has the same number
6736 // of elements), dig out the input value and use it.
6737 SDValue V = M < Size ? V1 : V2;
6738 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6741 SDValue Input = V.getOperand(M % Size);
6742 // The UNDEF opcode check really should be dead code here, but not quite
6743 // worth asserting on (it isn't invalid, just unexpected).
6744 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6751 // X86 has dedicated unpack instructions that can handle specific blend
6752 // operations: UNPCKH and UNPCKL.
6753 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6754 SDValue V1, SDValue V2,
6755 SelectionDAG &DAG) {
6756 int NumElts = VT.getVectorNumElements();
6757 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6758 SmallVector<int, 8> Unpckl;
6759 SmallVector<int, 8> Unpckh;
6761 for (int i = 0; i < NumElts; ++i) {
6762 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6763 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6764 int HiPos = LoPos + NumEltsInLane / 2;
6765 Unpckl.push_back(LoPos);
6766 Unpckh.push_back(HiPos);
6769 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6770 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6771 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6772 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6774 // Commute and try again.
6775 ShuffleVectorSDNode::commuteMask(Unpckl);
6776 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6777 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6779 ShuffleVectorSDNode::commuteMask(Unpckh);
6780 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6781 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6786 /// \brief Try to emit a bitmask instruction for a shuffle.
6788 /// This handles cases where we can model a blend exactly as a bitmask due to
6789 /// one of the inputs being zeroable.
6790 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6791 SDValue V2, ArrayRef<int> Mask,
6792 SelectionDAG &DAG) {
6793 MVT EltVT = VT.getScalarType();
6794 int NumEltBits = EltVT.getSizeInBits();
6795 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6796 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6797 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6799 if (EltVT.isFloatingPoint()) {
6800 Zero = DAG.getBitcast(EltVT, Zero);
6801 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6803 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6804 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6806 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6809 if (Mask[i] % Size != i)
6810 return SDValue(); // Not a blend.
6812 V = Mask[i] < Size ? V1 : V2;
6813 else if (V != (Mask[i] < Size ? V1 : V2))
6814 return SDValue(); // Can only let one input through the mask.
6816 VMaskOps[i] = AllOnes;
6819 return SDValue(); // No non-zeroable elements!
6821 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6822 V = DAG.getNode(VT.isFloatingPoint()
6823 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6828 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6830 /// This is used as a fallback approach when first class blend instructions are
6831 /// unavailable. Currently it is only suitable for integer vectors, but could
6832 /// be generalized for floating point vectors if desirable.
6833 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6834 SDValue V2, ArrayRef<int> Mask,
6835 SelectionDAG &DAG) {
6836 assert(VT.isInteger() && "Only supports integer vector types!");
6837 MVT EltVT = VT.getScalarType();
6838 int NumEltBits = EltVT.getSizeInBits();
6839 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6840 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6842 SmallVector<SDValue, 16> MaskOps;
6843 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6844 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6845 return SDValue(); // Shuffled input!
6846 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6849 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6850 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6851 // We have to cast V2 around.
6852 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6853 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6854 DAG.getBitcast(MaskVT, V1Mask),
6855 DAG.getBitcast(MaskVT, V2)));
6856 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6859 /// \brief Try to emit a blend instruction for a shuffle.
6861 /// This doesn't do any checks for the availability of instructions for blending
6862 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6863 /// be matched in the backend with the type given. What it does check for is
6864 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6865 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6866 SDValue V2, ArrayRef<int> Original,
6867 const X86Subtarget *Subtarget,
6868 SelectionDAG &DAG) {
6869 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6870 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6871 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6872 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6873 bool ForceV1Zero = false, ForceV2Zero = false;
6875 // Attempt to generate the binary blend mask. If an input is zero then
6876 // we can use any lane.
6877 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
6878 unsigned BlendMask = 0;
6879 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6885 if (M == i + Size) {
6886 BlendMask |= 1u << i;
6897 BlendMask |= 1u << i;
6902 return SDValue(); // Shuffled input!
6905 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
6907 V1 = getZeroVector(VT, Subtarget, DAG, DL);
6909 V2 = getZeroVector(VT, Subtarget, DAG, DL);
6911 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
6912 unsigned ScaledMask = 0;
6913 for (int i = 0; i != Size; ++i)
6914 if (BlendMask & (1u << i))
6915 for (int j = 0; j != Scale; ++j)
6916 ScaledMask |= 1u << (i * Scale + j);
6920 switch (VT.SimpleTy) {
6925 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6926 DAG.getConstant(BlendMask, DL, MVT::i8));
6930 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6934 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6935 // that instruction.
6936 if (Subtarget->hasAVX2()) {
6937 // Scale the blend by the number of 32-bit dwords per element.
6938 int Scale = VT.getScalarSizeInBits() / 32;
6939 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6940 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6941 V1 = DAG.getBitcast(BlendVT, V1);
6942 V2 = DAG.getBitcast(BlendVT, V2);
6943 return DAG.getBitcast(
6944 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6945 DAG.getConstant(BlendMask, DL, MVT::i8)));
6949 // For integer shuffles we need to expand the mask and cast the inputs to
6950 // v8i16s prior to blending.
6951 int Scale = 8 / VT.getVectorNumElements();
6952 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6953 V1 = DAG.getBitcast(MVT::v8i16, V1);
6954 V2 = DAG.getBitcast(MVT::v8i16, V2);
6955 return DAG.getBitcast(VT,
6956 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6957 DAG.getConstant(BlendMask, DL, MVT::i8)));
6961 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6962 SmallVector<int, 8> RepeatedMask;
6963 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6964 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6965 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6967 for (int i = 0; i < 8; ++i)
6968 if (RepeatedMask[i] >= 16)
6969 BlendMask |= 1u << i;
6970 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6971 DAG.getConstant(BlendMask, DL, MVT::i8));
6977 assert((VT.is128BitVector() || Subtarget->hasAVX2()) &&
6978 "256-bit byte-blends require AVX2 support!");
6980 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6981 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6984 // Scale the blend by the number of bytes per element.
6985 int Scale = VT.getScalarSizeInBits() / 8;
6987 // This form of blend is always done on bytes. Compute the byte vector
6989 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6991 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6992 // mix of LLVM's code generator and the x86 backend. We tell the code
6993 // generator that boolean values in the elements of an x86 vector register
6994 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6995 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6996 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6997 // of the element (the remaining are ignored) and 0 in that high bit would
6998 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6999 // the LLVM model for boolean values in vector elements gets the relevant
7000 // bit set, it is set backwards and over constrained relative to x86's
7002 SmallVector<SDValue, 32> VSELECTMask;
7003 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7004 for (int j = 0; j < Scale; ++j)
7005 VSELECTMask.push_back(
7006 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7007 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7010 V1 = DAG.getBitcast(BlendVT, V1);
7011 V2 = DAG.getBitcast(BlendVT, V2);
7012 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7013 DAG.getNode(ISD::BUILD_VECTOR, DL,
7014 BlendVT, VSELECTMask),
7019 llvm_unreachable("Not a supported integer vector type!");
7023 /// \brief Try to lower as a blend of elements from two inputs followed by
7024 /// a single-input permutation.
7026 /// This matches the pattern where we can blend elements from two inputs and
7027 /// then reduce the shuffle to a single-input permutation.
7028 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7031 SelectionDAG &DAG) {
7032 // We build up the blend mask while checking whether a blend is a viable way
7033 // to reduce the shuffle.
7034 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7035 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7037 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7041 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7043 if (BlendMask[Mask[i] % Size] == -1)
7044 BlendMask[Mask[i] % Size] = Mask[i];
7045 else if (BlendMask[Mask[i] % Size] != Mask[i])
7046 return SDValue(); // Can't blend in the needed input!
7048 PermuteMask[i] = Mask[i] % Size;
7051 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7052 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7055 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7056 /// blends and permutes.
7058 /// This matches the extremely common pattern for handling combined
7059 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7060 /// operations. It will try to pick the best arrangement of shuffles and
7062 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7066 SelectionDAG &DAG) {
7067 // Shuffle the input elements into the desired positions in V1 and V2 and
7068 // blend them together.
7069 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7070 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7071 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7072 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7073 if (Mask[i] >= 0 && Mask[i] < Size) {
7074 V1Mask[i] = Mask[i];
7076 } else if (Mask[i] >= Size) {
7077 V2Mask[i] = Mask[i] - Size;
7078 BlendMask[i] = i + Size;
7081 // Try to lower with the simpler initial blend strategy unless one of the
7082 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7083 // shuffle may be able to fold with a load or other benefit. However, when
7084 // we'll have to do 2x as many shuffles in order to achieve this, blending
7085 // first is a better strategy.
7086 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7087 if (SDValue BlendPerm =
7088 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7091 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7092 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7093 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7096 /// \brief Try to lower a vector shuffle as a byte rotation.
7098 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7099 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7100 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7101 /// try to generically lower a vector shuffle through such an pattern. It
7102 /// does not check for the profitability of lowering either as PALIGNR or
7103 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7104 /// This matches shuffle vectors that look like:
7106 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7108 /// Essentially it concatenates V1 and V2, shifts right by some number of
7109 /// elements, and takes the low elements as the result. Note that while this is
7110 /// specified as a *right shift* because x86 is little-endian, it is a *left
7111 /// rotate* of the vector lanes.
7112 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7115 const X86Subtarget *Subtarget,
7116 SelectionDAG &DAG) {
7117 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7119 int NumElts = Mask.size();
7120 int NumLanes = VT.getSizeInBits() / 128;
7121 int NumLaneElts = NumElts / NumLanes;
7123 // We need to detect various ways of spelling a rotation:
7124 // [11, 12, 13, 14, 15, 0, 1, 2]
7125 // [-1, 12, 13, 14, -1, -1, 1, -1]
7126 // [-1, -1, -1, -1, -1, -1, 1, 2]
7127 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7128 // [-1, 4, 5, 6, -1, -1, 9, -1]
7129 // [-1, 4, 5, 6, -1, -1, -1, -1]
7132 for (int l = 0; l < NumElts; l += NumLaneElts) {
7133 for (int i = 0; i < NumLaneElts; ++i) {
7134 if (Mask[l + i] == -1)
7136 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7138 // Get the mod-Size index and lane correct it.
7139 int LaneIdx = (Mask[l + i] % NumElts) - l;
7140 // Make sure it was in this lane.
7141 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7144 // Determine where a rotated vector would have started.
7145 int StartIdx = i - LaneIdx;
7147 // The identity rotation isn't interesting, stop.
7150 // If we found the tail of a vector the rotation must be the missing
7151 // front. If we found the head of a vector, it must be how much of the
7153 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7156 Rotation = CandidateRotation;
7157 else if (Rotation != CandidateRotation)
7158 // The rotations don't match, so we can't match this mask.
7161 // Compute which value this mask is pointing at.
7162 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7164 // Compute which of the two target values this index should be assigned
7165 // to. This reflects whether the high elements are remaining or the low
7166 // elements are remaining.
7167 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7169 // Either set up this value if we've not encountered it before, or check
7170 // that it remains consistent.
7173 else if (TargetV != MaskV)
7174 // This may be a rotation, but it pulls from the inputs in some
7175 // unsupported interleaving.
7180 // Check that we successfully analyzed the mask, and normalize the results.
7181 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7182 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7188 // The actual rotate instruction rotates bytes, so we need to scale the
7189 // rotation based on how many bytes are in the vector lane.
7190 int Scale = 16 / NumLaneElts;
7192 // SSSE3 targets can use the palignr instruction.
7193 if (Subtarget->hasSSSE3()) {
7194 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7195 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7196 Lo = DAG.getBitcast(AlignVT, Lo);
7197 Hi = DAG.getBitcast(AlignVT, Hi);
7199 return DAG.getBitcast(
7200 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7201 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7204 assert(VT.is128BitVector() &&
7205 "Rotate-based lowering only supports 128-bit lowering!");
7206 assert(Mask.size() <= 16 &&
7207 "Can shuffle at most 16 bytes in a 128-bit vector!");
7209 // Default SSE2 implementation
7210 int LoByteShift = 16 - Rotation * Scale;
7211 int HiByteShift = Rotation * Scale;
7213 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7214 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7215 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7217 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7218 DAG.getConstant(LoByteShift, DL, MVT::i8));
7219 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7220 DAG.getConstant(HiByteShift, DL, MVT::i8));
7221 return DAG.getBitcast(VT,
7222 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7225 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7227 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7228 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7229 /// matches elements from one of the input vectors shuffled to the left or
7230 /// right with zeroable elements 'shifted in'. It handles both the strictly
7231 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7234 /// PSHL : (little-endian) left bit shift.
7235 /// [ zz, 0, zz, 2 ]
7236 /// [ -1, 4, zz, -1 ]
7237 /// PSRL : (little-endian) right bit shift.
7239 /// [ -1, -1, 7, zz]
7240 /// PSLLDQ : (little-endian) left byte shift
7241 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7242 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7243 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7244 /// PSRLDQ : (little-endian) right byte shift
7245 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7246 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7247 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7248 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7249 SDValue V2, ArrayRef<int> Mask,
7250 SelectionDAG &DAG) {
7251 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7253 int Size = Mask.size();
7254 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7256 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7257 for (int i = 0; i < Size; i += Scale)
7258 for (int j = 0; j < Shift; ++j)
7259 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7265 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7266 for (int i = 0; i != Size; i += Scale) {
7267 unsigned Pos = Left ? i + Shift : i;
7268 unsigned Low = Left ? i : i + Shift;
7269 unsigned Len = Scale - Shift;
7270 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7271 Low + (V == V1 ? 0 : Size)))
7275 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7276 bool ByteShift = ShiftEltBits > 64;
7277 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7278 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7279 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7281 // Normalize the scale for byte shifts to still produce an i64 element
7283 Scale = ByteShift ? Scale / 2 : Scale;
7285 // We need to round trip through the appropriate type for the shift.
7286 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7287 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7288 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7289 "Illegal integer vector type");
7290 V = DAG.getBitcast(ShiftVT, V);
7292 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7293 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7294 return DAG.getBitcast(VT, V);
7297 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7298 // keep doubling the size of the integer elements up to that. We can
7299 // then shift the elements of the integer vector by whole multiples of
7300 // their width within the elements of the larger integer vector. Test each
7301 // multiple to see if we can find a match with the moved element indices
7302 // and that the shifted in elements are all zeroable.
7303 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7304 for (int Shift = 1; Shift != Scale; ++Shift)
7305 for (bool Left : {true, false})
7306 if (CheckZeros(Shift, Scale, Left))
7307 for (SDValue V : {V1, V2})
7308 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7315 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7316 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7317 SDValue V2, ArrayRef<int> Mask,
7318 SelectionDAG &DAG) {
7319 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7320 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7322 int Size = Mask.size();
7323 int HalfSize = Size / 2;
7324 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7326 // Upper half must be undefined.
7327 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7330 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7331 // Remainder of lower half result is zero and upper half is all undef.
7332 auto LowerAsEXTRQ = [&]() {
7333 // Determine the extraction length from the part of the
7334 // lower half that isn't zeroable.
7336 for (; Len > 0; --Len)
7337 if (!Zeroable[Len - 1])
7339 assert(Len > 0 && "Zeroable shuffle mask");
7341 // Attempt to match first Len sequential elements from the lower half.
7344 for (int i = 0; i != Len; ++i) {
7348 SDValue &V = (M < Size ? V1 : V2);
7351 // All mask elements must be in the lower half.
7355 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7366 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7367 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7368 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7369 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7370 DAG.getConstant(BitLen, DL, MVT::i8),
7371 DAG.getConstant(BitIdx, DL, MVT::i8));
7374 if (SDValue ExtrQ = LowerAsEXTRQ())
7377 // INSERTQ: Extract lowest Len elements from lower half of second source and
7378 // insert over first source, starting at Idx.
7379 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7380 auto LowerAsInsertQ = [&]() {
7381 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7384 // Attempt to match first source from mask before insertion point.
7385 if (isUndefInRange(Mask, 0, Idx)) {
7387 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7389 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7395 // Extend the extraction length looking to match both the insertion of
7396 // the second source and the remaining elements of the first.
7397 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7402 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7404 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7410 // Match the remaining elements of the lower half.
7411 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7413 } else if ((!Base || (Base == V1)) &&
7414 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7416 } else if ((!Base || (Base == V2)) &&
7417 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7424 // We may not have a base (first source) - this can safely be undefined.
7426 Base = DAG.getUNDEF(VT);
7428 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7429 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7430 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7431 DAG.getConstant(BitLen, DL, MVT::i8),
7432 DAG.getConstant(BitIdx, DL, MVT::i8));
7439 if (SDValue InsertQ = LowerAsInsertQ())
7445 /// \brief Lower a vector shuffle as a zero or any extension.
7447 /// Given a specific number of elements, element bit width, and extension
7448 /// stride, produce either a zero or any extension based on the available
7449 /// features of the subtarget. The extended elements are consecutive and
7450 /// begin and can start from an offseted element index in the input; to
7451 /// avoid excess shuffling the offset must either being in the bottom lane
7452 /// or at the start of a higher lane. All extended elements must be from
7454 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7455 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7456 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7457 assert(Scale > 1 && "Need a scale to extend.");
7458 int EltBits = VT.getScalarSizeInBits();
7459 int NumElements = VT.getVectorNumElements();
7460 int NumEltsPerLane = 128 / EltBits;
7461 int OffsetLane = Offset / NumEltsPerLane;
7462 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7463 "Only 8, 16, and 32 bit elements can be extended.");
7464 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7465 assert(0 <= Offset && "Extension offset must be positive.");
7466 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7467 "Extension offset must be in the first lane or start an upper lane.");
7469 // Check that an index is in same lane as the base offset.
7470 auto SafeOffset = [&](int Idx) {
7471 return OffsetLane == (Idx / NumEltsPerLane);
7474 // Shift along an input so that the offset base moves to the first element.
7475 auto ShuffleOffset = [&](SDValue V) {
7479 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7480 for (int i = 0; i * Scale < NumElements; ++i) {
7481 int SrcIdx = i + Offset;
7482 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7484 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7487 // Found a valid zext mask! Try various lowering strategies based on the
7488 // input type and available ISA extensions.
7489 if (Subtarget->hasSSE41()) {
7490 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7491 // PUNPCK will catch this in a later shuffle match.
7492 if (Offset && Scale == 2 && VT.is128BitVector())
7494 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7495 NumElements / Scale);
7496 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7497 return DAG.getBitcast(VT, InputV);
7500 assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
7502 // For any extends we can cheat for larger element sizes and use shuffle
7503 // instructions that can fold with a load and/or copy.
7504 if (AnyExt && EltBits == 32) {
7505 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7507 return DAG.getBitcast(
7508 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7509 DAG.getBitcast(MVT::v4i32, InputV),
7510 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7512 if (AnyExt && EltBits == 16 && Scale > 2) {
7513 int PSHUFDMask[4] = {Offset / 2, -1,
7514 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7515 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7516 DAG.getBitcast(MVT::v4i32, InputV),
7517 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7518 int PSHUFWMask[4] = {1, -1, -1, -1};
7519 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7520 return DAG.getBitcast(
7521 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7522 DAG.getBitcast(MVT::v8i16, InputV),
7523 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7526 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7528 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7529 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7530 assert(VT.is128BitVector() && "Unexpected vector width!");
7532 int LoIdx = Offset * EltBits;
7533 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7534 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7535 DAG.getConstant(EltBits, DL, MVT::i8),
7536 DAG.getConstant(LoIdx, DL, MVT::i8)));
7538 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7539 !SafeOffset(Offset + 1))
7540 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7542 int HiIdx = (Offset + 1) * EltBits;
7543 SDValue Hi = 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(HiIdx, DL, MVT::i8)));
7547 return DAG.getNode(ISD::BITCAST, DL, VT,
7548 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7551 // If this would require more than 2 unpack instructions to expand, use
7552 // pshufb when available. We can only use more than 2 unpack instructions
7553 // when zero extending i8 elements which also makes it easier to use pshufb.
7554 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7555 assert(NumElements == 16 && "Unexpected byte vector width!");
7556 SDValue PSHUFBMask[16];
7557 for (int i = 0; i < 16; ++i) {
7558 int Idx = Offset + (i / Scale);
7559 PSHUFBMask[i] = DAG.getConstant(
7560 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7562 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7563 return DAG.getBitcast(VT,
7564 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7565 DAG.getNode(ISD::BUILD_VECTOR, DL,
7566 MVT::v16i8, PSHUFBMask)));
7569 // If we are extending from an offset, ensure we start on a boundary that
7570 // we can unpack from.
7571 int AlignToUnpack = Offset % (NumElements / Scale);
7572 if (AlignToUnpack) {
7573 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7574 for (int i = AlignToUnpack; i < NumElements; ++i)
7575 ShMask[i - AlignToUnpack] = i;
7576 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7577 Offset -= AlignToUnpack;
7580 // Otherwise emit a sequence of unpacks.
7582 unsigned UnpackLoHi = X86ISD::UNPCKL;
7583 if (Offset >= (NumElements / 2)) {
7584 UnpackLoHi = X86ISD::UNPCKH;
7585 Offset -= (NumElements / 2);
7588 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7589 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7590 : getZeroVector(InputVT, Subtarget, DAG, DL);
7591 InputV = DAG.getBitcast(InputVT, InputV);
7592 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7596 } while (Scale > 1);
7597 return DAG.getBitcast(VT, InputV);
7600 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7602 /// This routine will try to do everything in its power to cleverly lower
7603 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7604 /// check for the profitability of this lowering, it tries to aggressively
7605 /// match this pattern. It will use all of the micro-architectural details it
7606 /// can to emit an efficient lowering. It handles both blends with all-zero
7607 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7608 /// masking out later).
7610 /// The reason we have dedicated lowering for zext-style shuffles is that they
7611 /// are both incredibly common and often quite performance sensitive.
7612 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7613 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7614 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7615 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7617 int Bits = VT.getSizeInBits();
7618 int NumLanes = Bits / 128;
7619 int NumElements = VT.getVectorNumElements();
7620 int NumEltsPerLane = NumElements / NumLanes;
7621 assert(VT.getScalarSizeInBits() <= 32 &&
7622 "Exceeds 32-bit integer zero extension limit");
7623 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7625 // Define a helper function to check a particular ext-scale and lower to it if
7627 auto Lower = [&](int Scale) -> SDValue {
7632 for (int i = 0; i < NumElements; ++i) {
7635 continue; // Valid anywhere but doesn't tell us anything.
7636 if (i % Scale != 0) {
7637 // Each of the extended elements need to be zeroable.
7641 // We no longer are in the anyext case.
7646 // Each of the base elements needs to be consecutive indices into the
7647 // same input vector.
7648 SDValue V = M < NumElements ? V1 : V2;
7649 M = M % NumElements;
7652 Offset = M - (i / Scale);
7653 } else if (InputV != V)
7654 return SDValue(); // Flip-flopping inputs.
7656 // Offset must start in the lowest 128-bit lane or at the start of an
7658 // FIXME: Is it ever worth allowing a negative base offset?
7659 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7660 (Offset % NumEltsPerLane) == 0))
7663 // If we are offsetting, all referenced entries must come from the same
7665 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7668 if ((M % NumElements) != (Offset + (i / Scale)))
7669 return SDValue(); // Non-consecutive strided elements.
7673 // If we fail to find an input, we have a zero-shuffle which should always
7674 // have already been handled.
7675 // FIXME: Maybe handle this here in case during blending we end up with one?
7679 // If we are offsetting, don't extend if we only match a single input, we
7680 // can always do better by using a basic PSHUF or PUNPCK.
7681 if (Offset != 0 && Matches < 2)
7684 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7685 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7688 // The widest scale possible for extending is to a 64-bit integer.
7689 assert(Bits % 64 == 0 &&
7690 "The number of bits in a vector must be divisible by 64 on x86!");
7691 int NumExtElements = Bits / 64;
7693 // Each iteration, try extending the elements half as much, but into twice as
7695 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7696 assert(NumElements % NumExtElements == 0 &&
7697 "The input vector size must be divisible by the extended size.");
7698 if (SDValue V = Lower(NumElements / NumExtElements))
7702 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7706 // Returns one of the source operands if the shuffle can be reduced to a
7707 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7708 auto CanZExtLowHalf = [&]() {
7709 for (int i = NumElements / 2; i != NumElements; ++i)
7712 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7714 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7719 if (SDValue V = CanZExtLowHalf()) {
7720 V = DAG.getBitcast(MVT::v2i64, V);
7721 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7722 return DAG.getBitcast(VT, V);
7725 // No viable ext lowering found.
7729 /// \brief Try to get a scalar value for a specific element of a vector.
7731 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7732 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7733 SelectionDAG &DAG) {
7734 MVT VT = V.getSimpleValueType();
7735 MVT EltVT = VT.getVectorElementType();
7736 while (V.getOpcode() == ISD::BITCAST)
7737 V = V.getOperand(0);
7738 // If the bitcasts shift the element size, we can't extract an equivalent
7740 MVT NewVT = V.getSimpleValueType();
7741 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7744 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7745 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7746 // Ensure the scalar operand is the same size as the destination.
7747 // FIXME: Add support for scalar truncation where possible.
7748 SDValue S = V.getOperand(Idx);
7749 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7750 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7756 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7758 /// This is particularly important because the set of instructions varies
7759 /// significantly based on whether the operand is a load or not.
7760 static bool isShuffleFoldableLoad(SDValue V) {
7761 while (V.getOpcode() == ISD::BITCAST)
7762 V = V.getOperand(0);
7764 return ISD::isNON_EXTLoad(V.getNode());
7767 /// \brief Try to lower insertion of a single element into a zero vector.
7769 /// This is a common pattern that we have especially efficient patterns to lower
7770 /// across all subtarget feature sets.
7771 static SDValue lowerVectorShuffleAsElementInsertion(
7772 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7773 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7774 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7776 MVT EltVT = VT.getVectorElementType();
7778 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7779 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7781 bool IsV1Zeroable = true;
7782 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7783 if (i != V2Index && !Zeroable[i]) {
7784 IsV1Zeroable = false;
7788 // Check for a single input from a SCALAR_TO_VECTOR node.
7789 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7790 // all the smarts here sunk into that routine. However, the current
7791 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7792 // vector shuffle lowering is dead.
7793 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7795 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7796 // We need to zext the scalar if it is smaller than an i32.
7797 V2S = DAG.getBitcast(EltVT, V2S);
7798 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7799 // Using zext to expand a narrow element won't work for non-zero
7804 // Zero-extend directly to i32.
7806 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7808 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7809 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7810 EltVT == MVT::i16) {
7811 // Either not inserting from the low element of the input or the input
7812 // element size is too small to use VZEXT_MOVL to clear the high bits.
7816 if (!IsV1Zeroable) {
7817 // If V1 can't be treated as a zero vector we have fewer options to lower
7818 // this. We can't support integer vectors or non-zero targets cheaply, and
7819 // the V1 elements can't be permuted in any way.
7820 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7821 if (!VT.isFloatingPoint() || V2Index != 0)
7823 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7824 V1Mask[V2Index] = -1;
7825 if (!isNoopShuffleMask(V1Mask))
7827 // This is essentially a special case blend operation, but if we have
7828 // general purpose blend operations, they are always faster. Bail and let
7829 // the rest of the lowering handle these as blends.
7830 if (Subtarget->hasSSE41())
7833 // Otherwise, use MOVSD or MOVSS.
7834 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7835 "Only two types of floating point element types to handle!");
7836 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7840 // This lowering only works for the low element with floating point vectors.
7841 if (VT.isFloatingPoint() && V2Index != 0)
7844 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7846 V2 = DAG.getBitcast(VT, V2);
7849 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7850 // the desired position. Otherwise it is more efficient to do a vector
7851 // shift left. We know that we can do a vector shift left because all
7852 // the inputs are zero.
7853 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7854 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7855 V2Shuffle[V2Index] = 0;
7856 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7858 V2 = DAG.getBitcast(MVT::v2i64, V2);
7860 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7861 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7862 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7863 DAG.getDataLayout(), VT)));
7864 V2 = DAG.getBitcast(VT, V2);
7870 /// \brief Try to lower broadcast of a single element.
7872 /// For convenience, this code also bundles all of the subtarget feature set
7873 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7874 /// a convenient way to factor it out.
7875 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7877 const X86Subtarget *Subtarget,
7878 SelectionDAG &DAG) {
7879 if (!Subtarget->hasAVX())
7881 if (VT.isInteger() && !Subtarget->hasAVX2())
7884 // Check that the mask is a broadcast.
7885 int BroadcastIdx = -1;
7887 if (M >= 0 && BroadcastIdx == -1)
7889 else if (M >= 0 && M != BroadcastIdx)
7892 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7893 "a sorted mask where the broadcast "
7896 // Go up the chain of (vector) values to find a scalar load that we can
7897 // combine with the broadcast.
7899 switch (V.getOpcode()) {
7900 case ISD::CONCAT_VECTORS: {
7901 int OperandSize = Mask.size() / V.getNumOperands();
7902 V = V.getOperand(BroadcastIdx / OperandSize);
7903 BroadcastIdx %= OperandSize;
7907 case ISD::INSERT_SUBVECTOR: {
7908 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7909 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7913 int BeginIdx = (int)ConstantIdx->getZExtValue();
7915 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7916 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7917 BroadcastIdx -= BeginIdx;
7928 // Check if this is a broadcast of a scalar. We special case lowering
7929 // for scalars so that we can more effectively fold with loads.
7930 // First, look through bitcast: if the original value has a larger element
7931 // type than the shuffle, the broadcast element is in essence truncated.
7932 // Make that explicit to ease folding.
7933 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7934 EVT EltVT = VT.getVectorElementType();
7935 SDValue V0 = V.getOperand(0);
7936 EVT V0VT = V0.getValueType();
7938 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7939 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7940 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7941 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7946 // Also check the simpler case, where we can directly reuse the scalar.
7947 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7948 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7949 V = V.getOperand(BroadcastIdx);
7951 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7952 // Only AVX2 has register broadcasts.
7953 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7955 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7956 // We can't broadcast from a vector register without AVX2, and we can only
7957 // broadcast from the zero-element of a vector register.
7961 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7964 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7965 // INSERTPS when the V1 elements are already in the correct locations
7966 // because otherwise we can just always use two SHUFPS instructions which
7967 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7968 // perform INSERTPS if a single V1 element is out of place and all V2
7969 // elements are zeroable.
7970 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7972 SelectionDAG &DAG) {
7973 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7974 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7975 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7976 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7978 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7981 int V1DstIndex = -1;
7982 int V2DstIndex = -1;
7983 bool V1UsedInPlace = false;
7985 for (int i = 0; i < 4; ++i) {
7986 // Synthesize a zero mask from the zeroable elements (includes undefs).
7992 // Flag if we use any V1 inputs in place.
7994 V1UsedInPlace = true;
7998 // We can only insert a single non-zeroable element.
7999 if (V1DstIndex != -1 || V2DstIndex != -1)
8003 // V1 input out of place for insertion.
8006 // V2 input for insertion.
8011 // Don't bother if we have no (non-zeroable) element for insertion.
8012 if (V1DstIndex == -1 && V2DstIndex == -1)
8015 // Determine element insertion src/dst indices. The src index is from the
8016 // start of the inserted vector, not the start of the concatenated vector.
8017 unsigned V2SrcIndex = 0;
8018 if (V1DstIndex != -1) {
8019 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8020 // and don't use the original V2 at all.
8021 V2SrcIndex = Mask[V1DstIndex];
8022 V2DstIndex = V1DstIndex;
8025 V2SrcIndex = Mask[V2DstIndex] - 4;
8028 // If no V1 inputs are used in place, then the result is created only from
8029 // the zero mask and the V2 insertion - so remove V1 dependency.
8031 V1 = DAG.getUNDEF(MVT::v4f32);
8033 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8034 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8036 // Insert the V2 element into the desired position.
8038 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8039 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8042 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8043 /// UNPCK instruction.
8045 /// This specifically targets cases where we end up with alternating between
8046 /// the two inputs, and so can permute them into something that feeds a single
8047 /// UNPCK instruction. Note that this routine only targets integer vectors
8048 /// because for floating point vectors we have a generalized SHUFPS lowering
8049 /// strategy that handles everything that doesn't *exactly* match an unpack,
8050 /// making this clever lowering unnecessary.
8051 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8052 SDValue V1, SDValue V2,
8054 SelectionDAG &DAG) {
8055 assert(!VT.isFloatingPoint() &&
8056 "This routine only supports integer vectors.");
8057 assert(!isSingleInputShuffleMask(Mask) &&
8058 "This routine should only be used when blending two inputs.");
8059 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8061 int Size = Mask.size();
8063 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8064 return M >= 0 && M % Size < Size / 2;
8066 int NumHiInputs = std::count_if(
8067 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8069 bool UnpackLo = NumLoInputs >= NumHiInputs;
8071 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8072 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8073 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8075 for (int i = 0; i < Size; ++i) {
8079 // Each element of the unpack contains Scale elements from this mask.
8080 int UnpackIdx = i / Scale;
8082 // We only handle the case where V1 feeds the first slots of the unpack.
8083 // We rely on canonicalization to ensure this is the case.
8084 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8087 // Setup the mask for this input. The indexing is tricky as we have to
8088 // handle the unpack stride.
8089 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8090 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8094 // If we will have to shuffle both inputs to use the unpack, check whether
8095 // we can just unpack first and shuffle the result. If so, skip this unpack.
8096 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8097 !isNoopShuffleMask(V2Mask))
8100 // Shuffle the inputs into place.
8101 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8102 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8104 // Cast the inputs to the type we will use to unpack them.
8105 V1 = DAG.getBitcast(UnpackVT, V1);
8106 V2 = DAG.getBitcast(UnpackVT, V2);
8108 // Unpack the inputs and cast the result back to the desired type.
8109 return DAG.getBitcast(
8110 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8114 // We try each unpack from the largest to the smallest to try and find one
8115 // that fits this mask.
8116 int OrigNumElements = VT.getVectorNumElements();
8117 int OrigScalarSize = VT.getScalarSizeInBits();
8118 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8119 int Scale = ScalarSize / OrigScalarSize;
8120 int NumElements = OrigNumElements / Scale;
8121 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8122 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8126 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8128 if (NumLoInputs == 0 || NumHiInputs == 0) {
8129 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8130 "We have to have *some* inputs!");
8131 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8133 // FIXME: We could consider the total complexity of the permute of each
8134 // possible unpacking. Or at the least we should consider how many
8135 // half-crossings are created.
8136 // FIXME: We could consider commuting the unpacks.
8138 SmallVector<int, 32> PermMask;
8139 PermMask.assign(Size, -1);
8140 for (int i = 0; i < Size; ++i) {
8144 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8147 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8149 return DAG.getVectorShuffle(
8150 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8152 DAG.getUNDEF(VT), PermMask);
8158 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8160 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8161 /// support for floating point shuffles but not integer shuffles. These
8162 /// instructions will incur a domain crossing penalty on some chips though so
8163 /// it is better to avoid lowering through this for integer vectors where
8165 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8166 const X86Subtarget *Subtarget,
8167 SelectionDAG &DAG) {
8169 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8170 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8171 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8172 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8173 ArrayRef<int> Mask = SVOp->getMask();
8174 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8176 if (isSingleInputShuffleMask(Mask)) {
8177 // Use low duplicate instructions for masks that match their pattern.
8178 if (Subtarget->hasSSE3())
8179 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8180 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8182 // Straight shuffle of a single input vector. Simulate this by using the
8183 // single input as both of the "inputs" to this instruction..
8184 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8186 if (Subtarget->hasAVX()) {
8187 // If we have AVX, we can use VPERMILPS which will allow folding a load
8188 // into the shuffle.
8189 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8190 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8193 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8194 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8196 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8197 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8199 // If we have a single input, insert that into V1 if we can do so cheaply.
8200 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8201 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8202 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8204 // Try inverting the insertion since for v2 masks it is easy to do and we
8205 // can't reliably sort the mask one way or the other.
8206 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8207 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8208 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8209 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8213 // Try to use one of the special instruction patterns to handle two common
8214 // blend patterns if a zero-blend above didn't work.
8215 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8216 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8217 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8218 // We can either use a special instruction to load over the low double or
8219 // to move just the low double.
8221 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8223 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8225 if (Subtarget->hasSSE41())
8226 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8230 // Use dedicated unpack instructions for masks that match their pattern.
8232 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8235 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8236 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8237 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8240 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8242 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8243 /// the integer unit to minimize domain crossing penalties. However, for blends
8244 /// it falls back to the floating point shuffle operation with appropriate bit
8246 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8247 const X86Subtarget *Subtarget,
8248 SelectionDAG &DAG) {
8250 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8251 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8252 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8253 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8254 ArrayRef<int> Mask = SVOp->getMask();
8255 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8257 if (isSingleInputShuffleMask(Mask)) {
8258 // Check for being able to broadcast a single element.
8259 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8260 Mask, Subtarget, DAG))
8263 // Straight shuffle of a single input vector. For everything from SSE2
8264 // onward this has a single fast instruction with no scary immediates.
8265 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8266 V1 = DAG.getBitcast(MVT::v4i32, V1);
8267 int WidenedMask[4] = {
8268 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8269 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8270 return DAG.getBitcast(
8272 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8273 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8275 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8276 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8277 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8278 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8280 // If we have a blend of two PACKUS operations an the blend aligns with the
8281 // low and half halves, we can just merge the PACKUS operations. This is
8282 // particularly important as it lets us merge shuffles that this routine itself
8284 auto GetPackNode = [](SDValue V) {
8285 while (V.getOpcode() == ISD::BITCAST)
8286 V = V.getOperand(0);
8288 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8290 if (SDValue V1Pack = GetPackNode(V1))
8291 if (SDValue V2Pack = GetPackNode(V2))
8292 return DAG.getBitcast(MVT::v2i64,
8293 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8294 Mask[0] == 0 ? V1Pack.getOperand(0)
8295 : V1Pack.getOperand(1),
8296 Mask[1] == 2 ? V2Pack.getOperand(0)
8297 : V2Pack.getOperand(1)));
8299 // Try to use shift instructions.
8301 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8304 // When loading a scalar and then shuffling it into a vector we can often do
8305 // the insertion cheaply.
8306 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8307 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8309 // Try inverting the insertion since for v2 masks it is easy to do and we
8310 // can't reliably sort the mask one way or the other.
8311 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8312 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8313 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8316 // We have different paths for blend lowering, but they all must use the
8317 // *exact* same predicate.
8318 bool IsBlendSupported = Subtarget->hasSSE41();
8319 if (IsBlendSupported)
8320 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8324 // Use dedicated unpack instructions for masks that match their pattern.
8326 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8329 // Try to use byte rotation instructions.
8330 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8331 if (Subtarget->hasSSSE3())
8332 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8333 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8336 // If we have direct support for blends, we should lower by decomposing into
8337 // a permute. That will be faster than the domain cross.
8338 if (IsBlendSupported)
8339 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8342 // We implement this with SHUFPD which is pretty lame because it will likely
8343 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8344 // However, all the alternatives are still more cycles and newer chips don't
8345 // have this problem. It would be really nice if x86 had better shuffles here.
8346 V1 = DAG.getBitcast(MVT::v2f64, V1);
8347 V2 = DAG.getBitcast(MVT::v2f64, V2);
8348 return DAG.getBitcast(MVT::v2i64,
8349 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8352 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8354 /// This is used to disable more specialized lowerings when the shufps lowering
8355 /// will happen to be efficient.
8356 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8357 // This routine only handles 128-bit shufps.
8358 assert(Mask.size() == 4 && "Unsupported mask size!");
8360 // To lower with a single SHUFPS we need to have the low half and high half
8361 // each requiring a single input.
8362 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8364 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8370 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8372 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8373 /// It makes no assumptions about whether this is the *best* lowering, it simply
8375 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8376 ArrayRef<int> Mask, SDValue V1,
8377 SDValue V2, SelectionDAG &DAG) {
8378 SDValue LowV = V1, HighV = V2;
8379 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8382 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8384 if (NumV2Elements == 1) {
8386 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8389 // Compute the index adjacent to V2Index and in the same half by toggling
8391 int V2AdjIndex = V2Index ^ 1;
8393 if (Mask[V2AdjIndex] == -1) {
8394 // Handles all the cases where we have a single V2 element and an undef.
8395 // This will only ever happen in the high lanes because we commute the
8396 // vector otherwise.
8398 std::swap(LowV, HighV);
8399 NewMask[V2Index] -= 4;
8401 // Handle the case where the V2 element ends up adjacent to a V1 element.
8402 // To make this work, blend them together as the first step.
8403 int V1Index = V2AdjIndex;
8404 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8405 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8406 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8408 // Now proceed to reconstruct the final blend as we have the necessary
8409 // high or low half formed.
8416 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8417 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8419 } else if (NumV2Elements == 2) {
8420 if (Mask[0] < 4 && Mask[1] < 4) {
8421 // Handle the easy case where we have V1 in the low lanes and V2 in the
8425 } else if (Mask[2] < 4 && Mask[3] < 4) {
8426 // We also handle the reversed case because this utility may get called
8427 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8428 // arrange things in the right direction.
8434 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8435 // trying to place elements directly, just blend them and set up the final
8436 // shuffle to place them.
8438 // The first two blend mask elements are for V1, the second two are for
8440 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8441 Mask[2] < 4 ? Mask[2] : Mask[3],
8442 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8443 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8444 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8445 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8447 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8450 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8451 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8452 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8453 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8456 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8457 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8460 /// \brief Lower 4-lane 32-bit floating point shuffles.
8462 /// Uses instructions exclusively from the floating point unit to minimize
8463 /// domain crossing penalties, as these are sufficient to implement all v4f32
8465 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8466 const X86Subtarget *Subtarget,
8467 SelectionDAG &DAG) {
8469 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8470 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8471 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8472 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8473 ArrayRef<int> Mask = SVOp->getMask();
8474 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8477 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8479 if (NumV2Elements == 0) {
8480 // Check for being able to broadcast a single element.
8481 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8482 Mask, Subtarget, DAG))
8485 // Use even/odd duplicate instructions for masks that match their pattern.
8486 if (Subtarget->hasSSE3()) {
8487 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8488 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8489 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8490 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8493 if (Subtarget->hasAVX()) {
8494 // If we have AVX, we can use VPERMILPS which will allow folding a load
8495 // into the shuffle.
8496 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8497 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8500 // Otherwise, use a straight shuffle of a single input vector. We pass the
8501 // input vector to both operands to simulate this with a SHUFPS.
8502 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8503 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8506 // There are special ways we can lower some single-element blends. However, we
8507 // have custom ways we can lower more complex single-element blends below that
8508 // we defer to if both this and BLENDPS fail to match, so restrict this to
8509 // when the V2 input is targeting element 0 of the mask -- that is the fast
8511 if (NumV2Elements == 1 && Mask[0] >= 4)
8512 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8513 Mask, Subtarget, DAG))
8516 if (Subtarget->hasSSE41()) {
8517 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8521 // Use INSERTPS if we can complete the shuffle efficiently.
8522 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8525 if (!isSingleSHUFPSMask(Mask))
8526 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8527 DL, MVT::v4f32, V1, V2, Mask, DAG))
8531 // Use dedicated unpack instructions for masks that match their pattern.
8533 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8536 // Otherwise fall back to a SHUFPS lowering strategy.
8537 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8540 /// \brief Lower 4-lane i32 vector shuffles.
8542 /// We try to handle these with integer-domain shuffles where we can, but for
8543 /// blends we use the floating point domain blend instructions.
8544 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8545 const X86Subtarget *Subtarget,
8546 SelectionDAG &DAG) {
8548 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8549 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8550 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8551 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8552 ArrayRef<int> Mask = SVOp->getMask();
8553 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8555 // Whenever we can lower this as a zext, that instruction is strictly faster
8556 // than any alternative. It also allows us to fold memory operands into the
8557 // shuffle in many cases.
8558 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8559 Mask, Subtarget, DAG))
8563 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8565 if (NumV2Elements == 0) {
8566 // Check for being able to broadcast a single element.
8567 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8568 Mask, Subtarget, DAG))
8571 // Straight shuffle of a single input vector. For everything from SSE2
8572 // onward this has a single fast instruction with no scary immediates.
8573 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8574 // but we aren't actually going to use the UNPCK instruction because doing
8575 // so prevents folding a load into this instruction or making a copy.
8576 const int UnpackLoMask[] = {0, 0, 1, 1};
8577 const int UnpackHiMask[] = {2, 2, 3, 3};
8578 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8579 Mask = UnpackLoMask;
8580 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8581 Mask = UnpackHiMask;
8583 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8584 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8587 // Try to use shift instructions.
8589 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8592 // There are special ways we can lower some single-element blends.
8593 if (NumV2Elements == 1)
8594 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8595 Mask, Subtarget, DAG))
8598 // We have different paths for blend lowering, but they all must use the
8599 // *exact* same predicate.
8600 bool IsBlendSupported = Subtarget->hasSSE41();
8601 if (IsBlendSupported)
8602 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8606 if (SDValue Masked =
8607 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8610 // Use dedicated unpack instructions for masks that match their pattern.
8612 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8615 // Try to use byte rotation instructions.
8616 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8617 if (Subtarget->hasSSSE3())
8618 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8619 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8622 // If we have direct support for blends, we should lower by decomposing into
8623 // a permute. That will be faster than the domain cross.
8624 if (IsBlendSupported)
8625 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8628 // Try to lower by permuting the inputs into an unpack instruction.
8629 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8633 // We implement this with SHUFPS because it can blend from two vectors.
8634 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8635 // up the inputs, bypassing domain shift penalties that we would encur if we
8636 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8638 return DAG.getBitcast(
8640 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8641 DAG.getBitcast(MVT::v4f32, V2), Mask));
8644 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8645 /// shuffle lowering, and the most complex part.
8647 /// The lowering strategy is to try to form pairs of input lanes which are
8648 /// targeted at the same half of the final vector, and then use a dword shuffle
8649 /// to place them onto the right half, and finally unpack the paired lanes into
8650 /// their final position.
8652 /// The exact breakdown of how to form these dword pairs and align them on the
8653 /// correct sides is really tricky. See the comments within the function for
8654 /// more of the details.
8656 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8657 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8658 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8659 /// vector, form the analogous 128-bit 8-element Mask.
8660 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8661 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8662 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8663 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8664 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8666 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8667 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8668 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8670 SmallVector<int, 4> LoInputs;
8671 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8672 [](int M) { return M >= 0; });
8673 std::sort(LoInputs.begin(), LoInputs.end());
8674 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8675 SmallVector<int, 4> HiInputs;
8676 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8677 [](int M) { return M >= 0; });
8678 std::sort(HiInputs.begin(), HiInputs.end());
8679 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8681 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8682 int NumHToL = LoInputs.size() - NumLToL;
8684 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8685 int NumHToH = HiInputs.size() - NumLToH;
8686 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8687 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8688 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8689 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8691 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8692 // such inputs we can swap two of the dwords across the half mark and end up
8693 // with <=2 inputs to each half in each half. Once there, we can fall through
8694 // to the generic code below. For example:
8696 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8697 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8699 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8700 // and an existing 2-into-2 on the other half. In this case we may have to
8701 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8702 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8703 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8704 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8705 // half than the one we target for fixing) will be fixed when we re-enter this
8706 // path. We will also combine away any sequence of PSHUFD instructions that
8707 // result into a single instruction. Here is an example of the tricky case:
8709 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8710 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8712 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8714 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8715 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8717 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8718 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8720 // The result is fine to be handled by the generic logic.
8721 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8722 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8723 int AOffset, int BOffset) {
8724 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8725 "Must call this with A having 3 or 1 inputs from the A half.");
8726 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8727 "Must call this with B having 1 or 3 inputs from the B half.");
8728 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8729 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8731 bool ThreeAInputs = AToAInputs.size() == 3;
8733 // Compute the index of dword with only one word among the three inputs in
8734 // a half by taking the sum of the half with three inputs and subtracting
8735 // the sum of the actual three inputs. The difference is the remaining
8738 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8739 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8740 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8741 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8742 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8743 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8744 int TripleNonInputIdx =
8745 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8746 TripleDWord = TripleNonInputIdx / 2;
8748 // We use xor with one to compute the adjacent DWord to whichever one the
8750 OneInputDWord = (OneInput / 2) ^ 1;
8752 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8753 // and BToA inputs. If there is also such a problem with the BToB and AToB
8754 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8755 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8756 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8757 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8758 // Compute how many inputs will be flipped by swapping these DWords. We
8760 // to balance this to ensure we don't form a 3-1 shuffle in the other
8762 int NumFlippedAToBInputs =
8763 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8764 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8765 int NumFlippedBToBInputs =
8766 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8767 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8768 if ((NumFlippedAToBInputs == 1 &&
8769 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8770 (NumFlippedBToBInputs == 1 &&
8771 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8772 // We choose whether to fix the A half or B half based on whether that
8773 // half has zero flipped inputs. At zero, we may not be able to fix it
8774 // with that half. We also bias towards fixing the B half because that
8775 // will more commonly be the high half, and we have to bias one way.
8776 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8777 ArrayRef<int> Inputs) {
8778 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8779 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8780 PinnedIdx ^ 1) != Inputs.end();
8781 // Determine whether the free index is in the flipped dword or the
8782 // unflipped dword based on where the pinned index is. We use this bit
8783 // in an xor to conditionally select the adjacent dword.
8784 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8785 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8786 FixFreeIdx) != Inputs.end();
8787 if (IsFixIdxInput == IsFixFreeIdxInput)
8789 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8790 FixFreeIdx) != Inputs.end();
8791 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8792 "We need to be changing the number of flipped inputs!");
8793 int PSHUFHalfMask[] = {0, 1, 2, 3};
8794 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8795 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8797 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8800 if (M != -1 && M == FixIdx)
8802 else if (M != -1 && M == FixFreeIdx)
8805 if (NumFlippedBToBInputs != 0) {
8807 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8808 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8810 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8811 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8812 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8817 int PSHUFDMask[] = {0, 1, 2, 3};
8818 PSHUFDMask[ADWord] = BDWord;
8819 PSHUFDMask[BDWord] = ADWord;
8822 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8823 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8825 // Adjust the mask to match the new locations of A and B.
8827 if (M != -1 && M/2 == ADWord)
8828 M = 2 * BDWord + M % 2;
8829 else if (M != -1 && M/2 == BDWord)
8830 M = 2 * ADWord + M % 2;
8832 // Recurse back into this routine to re-compute state now that this isn't
8833 // a 3 and 1 problem.
8834 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8837 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8838 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8839 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8840 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8842 // At this point there are at most two inputs to the low and high halves from
8843 // each half. That means the inputs can always be grouped into dwords and
8844 // those dwords can then be moved to the correct half with a dword shuffle.
8845 // We use at most one low and one high word shuffle to collect these paired
8846 // inputs into dwords, and finally a dword shuffle to place them.
8847 int PSHUFLMask[4] = {-1, -1, -1, -1};
8848 int PSHUFHMask[4] = {-1, -1, -1, -1};
8849 int PSHUFDMask[4] = {-1, -1, -1, -1};
8851 // First fix the masks for all the inputs that are staying in their
8852 // original halves. This will then dictate the targets of the cross-half
8854 auto fixInPlaceInputs =
8855 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8856 MutableArrayRef<int> SourceHalfMask,
8857 MutableArrayRef<int> HalfMask, int HalfOffset) {
8858 if (InPlaceInputs.empty())
8860 if (InPlaceInputs.size() == 1) {
8861 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8862 InPlaceInputs[0] - HalfOffset;
8863 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8866 if (IncomingInputs.empty()) {
8867 // Just fix all of the in place inputs.
8868 for (int Input : InPlaceInputs) {
8869 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8870 PSHUFDMask[Input / 2] = Input / 2;
8875 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8876 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8877 InPlaceInputs[0] - HalfOffset;
8878 // Put the second input next to the first so that they are packed into
8879 // a dword. We find the adjacent index by toggling the low bit.
8880 int AdjIndex = InPlaceInputs[0] ^ 1;
8881 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8882 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8883 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8885 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8886 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8888 // Now gather the cross-half inputs and place them into a free dword of
8889 // their target half.
8890 // FIXME: This operation could almost certainly be simplified dramatically to
8891 // look more like the 3-1 fixing operation.
8892 auto moveInputsToRightHalf = [&PSHUFDMask](
8893 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8894 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8895 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8897 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8898 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8900 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8902 int LowWord = Word & ~1;
8903 int HighWord = Word | 1;
8904 return isWordClobbered(SourceHalfMask, LowWord) ||
8905 isWordClobbered(SourceHalfMask, HighWord);
8908 if (IncomingInputs.empty())
8911 if (ExistingInputs.empty()) {
8912 // Map any dwords with inputs from them into the right half.
8913 for (int Input : IncomingInputs) {
8914 // If the source half mask maps over the inputs, turn those into
8915 // swaps and use the swapped lane.
8916 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8917 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8918 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8919 Input - SourceOffset;
8920 // We have to swap the uses in our half mask in one sweep.
8921 for (int &M : HalfMask)
8922 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8924 else if (M == Input)
8925 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8927 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8928 Input - SourceOffset &&
8929 "Previous placement doesn't match!");
8931 // Note that this correctly re-maps both when we do a swap and when
8932 // we observe the other side of the swap above. We rely on that to
8933 // avoid swapping the members of the input list directly.
8934 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8937 // Map the input's dword into the correct half.
8938 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8939 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8941 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8943 "Previous placement doesn't match!");
8946 // And just directly shift any other-half mask elements to be same-half
8947 // as we will have mirrored the dword containing the element into the
8948 // same position within that half.
8949 for (int &M : HalfMask)
8950 if (M >= SourceOffset && M < SourceOffset + 4) {
8951 M = M - SourceOffset + DestOffset;
8952 assert(M >= 0 && "This should never wrap below zero!");
8957 // Ensure we have the input in a viable dword of its current half. This
8958 // is particularly tricky because the original position may be clobbered
8959 // by inputs being moved and *staying* in that half.
8960 if (IncomingInputs.size() == 1) {
8961 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8962 int InputFixed = std::find(std::begin(SourceHalfMask),
8963 std::end(SourceHalfMask), -1) -
8964 std::begin(SourceHalfMask) + SourceOffset;
8965 SourceHalfMask[InputFixed - SourceOffset] =
8966 IncomingInputs[0] - SourceOffset;
8967 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8969 IncomingInputs[0] = InputFixed;
8971 } else if (IncomingInputs.size() == 2) {
8972 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8973 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8974 // We have two non-adjacent or clobbered inputs we need to extract from
8975 // the source half. To do this, we need to map them into some adjacent
8976 // dword slot in the source mask.
8977 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8978 IncomingInputs[1] - SourceOffset};
8980 // If there is a free slot in the source half mask adjacent to one of
8981 // the inputs, place the other input in it. We use (Index XOR 1) to
8982 // compute an adjacent index.
8983 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8984 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8985 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8986 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8987 InputsFixed[1] = InputsFixed[0] ^ 1;
8988 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8989 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8990 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8991 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8992 InputsFixed[0] = InputsFixed[1] ^ 1;
8993 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8994 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8995 // The two inputs are in the same DWord but it is clobbered and the
8996 // adjacent DWord isn't used at all. Move both inputs to the free
8998 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8999 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9000 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9001 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9003 // The only way we hit this point is if there is no clobbering
9004 // (because there are no off-half inputs to this half) and there is no
9005 // free slot adjacent to one of the inputs. In this case, we have to
9006 // swap an input with a non-input.
9007 for (int i = 0; i < 4; ++i)
9008 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9009 "We can't handle any clobbers here!");
9010 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9011 "Cannot have adjacent inputs here!");
9013 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9014 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9016 // We also have to update the final source mask in this case because
9017 // it may need to undo the above swap.
9018 for (int &M : FinalSourceHalfMask)
9019 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9020 M = InputsFixed[1] + SourceOffset;
9021 else if (M == InputsFixed[1] + SourceOffset)
9022 M = (InputsFixed[0] ^ 1) + SourceOffset;
9024 InputsFixed[1] = InputsFixed[0] ^ 1;
9027 // Point everything at the fixed inputs.
9028 for (int &M : HalfMask)
9029 if (M == IncomingInputs[0])
9030 M = InputsFixed[0] + SourceOffset;
9031 else if (M == IncomingInputs[1])
9032 M = InputsFixed[1] + SourceOffset;
9034 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9035 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9038 llvm_unreachable("Unhandled input size!");
9041 // Now hoist the DWord down to the right half.
9042 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9043 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9044 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9045 for (int &M : HalfMask)
9046 for (int Input : IncomingInputs)
9048 M = FreeDWord * 2 + Input % 2;
9050 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9051 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9052 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9053 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9055 // Now enact all the shuffles we've computed to move the inputs into their
9057 if (!isNoopShuffleMask(PSHUFLMask))
9058 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9059 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9060 if (!isNoopShuffleMask(PSHUFHMask))
9061 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9062 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9063 if (!isNoopShuffleMask(PSHUFDMask))
9066 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9067 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9069 // At this point, each half should contain all its inputs, and we can then
9070 // just shuffle them into their final position.
9071 assert(std::count_if(LoMask.begin(), LoMask.end(),
9072 [](int M) { return M >= 4; }) == 0 &&
9073 "Failed to lift all the high half inputs to the low mask!");
9074 assert(std::count_if(HiMask.begin(), HiMask.end(),
9075 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9076 "Failed to lift all the low half inputs to the high mask!");
9078 // Do a half shuffle for the low mask.
9079 if (!isNoopShuffleMask(LoMask))
9080 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9081 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9083 // Do a half shuffle with the high mask after shifting its values down.
9084 for (int &M : HiMask)
9087 if (!isNoopShuffleMask(HiMask))
9088 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9089 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9094 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9095 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9096 SDValue V2, ArrayRef<int> Mask,
9097 SelectionDAG &DAG, bool &V1InUse,
9099 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9105 int Size = Mask.size();
9106 int Scale = 16 / Size;
9107 for (int i = 0; i < 16; ++i) {
9108 if (Mask[i / Scale] == -1) {
9109 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9111 const int ZeroMask = 0x80;
9112 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9114 int V2Idx = Mask[i / Scale] < Size
9116 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9117 if (Zeroable[i / Scale])
9118 V1Idx = V2Idx = ZeroMask;
9119 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9120 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9121 V1InUse |= (ZeroMask != V1Idx);
9122 V2InUse |= (ZeroMask != V2Idx);
9127 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9128 DAG.getBitcast(MVT::v16i8, V1),
9129 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9131 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9132 DAG.getBitcast(MVT::v16i8, V2),
9133 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9135 // If we need shuffled inputs from both, blend the two.
9137 if (V1InUse && V2InUse)
9138 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9140 V = V1InUse ? V1 : V2;
9142 // Cast the result back to the correct type.
9143 return DAG.getBitcast(VT, V);
9146 /// \brief Generic lowering of 8-lane i16 shuffles.
9148 /// This handles both single-input shuffles and combined shuffle/blends with
9149 /// two inputs. The single input shuffles are immediately delegated to
9150 /// a dedicated lowering routine.
9152 /// The blends are lowered in one of three fundamental ways. If there are few
9153 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9154 /// of the input is significantly cheaper when lowered as an interleaving of
9155 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9156 /// halves of the inputs separately (making them have relatively few inputs)
9157 /// and then concatenate them.
9158 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9159 const X86Subtarget *Subtarget,
9160 SelectionDAG &DAG) {
9162 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9163 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9164 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9165 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9166 ArrayRef<int> OrigMask = SVOp->getMask();
9167 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9168 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9169 MutableArrayRef<int> Mask(MaskStorage);
9171 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9173 // Whenever we can lower this as a zext, that instruction is strictly faster
9174 // than any alternative.
9175 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9176 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9179 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9181 auto isV2 = [](int M) { return M >= 8; };
9183 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9185 if (NumV2Inputs == 0) {
9186 // Check for being able to broadcast a single element.
9187 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9188 Mask, Subtarget, DAG))
9191 // Try to use shift instructions.
9193 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9196 // Use dedicated unpack instructions for masks that match their pattern.
9198 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9201 // Try to use byte rotation instructions.
9202 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9203 Mask, Subtarget, DAG))
9206 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9210 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9211 "All single-input shuffles should be canonicalized to be V1-input "
9214 // Try to use shift instructions.
9216 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9219 // See if we can use SSE4A Extraction / Insertion.
9220 if (Subtarget->hasSSE4A())
9221 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9224 // There are special ways we can lower some single-element blends.
9225 if (NumV2Inputs == 1)
9226 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9227 Mask, Subtarget, DAG))
9230 // We have different paths for blend lowering, but they all must use the
9231 // *exact* same predicate.
9232 bool IsBlendSupported = Subtarget->hasSSE41();
9233 if (IsBlendSupported)
9234 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9238 if (SDValue Masked =
9239 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9242 // Use dedicated unpack instructions for masks that match their pattern.
9244 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9247 // Try to use byte rotation instructions.
9248 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9249 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9252 if (SDValue BitBlend =
9253 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9256 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9260 // If we can't directly blend but can use PSHUFB, that will be better as it
9261 // can both shuffle and set up the inefficient blend.
9262 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9263 bool V1InUse, V2InUse;
9264 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9268 // We can always bit-blend if we have to so the fallback strategy is to
9269 // decompose into single-input permutes and blends.
9270 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9274 /// \brief Check whether a compaction lowering can be done by dropping even
9275 /// elements and compute how many times even elements must be dropped.
9277 /// This handles shuffles which take every Nth element where N is a power of
9278 /// two. Example shuffle masks:
9280 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9281 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9282 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9283 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9284 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9285 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9287 /// Any of these lanes can of course be undef.
9289 /// This routine only supports N <= 3.
9290 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9293 /// \returns N above, or the number of times even elements must be dropped if
9294 /// there is such a number. Otherwise returns zero.
9295 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9296 // Figure out whether we're looping over two inputs or just one.
9297 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9299 // The modulus for the shuffle vector entries is based on whether this is
9300 // a single input or not.
9301 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9302 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9303 "We should only be called with masks with a power-of-2 size!");
9305 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9307 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9308 // and 2^3 simultaneously. This is because we may have ambiguity with
9309 // partially undef inputs.
9310 bool ViableForN[3] = {true, true, true};
9312 for (int i = 0, e = Mask.size(); i < e; ++i) {
9313 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9318 bool IsAnyViable = false;
9319 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9320 if (ViableForN[j]) {
9323 // The shuffle mask must be equal to (i * 2^N) % M.
9324 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9327 ViableForN[j] = false;
9329 // Early exit if we exhaust the possible powers of two.
9334 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9338 // Return 0 as there is no viable power of two.
9342 /// \brief Generic lowering of v16i8 shuffles.
9344 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9345 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9346 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9347 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9349 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9350 const X86Subtarget *Subtarget,
9351 SelectionDAG &DAG) {
9353 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9354 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9355 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9356 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9357 ArrayRef<int> Mask = SVOp->getMask();
9358 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9360 // Try to use shift instructions.
9362 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9365 // Try to use byte rotation instructions.
9366 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9367 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9370 // Try to use a zext lowering.
9371 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9372 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9375 // See if we can use SSE4A Extraction / Insertion.
9376 if (Subtarget->hasSSE4A())
9377 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9381 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9383 // For single-input shuffles, there are some nicer lowering tricks we can use.
9384 if (NumV2Elements == 0) {
9385 // Check for being able to broadcast a single element.
9386 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9387 Mask, Subtarget, DAG))
9390 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9391 // Notably, this handles splat and partial-splat shuffles more efficiently.
9392 // However, it only makes sense if the pre-duplication shuffle simplifies
9393 // things significantly. Currently, this means we need to be able to
9394 // express the pre-duplication shuffle as an i16 shuffle.
9396 // FIXME: We should check for other patterns which can be widened into an
9397 // i16 shuffle as well.
9398 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9399 for (int i = 0; i < 16; i += 2)
9400 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9405 auto tryToWidenViaDuplication = [&]() -> SDValue {
9406 if (!canWidenViaDuplication(Mask))
9408 SmallVector<int, 4> LoInputs;
9409 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9410 [](int M) { return M >= 0 && M < 8; });
9411 std::sort(LoInputs.begin(), LoInputs.end());
9412 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9414 SmallVector<int, 4> HiInputs;
9415 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9416 [](int M) { return M >= 8; });
9417 std::sort(HiInputs.begin(), HiInputs.end());
9418 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9421 bool TargetLo = LoInputs.size() >= HiInputs.size();
9422 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9423 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9425 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9426 SmallDenseMap<int, int, 8> LaneMap;
9427 for (int I : InPlaceInputs) {
9428 PreDupI16Shuffle[I/2] = I/2;
9431 int j = TargetLo ? 0 : 4, je = j + 4;
9432 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9433 // Check if j is already a shuffle of this input. This happens when
9434 // there are two adjacent bytes after we move the low one.
9435 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9436 // If we haven't yet mapped the input, search for a slot into which
9438 while (j < je && PreDupI16Shuffle[j] != -1)
9442 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9445 // Map this input with the i16 shuffle.
9446 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9449 // Update the lane map based on the mapping we ended up with.
9450 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9452 V1 = DAG.getBitcast(
9454 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9455 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9457 // Unpack the bytes to form the i16s that will be shuffled into place.
9458 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9459 MVT::v16i8, V1, V1);
9461 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9462 for (int i = 0; i < 16; ++i)
9463 if (Mask[i] != -1) {
9464 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9465 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9466 if (PostDupI16Shuffle[i / 2] == -1)
9467 PostDupI16Shuffle[i / 2] = MappedMask;
9469 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9470 "Conflicting entrties in the original shuffle!");
9472 return DAG.getBitcast(
9474 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9475 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9477 if (SDValue V = tryToWidenViaDuplication())
9481 if (SDValue Masked =
9482 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9485 // Use dedicated unpack instructions for masks that match their pattern.
9487 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9490 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9491 // with PSHUFB. It is important to do this before we attempt to generate any
9492 // blends but after all of the single-input lowerings. If the single input
9493 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9494 // want to preserve that and we can DAG combine any longer sequences into
9495 // a PSHUFB in the end. But once we start blending from multiple inputs,
9496 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9497 // and there are *very* few patterns that would actually be faster than the
9498 // PSHUFB approach because of its ability to zero lanes.
9500 // FIXME: The only exceptions to the above are blends which are exact
9501 // interleavings with direct instructions supporting them. We currently don't
9502 // handle those well here.
9503 if (Subtarget->hasSSSE3()) {
9504 bool V1InUse = false;
9505 bool V2InUse = false;
9507 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9508 DAG, V1InUse, V2InUse);
9510 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9511 // do so. This avoids using them to handle blends-with-zero which is
9512 // important as a single pshufb is significantly faster for that.
9513 if (V1InUse && V2InUse) {
9514 if (Subtarget->hasSSE41())
9515 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9516 Mask, Subtarget, DAG))
9519 // We can use an unpack to do the blending rather than an or in some
9520 // cases. Even though the or may be (very minorly) more efficient, we
9521 // preference this lowering because there are common cases where part of
9522 // the complexity of the shuffles goes away when we do the final blend as
9524 // FIXME: It might be worth trying to detect if the unpack-feeding
9525 // shuffles will both be pshufb, in which case we shouldn't bother with
9527 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9528 DL, MVT::v16i8, V1, V2, Mask, DAG))
9535 // There are special ways we can lower some single-element blends.
9536 if (NumV2Elements == 1)
9537 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9538 Mask, Subtarget, DAG))
9541 if (SDValue BitBlend =
9542 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9545 // Check whether a compaction lowering can be done. This handles shuffles
9546 // which take every Nth element for some even N. See the helper function for
9549 // We special case these as they can be particularly efficiently handled with
9550 // the PACKUSB instruction on x86 and they show up in common patterns of
9551 // rearranging bytes to truncate wide elements.
9552 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9553 // NumEvenDrops is the power of two stride of the elements. Another way of
9554 // thinking about it is that we need to drop the even elements this many
9555 // times to get the original input.
9556 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9558 // First we need to zero all the dropped bytes.
9559 assert(NumEvenDrops <= 3 &&
9560 "No support for dropping even elements more than 3 times.");
9561 // We use the mask type to pick which bytes are preserved based on how many
9562 // elements are dropped.
9563 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9564 SDValue ByteClearMask = DAG.getBitcast(
9565 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9566 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9568 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9570 // Now pack things back together.
9571 V1 = DAG.getBitcast(MVT::v8i16, V1);
9572 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9573 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9574 for (int i = 1; i < NumEvenDrops; ++i) {
9575 Result = DAG.getBitcast(MVT::v8i16, Result);
9576 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9582 // Handle multi-input cases by blending single-input shuffles.
9583 if (NumV2Elements > 0)
9584 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9587 // The fallback path for single-input shuffles widens this into two v8i16
9588 // vectors with unpacks, shuffles those, and then pulls them back together
9592 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9593 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9594 for (int i = 0; i < 16; ++i)
9596 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9598 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9600 SDValue VLoHalf, VHiHalf;
9601 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9602 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9604 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9605 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9606 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9607 [](int M) { return M >= 0 && M % 2 == 1; })) {
9608 // Use a mask to drop the high bytes.
9609 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9610 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9611 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9613 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9614 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9616 // Squash the masks to point directly into VLoHalf.
9617 for (int &M : LoBlendMask)
9620 for (int &M : HiBlendMask)
9624 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9625 // VHiHalf so that we can blend them as i16s.
9626 VLoHalf = DAG.getBitcast(
9627 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9628 VHiHalf = DAG.getBitcast(
9629 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9632 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9633 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9635 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9638 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9640 /// This routine breaks down the specific type of 128-bit shuffle and
9641 /// dispatches to the lowering routines accordingly.
9642 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9643 MVT VT, const X86Subtarget *Subtarget,
9644 SelectionDAG &DAG) {
9645 switch (VT.SimpleTy) {
9647 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9649 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9651 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9653 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9655 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9657 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9660 llvm_unreachable("Unimplemented!");
9664 /// \brief Helper function to test whether a shuffle mask could be
9665 /// simplified by widening the elements being shuffled.
9667 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9668 /// leaves it in an unspecified state.
9670 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9671 /// shuffle masks. The latter have the special property of a '-2' representing
9672 /// a zero-ed lane of a vector.
9673 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9674 SmallVectorImpl<int> &WidenedMask) {
9675 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9676 // If both elements are undef, its trivial.
9677 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9678 WidenedMask.push_back(SM_SentinelUndef);
9682 // Check for an undef mask and a mask value properly aligned to fit with
9683 // a pair of values. If we find such a case, use the non-undef mask's value.
9684 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9685 WidenedMask.push_back(Mask[i + 1] / 2);
9688 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9689 WidenedMask.push_back(Mask[i] / 2);
9693 // When zeroing, we need to spread the zeroing across both lanes to widen.
9694 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9695 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9696 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9697 WidenedMask.push_back(SM_SentinelZero);
9703 // Finally check if the two mask values are adjacent and aligned with
9705 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9706 WidenedMask.push_back(Mask[i] / 2);
9710 // Otherwise we can't safely widen the elements used in this shuffle.
9713 assert(WidenedMask.size() == Mask.size() / 2 &&
9714 "Incorrect size of mask after widening the elements!");
9719 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9721 /// This routine just extracts two subvectors, shuffles them independently, and
9722 /// then concatenates them back together. This should work effectively with all
9723 /// AVX vector shuffle types.
9724 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9725 SDValue V2, ArrayRef<int> Mask,
9726 SelectionDAG &DAG) {
9727 assert(VT.getSizeInBits() >= 256 &&
9728 "Only for 256-bit or wider vector shuffles!");
9729 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9730 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9732 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9733 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9735 int NumElements = VT.getVectorNumElements();
9736 int SplitNumElements = NumElements / 2;
9737 MVT ScalarVT = VT.getScalarType();
9738 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9740 // Rather than splitting build-vectors, just build two narrower build
9741 // vectors. This helps shuffling with splats and zeros.
9742 auto SplitVector = [&](SDValue V) {
9743 while (V.getOpcode() == ISD::BITCAST)
9744 V = V->getOperand(0);
9746 MVT OrigVT = V.getSimpleValueType();
9747 int OrigNumElements = OrigVT.getVectorNumElements();
9748 int OrigSplitNumElements = OrigNumElements / 2;
9749 MVT OrigScalarVT = OrigVT.getScalarType();
9750 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9754 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9756 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9757 DAG.getIntPtrConstant(0, DL));
9758 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9759 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9762 SmallVector<SDValue, 16> LoOps, HiOps;
9763 for (int i = 0; i < OrigSplitNumElements; ++i) {
9764 LoOps.push_back(BV->getOperand(i));
9765 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9767 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9768 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9770 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9771 DAG.getBitcast(SplitVT, HiV));
9774 SDValue LoV1, HiV1, LoV2, HiV2;
9775 std::tie(LoV1, HiV1) = SplitVector(V1);
9776 std::tie(LoV2, HiV2) = SplitVector(V2);
9778 // Now create two 4-way blends of these half-width vectors.
9779 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9780 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9781 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9782 for (int i = 0; i < SplitNumElements; ++i) {
9783 int M = HalfMask[i];
9784 if (M >= NumElements) {
9785 if (M >= NumElements + SplitNumElements)
9789 V2BlendMask.push_back(M - NumElements);
9790 V1BlendMask.push_back(-1);
9791 BlendMask.push_back(SplitNumElements + i);
9792 } else if (M >= 0) {
9793 if (M >= SplitNumElements)
9797 V2BlendMask.push_back(-1);
9798 V1BlendMask.push_back(M);
9799 BlendMask.push_back(i);
9801 V2BlendMask.push_back(-1);
9802 V1BlendMask.push_back(-1);
9803 BlendMask.push_back(-1);
9807 // Because the lowering happens after all combining takes place, we need to
9808 // manually combine these blend masks as much as possible so that we create
9809 // a minimal number of high-level vector shuffle nodes.
9811 // First try just blending the halves of V1 or V2.
9812 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9813 return DAG.getUNDEF(SplitVT);
9814 if (!UseLoV2 && !UseHiV2)
9815 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9816 if (!UseLoV1 && !UseHiV1)
9817 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9819 SDValue V1Blend, V2Blend;
9820 if (UseLoV1 && UseHiV1) {
9822 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9824 // We only use half of V1 so map the usage down into the final blend mask.
9825 V1Blend = UseLoV1 ? LoV1 : HiV1;
9826 for (int i = 0; i < SplitNumElements; ++i)
9827 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9828 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9830 if (UseLoV2 && UseHiV2) {
9832 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9834 // We only use half of V2 so map the usage down into the final blend mask.
9835 V2Blend = UseLoV2 ? LoV2 : HiV2;
9836 for (int i = 0; i < SplitNumElements; ++i)
9837 if (BlendMask[i] >= SplitNumElements)
9838 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9840 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9842 SDValue Lo = HalfBlend(LoMask);
9843 SDValue Hi = HalfBlend(HiMask);
9844 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9847 /// \brief Either split a vector in halves or decompose the shuffles and the
9850 /// This is provided as a good fallback for many lowerings of non-single-input
9851 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9852 /// between splitting the shuffle into 128-bit components and stitching those
9853 /// back together vs. extracting the single-input shuffles and blending those
9855 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9856 SDValue V2, ArrayRef<int> Mask,
9857 SelectionDAG &DAG) {
9858 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9859 "lower single-input shuffles as it "
9860 "could then recurse on itself.");
9861 int Size = Mask.size();
9863 // If this can be modeled as a broadcast of two elements followed by a blend,
9864 // prefer that lowering. This is especially important because broadcasts can
9865 // often fold with memory operands.
9866 auto DoBothBroadcast = [&] {
9867 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9870 if (V2BroadcastIdx == -1)
9871 V2BroadcastIdx = M - Size;
9872 else if (M - Size != V2BroadcastIdx)
9874 } else if (M >= 0) {
9875 if (V1BroadcastIdx == -1)
9877 else if (M != V1BroadcastIdx)
9882 if (DoBothBroadcast())
9883 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9886 // If the inputs all stem from a single 128-bit lane of each input, then we
9887 // split them rather than blending because the split will decompose to
9888 // unusually few instructions.
9889 int LaneCount = VT.getSizeInBits() / 128;
9890 int LaneSize = Size / LaneCount;
9891 SmallBitVector LaneInputs[2];
9892 LaneInputs[0].resize(LaneCount, false);
9893 LaneInputs[1].resize(LaneCount, false);
9894 for (int i = 0; i < Size; ++i)
9896 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9897 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9898 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9900 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9901 // that the decomposed single-input shuffles don't end up here.
9902 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9905 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9906 /// a permutation and blend of those lanes.
9908 /// This essentially blends the out-of-lane inputs to each lane into the lane
9909 /// from a permuted copy of the vector. This lowering strategy results in four
9910 /// instructions in the worst case for a single-input cross lane shuffle which
9911 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9912 /// of. Special cases for each particular shuffle pattern should be handled
9913 /// prior to trying this lowering.
9914 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9915 SDValue V1, SDValue V2,
9917 SelectionDAG &DAG) {
9918 // FIXME: This should probably be generalized for 512-bit vectors as well.
9919 assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
9920 int LaneSize = Mask.size() / 2;
9922 // If there are only inputs from one 128-bit lane, splitting will in fact be
9923 // less expensive. The flags track whether the given lane contains an element
9924 // that crosses to another lane.
9925 bool LaneCrossing[2] = {false, false};
9926 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9927 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9928 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9929 if (!LaneCrossing[0] || !LaneCrossing[1])
9930 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9932 if (isSingleInputShuffleMask(Mask)) {
9933 SmallVector<int, 32> FlippedBlendMask;
9934 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9935 FlippedBlendMask.push_back(
9936 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9938 : Mask[i] % LaneSize +
9939 (i / LaneSize) * LaneSize + Size));
9941 // Flip the vector, and blend the results which should now be in-lane. The
9942 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9943 // 5 for the high source. The value 3 selects the high half of source 2 and
9944 // the value 2 selects the low half of source 2. We only use source 2 to
9945 // allow folding it into a memory operand.
9946 unsigned PERMMask = 3 | 2 << 4;
9947 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9948 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9949 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9952 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9953 // will be handled by the above logic and a blend of the results, much like
9954 // other patterns in AVX.
9955 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9958 /// \brief Handle lowering 2-lane 128-bit shuffles.
9959 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9960 SDValue V2, ArrayRef<int> Mask,
9961 const X86Subtarget *Subtarget,
9962 SelectionDAG &DAG) {
9963 // TODO: If minimizing size and one of the inputs is a zero vector and the
9964 // the zero vector has only one use, we could use a VPERM2X128 to save the
9965 // instruction bytes needed to explicitly generate the zero vector.
9967 // Blends are faster and handle all the non-lane-crossing cases.
9968 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9972 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9973 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9975 // If either input operand is a zero vector, use VPERM2X128 because its mask
9976 // allows us to replace the zero input with an implicit zero.
9977 if (!IsV1Zero && !IsV2Zero) {
9978 // Check for patterns which can be matched with a single insert of a 128-bit
9980 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9981 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9982 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9983 VT.getVectorNumElements() / 2);
9984 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9985 DAG.getIntPtrConstant(0, DL));
9986 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9987 OnlyUsesV1 ? V1 : V2,
9988 DAG.getIntPtrConstant(0, DL));
9989 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9993 // Otherwise form a 128-bit permutation. After accounting for undefs,
9994 // convert the 64-bit shuffle mask selection values into 128-bit
9995 // selection bits by dividing the indexes by 2 and shifting into positions
9996 // defined by a vperm2*128 instruction's immediate control byte.
9998 // The immediate permute control byte looks like this:
9999 // [1:0] - select 128 bits from sources for low half of destination
10001 // [3] - zero low half of destination
10002 // [5:4] - select 128 bits from sources for high half of destination
10004 // [7] - zero high half of destination
10006 int MaskLO = Mask[0];
10007 if (MaskLO == SM_SentinelUndef)
10008 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10010 int MaskHI = Mask[2];
10011 if (MaskHI == SM_SentinelUndef)
10012 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10014 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10016 // If either input is a zero vector, replace it with an undef input.
10017 // Shuffle mask values < 4 are selecting elements of V1.
10018 // Shuffle mask values >= 4 are selecting elements of V2.
10019 // Adjust each half of the permute mask by clearing the half that was
10020 // selecting the zero vector and setting the zero mask bit.
10022 V1 = DAG.getUNDEF(VT);
10024 PermMask = (PermMask & 0xf0) | 0x08;
10026 PermMask = (PermMask & 0x0f) | 0x80;
10029 V2 = DAG.getUNDEF(VT);
10031 PermMask = (PermMask & 0xf0) | 0x08;
10033 PermMask = (PermMask & 0x0f) | 0x80;
10036 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10037 DAG.getConstant(PermMask, DL, MVT::i8));
10040 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10041 /// shuffling each lane.
10043 /// This will only succeed when the result of fixing the 128-bit lanes results
10044 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10045 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10046 /// the lane crosses early and then use simpler shuffles within each lane.
10048 /// FIXME: It might be worthwhile at some point to support this without
10049 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10050 /// in x86 only floating point has interesting non-repeating shuffles, and even
10051 /// those are still *marginally* more expensive.
10052 static SDValue lowerVectorShuffleByMerging128BitLanes(
10053 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10054 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10055 assert(!isSingleInputShuffleMask(Mask) &&
10056 "This is only useful with multiple inputs.");
10058 int Size = Mask.size();
10059 int LaneSize = 128 / VT.getScalarSizeInBits();
10060 int NumLanes = Size / LaneSize;
10061 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10063 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10064 // check whether the in-128-bit lane shuffles share a repeating pattern.
10065 SmallVector<int, 4> Lanes;
10066 Lanes.resize(NumLanes, -1);
10067 SmallVector<int, 4> InLaneMask;
10068 InLaneMask.resize(LaneSize, -1);
10069 for (int i = 0; i < Size; ++i) {
10073 int j = i / LaneSize;
10075 if (Lanes[j] < 0) {
10076 // First entry we've seen for this lane.
10077 Lanes[j] = Mask[i] / LaneSize;
10078 } else if (Lanes[j] != Mask[i] / LaneSize) {
10079 // This doesn't match the lane selected previously!
10083 // Check that within each lane we have a consistent shuffle mask.
10084 int k = i % LaneSize;
10085 if (InLaneMask[k] < 0) {
10086 InLaneMask[k] = Mask[i] % LaneSize;
10087 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10088 // This doesn't fit a repeating in-lane mask.
10093 // First shuffle the lanes into place.
10094 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10095 VT.getSizeInBits() / 64);
10096 SmallVector<int, 8> LaneMask;
10097 LaneMask.resize(NumLanes * 2, -1);
10098 for (int i = 0; i < NumLanes; ++i)
10099 if (Lanes[i] >= 0) {
10100 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10101 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10104 V1 = DAG.getBitcast(LaneVT, V1);
10105 V2 = DAG.getBitcast(LaneVT, V2);
10106 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10108 // Cast it back to the type we actually want.
10109 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10111 // Now do a simple shuffle that isn't lane crossing.
10112 SmallVector<int, 8> NewMask;
10113 NewMask.resize(Size, -1);
10114 for (int i = 0; i < Size; ++i)
10116 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10117 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10118 "Must not introduce lane crosses at this point!");
10120 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10123 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10126 /// This returns true if the elements from a particular input are already in the
10127 /// slot required by the given mask and require no permutation.
10128 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10129 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10130 int Size = Mask.size();
10131 for (int i = 0; i < Size; ++i)
10132 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10138 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10139 ArrayRef<int> Mask, SDValue V1,
10140 SDValue V2, SelectionDAG &DAG) {
10142 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10143 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10144 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10145 int NumElts = VT.getVectorNumElements();
10146 bool ShufpdMask = true;
10147 bool CommutableMask = true;
10148 unsigned Immediate = 0;
10149 for (int i = 0; i < NumElts; ++i) {
10152 int Val = (i & 6) + NumElts * (i & 1);
10153 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10154 if (Mask[i] < Val || Mask[i] > Val + 1)
10155 ShufpdMask = false;
10156 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10157 CommutableMask = false;
10158 Immediate |= (Mask[i] % 2) << i;
10161 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10162 DAG.getConstant(Immediate, DL, MVT::i8));
10163 if (CommutableMask)
10164 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10165 DAG.getConstant(Immediate, DL, MVT::i8));
10169 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10171 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10172 /// isn't available.
10173 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10174 const X86Subtarget *Subtarget,
10175 SelectionDAG &DAG) {
10177 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10178 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10179 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10180 ArrayRef<int> Mask = SVOp->getMask();
10181 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10183 SmallVector<int, 4> WidenedMask;
10184 if (canWidenShuffleElements(Mask, WidenedMask))
10185 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10188 if (isSingleInputShuffleMask(Mask)) {
10189 // Check for being able to broadcast a single element.
10190 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10191 Mask, Subtarget, DAG))
10194 // Use low duplicate instructions for masks that match their pattern.
10195 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10196 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10198 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10199 // Non-half-crossing single input shuffles can be lowerid with an
10200 // interleaved permutation.
10201 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10202 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10203 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10204 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10207 // With AVX2 we have direct support for this permutation.
10208 if (Subtarget->hasAVX2())
10209 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10210 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10212 // Otherwise, fall back.
10213 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10217 // Use dedicated unpack instructions for masks that match their pattern.
10219 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10222 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10226 // Check if the blend happens to exactly fit that of SHUFPD.
10228 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10231 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10232 // shuffle. However, if we have AVX2 and either inputs are already in place,
10233 // we will be able to shuffle even across lanes the other input in a single
10234 // instruction so skip this pattern.
10235 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10236 isShuffleMaskInputInPlace(1, Mask))))
10237 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10238 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10241 // If we have AVX2 then we always want to lower with a blend because an v4 we
10242 // can fully permute the elements.
10243 if (Subtarget->hasAVX2())
10244 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10247 // Otherwise fall back on generic lowering.
10248 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10251 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10253 /// This routine is only called when we have AVX2 and thus a reasonable
10254 /// instruction set for v4i64 shuffling..
10255 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10256 const X86Subtarget *Subtarget,
10257 SelectionDAG &DAG) {
10259 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10260 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10261 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10262 ArrayRef<int> Mask = SVOp->getMask();
10263 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10264 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10266 SmallVector<int, 4> WidenedMask;
10267 if (canWidenShuffleElements(Mask, WidenedMask))
10268 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10271 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10275 // Check for being able to broadcast a single element.
10276 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10277 Mask, Subtarget, DAG))
10280 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10281 // use lower latency instructions that will operate on both 128-bit lanes.
10282 SmallVector<int, 2> RepeatedMask;
10283 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10284 if (isSingleInputShuffleMask(Mask)) {
10285 int PSHUFDMask[] = {-1, -1, -1, -1};
10286 for (int i = 0; i < 2; ++i)
10287 if (RepeatedMask[i] >= 0) {
10288 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10289 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10291 return DAG.getBitcast(
10293 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10294 DAG.getBitcast(MVT::v8i32, V1),
10295 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10299 // AVX2 provides a direct instruction for permuting a single input across
10301 if (isSingleInputShuffleMask(Mask))
10302 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10303 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10305 // Try to use shift instructions.
10306 if (SDValue Shift =
10307 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10310 // Use dedicated unpack instructions for masks that match their pattern.
10312 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10315 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10316 // shuffle. However, if we have AVX2 and either inputs are already in place,
10317 // we will be able to shuffle even across lanes the other input in a single
10318 // instruction so skip this pattern.
10319 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10320 isShuffleMaskInputInPlace(1, Mask))))
10321 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10322 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10325 // Otherwise fall back on generic blend lowering.
10326 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10330 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10332 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10333 /// isn't available.
10334 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10335 const X86Subtarget *Subtarget,
10336 SelectionDAG &DAG) {
10338 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10339 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10340 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10341 ArrayRef<int> Mask = SVOp->getMask();
10342 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10344 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10348 // Check for being able to broadcast a single element.
10349 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10350 Mask, Subtarget, DAG))
10353 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10354 // options to efficiently lower the shuffle.
10355 SmallVector<int, 4> RepeatedMask;
10356 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10357 assert(RepeatedMask.size() == 4 &&
10358 "Repeated masks must be half the mask width!");
10360 // Use even/odd duplicate instructions for masks that match their pattern.
10361 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10362 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10363 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10364 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10366 if (isSingleInputShuffleMask(Mask))
10367 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10368 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10370 // Use dedicated unpack instructions for masks that match their pattern.
10372 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10375 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10376 // have already handled any direct blends. We also need to squash the
10377 // repeated mask into a simulated v4f32 mask.
10378 for (int i = 0; i < 4; ++i)
10379 if (RepeatedMask[i] >= 8)
10380 RepeatedMask[i] -= 4;
10381 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10384 // If we have a single input shuffle with different shuffle patterns in the
10385 // two 128-bit lanes use the variable mask to VPERMILPS.
10386 if (isSingleInputShuffleMask(Mask)) {
10387 SDValue VPermMask[8];
10388 for (int i = 0; i < 8; ++i)
10389 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10390 : DAG.getConstant(Mask[i], DL, MVT::i32);
10391 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10392 return DAG.getNode(
10393 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10394 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10396 if (Subtarget->hasAVX2())
10397 return DAG.getNode(
10398 X86ISD::VPERMV, DL, MVT::v8f32,
10399 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10400 MVT::v8i32, VPermMask)),
10403 // Otherwise, fall back.
10404 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10408 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10410 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10411 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10414 // If we have AVX2 then we always want to lower with a blend because at v8 we
10415 // can fully permute the elements.
10416 if (Subtarget->hasAVX2())
10417 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10420 // Otherwise fall back on generic lowering.
10421 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10424 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10426 /// This routine is only called when we have AVX2 and thus a reasonable
10427 /// instruction set for v8i32 shuffling..
10428 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10429 const X86Subtarget *Subtarget,
10430 SelectionDAG &DAG) {
10432 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10433 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10434 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10435 ArrayRef<int> Mask = SVOp->getMask();
10436 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10437 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10439 // Whenever we can lower this as a zext, that instruction is strictly faster
10440 // than any alternative. It also allows us to fold memory operands into the
10441 // shuffle in many cases.
10442 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10443 Mask, Subtarget, DAG))
10446 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10450 // Check for being able to broadcast a single element.
10451 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10452 Mask, Subtarget, DAG))
10455 // If the shuffle mask is repeated in each 128-bit lane we can use more
10456 // efficient instructions that mirror the shuffles across the two 128-bit
10458 SmallVector<int, 4> RepeatedMask;
10459 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10460 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10461 if (isSingleInputShuffleMask(Mask))
10462 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10463 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10465 // Use dedicated unpack instructions for masks that match their pattern.
10467 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10471 // Try to use shift instructions.
10472 if (SDValue Shift =
10473 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10476 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10477 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10480 // If the shuffle patterns aren't repeated but it is a single input, directly
10481 // generate a cross-lane VPERMD instruction.
10482 if (isSingleInputShuffleMask(Mask)) {
10483 SDValue VPermMask[8];
10484 for (int i = 0; i < 8; ++i)
10485 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10486 : DAG.getConstant(Mask[i], DL, MVT::i32);
10487 return DAG.getNode(
10488 X86ISD::VPERMV, DL, MVT::v8i32,
10489 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10492 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10494 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10495 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10498 // Otherwise fall back on generic blend lowering.
10499 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10503 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10505 /// This routine is only called when we have AVX2 and thus a reasonable
10506 /// instruction set for v16i16 shuffling..
10507 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10508 const X86Subtarget *Subtarget,
10509 SelectionDAG &DAG) {
10511 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10512 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10513 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10514 ArrayRef<int> Mask = SVOp->getMask();
10515 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10516 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10518 // Whenever we can lower this as a zext, that instruction is strictly faster
10519 // than any alternative. It also allows us to fold memory operands into the
10520 // shuffle in many cases.
10521 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10522 Mask, Subtarget, DAG))
10525 // Check for being able to broadcast a single element.
10526 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10527 Mask, Subtarget, DAG))
10530 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10534 // Use dedicated unpack instructions for masks that match their pattern.
10536 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10539 // Try to use shift instructions.
10540 if (SDValue Shift =
10541 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10544 // Try to use byte rotation instructions.
10545 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10546 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10549 if (isSingleInputShuffleMask(Mask)) {
10550 // There are no generalized cross-lane shuffle operations available on i16
10552 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10553 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10556 SmallVector<int, 8> RepeatedMask;
10557 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10558 // As this is a single-input shuffle, the repeated mask should be
10559 // a strictly valid v8i16 mask that we can pass through to the v8i16
10560 // lowering to handle even the v16 case.
10561 return lowerV8I16GeneralSingleInputVectorShuffle(
10562 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10565 SDValue PSHUFBMask[32];
10566 for (int i = 0; i < 16; ++i) {
10567 if (Mask[i] == -1) {
10568 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10572 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10573 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10574 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10575 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10577 return DAG.getBitcast(MVT::v16i16,
10578 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10579 DAG.getBitcast(MVT::v32i8, V1),
10580 DAG.getNode(ISD::BUILD_VECTOR, DL,
10581 MVT::v32i8, PSHUFBMask)));
10584 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10586 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10587 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10590 // Otherwise fall back on generic lowering.
10591 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10594 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10596 /// This routine is only called when we have AVX2 and thus a reasonable
10597 /// instruction set for v32i8 shuffling..
10598 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10599 const X86Subtarget *Subtarget,
10600 SelectionDAG &DAG) {
10602 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10603 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10604 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10605 ArrayRef<int> Mask = SVOp->getMask();
10606 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10607 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10609 // Whenever we can lower this as a zext, that instruction is strictly faster
10610 // than any alternative. It also allows us to fold memory operands into the
10611 // shuffle in many cases.
10612 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10613 Mask, Subtarget, DAG))
10616 // Check for being able to broadcast a single element.
10617 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10618 Mask, Subtarget, DAG))
10621 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10625 // Use dedicated unpack instructions for masks that match their pattern.
10627 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10630 // Try to use shift instructions.
10631 if (SDValue Shift =
10632 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10635 // Try to use byte rotation instructions.
10636 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10637 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10640 if (isSingleInputShuffleMask(Mask)) {
10641 // There are no generalized cross-lane shuffle operations available on i8
10643 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10644 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10647 SDValue PSHUFBMask[32];
10648 for (int i = 0; i < 32; ++i)
10651 ? DAG.getUNDEF(MVT::i8)
10652 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10655 return DAG.getNode(
10656 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10657 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10660 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10662 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10663 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10666 // Otherwise fall back on generic lowering.
10667 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10670 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10672 /// This routine either breaks down the specific type of a 256-bit x86 vector
10673 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10674 /// together based on the available instructions.
10675 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10676 MVT VT, const X86Subtarget *Subtarget,
10677 SelectionDAG &DAG) {
10679 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10680 ArrayRef<int> Mask = SVOp->getMask();
10682 // If we have a single input to the zero element, insert that into V1 if we
10683 // can do so cheaply.
10684 int NumElts = VT.getVectorNumElements();
10685 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10686 return M >= NumElts;
10689 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10690 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10691 DL, VT, V1, V2, Mask, Subtarget, DAG))
10694 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10695 // can check for those subtargets here and avoid much of the subtarget
10696 // querying in the per-vector-type lowering routines. With AVX1 we have
10697 // essentially *zero* ability to manipulate a 256-bit vector with integer
10698 // types. Since we'll use floating point types there eventually, just
10699 // immediately cast everything to a float and operate entirely in that domain.
10700 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10701 int ElementBits = VT.getScalarSizeInBits();
10702 if (ElementBits < 32)
10703 // No floating point type available, decompose into 128-bit vectors.
10704 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10706 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10707 VT.getVectorNumElements());
10708 V1 = DAG.getBitcast(FpVT, V1);
10709 V2 = DAG.getBitcast(FpVT, V2);
10710 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10713 switch (VT.SimpleTy) {
10715 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10717 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10719 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10721 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10723 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10725 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10728 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10732 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10733 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10734 ArrayRef<int> Mask,
10735 SDValue V1, SDValue V2,
10736 SelectionDAG &DAG) {
10737 assert(VT.getScalarSizeInBits() == 64 &&
10738 "Unexpected element type size for 128bit shuffle.");
10740 // To handle 256 bit vector requires VLX and most probably
10741 // function lowerV2X128VectorShuffle() is better solution.
10742 assert(VT.is512BitVector() && "Unexpected vector size for 128bit shuffle.");
10744 SmallVector<int, 4> WidenedMask;
10745 if (!canWidenShuffleElements(Mask, WidenedMask))
10748 // Form a 128-bit permutation.
10749 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10750 // bits defined by a vshuf64x2 instruction's immediate control byte.
10751 unsigned PermMask = 0, Imm = 0;
10752 unsigned ControlBitsNum = WidenedMask.size() / 2;
10754 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10755 if (WidenedMask[i] == SM_SentinelZero)
10758 // Use first element in place of undef mask.
10759 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10760 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10763 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10764 DAG.getConstant(PermMask, DL, MVT::i8));
10767 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10768 ArrayRef<int> Mask, SDValue V1,
10769 SDValue V2, SelectionDAG &DAG) {
10771 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10773 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10774 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10776 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10777 if (isSingleInputShuffleMask(Mask))
10778 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10780 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10783 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10784 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10785 const X86Subtarget *Subtarget,
10786 SelectionDAG &DAG) {
10788 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10789 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10790 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10791 ArrayRef<int> Mask = SVOp->getMask();
10792 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10794 if (SDValue Shuf128 =
10795 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10798 if (SDValue Unpck =
10799 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10802 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10805 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10806 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10807 const X86Subtarget *Subtarget,
10808 SelectionDAG &DAG) {
10810 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10811 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10812 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10813 ArrayRef<int> Mask = SVOp->getMask();
10814 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10816 if (SDValue Unpck =
10817 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10820 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10823 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10824 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10825 const X86Subtarget *Subtarget,
10826 SelectionDAG &DAG) {
10828 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10829 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10830 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10831 ArrayRef<int> Mask = SVOp->getMask();
10832 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10834 if (SDValue Shuf128 =
10835 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10838 if (SDValue Unpck =
10839 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10842 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10845 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10846 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10847 const X86Subtarget *Subtarget,
10848 SelectionDAG &DAG) {
10850 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10851 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10852 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10853 ArrayRef<int> Mask = SVOp->getMask();
10854 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10856 if (SDValue Unpck =
10857 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10860 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10863 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10864 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10865 const X86Subtarget *Subtarget,
10866 SelectionDAG &DAG) {
10868 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10869 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10870 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10871 ArrayRef<int> Mask = SVOp->getMask();
10872 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10873 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10875 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10878 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10879 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10880 const X86Subtarget *Subtarget,
10881 SelectionDAG &DAG) {
10883 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10884 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10885 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10886 ArrayRef<int> Mask = SVOp->getMask();
10887 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10888 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10890 // FIXME: Implement direct support for this type!
10891 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10894 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10896 /// This routine either breaks down the specific type of a 512-bit x86 vector
10897 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10898 /// together based on the available instructions.
10899 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10900 MVT VT, const X86Subtarget *Subtarget,
10901 SelectionDAG &DAG) {
10903 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10904 ArrayRef<int> Mask = SVOp->getMask();
10905 assert(Subtarget->hasAVX512() &&
10906 "Cannot lower 512-bit vectors w/ basic ISA!");
10908 // Check for being able to broadcast a single element.
10909 if (SDValue Broadcast =
10910 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10913 // Dispatch to each element type for lowering. If we don't have supprot for
10914 // specific element type shuffles at 512 bits, immediately split them and
10915 // lower them. Each lowering routine of a given type is allowed to assume that
10916 // the requisite ISA extensions for that element type are available.
10917 switch (VT.SimpleTy) {
10919 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10921 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10923 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10925 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10927 if (Subtarget->hasBWI())
10928 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10931 if (Subtarget->hasBWI())
10932 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10936 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10939 // Otherwise fall back on splitting.
10940 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10943 // Lower vXi1 vector shuffles.
10944 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10945 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10946 // vector, shuffle and then truncate it back.
10947 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10948 MVT VT, const X86Subtarget *Subtarget,
10949 SelectionDAG &DAG) {
10951 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10952 ArrayRef<int> Mask = SVOp->getMask();
10953 assert(Subtarget->hasAVX512() &&
10954 "Cannot lower 512-bit vectors w/o basic ISA!");
10956 switch (VT.SimpleTy) {
10958 llvm_unreachable("Expected a vector of i1 elements");
10960 ExtVT = MVT::v2i64;
10963 ExtVT = MVT::v4i32;
10966 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10969 ExtVT = MVT::v16i32;
10972 ExtVT = MVT::v32i16;
10975 ExtVT = MVT::v64i8;
10979 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10980 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10981 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10982 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10984 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10987 V2 = DAG.getUNDEF(ExtVT);
10988 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10989 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10990 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10991 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10993 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10994 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10995 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
10997 /// \brief Top-level lowering for x86 vector shuffles.
10999 /// This handles decomposition, canonicalization, and lowering of all x86
11000 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11001 /// above in helper routines. The canonicalization attempts to widen shuffles
11002 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11003 /// s.t. only one of the two inputs needs to be tested, etc.
11004 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11005 SelectionDAG &DAG) {
11006 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11007 ArrayRef<int> Mask = SVOp->getMask();
11008 SDValue V1 = Op.getOperand(0);
11009 SDValue V2 = Op.getOperand(1);
11010 MVT VT = Op.getSimpleValueType();
11011 int NumElements = VT.getVectorNumElements();
11013 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
11015 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11016 "Can't lower MMX shuffles");
11018 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11019 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11020 if (V1IsUndef && V2IsUndef)
11021 return DAG.getUNDEF(VT);
11023 // When we create a shuffle node we put the UNDEF node to second operand,
11024 // but in some cases the first operand may be transformed to UNDEF.
11025 // In this case we should just commute the node.
11027 return DAG.getCommutedVectorShuffle(*SVOp);
11029 // Check for non-undef masks pointing at an undef vector and make the masks
11030 // undef as well. This makes it easier to match the shuffle based solely on
11034 if (M >= NumElements) {
11035 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11036 for (int &M : NewMask)
11037 if (M >= NumElements)
11039 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11042 // We actually see shuffles that are entirely re-arrangements of a set of
11043 // zero inputs. This mostly happens while decomposing complex shuffles into
11044 // simple ones. Directly lower these as a buildvector of zeros.
11045 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11046 if (Zeroable.all())
11047 return getZeroVector(VT, Subtarget, DAG, dl);
11049 // Try to collapse shuffles into using a vector type with fewer elements but
11050 // wider element types. We cap this to not form integers or floating point
11051 // elements wider than 64 bits, but it might be interesting to form i128
11052 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11053 SmallVector<int, 16> WidenedMask;
11054 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11055 canWidenShuffleElements(Mask, WidenedMask)) {
11056 MVT NewEltVT = VT.isFloatingPoint()
11057 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11058 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11059 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11060 // Make sure that the new vector type is legal. For example, v2f64 isn't
11062 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11063 V1 = DAG.getBitcast(NewVT, V1);
11064 V2 = DAG.getBitcast(NewVT, V2);
11065 return DAG.getBitcast(
11066 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11070 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11071 for (int M : SVOp->getMask())
11073 ++NumUndefElements;
11074 else if (M < NumElements)
11079 // Commute the shuffle as needed such that more elements come from V1 than
11080 // V2. This allows us to match the shuffle pattern strictly on how many
11081 // elements come from V1 without handling the symmetric cases.
11082 if (NumV2Elements > NumV1Elements)
11083 return DAG.getCommutedVectorShuffle(*SVOp);
11085 // When the number of V1 and V2 elements are the same, try to minimize the
11086 // number of uses of V2 in the low half of the vector. When that is tied,
11087 // ensure that the sum of indices for V1 is equal to or lower than the sum
11088 // indices for V2. When those are equal, try to ensure that the number of odd
11089 // indices for V1 is lower than the number of odd indices for V2.
11090 if (NumV1Elements == NumV2Elements) {
11091 int LowV1Elements = 0, LowV2Elements = 0;
11092 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11093 if (M >= NumElements)
11097 if (LowV2Elements > LowV1Elements) {
11098 return DAG.getCommutedVectorShuffle(*SVOp);
11099 } else if (LowV2Elements == LowV1Elements) {
11100 int SumV1Indices = 0, SumV2Indices = 0;
11101 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11102 if (SVOp->getMask()[i] >= NumElements)
11104 else if (SVOp->getMask()[i] >= 0)
11106 if (SumV2Indices < SumV1Indices) {
11107 return DAG.getCommutedVectorShuffle(*SVOp);
11108 } else if (SumV2Indices == SumV1Indices) {
11109 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11110 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11111 if (SVOp->getMask()[i] >= NumElements)
11112 NumV2OddIndices += i % 2;
11113 else if (SVOp->getMask()[i] >= 0)
11114 NumV1OddIndices += i % 2;
11115 if (NumV2OddIndices < NumV1OddIndices)
11116 return DAG.getCommutedVectorShuffle(*SVOp);
11121 // For each vector width, delegate to a specialized lowering routine.
11122 if (VT.is128BitVector())
11123 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11125 if (VT.is256BitVector())
11126 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11128 if (VT.is512BitVector())
11129 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11132 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11133 llvm_unreachable("Unimplemented!");
11136 // This function assumes its argument is a BUILD_VECTOR of constants or
11137 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11139 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11140 unsigned &MaskValue) {
11142 unsigned NumElems = BuildVector->getNumOperands();
11144 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11145 // We don't handle the >2 lanes case right now.
11146 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11150 unsigned NumElemsInLane = NumElems / NumLanes;
11152 // Blend for v16i16 should be symmetric for the both lanes.
11153 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11154 SDValue EltCond = BuildVector->getOperand(i);
11155 SDValue SndLaneEltCond =
11156 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11158 int Lane1Cond = -1, Lane2Cond = -1;
11159 if (isa<ConstantSDNode>(EltCond))
11160 Lane1Cond = !isZero(EltCond);
11161 if (isa<ConstantSDNode>(SndLaneEltCond))
11162 Lane2Cond = !isZero(SndLaneEltCond);
11164 unsigned LaneMask = 0;
11165 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11166 // Lane1Cond != 0, means we want the first argument.
11167 // Lane1Cond == 0, means we want the second argument.
11168 // The encoding of this argument is 0 for the first argument, 1
11169 // for the second. Therefore, invert the condition.
11170 LaneMask = !Lane1Cond << i;
11171 else if (Lane1Cond < 0)
11172 LaneMask = !Lane2Cond << i;
11176 MaskValue |= LaneMask;
11178 MaskValue |= LaneMask << NumElemsInLane;
11183 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11184 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11185 const X86Subtarget *Subtarget,
11186 SelectionDAG &DAG) {
11187 SDValue Cond = Op.getOperand(0);
11188 SDValue LHS = Op.getOperand(1);
11189 SDValue RHS = Op.getOperand(2);
11191 MVT VT = Op.getSimpleValueType();
11193 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11195 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11197 // Only non-legal VSELECTs reach this lowering, convert those into generic
11198 // shuffles and re-use the shuffle lowering path for blends.
11199 SmallVector<int, 32> Mask;
11200 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11201 SDValue CondElt = CondBV->getOperand(i);
11203 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11205 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11208 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11209 // A vselect where all conditions and data are constants can be optimized into
11210 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11211 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11212 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11213 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11216 // Try to lower this to a blend-style vector shuffle. This can handle all
11217 // constant condition cases.
11218 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11221 // Variable blends are only legal from SSE4.1 onward.
11222 if (!Subtarget->hasSSE41())
11225 // Only some types will be legal on some subtargets. If we can emit a legal
11226 // VSELECT-matching blend, return Op, and but if we need to expand, return
11228 switch (Op.getSimpleValueType().SimpleTy) {
11230 // Most of the vector types have blends past SSE4.1.
11234 // The byte blends for AVX vectors were introduced only in AVX2.
11235 if (Subtarget->hasAVX2())
11242 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11243 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11246 // FIXME: We should custom lower this by fixing the condition and using i8
11252 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11253 MVT VT = Op.getSimpleValueType();
11256 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11259 if (VT.getSizeInBits() == 8) {
11260 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11261 Op.getOperand(0), Op.getOperand(1));
11262 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11263 DAG.getValueType(VT));
11264 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11267 if (VT.getSizeInBits() == 16) {
11268 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11269 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11271 return DAG.getNode(
11272 ISD::TRUNCATE, dl, MVT::i16,
11273 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11274 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11275 Op.getOperand(1)));
11276 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11277 Op.getOperand(0), Op.getOperand(1));
11278 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11279 DAG.getValueType(VT));
11280 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11283 if (VT == MVT::f32) {
11284 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11285 // the result back to FR32 register. It's only worth matching if the
11286 // result has a single use which is a store or a bitcast to i32. And in
11287 // the case of a store, it's not worth it if the index is a constant 0,
11288 // because a MOVSSmr can be used instead, which is smaller and faster.
11289 if (!Op.hasOneUse())
11291 SDNode *User = *Op.getNode()->use_begin();
11292 if ((User->getOpcode() != ISD::STORE ||
11293 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11294 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11295 (User->getOpcode() != ISD::BITCAST ||
11296 User->getValueType(0) != MVT::i32))
11298 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11299 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11301 return DAG.getBitcast(MVT::f32, Extract);
11304 if (VT == MVT::i32 || VT == MVT::i64) {
11305 // ExtractPS/pextrq works with constant index.
11306 if (isa<ConstantSDNode>(Op.getOperand(1)))
11312 /// Extract one bit from mask vector, like v16i1 or v8i1.
11313 /// AVX-512 feature.
11315 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11316 SDValue Vec = Op.getOperand(0);
11318 MVT VecVT = Vec.getSimpleValueType();
11319 SDValue Idx = Op.getOperand(1);
11320 MVT EltVT = Op.getSimpleValueType();
11322 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11323 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11324 "Unexpected vector type in ExtractBitFromMaskVector");
11326 // variable index can't be handled in mask registers,
11327 // extend vector to VR512
11328 if (!isa<ConstantSDNode>(Idx)) {
11329 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11330 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11331 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11332 ExtVT.getVectorElementType(), Ext, Idx);
11333 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11336 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11337 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11338 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11339 rc = getRegClassFor(MVT::v16i1);
11340 unsigned MaxSift = rc->getSize()*8 - 1;
11341 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11342 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11343 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11344 DAG.getConstant(MaxSift, dl, MVT::i8));
11345 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11346 DAG.getIntPtrConstant(0, dl));
11350 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11351 SelectionDAG &DAG) const {
11353 SDValue Vec = Op.getOperand(0);
11354 MVT VecVT = Vec.getSimpleValueType();
11355 SDValue Idx = Op.getOperand(1);
11357 if (Op.getSimpleValueType() == MVT::i1)
11358 return ExtractBitFromMaskVector(Op, DAG);
11360 if (!isa<ConstantSDNode>(Idx)) {
11361 if (VecVT.is512BitVector() ||
11362 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11363 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11366 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11367 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11368 MaskEltVT.getSizeInBits());
11370 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11371 auto PtrVT = getPointerTy(DAG.getDataLayout());
11372 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11373 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11374 DAG.getConstant(0, dl, PtrVT));
11375 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11376 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11377 DAG.getConstant(0, dl, PtrVT));
11382 // If this is a 256-bit vector result, first extract the 128-bit vector and
11383 // then extract the element from the 128-bit vector.
11384 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11386 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11387 // Get the 128-bit vector.
11388 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11389 MVT EltVT = VecVT.getVectorElementType();
11391 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11392 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
11394 // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
11395 // this can be done with a mask.
11396 IdxVal &= ElemsPerChunk - 1;
11397 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11398 DAG.getConstant(IdxVal, dl, MVT::i32));
11401 assert(VecVT.is128BitVector() && "Unexpected vector length");
11403 if (Subtarget->hasSSE41())
11404 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11407 MVT VT = Op.getSimpleValueType();
11408 // TODO: handle v16i8.
11409 if (VT.getSizeInBits() == 16) {
11410 SDValue Vec = Op.getOperand(0);
11411 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11413 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11414 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11415 DAG.getBitcast(MVT::v4i32, Vec),
11416 Op.getOperand(1)));
11417 // Transform it so it match pextrw which produces a 32-bit result.
11418 MVT EltVT = MVT::i32;
11419 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11420 Op.getOperand(0), Op.getOperand(1));
11421 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11422 DAG.getValueType(VT));
11423 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11426 if (VT.getSizeInBits() == 32) {
11427 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11431 // SHUFPS the element to the lowest double word, then movss.
11432 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11433 MVT VVT = Op.getOperand(0).getSimpleValueType();
11434 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11435 DAG.getUNDEF(VVT), Mask);
11436 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11437 DAG.getIntPtrConstant(0, dl));
11440 if (VT.getSizeInBits() == 64) {
11441 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11442 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11443 // to match extract_elt for f64.
11444 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11448 // UNPCKHPD the element to the lowest double word, then movsd.
11449 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11450 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11451 int Mask[2] = { 1, -1 };
11452 MVT VVT = Op.getOperand(0).getSimpleValueType();
11453 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11454 DAG.getUNDEF(VVT), Mask);
11455 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11456 DAG.getIntPtrConstant(0, dl));
11462 /// Insert one bit to mask vector, like v16i1 or v8i1.
11463 /// AVX-512 feature.
11465 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11467 SDValue Vec = Op.getOperand(0);
11468 SDValue Elt = Op.getOperand(1);
11469 SDValue Idx = Op.getOperand(2);
11470 MVT VecVT = Vec.getSimpleValueType();
11472 if (!isa<ConstantSDNode>(Idx)) {
11473 // Non constant index. Extend source and destination,
11474 // insert element and then truncate the result.
11475 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11476 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11477 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11478 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11479 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11480 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11483 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11484 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11486 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11487 DAG.getConstant(IdxVal, dl, MVT::i8));
11488 if (Vec.getOpcode() == ISD::UNDEF)
11490 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11493 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11494 SelectionDAG &DAG) const {
11495 MVT VT = Op.getSimpleValueType();
11496 MVT EltVT = VT.getVectorElementType();
11498 if (EltVT == MVT::i1)
11499 return InsertBitToMaskVector(Op, DAG);
11502 SDValue N0 = Op.getOperand(0);
11503 SDValue N1 = Op.getOperand(1);
11504 SDValue N2 = Op.getOperand(2);
11505 if (!isa<ConstantSDNode>(N2))
11507 auto *N2C = cast<ConstantSDNode>(N2);
11508 unsigned IdxVal = N2C->getZExtValue();
11510 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11511 // into that, and then insert the subvector back into the result.
11512 if (VT.is256BitVector() || VT.is512BitVector()) {
11513 // With a 256-bit vector, we can insert into the zero element efficiently
11514 // using a blend if we have AVX or AVX2 and the right data type.
11515 if (VT.is256BitVector() && IdxVal == 0) {
11516 // TODO: It is worthwhile to cast integer to floating point and back
11517 // and incur a domain crossing penalty if that's what we'll end up
11518 // doing anyway after extracting to a 128-bit vector.
11519 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11520 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11521 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11522 N2 = DAG.getIntPtrConstant(1, dl);
11523 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11527 // Get the desired 128-bit vector chunk.
11528 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11530 // Insert the element into the desired chunk.
11531 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11532 assert(isPowerOf2_32(NumEltsIn128));
11533 // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
11534 unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
11536 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11537 DAG.getConstant(IdxIn128, dl, MVT::i32));
11539 // Insert the changed part back into the bigger vector
11540 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11542 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11544 if (Subtarget->hasSSE41()) {
11545 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11547 if (VT == MVT::v8i16) {
11548 Opc = X86ISD::PINSRW;
11550 assert(VT == MVT::v16i8);
11551 Opc = X86ISD::PINSRB;
11554 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11556 if (N1.getValueType() != MVT::i32)
11557 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11558 if (N2.getValueType() != MVT::i32)
11559 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11560 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11563 if (EltVT == MVT::f32) {
11564 // Bits [7:6] of the constant are the source select. This will always be
11565 // zero here. The DAG Combiner may combine an extract_elt index into
11566 // these bits. For example (insert (extract, 3), 2) could be matched by
11567 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11568 // Bits [5:4] of the constant are the destination select. This is the
11569 // value of the incoming immediate.
11570 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11571 // combine either bitwise AND or insert of float 0.0 to set these bits.
11573 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11574 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11575 // If this is an insertion of 32-bits into the low 32-bits of
11576 // a vector, we prefer to generate a blend with immediate rather
11577 // than an insertps. Blends are simpler operations in hardware and so
11578 // will always have equal or better performance than insertps.
11579 // But if optimizing for size and there's a load folding opportunity,
11580 // generate insertps because blendps does not have a 32-bit memory
11582 N2 = DAG.getIntPtrConstant(1, dl);
11583 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11584 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11586 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11587 // Create this as a scalar to vector..
11588 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11589 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11592 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11593 // PINSR* works with constant index.
11598 if (EltVT == MVT::i8)
11601 if (EltVT.getSizeInBits() == 16) {
11602 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11603 // as its second argument.
11604 if (N1.getValueType() != MVT::i32)
11605 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11606 if (N2.getValueType() != MVT::i32)
11607 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11608 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11613 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11615 MVT OpVT = Op.getSimpleValueType();
11617 // If this is a 256-bit vector result, first insert into a 128-bit
11618 // vector and then insert into the 256-bit vector.
11619 if (!OpVT.is128BitVector()) {
11620 // Insert into a 128-bit vector.
11621 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11622 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11623 OpVT.getVectorNumElements() / SizeFactor);
11625 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11627 // Insert the 128-bit vector.
11628 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11631 if (OpVT == MVT::v1i64 &&
11632 Op.getOperand(0).getValueType() == MVT::i64)
11633 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11635 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11636 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11637 return DAG.getBitcast(
11638 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11641 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11642 // a simple subregister reference or explicit instructions to grab
11643 // upper bits of a vector.
11644 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11645 SelectionDAG &DAG) {
11647 SDValue In = Op.getOperand(0);
11648 SDValue Idx = Op.getOperand(1);
11649 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11650 MVT ResVT = Op.getSimpleValueType();
11651 MVT InVT = In.getSimpleValueType();
11653 if (Subtarget->hasFp256()) {
11654 if (ResVT.is128BitVector() &&
11655 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11656 isa<ConstantSDNode>(Idx)) {
11657 return Extract128BitVector(In, IdxVal, DAG, dl);
11659 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11660 isa<ConstantSDNode>(Idx)) {
11661 return Extract256BitVector(In, IdxVal, DAG, dl);
11667 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11668 // simple superregister reference or explicit instructions to insert
11669 // the upper bits of a vector.
11670 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11671 SelectionDAG &DAG) {
11672 if (!Subtarget->hasAVX())
11676 SDValue Vec = Op.getOperand(0);
11677 SDValue SubVec = Op.getOperand(1);
11678 SDValue Idx = Op.getOperand(2);
11680 if (!isa<ConstantSDNode>(Idx))
11683 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11684 MVT OpVT = Op.getSimpleValueType();
11685 MVT SubVecVT = SubVec.getSimpleValueType();
11687 // Fold two 16-byte subvector loads into one 32-byte load:
11688 // (insert_subvector (insert_subvector undef, (load addr), 0),
11689 // (load addr + 16), Elts/2)
11691 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11692 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11693 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11694 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11695 if (Idx2 && Idx2->getZExtValue() == 0) {
11696 SDValue SubVec2 = Vec.getOperand(1);
11697 // If needed, look through a bitcast to get to the load.
11698 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11699 SubVec2 = SubVec2.getOperand(0);
11701 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11703 unsigned Alignment = FirstLd->getAlignment();
11704 unsigned AS = FirstLd->getAddressSpace();
11705 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11706 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11707 OpVT, AS, Alignment, &Fast) && Fast) {
11708 SDValue Ops[] = { SubVec2, SubVec };
11709 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11716 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11717 SubVecVT.is128BitVector())
11718 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11720 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11721 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11723 if (OpVT.getVectorElementType() == MVT::i1) {
11724 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11726 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11727 SDValue Undef = DAG.getUNDEF(OpVT);
11728 unsigned NumElems = OpVT.getVectorNumElements();
11729 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11731 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11732 // Zero upper bits of the Vec
11733 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11734 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11736 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11738 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11739 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11742 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11744 // Zero upper bits of the Vec2
11745 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11746 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11747 // Zero lower bits of the Vec
11748 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11749 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11750 // Merge them together
11751 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11757 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11758 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11759 // one of the above mentioned nodes. It has to be wrapped because otherwise
11760 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11761 // be used to form addressing mode. These wrapped nodes will be selected
11764 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11765 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11767 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11768 // global base reg.
11769 unsigned char OpFlag = 0;
11770 unsigned WrapperKind = X86ISD::Wrapper;
11771 CodeModel::Model M = DAG.getTarget().getCodeModel();
11773 if (Subtarget->isPICStyleRIPRel() &&
11774 (M == CodeModel::Small || M == CodeModel::Kernel))
11775 WrapperKind = X86ISD::WrapperRIP;
11776 else if (Subtarget->isPICStyleGOT())
11777 OpFlag = X86II::MO_GOTOFF;
11778 else if (Subtarget->isPICStyleStubPIC())
11779 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11781 auto PtrVT = getPointerTy(DAG.getDataLayout());
11782 SDValue Result = DAG.getTargetConstantPool(
11783 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11785 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11786 // With PIC, the address is actually $g + Offset.
11789 DAG.getNode(ISD::ADD, DL, PtrVT,
11790 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11796 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11797 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11799 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11800 // global base reg.
11801 unsigned char OpFlag = 0;
11802 unsigned WrapperKind = X86ISD::Wrapper;
11803 CodeModel::Model M = DAG.getTarget().getCodeModel();
11805 if (Subtarget->isPICStyleRIPRel() &&
11806 (M == CodeModel::Small || M == CodeModel::Kernel))
11807 WrapperKind = X86ISD::WrapperRIP;
11808 else if (Subtarget->isPICStyleGOT())
11809 OpFlag = X86II::MO_GOTOFF;
11810 else if (Subtarget->isPICStyleStubPIC())
11811 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11813 auto PtrVT = getPointerTy(DAG.getDataLayout());
11814 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11816 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11818 // With PIC, the address is actually $g + Offset.
11821 DAG.getNode(ISD::ADD, DL, PtrVT,
11822 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11828 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11829 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11831 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11832 // global base reg.
11833 unsigned char OpFlag = 0;
11834 unsigned WrapperKind = X86ISD::Wrapper;
11835 CodeModel::Model M = DAG.getTarget().getCodeModel();
11837 if (Subtarget->isPICStyleRIPRel() &&
11838 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11839 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11840 OpFlag = X86II::MO_GOTPCREL;
11841 WrapperKind = X86ISD::WrapperRIP;
11842 } else if (Subtarget->isPICStyleGOT()) {
11843 OpFlag = X86II::MO_GOT;
11844 } else if (Subtarget->isPICStyleStubPIC()) {
11845 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11846 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11847 OpFlag = X86II::MO_DARWIN_NONLAZY;
11850 auto PtrVT = getPointerTy(DAG.getDataLayout());
11851 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11854 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11856 // With PIC, the address is actually $g + Offset.
11857 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11858 !Subtarget->is64Bit()) {
11860 DAG.getNode(ISD::ADD, DL, PtrVT,
11861 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11864 // For symbols that require a load from a stub to get the address, emit the
11866 if (isGlobalStubReference(OpFlag))
11867 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11868 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11869 false, false, false, 0);
11875 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11876 // Create the TargetBlockAddressAddress node.
11877 unsigned char OpFlags =
11878 Subtarget->ClassifyBlockAddressReference();
11879 CodeModel::Model M = DAG.getTarget().getCodeModel();
11880 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11881 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11883 auto PtrVT = getPointerTy(DAG.getDataLayout());
11884 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11886 if (Subtarget->isPICStyleRIPRel() &&
11887 (M == CodeModel::Small || M == CodeModel::Kernel))
11888 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11890 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11892 // With PIC, the address is actually $g + Offset.
11893 if (isGlobalRelativeToPICBase(OpFlags)) {
11894 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11895 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11902 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11903 int64_t Offset, SelectionDAG &DAG) const {
11904 // Create the TargetGlobalAddress node, folding in the constant
11905 // offset if it is legal.
11906 unsigned char OpFlags =
11907 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11908 CodeModel::Model M = DAG.getTarget().getCodeModel();
11909 auto PtrVT = getPointerTy(DAG.getDataLayout());
11911 if (OpFlags == X86II::MO_NO_FLAG &&
11912 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11913 // A direct static reference to a global.
11914 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11917 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11920 if (Subtarget->isPICStyleRIPRel() &&
11921 (M == CodeModel::Small || M == CodeModel::Kernel))
11922 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11924 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11926 // With PIC, the address is actually $g + Offset.
11927 if (isGlobalRelativeToPICBase(OpFlags)) {
11928 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11929 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11932 // For globals that require a load from a stub to get the address, emit the
11934 if (isGlobalStubReference(OpFlags))
11935 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11936 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11937 false, false, false, 0);
11939 // If there was a non-zero offset that we didn't fold, create an explicit
11940 // addition for it.
11942 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11943 DAG.getConstant(Offset, dl, PtrVT));
11949 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11950 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11951 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11952 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11956 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11957 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11958 unsigned char OperandFlags, bool LocalDynamic = false) {
11959 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11960 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11962 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11963 GA->getValueType(0),
11967 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11971 SDValue Ops[] = { Chain, TGA, *InFlag };
11972 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11974 SDValue Ops[] = { Chain, TGA };
11975 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11978 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11979 MFI->setAdjustsStack(true);
11980 MFI->setHasCalls(true);
11982 SDValue Flag = Chain.getValue(1);
11983 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11986 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11988 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11991 SDLoc dl(GA); // ? function entry point might be better
11992 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11993 DAG.getNode(X86ISD::GlobalBaseReg,
11994 SDLoc(), PtrVT), InFlag);
11995 InFlag = Chain.getValue(1);
11997 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12000 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12002 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12004 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12005 X86::RAX, X86II::MO_TLSGD);
12008 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12014 // Get the start address of the TLS block for this module.
12015 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12016 .getInfo<X86MachineFunctionInfo>();
12017 MFI->incNumLocalDynamicTLSAccesses();
12021 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12022 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12025 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12026 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12027 InFlag = Chain.getValue(1);
12028 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12029 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12032 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12036 unsigned char OperandFlags = X86II::MO_DTPOFF;
12037 unsigned WrapperKind = X86ISD::Wrapper;
12038 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12039 GA->getValueType(0),
12040 GA->getOffset(), OperandFlags);
12041 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12043 // Add x@dtpoff with the base.
12044 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12047 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12048 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12049 const EVT PtrVT, TLSModel::Model model,
12050 bool is64Bit, bool isPIC) {
12053 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12054 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12055 is64Bit ? 257 : 256));
12057 SDValue ThreadPointer =
12058 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12059 MachinePointerInfo(Ptr), false, false, false, 0);
12061 unsigned char OperandFlags = 0;
12062 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12064 unsigned WrapperKind = X86ISD::Wrapper;
12065 if (model == TLSModel::LocalExec) {
12066 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12067 } else if (model == TLSModel::InitialExec) {
12069 OperandFlags = X86II::MO_GOTTPOFF;
12070 WrapperKind = X86ISD::WrapperRIP;
12072 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12075 llvm_unreachable("Unexpected model");
12078 // emit "addl x@ntpoff,%eax" (local exec)
12079 // or "addl x@indntpoff,%eax" (initial exec)
12080 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12082 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12083 GA->getOffset(), OperandFlags);
12084 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12086 if (model == TLSModel::InitialExec) {
12087 if (isPIC && !is64Bit) {
12088 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12089 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12093 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12094 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12095 false, false, false, 0);
12098 // The address of the thread local variable is the add of the thread
12099 // pointer with the offset of the variable.
12100 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12104 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12106 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12107 const GlobalValue *GV = GA->getGlobal();
12108 auto PtrVT = getPointerTy(DAG.getDataLayout());
12110 if (Subtarget->isTargetELF()) {
12111 if (DAG.getTarget().Options.EmulatedTLS)
12112 return LowerToTLSEmulatedModel(GA, DAG);
12113 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12115 case TLSModel::GeneralDynamic:
12116 if (Subtarget->is64Bit())
12117 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12118 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12119 case TLSModel::LocalDynamic:
12120 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12121 Subtarget->is64Bit());
12122 case TLSModel::InitialExec:
12123 case TLSModel::LocalExec:
12124 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12125 DAG.getTarget().getRelocationModel() ==
12128 llvm_unreachable("Unknown TLS model.");
12131 if (Subtarget->isTargetDarwin()) {
12132 // Darwin only has one model of TLS. Lower to that.
12133 unsigned char OpFlag = 0;
12134 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12135 X86ISD::WrapperRIP : X86ISD::Wrapper;
12137 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12138 // global base reg.
12139 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12140 !Subtarget->is64Bit();
12142 OpFlag = X86II::MO_TLVP_PIC_BASE;
12144 OpFlag = X86II::MO_TLVP;
12146 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12147 GA->getValueType(0),
12148 GA->getOffset(), OpFlag);
12149 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12151 // With PIC32, the address is actually $g + Offset.
12153 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12154 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12157 // Lowering the machine isd will make sure everything is in the right
12159 SDValue Chain = DAG.getEntryNode();
12160 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12161 SDValue Args[] = { Chain, Offset };
12162 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12164 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12165 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12166 MFI->setAdjustsStack(true);
12168 // And our return value (tls address) is in the standard call return value
12170 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12171 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12174 if (Subtarget->isTargetKnownWindowsMSVC() ||
12175 Subtarget->isTargetWindowsGNU()) {
12176 // Just use the implicit TLS architecture
12177 // Need to generate someting similar to:
12178 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12180 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12181 // mov rcx, qword [rdx+rcx*8]
12182 // mov eax, .tls$:tlsvar
12183 // [rax+rcx] contains the address
12184 // Windows 64bit: gs:0x58
12185 // Windows 32bit: fs:__tls_array
12188 SDValue Chain = DAG.getEntryNode();
12190 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12191 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12192 // use its literal value of 0x2C.
12193 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12194 ? Type::getInt8PtrTy(*DAG.getContext(),
12196 : Type::getInt32PtrTy(*DAG.getContext(),
12199 SDValue TlsArray = Subtarget->is64Bit()
12200 ? DAG.getIntPtrConstant(0x58, dl)
12201 : (Subtarget->isTargetWindowsGNU()
12202 ? DAG.getIntPtrConstant(0x2C, dl)
12203 : DAG.getExternalSymbol("_tls_array", PtrVT));
12205 SDValue ThreadPointer =
12206 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12210 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12211 res = ThreadPointer;
12213 // Load the _tls_index variable
12214 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12215 if (Subtarget->is64Bit())
12216 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12217 MachinePointerInfo(), MVT::i32, false, false,
12220 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12223 auto &DL = DAG.getDataLayout();
12225 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12226 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12228 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12231 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12234 // Get the offset of start of .tls section
12235 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12236 GA->getValueType(0),
12237 GA->getOffset(), X86II::MO_SECREL);
12238 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12240 // The address of the thread local variable is the add of the thread
12241 // pointer with the offset of the variable.
12242 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12245 llvm_unreachable("TLS not implemented for this target.");
12248 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12249 /// and take a 2 x i32 value to shift plus a shift amount.
12250 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12251 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12252 MVT VT = Op.getSimpleValueType();
12253 unsigned VTBits = VT.getSizeInBits();
12255 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12256 SDValue ShOpLo = Op.getOperand(0);
12257 SDValue ShOpHi = Op.getOperand(1);
12258 SDValue ShAmt = Op.getOperand(2);
12259 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12260 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12262 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12263 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12264 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12265 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12266 : DAG.getConstant(0, dl, VT);
12268 SDValue Tmp2, Tmp3;
12269 if (Op.getOpcode() == ISD::SHL_PARTS) {
12270 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12271 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12273 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12274 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12277 // If the shift amount is larger or equal than the width of a part we can't
12278 // rely on the results of shld/shrd. Insert a test and select the appropriate
12279 // values for large shift amounts.
12280 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12281 DAG.getConstant(VTBits, dl, MVT::i8));
12282 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12283 AndNode, DAG.getConstant(0, dl, MVT::i8));
12286 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12287 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12288 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12290 if (Op.getOpcode() == ISD::SHL_PARTS) {
12291 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12292 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12294 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12295 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12298 SDValue Ops[2] = { Lo, Hi };
12299 return DAG.getMergeValues(Ops, dl);
12302 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12303 SelectionDAG &DAG) const {
12304 SDValue Src = Op.getOperand(0);
12305 MVT SrcVT = Src.getSimpleValueType();
12306 MVT VT = Op.getSimpleValueType();
12309 if (SrcVT.isVector()) {
12310 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12311 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12312 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12313 DAG.getUNDEF(SrcVT)));
12315 if (SrcVT.getVectorElementType() == MVT::i1) {
12316 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12317 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12318 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12323 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12324 "Unknown SINT_TO_FP to lower!");
12326 // These are really Legal; return the operand so the caller accepts it as
12328 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12330 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12331 Subtarget->is64Bit()) {
12335 unsigned Size = SrcVT.getSizeInBits()/8;
12336 MachineFunction &MF = DAG.getMachineFunction();
12337 auto PtrVT = getPointerTy(MF.getDataLayout());
12338 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12339 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12340 SDValue Chain = DAG.getStore(
12341 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12342 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12344 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12347 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12349 SelectionDAG &DAG) const {
12353 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12355 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12357 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12359 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12361 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12362 MachineMemOperand *MMO;
12364 int SSFI = FI->getIndex();
12365 MMO = DAG.getMachineFunction().getMachineMemOperand(
12366 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12367 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12369 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12370 StackSlot = StackSlot.getOperand(1);
12372 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12373 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12375 Tys, Ops, SrcVT, MMO);
12378 Chain = Result.getValue(1);
12379 SDValue InFlag = Result.getValue(2);
12381 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12382 // shouldn't be necessary except that RFP cannot be live across
12383 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12384 MachineFunction &MF = DAG.getMachineFunction();
12385 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12386 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12387 auto PtrVT = getPointerTy(MF.getDataLayout());
12388 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12389 Tys = DAG.getVTList(MVT::Other);
12391 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12393 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12394 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12395 MachineMemOperand::MOStore, SSFISize, SSFISize);
12397 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12398 Ops, Op.getValueType(), MMO);
12399 Result = DAG.getLoad(
12400 Op.getValueType(), DL, Chain, StackSlot,
12401 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12402 false, false, false, 0);
12408 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12409 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12410 SelectionDAG &DAG) const {
12411 // This algorithm is not obvious. Here it is what we're trying to output:
12414 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12415 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12417 haddpd %xmm0, %xmm0
12419 pshufd $0x4e, %xmm0, %xmm1
12425 LLVMContext *Context = DAG.getContext();
12427 // Build some magic constants.
12428 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12429 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12430 auto PtrVT = getPointerTy(DAG.getDataLayout());
12431 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12433 SmallVector<Constant*,2> CV1;
12435 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12436 APInt(64, 0x4330000000000000ULL))));
12438 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12439 APInt(64, 0x4530000000000000ULL))));
12440 Constant *C1 = ConstantVector::get(CV1);
12441 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12443 // Load the 64-bit value into an XMM register.
12444 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12447 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12448 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12449 false, false, false, 16);
12451 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12454 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12455 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12456 false, false, false, 16);
12457 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12458 // TODO: Are there any fast-math-flags to propagate here?
12459 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12462 if (Subtarget->hasSSE3()) {
12463 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12464 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12466 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12467 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12469 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12470 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12473 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12474 DAG.getIntPtrConstant(0, dl));
12477 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12478 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12479 SelectionDAG &DAG) const {
12481 // FP constant to bias correct the final result.
12482 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12485 // Load the 32-bit value into an XMM register.
12486 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12489 // Zero out the upper parts of the register.
12490 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12492 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12493 DAG.getBitcast(MVT::v2f64, Load),
12494 DAG.getIntPtrConstant(0, dl));
12496 // Or the load with the bias.
12497 SDValue Or = DAG.getNode(
12498 ISD::OR, dl, MVT::v2i64,
12499 DAG.getBitcast(MVT::v2i64,
12500 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12501 DAG.getBitcast(MVT::v2i64,
12502 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12504 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12505 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12507 // Subtract the bias.
12508 // TODO: Are there any fast-math-flags to propagate here?
12509 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12511 // Handle final rounding.
12512 EVT DestVT = Op.getValueType();
12514 if (DestVT.bitsLT(MVT::f64))
12515 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12516 DAG.getIntPtrConstant(0, dl));
12517 if (DestVT.bitsGT(MVT::f64))
12518 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12520 // Handle final rounding.
12524 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12525 const X86Subtarget &Subtarget) {
12526 // The algorithm is the following:
12527 // #ifdef __SSE4_1__
12528 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12529 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12530 // (uint4) 0x53000000, 0xaa);
12532 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12533 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12535 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12536 // return (float4) lo + fhi;
12538 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12539 // reassociate the two FADDs, and if we do that, the algorithm fails
12540 // spectacularly (PR24512).
12541 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12542 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12543 // there's also the MachineCombiner reassociations happening on Machine IR.
12544 if (DAG.getTarget().Options.UnsafeFPMath)
12548 SDValue V = Op->getOperand(0);
12549 EVT VecIntVT = V.getValueType();
12550 bool Is128 = VecIntVT == MVT::v4i32;
12551 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12552 // If we convert to something else than the supported type, e.g., to v4f64,
12554 if (VecFloatVT != Op->getValueType(0))
12557 unsigned NumElts = VecIntVT.getVectorNumElements();
12558 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12559 "Unsupported custom type");
12560 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12562 // In the #idef/#else code, we have in common:
12563 // - The vector of constants:
12569 // Create the splat vector for 0x4b000000.
12570 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12571 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12572 CstLow, CstLow, CstLow, CstLow};
12573 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12574 makeArrayRef(&CstLowArray[0], NumElts));
12575 // Create the splat vector for 0x53000000.
12576 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12577 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12578 CstHigh, CstHigh, CstHigh, CstHigh};
12579 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12580 makeArrayRef(&CstHighArray[0], NumElts));
12582 // Create the right shift.
12583 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12584 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12585 CstShift, CstShift, CstShift, CstShift};
12586 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12587 makeArrayRef(&CstShiftArray[0], NumElts));
12588 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12591 if (Subtarget.hasSSE41()) {
12592 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12593 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12594 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12595 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12596 // Low will be bitcasted right away, so do not bother bitcasting back to its
12598 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12599 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12600 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12601 // (uint4) 0x53000000, 0xaa);
12602 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12603 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12604 // High will be bitcasted right away, so do not bother bitcasting back to
12605 // its original type.
12606 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12607 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12609 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12610 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12611 CstMask, CstMask, CstMask);
12612 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12613 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12614 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12616 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12617 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12620 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12621 SDValue CstFAdd = DAG.getConstantFP(
12622 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12623 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12624 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12625 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12626 makeArrayRef(&CstFAddArray[0], NumElts));
12628 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12629 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12630 // TODO: Are there any fast-math-flags to propagate here?
12632 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12633 // return (float4) lo + fhi;
12634 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12635 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12638 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12639 SelectionDAG &DAG) const {
12640 SDValue N0 = Op.getOperand(0);
12641 MVT SVT = N0.getSimpleValueType();
12644 switch (SVT.SimpleTy) {
12646 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12651 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12652 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12653 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12657 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12660 if (Subtarget->hasAVX512())
12661 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12662 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12664 llvm_unreachable(nullptr);
12667 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12668 SelectionDAG &DAG) const {
12669 SDValue N0 = Op.getOperand(0);
12671 auto PtrVT = getPointerTy(DAG.getDataLayout());
12673 if (Op.getValueType().isVector())
12674 return lowerUINT_TO_FP_vec(Op, DAG);
12676 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12677 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12678 // the optimization here.
12679 if (DAG.SignBitIsZero(N0))
12680 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12682 MVT SrcVT = N0.getSimpleValueType();
12683 MVT DstVT = Op.getSimpleValueType();
12685 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12686 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12687 // Conversions from unsigned i32 to f32/f64 are legal,
12688 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12692 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12693 return LowerUINT_TO_FP_i64(Op, DAG);
12694 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12695 return LowerUINT_TO_FP_i32(Op, DAG);
12696 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12699 // Make a 64-bit buffer, and use it to build an FILD.
12700 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12701 if (SrcVT == MVT::i32) {
12702 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12703 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12704 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12705 StackSlot, MachinePointerInfo(),
12707 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12708 OffsetSlot, MachinePointerInfo(),
12710 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12714 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12715 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12716 StackSlot, MachinePointerInfo(),
12718 // For i64 source, we need to add the appropriate power of 2 if the input
12719 // was negative. This is the same as the optimization in
12720 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12721 // we must be careful to do the computation in x87 extended precision, not
12722 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12723 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12724 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12725 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12726 MachineMemOperand::MOLoad, 8, 8);
12728 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12729 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12730 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12733 APInt FF(32, 0x5F800000ULL);
12735 // Check whether the sign bit is set.
12736 SDValue SignSet = DAG.getSetCC(
12737 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12738 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12740 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12741 SDValue FudgePtr = DAG.getConstantPool(
12742 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12744 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12745 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12746 SDValue Four = DAG.getIntPtrConstant(4, dl);
12747 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12749 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12751 // Load the value out, extending it from f32 to f80.
12752 // FIXME: Avoid the extend by constructing the right constant pool?
12753 SDValue Fudge = DAG.getExtLoad(
12754 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12755 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12756 false, false, false, 4);
12757 // Extend everything to 80 bits to force it to be done on x87.
12758 // TODO: Are there any fast-math-flags to propagate here?
12759 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12760 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12761 DAG.getIntPtrConstant(0, dl));
12764 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12765 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12766 // just return an <SDValue(), SDValue()> pair.
12767 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12768 // to i16, i32 or i64, and we lower it to a legal sequence.
12769 // If lowered to the final integer result we return a <result, SDValue()> pair.
12770 // Otherwise we lower it to a sequence ending with a FIST, return a
12771 // <FIST, StackSlot> pair, and the caller is responsible for loading
12772 // the final integer result from StackSlot.
12773 std::pair<SDValue,SDValue>
12774 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12775 bool IsSigned, bool IsReplace) const {
12778 EVT DstTy = Op.getValueType();
12779 EVT TheVT = Op.getOperand(0).getValueType();
12780 auto PtrVT = getPointerTy(DAG.getDataLayout());
12782 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12783 // f16 must be promoted before using the lowering in this routine.
12784 // fp128 does not use this lowering.
12785 return std::make_pair(SDValue(), SDValue());
12788 // If using FIST to compute an unsigned i64, we'll need some fixup
12789 // to handle values above the maximum signed i64. A FIST is always
12790 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12791 bool UnsignedFixup = !IsSigned &&
12792 DstTy == MVT::i64 &&
12793 (!Subtarget->is64Bit() ||
12794 !isScalarFPTypeInSSEReg(TheVT));
12796 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12797 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12798 // The low 32 bits of the fist result will have the correct uint32 result.
12799 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12803 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12804 DstTy.getSimpleVT() >= MVT::i16 &&
12805 "Unknown FP_TO_INT to lower!");
12807 // These are really Legal.
12808 if (DstTy == MVT::i32 &&
12809 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12810 return std::make_pair(SDValue(), SDValue());
12811 if (Subtarget->is64Bit() &&
12812 DstTy == MVT::i64 &&
12813 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12814 return std::make_pair(SDValue(), SDValue());
12816 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12818 MachineFunction &MF = DAG.getMachineFunction();
12819 unsigned MemSize = DstTy.getSizeInBits()/8;
12820 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12821 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12824 switch (DstTy.getSimpleVT().SimpleTy) {
12825 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12826 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12827 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12828 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12831 SDValue Chain = DAG.getEntryNode();
12832 SDValue Value = Op.getOperand(0);
12833 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12835 if (UnsignedFixup) {
12837 // Conversion to unsigned i64 is implemented with a select,
12838 // depending on whether the source value fits in the range
12839 // of a signed i64. Let Thresh be the FP equivalent of
12840 // 0x8000000000000000ULL.
12842 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12843 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12844 // Fist-to-mem64 FistSrc
12845 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12846 // to XOR'ing the high 32 bits with Adjust.
12848 // Being a power of 2, Thresh is exactly representable in all FP formats.
12849 // For X87 we'd like to use the smallest FP type for this constant, but
12850 // for DAG type consistency we have to match the FP operand type.
12852 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12853 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12854 bool LosesInfo = false;
12855 if (TheVT == MVT::f64)
12856 // The rounding mode is irrelevant as the conversion should be exact.
12857 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12859 else if (TheVT == MVT::f80)
12860 Status = Thresh.convert(APFloat::x87DoubleExtended,
12861 APFloat::rmNearestTiesToEven, &LosesInfo);
12863 assert(Status == APFloat::opOK && !LosesInfo &&
12864 "FP conversion should have been exact");
12866 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12868 SDValue Cmp = DAG.getSetCC(DL,
12869 getSetCCResultType(DAG.getDataLayout(),
12870 *DAG.getContext(), TheVT),
12871 Value, ThreshVal, ISD::SETLT);
12872 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12873 DAG.getConstant(0, DL, MVT::i32),
12874 DAG.getConstant(0x80000000, DL, MVT::i32));
12875 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12876 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12877 *DAG.getContext(), TheVT),
12878 Value, ThreshVal, ISD::SETLT);
12879 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12882 // FIXME This causes a redundant load/store if the SSE-class value is already
12883 // in memory, such as if it is on the callstack.
12884 if (isScalarFPTypeInSSEReg(TheVT)) {
12885 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12886 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12887 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12889 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12891 Chain, StackSlot, DAG.getValueType(TheVT)
12894 MachineMemOperand *MMO =
12895 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12896 MachineMemOperand::MOLoad, MemSize, MemSize);
12897 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12898 Chain = Value.getValue(1);
12899 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12900 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12903 MachineMemOperand *MMO =
12904 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12905 MachineMemOperand::MOStore, MemSize, MemSize);
12907 if (UnsignedFixup) {
12909 // Insert the FIST, load its result as two i32's,
12910 // and XOR the high i32 with Adjust.
12912 SDValue FistOps[] = { Chain, Value, StackSlot };
12913 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12914 FistOps, DstTy, MMO);
12916 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12917 MachinePointerInfo(),
12918 false, false, false, 0);
12919 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12920 DAG.getConstant(4, DL, PtrVT));
12922 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12923 MachinePointerInfo(),
12924 false, false, false, 0);
12925 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12927 if (Subtarget->is64Bit()) {
12928 // Join High32 and Low32 into a 64-bit result.
12929 // (High32 << 32) | Low32
12930 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12931 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12932 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12933 DAG.getConstant(32, DL, MVT::i8));
12934 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12935 return std::make_pair(Result, SDValue());
12938 SDValue ResultOps[] = { Low32, High32 };
12940 SDValue pair = IsReplace
12941 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12942 : DAG.getMergeValues(ResultOps, DL);
12943 return std::make_pair(pair, SDValue());
12945 // Build the FP_TO_INT*_IN_MEM
12946 SDValue Ops[] = { Chain, Value, StackSlot };
12947 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12949 return std::make_pair(FIST, StackSlot);
12953 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12954 const X86Subtarget *Subtarget) {
12955 MVT VT = Op->getSimpleValueType(0);
12956 SDValue In = Op->getOperand(0);
12957 MVT InVT = In.getSimpleValueType();
12960 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12961 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12963 // Optimize vectors in AVX mode:
12966 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12967 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12968 // Concat upper and lower parts.
12971 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12972 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12973 // Concat upper and lower parts.
12976 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12977 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12978 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12981 if (Subtarget->hasInt256())
12982 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12984 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12985 SDValue Undef = DAG.getUNDEF(InVT);
12986 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12987 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12988 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12990 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12991 VT.getVectorNumElements()/2);
12993 OpLo = DAG.getBitcast(HVT, OpLo);
12994 OpHi = DAG.getBitcast(HVT, OpHi);
12996 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12999 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13000 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13001 MVT VT = Op->getSimpleValueType(0);
13002 SDValue In = Op->getOperand(0);
13003 MVT InVT = In.getSimpleValueType();
13005 unsigned int NumElts = VT.getVectorNumElements();
13006 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13009 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13010 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13012 assert(InVT.getVectorElementType() == MVT::i1);
13013 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13015 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13017 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13019 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13020 if (VT.is512BitVector())
13022 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13025 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13026 SelectionDAG &DAG) {
13027 if (Subtarget->hasFp256())
13028 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13034 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13035 SelectionDAG &DAG) {
13037 MVT VT = Op.getSimpleValueType();
13038 SDValue In = Op.getOperand(0);
13039 MVT SVT = In.getSimpleValueType();
13041 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13042 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13044 if (Subtarget->hasFp256())
13045 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13048 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13049 VT.getVectorNumElements() != SVT.getVectorNumElements());
13053 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13055 MVT VT = Op.getSimpleValueType();
13056 SDValue In = Op.getOperand(0);
13057 MVT InVT = In.getSimpleValueType();
13059 if (VT == MVT::i1) {
13060 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13061 "Invalid scalar TRUNCATE operation");
13062 if (InVT.getSizeInBits() >= 32)
13064 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13065 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13067 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13068 "Invalid TRUNCATE operation");
13070 // move vector to mask - truncate solution for SKX
13071 if (VT.getVectorElementType() == MVT::i1) {
13072 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13073 Subtarget->hasBWI())
13074 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13075 if ((InVT.is256BitVector() || InVT.is128BitVector())
13076 && InVT.getScalarSizeInBits() <= 16 &&
13077 Subtarget->hasBWI() && Subtarget->hasVLX())
13078 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13079 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13080 Subtarget->hasDQI())
13081 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13082 if ((InVT.is256BitVector() || InVT.is128BitVector())
13083 && InVT.getScalarSizeInBits() >= 32 &&
13084 Subtarget->hasDQI() && Subtarget->hasVLX())
13085 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13088 if (VT.getVectorElementType() == MVT::i1) {
13089 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13090 unsigned NumElts = InVT.getVectorNumElements();
13091 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13092 if (InVT.getSizeInBits() < 512) {
13093 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13094 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13099 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13100 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13101 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13104 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13105 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
13106 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
13107 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13109 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13110 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13111 if (Subtarget->hasInt256()) {
13112 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13113 In = DAG.getBitcast(MVT::v8i32, In);
13114 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13116 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13117 DAG.getIntPtrConstant(0, DL));
13120 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13121 DAG.getIntPtrConstant(0, DL));
13122 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13123 DAG.getIntPtrConstant(2, DL));
13124 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13125 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13126 static const int ShufMask[] = {0, 2, 4, 6};
13127 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13130 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13131 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13132 if (Subtarget->hasInt256()) {
13133 In = DAG.getBitcast(MVT::v32i8, In);
13135 SmallVector<SDValue,32> pshufbMask;
13136 for (unsigned i = 0; i < 2; ++i) {
13137 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13138 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13139 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13140 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13141 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13142 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13143 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13144 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13145 for (unsigned j = 0; j < 8; ++j)
13146 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13148 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13149 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13150 In = DAG.getBitcast(MVT::v4i64, In);
13152 static const int ShufMask[] = {0, 2, -1, -1};
13153 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13155 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13156 DAG.getIntPtrConstant(0, DL));
13157 return DAG.getBitcast(VT, In);
13160 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13161 DAG.getIntPtrConstant(0, DL));
13163 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13164 DAG.getIntPtrConstant(4, DL));
13166 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13167 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13169 // The PSHUFB mask:
13170 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13171 -1, -1, -1, -1, -1, -1, -1, -1};
13173 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13174 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13175 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13177 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13178 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13180 // The MOVLHPS Mask:
13181 static const int ShufMask2[] = {0, 1, 4, 5};
13182 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13183 return DAG.getBitcast(MVT::v8i16, res);
13186 // Handle truncation of V256 to V128 using shuffles.
13187 if (!VT.is128BitVector() || !InVT.is256BitVector())
13190 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13192 unsigned NumElems = VT.getVectorNumElements();
13193 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13195 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13196 // Prepare truncation shuffle mask
13197 for (unsigned i = 0; i != NumElems; ++i)
13198 MaskVec[i] = i * 2;
13199 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13200 DAG.getUNDEF(NVT), &MaskVec[0]);
13201 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13202 DAG.getIntPtrConstant(0, DL));
13205 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13206 SelectionDAG &DAG) const {
13207 assert(!Op.getSimpleValueType().isVector());
13209 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13210 /*IsSigned=*/ true, /*IsReplace=*/ false);
13211 SDValue FIST = Vals.first, StackSlot = Vals.second;
13212 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13213 if (!FIST.getNode())
13216 if (StackSlot.getNode())
13217 // Load the result.
13218 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13219 FIST, StackSlot, MachinePointerInfo(),
13220 false, false, false, 0);
13222 // The node is the result.
13226 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13227 SelectionDAG &DAG) const {
13228 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13229 /*IsSigned=*/ false, /*IsReplace=*/ false);
13230 SDValue FIST = Vals.first, StackSlot = Vals.second;
13231 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13232 if (!FIST.getNode())
13235 if (StackSlot.getNode())
13236 // Load the result.
13237 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13238 FIST, StackSlot, MachinePointerInfo(),
13239 false, false, false, 0);
13241 // The node is the result.
13245 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13247 MVT VT = Op.getSimpleValueType();
13248 SDValue In = Op.getOperand(0);
13249 MVT SVT = In.getSimpleValueType();
13251 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13253 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13254 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13255 In, DAG.getUNDEF(SVT)));
13258 /// The only differences between FABS and FNEG are the mask and the logic op.
13259 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13260 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13261 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13262 "Wrong opcode for lowering FABS or FNEG.");
13264 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13266 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13267 // into an FNABS. We'll lower the FABS after that if it is still in use.
13269 for (SDNode *User : Op->uses())
13270 if (User->getOpcode() == ISD::FNEG)
13274 MVT VT = Op.getSimpleValueType();
13276 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13277 // decide if we should generate a 16-byte constant mask when we only need 4 or
13278 // 8 bytes for the scalar case.
13284 if (VT.isVector()) {
13286 EltVT = VT.getVectorElementType();
13287 NumElts = VT.getVectorNumElements();
13289 // There are no scalar bitwise logical SSE/AVX instructions, so we
13290 // generate a 16-byte vector constant and logic op even for the scalar case.
13291 // Using a 16-byte mask allows folding the load of the mask with
13292 // the logic op, so it can save (~4 bytes) on code size.
13293 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13295 NumElts = (VT == MVT::f64) ? 2 : 4;
13298 unsigned EltBits = EltVT.getSizeInBits();
13299 LLVMContext *Context = DAG.getContext();
13300 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13302 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13303 Constant *C = ConstantInt::get(*Context, MaskElt);
13304 C = ConstantVector::getSplat(NumElts, C);
13305 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13306 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13307 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13309 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13310 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13311 false, false, false, Alignment);
13313 SDValue Op0 = Op.getOperand(0);
13314 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13316 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13317 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13320 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13322 // For the scalar case extend to a 128-bit vector, perform the logic op,
13323 // and extract the scalar result back out.
13324 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13325 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13326 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13327 DAG.getIntPtrConstant(0, dl));
13330 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13331 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13332 LLVMContext *Context = DAG.getContext();
13333 SDValue Op0 = Op.getOperand(0);
13334 SDValue Op1 = Op.getOperand(1);
13336 MVT VT = Op.getSimpleValueType();
13337 MVT SrcVT = Op1.getSimpleValueType();
13339 // If second operand is smaller, extend it first.
13340 if (SrcVT.bitsLT(VT)) {
13341 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13344 // And if it is bigger, shrink it first.
13345 if (SrcVT.bitsGT(VT)) {
13346 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13350 // At this point the operands and the result should have the same
13351 // type, and that won't be f80 since that is not custom lowered.
13353 const fltSemantics &Sem =
13354 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13355 const unsigned SizeInBits = VT.getSizeInBits();
13357 SmallVector<Constant *, 4> CV(
13358 VT == MVT::f64 ? 2 : 4,
13359 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13361 // First, clear all bits but the sign bit from the second operand (sign).
13362 CV[0] = ConstantFP::get(*Context,
13363 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13364 Constant *C = ConstantVector::get(CV);
13365 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13366 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13368 // Perform all logic operations as 16-byte vectors because there are no
13369 // scalar FP logic instructions in SSE. This allows load folding of the
13370 // constants into the logic instructions.
13371 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13373 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13374 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13375 false, false, false, 16);
13376 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13377 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13379 // Next, clear the sign bit from the first operand (magnitude).
13380 // If it's a constant, we can clear it here.
13381 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13382 APFloat APF = Op0CN->getValueAPF();
13383 // If the magnitude is a positive zero, the sign bit alone is enough.
13384 if (APF.isPosZero())
13385 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13386 DAG.getIntPtrConstant(0, dl));
13388 CV[0] = ConstantFP::get(*Context, APF);
13390 CV[0] = ConstantFP::get(
13392 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13394 C = ConstantVector::get(CV);
13395 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13397 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13398 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13399 false, false, false, 16);
13400 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13401 if (!isa<ConstantFPSDNode>(Op0)) {
13402 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13403 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13405 // OR the magnitude value with the sign bit.
13406 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13407 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13408 DAG.getIntPtrConstant(0, dl));
13411 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13412 SDValue N0 = Op.getOperand(0);
13414 MVT VT = Op.getSimpleValueType();
13416 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13417 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13418 DAG.getConstant(1, dl, VT));
13419 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13422 // Check whether an OR'd tree is PTEST-able.
13423 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13424 SelectionDAG &DAG) {
13425 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13427 if (!Subtarget->hasSSE41())
13430 if (!Op->hasOneUse())
13433 SDNode *N = Op.getNode();
13436 SmallVector<SDValue, 8> Opnds;
13437 DenseMap<SDValue, unsigned> VecInMap;
13438 SmallVector<SDValue, 8> VecIns;
13439 EVT VT = MVT::Other;
13441 // Recognize a special case where a vector is casted into wide integer to
13443 Opnds.push_back(N->getOperand(0));
13444 Opnds.push_back(N->getOperand(1));
13446 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13447 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13448 // BFS traverse all OR'd operands.
13449 if (I->getOpcode() == ISD::OR) {
13450 Opnds.push_back(I->getOperand(0));
13451 Opnds.push_back(I->getOperand(1));
13452 // Re-evaluate the number of nodes to be traversed.
13453 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13457 // Quit if a non-EXTRACT_VECTOR_ELT
13458 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13461 // Quit if without a constant index.
13462 SDValue Idx = I->getOperand(1);
13463 if (!isa<ConstantSDNode>(Idx))
13466 SDValue ExtractedFromVec = I->getOperand(0);
13467 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13468 if (M == VecInMap.end()) {
13469 VT = ExtractedFromVec.getValueType();
13470 // Quit if not 128/256-bit vector.
13471 if (!VT.is128BitVector() && !VT.is256BitVector())
13473 // Quit if not the same type.
13474 if (VecInMap.begin() != VecInMap.end() &&
13475 VT != VecInMap.begin()->first.getValueType())
13477 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13478 VecIns.push_back(ExtractedFromVec);
13480 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13483 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13484 "Not extracted from 128-/256-bit vector.");
13486 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13488 for (DenseMap<SDValue, unsigned>::const_iterator
13489 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13490 // Quit if not all elements are used.
13491 if (I->second != FullMask)
13495 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13497 // Cast all vectors into TestVT for PTEST.
13498 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13499 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13501 // If more than one full vectors are evaluated, OR them first before PTEST.
13502 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13503 // Each iteration will OR 2 nodes and append the result until there is only
13504 // 1 node left, i.e. the final OR'd value of all vectors.
13505 SDValue LHS = VecIns[Slot];
13506 SDValue RHS = VecIns[Slot + 1];
13507 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13510 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13511 VecIns.back(), VecIns.back());
13514 /// \brief return true if \c Op has a use that doesn't just read flags.
13515 static bool hasNonFlagsUse(SDValue Op) {
13516 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13518 SDNode *User = *UI;
13519 unsigned UOpNo = UI.getOperandNo();
13520 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13521 // Look pass truncate.
13522 UOpNo = User->use_begin().getOperandNo();
13523 User = *User->use_begin();
13526 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13527 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13533 /// Emit nodes that will be selected as "test Op0,Op0", or something
13535 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13536 SelectionDAG &DAG) const {
13537 if (Op.getValueType() == MVT::i1) {
13538 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13539 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13540 DAG.getConstant(0, dl, MVT::i8));
13542 // CF and OF aren't always set the way we want. Determine which
13543 // of these we need.
13544 bool NeedCF = false;
13545 bool NeedOF = false;
13548 case X86::COND_A: case X86::COND_AE:
13549 case X86::COND_B: case X86::COND_BE:
13552 case X86::COND_G: case X86::COND_GE:
13553 case X86::COND_L: case X86::COND_LE:
13554 case X86::COND_O: case X86::COND_NO: {
13555 // Check if we really need to set the
13556 // Overflow flag. If NoSignedWrap is present
13557 // that is not actually needed.
13558 switch (Op->getOpcode()) {
13563 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13564 if (BinNode->Flags.hasNoSignedWrap())
13574 // See if we can use the EFLAGS value from the operand instead of
13575 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13576 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13577 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13578 // Emit a CMP with 0, which is the TEST pattern.
13579 //if (Op.getValueType() == MVT::i1)
13580 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13581 // DAG.getConstant(0, MVT::i1));
13582 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13583 DAG.getConstant(0, dl, Op.getValueType()));
13585 unsigned Opcode = 0;
13586 unsigned NumOperands = 0;
13588 // Truncate operations may prevent the merge of the SETCC instruction
13589 // and the arithmetic instruction before it. Attempt to truncate the operands
13590 // of the arithmetic instruction and use a reduced bit-width instruction.
13591 bool NeedTruncation = false;
13592 SDValue ArithOp = Op;
13593 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13594 SDValue Arith = Op->getOperand(0);
13595 // Both the trunc and the arithmetic op need to have one user each.
13596 if (Arith->hasOneUse())
13597 switch (Arith.getOpcode()) {
13604 NeedTruncation = true;
13610 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13611 // which may be the result of a CAST. We use the variable 'Op', which is the
13612 // non-casted variable when we check for possible users.
13613 switch (ArithOp.getOpcode()) {
13615 // Due to an isel shortcoming, be conservative if this add is likely to be
13616 // selected as part of a load-modify-store instruction. When the root node
13617 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13618 // uses of other nodes in the match, such as the ADD in this case. This
13619 // leads to the ADD being left around and reselected, with the result being
13620 // two adds in the output. Alas, even if none our users are stores, that
13621 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13622 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13623 // climbing the DAG back to the root, and it doesn't seem to be worth the
13625 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13626 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13627 if (UI->getOpcode() != ISD::CopyToReg &&
13628 UI->getOpcode() != ISD::SETCC &&
13629 UI->getOpcode() != ISD::STORE)
13632 if (ConstantSDNode *C =
13633 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13634 // An add of one will be selected as an INC.
13635 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13636 Opcode = X86ISD::INC;
13641 // An add of negative one (subtract of one) will be selected as a DEC.
13642 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13643 Opcode = X86ISD::DEC;
13649 // Otherwise use a regular EFLAGS-setting add.
13650 Opcode = X86ISD::ADD;
13655 // If we have a constant logical shift that's only used in a comparison
13656 // against zero turn it into an equivalent AND. This allows turning it into
13657 // a TEST instruction later.
13658 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13659 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13660 EVT VT = Op.getValueType();
13661 unsigned BitWidth = VT.getSizeInBits();
13662 unsigned ShAmt = Op->getConstantOperandVal(1);
13663 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13665 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13666 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13667 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13668 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13670 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13671 DAG.getConstant(Mask, dl, VT));
13672 DAG.ReplaceAllUsesWith(Op, New);
13678 // If the primary and result isn't used, don't bother using X86ISD::AND,
13679 // because a TEST instruction will be better.
13680 if (!hasNonFlagsUse(Op))
13686 // Due to the ISEL shortcoming noted above, be conservative if this op is
13687 // likely to be selected as part of a load-modify-store instruction.
13688 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13689 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13690 if (UI->getOpcode() == ISD::STORE)
13693 // Otherwise use a regular EFLAGS-setting instruction.
13694 switch (ArithOp.getOpcode()) {
13695 default: llvm_unreachable("unexpected operator!");
13696 case ISD::SUB: Opcode = X86ISD::SUB; break;
13697 case ISD::XOR: Opcode = X86ISD::XOR; break;
13698 case ISD::AND: Opcode = X86ISD::AND; break;
13700 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13701 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13702 if (EFLAGS.getNode())
13705 Opcode = X86ISD::OR;
13719 return SDValue(Op.getNode(), 1);
13725 // If we found that truncation is beneficial, perform the truncation and
13727 if (NeedTruncation) {
13728 EVT VT = Op.getValueType();
13729 SDValue WideVal = Op->getOperand(0);
13730 EVT WideVT = WideVal.getValueType();
13731 unsigned ConvertedOp = 0;
13732 // Use a target machine opcode to prevent further DAGCombine
13733 // optimizations that may separate the arithmetic operations
13734 // from the setcc node.
13735 switch (WideVal.getOpcode()) {
13737 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13738 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13739 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13740 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13741 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13745 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13746 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13747 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13748 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13749 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13755 // Emit a CMP with 0, which is the TEST pattern.
13756 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13757 DAG.getConstant(0, dl, Op.getValueType()));
13759 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13760 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13762 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13763 DAG.ReplaceAllUsesWith(Op, New);
13764 return SDValue(New.getNode(), 1);
13767 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13769 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13770 SDLoc dl, SelectionDAG &DAG) const {
13771 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13772 if (C->getAPIntValue() == 0)
13773 return EmitTest(Op0, X86CC, dl, DAG);
13775 if (Op0.getValueType() == MVT::i1)
13776 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13779 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13780 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13781 // Do the comparison at i32 if it's smaller, besides the Atom case.
13782 // This avoids subregister aliasing issues. Keep the smaller reference
13783 // if we're optimizing for size, however, as that'll allow better folding
13784 // of memory operations.
13785 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13786 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13787 !Subtarget->isAtom()) {
13788 unsigned ExtendOp =
13789 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13790 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13791 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13793 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13794 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13795 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13797 return SDValue(Sub.getNode(), 1);
13799 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13802 /// Convert a comparison if required by the subtarget.
13803 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13804 SelectionDAG &DAG) const {
13805 // If the subtarget does not support the FUCOMI instruction, floating-point
13806 // comparisons have to be converted.
13807 if (Subtarget->hasCMov() ||
13808 Cmp.getOpcode() != X86ISD::CMP ||
13809 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13810 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13813 // The instruction selector will select an FUCOM instruction instead of
13814 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13815 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13816 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13818 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13819 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13820 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13821 DAG.getConstant(8, dl, MVT::i8));
13822 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13823 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13826 /// The minimum architected relative accuracy is 2^-12. We need one
13827 /// Newton-Raphson step to have a good float result (24 bits of precision).
13828 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13829 DAGCombinerInfo &DCI,
13830 unsigned &RefinementSteps,
13831 bool &UseOneConstNR) const {
13832 EVT VT = Op.getValueType();
13833 const char *RecipOp;
13835 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13836 // TODO: Add support for AVX512 (v16f32).
13837 // It is likely not profitable to do this for f64 because a double-precision
13838 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13839 // instructions: convert to single, rsqrtss, convert back to double, refine
13840 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13841 // along with FMA, this could be a throughput win.
13842 if (VT == MVT::f32 && Subtarget->hasSSE1())
13844 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13845 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13846 RecipOp = "vec-sqrtf";
13850 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13851 if (!Recips.isEnabled(RecipOp))
13854 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13855 UseOneConstNR = false;
13856 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13859 /// The minimum architected relative accuracy is 2^-12. We need one
13860 /// Newton-Raphson step to have a good float result (24 bits of precision).
13861 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13862 DAGCombinerInfo &DCI,
13863 unsigned &RefinementSteps) const {
13864 EVT VT = Op.getValueType();
13865 const char *RecipOp;
13867 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13868 // TODO: Add support for AVX512 (v16f32).
13869 // It is likely not profitable to do this for f64 because a double-precision
13870 // reciprocal estimate with refinement on x86 prior to FMA requires
13871 // 15 instructions: convert to single, rcpss, convert back to double, refine
13872 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13873 // along with FMA, this could be a throughput win.
13874 if (VT == MVT::f32 && Subtarget->hasSSE1())
13876 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13877 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13878 RecipOp = "vec-divf";
13882 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13883 if (!Recips.isEnabled(RecipOp))
13886 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13887 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13890 /// If we have at least two divisions that use the same divisor, convert to
13891 /// multplication by a reciprocal. This may need to be adjusted for a given
13892 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13893 /// This is because we still need one division to calculate the reciprocal and
13894 /// then we need two multiplies by that reciprocal as replacements for the
13895 /// original divisions.
13896 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13900 static bool isAllOnes(SDValue V) {
13901 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13902 return C && C->isAllOnesValue();
13905 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13906 /// if it's possible.
13907 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13908 SDLoc dl, SelectionDAG &DAG) const {
13909 SDValue Op0 = And.getOperand(0);
13910 SDValue Op1 = And.getOperand(1);
13911 if (Op0.getOpcode() == ISD::TRUNCATE)
13912 Op0 = Op0.getOperand(0);
13913 if (Op1.getOpcode() == ISD::TRUNCATE)
13914 Op1 = Op1.getOperand(0);
13917 if (Op1.getOpcode() == ISD::SHL)
13918 std::swap(Op0, Op1);
13919 if (Op0.getOpcode() == ISD::SHL) {
13920 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13921 if (And00C->getZExtValue() == 1) {
13922 // If we looked past a truncate, check that it's only truncating away
13924 unsigned BitWidth = Op0.getValueSizeInBits();
13925 unsigned AndBitWidth = And.getValueSizeInBits();
13926 if (BitWidth > AndBitWidth) {
13928 DAG.computeKnownBits(Op0, Zeros, Ones);
13929 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13933 RHS = Op0.getOperand(1);
13935 } else if (Op1.getOpcode() == ISD::Constant) {
13936 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13937 uint64_t AndRHSVal = AndRHS->getZExtValue();
13938 SDValue AndLHS = Op0;
13940 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13941 LHS = AndLHS.getOperand(0);
13942 RHS = AndLHS.getOperand(1);
13945 // Use BT if the immediate can't be encoded in a TEST instruction.
13946 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13948 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13952 if (LHS.getNode()) {
13953 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13954 // instruction. Since the shift amount is in-range-or-undefined, we know
13955 // that doing a bittest on the i32 value is ok. We extend to i32 because
13956 // the encoding for the i16 version is larger than the i32 version.
13957 // Also promote i16 to i32 for performance / code size reason.
13958 if (LHS.getValueType() == MVT::i8 ||
13959 LHS.getValueType() == MVT::i16)
13960 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13962 // If the operand types disagree, extend the shift amount to match. Since
13963 // BT ignores high bits (like shifts) we can use anyextend.
13964 if (LHS.getValueType() != RHS.getValueType())
13965 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13967 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13968 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13969 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13970 DAG.getConstant(Cond, dl, MVT::i8), BT);
13976 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13978 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13983 // SSE Condition code mapping:
13992 switch (SetCCOpcode) {
13993 default: llvm_unreachable("Unexpected SETCC condition");
13995 case ISD::SETEQ: SSECC = 0; break;
13997 case ISD::SETGT: Swap = true; // Fallthrough
13999 case ISD::SETOLT: SSECC = 1; break;
14001 case ISD::SETGE: Swap = true; // Fallthrough
14003 case ISD::SETOLE: SSECC = 2; break;
14004 case ISD::SETUO: SSECC = 3; break;
14006 case ISD::SETNE: SSECC = 4; break;
14007 case ISD::SETULE: Swap = true; // Fallthrough
14008 case ISD::SETUGE: SSECC = 5; break;
14009 case ISD::SETULT: Swap = true; // Fallthrough
14010 case ISD::SETUGT: SSECC = 6; break;
14011 case ISD::SETO: SSECC = 7; break;
14013 case ISD::SETONE: SSECC = 8; break;
14016 std::swap(Op0, Op1);
14021 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14022 // ones, and then concatenate the result back.
14023 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14024 MVT VT = Op.getSimpleValueType();
14026 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14027 "Unsupported value type for operation");
14029 unsigned NumElems = VT.getVectorNumElements();
14031 SDValue CC = Op.getOperand(2);
14033 // Extract the LHS vectors
14034 SDValue LHS = Op.getOperand(0);
14035 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14036 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14038 // Extract the RHS vectors
14039 SDValue RHS = Op.getOperand(1);
14040 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14041 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14043 // Issue the operation on the smaller types and concatenate the result back
14044 MVT EltVT = VT.getVectorElementType();
14045 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14046 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14047 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14048 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14051 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14052 SDValue Op0 = Op.getOperand(0);
14053 SDValue Op1 = Op.getOperand(1);
14054 SDValue CC = Op.getOperand(2);
14055 MVT VT = Op.getSimpleValueType();
14058 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
14059 "Unexpected type for boolean compare operation");
14060 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14061 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14062 DAG.getConstant(-1, dl, VT));
14063 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14064 DAG.getConstant(-1, dl, VT));
14065 switch (SetCCOpcode) {
14066 default: llvm_unreachable("Unexpected SETCC condition");
14068 // (x == y) -> ~(x ^ y)
14069 return DAG.getNode(ISD::XOR, dl, VT,
14070 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14071 DAG.getConstant(-1, dl, VT));
14073 // (x != y) -> (x ^ y)
14074 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14077 // (x > y) -> (x & ~y)
14078 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14081 // (x < y) -> (~x & y)
14082 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14085 // (x <= y) -> (~x | y)
14086 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14089 // (x >=y) -> (x | ~y)
14090 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14094 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14095 const X86Subtarget *Subtarget) {
14096 SDValue Op0 = Op.getOperand(0);
14097 SDValue Op1 = Op.getOperand(1);
14098 SDValue CC = Op.getOperand(2);
14099 MVT VT = Op.getSimpleValueType();
14102 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14103 Op.getValueType().getScalarType() == MVT::i1 &&
14104 "Cannot set masked compare for this operation");
14106 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14108 bool Unsigned = false;
14111 switch (SetCCOpcode) {
14112 default: llvm_unreachable("Unexpected SETCC condition");
14113 case ISD::SETNE: SSECC = 4; break;
14114 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14115 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14116 case ISD::SETLT: Swap = true; //fall-through
14117 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14118 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14119 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14120 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14121 case ISD::SETULE: Unsigned = true; //fall-through
14122 case ISD::SETLE: SSECC = 2; break;
14126 std::swap(Op0, Op1);
14128 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14129 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14130 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14131 DAG.getConstant(SSECC, dl, MVT::i8));
14134 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14135 /// operand \p Op1. If non-trivial (for example because it's not constant)
14136 /// return an empty value.
14137 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14139 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14143 MVT VT = Op1.getSimpleValueType();
14144 MVT EVT = VT.getVectorElementType();
14145 unsigned n = VT.getVectorNumElements();
14146 SmallVector<SDValue, 8> ULTOp1;
14148 for (unsigned i = 0; i < n; ++i) {
14149 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14150 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14153 // Avoid underflow.
14154 APInt Val = Elt->getAPIntValue();
14158 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14161 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14164 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14165 SelectionDAG &DAG) {
14166 SDValue Op0 = Op.getOperand(0);
14167 SDValue Op1 = Op.getOperand(1);
14168 SDValue CC = Op.getOperand(2);
14169 MVT VT = Op.getSimpleValueType();
14170 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14171 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14176 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14177 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14180 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14181 unsigned Opc = X86ISD::CMPP;
14182 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14183 assert(VT.getVectorNumElements() <= 16);
14184 Opc = X86ISD::CMPM;
14186 // In the two special cases we can't handle, emit two comparisons.
14189 unsigned CombineOpc;
14190 if (SetCCOpcode == ISD::SETUEQ) {
14191 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14193 assert(SetCCOpcode == ISD::SETONE);
14194 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14197 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14198 DAG.getConstant(CC0, dl, MVT::i8));
14199 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14200 DAG.getConstant(CC1, dl, MVT::i8));
14201 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14203 // Handle all other FP comparisons here.
14204 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14205 DAG.getConstant(SSECC, dl, MVT::i8));
14208 MVT VTOp0 = Op0.getSimpleValueType();
14209 assert(VTOp0 == Op1.getSimpleValueType() &&
14210 "Expected operands with same type!");
14211 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14212 "Invalid number of packed elements for source and destination!");
14214 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14215 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14216 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14217 // legalizer firstly checks if the first operand in input to the setcc has
14218 // a legal type. If so, then it promotes the return type to that same type.
14219 // Otherwise, the return type is promoted to the 'next legal type' which,
14220 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14222 // We reach this code only if the following two conditions are met:
14223 // 1. Both return type and operand type have been promoted to wider types
14224 // by the type legalizer.
14225 // 2. The original operand type has been promoted to a 256-bit vector.
14227 // Note that condition 2. only applies for AVX targets.
14228 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14229 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14232 // The non-AVX512 code below works under the assumption that source and
14233 // destination types are the same.
14234 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14235 "Value types for source and destination must be the same!");
14237 // Break 256-bit integer vector compare into smaller ones.
14238 if (VT.is256BitVector() && !Subtarget->hasInt256())
14239 return Lower256IntVSETCC(Op, DAG);
14241 EVT OpVT = Op1.getValueType();
14242 if (OpVT.getVectorElementType() == MVT::i1)
14243 return LowerBoolVSETCC_AVX512(Op, DAG);
14245 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14246 if (Subtarget->hasAVX512()) {
14247 if (Op1.getValueType().is512BitVector() ||
14248 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14249 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14250 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14252 // In AVX-512 architecture setcc returns mask with i1 elements,
14253 // But there is no compare instruction for i8 and i16 elements in KNL.
14254 // We are not talking about 512-bit operands in this case, these
14255 // types are illegal.
14257 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14258 OpVT.getVectorElementType().getSizeInBits() >= 8))
14259 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14260 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14263 // Lower using XOP integer comparisons.
14264 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14265 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14266 // Translate compare code to XOP PCOM compare mode.
14267 unsigned CmpMode = 0;
14268 switch (SetCCOpcode) {
14269 default: llvm_unreachable("Unexpected SETCC condition");
14271 case ISD::SETLT: CmpMode = 0x00; break;
14273 case ISD::SETLE: CmpMode = 0x01; break;
14275 case ISD::SETGT: CmpMode = 0x02; break;
14277 case ISD::SETGE: CmpMode = 0x03; break;
14278 case ISD::SETEQ: CmpMode = 0x04; break;
14279 case ISD::SETNE: CmpMode = 0x05; break;
14282 // Are we comparing unsigned or signed integers?
14283 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14284 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14286 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14287 DAG.getConstant(CmpMode, dl, MVT::i8));
14290 // We are handling one of the integer comparisons here. Since SSE only has
14291 // GT and EQ comparisons for integer, swapping operands and multiple
14292 // operations may be required for some comparisons.
14294 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14295 bool Subus = false;
14297 switch (SetCCOpcode) {
14298 default: llvm_unreachable("Unexpected SETCC condition");
14299 case ISD::SETNE: Invert = true;
14300 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14301 case ISD::SETLT: Swap = true;
14302 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14303 case ISD::SETGE: Swap = true;
14304 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14305 Invert = true; break;
14306 case ISD::SETULT: Swap = true;
14307 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14308 FlipSigns = true; break;
14309 case ISD::SETUGE: Swap = true;
14310 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14311 FlipSigns = true; Invert = true; break;
14314 // Special case: Use min/max operations for SETULE/SETUGE
14315 MVT VET = VT.getVectorElementType();
14317 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14318 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14321 switch (SetCCOpcode) {
14323 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14324 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14327 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14330 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14331 if (!MinMax && hasSubus) {
14332 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14334 // t = psubus Op0, Op1
14335 // pcmpeq t, <0..0>
14336 switch (SetCCOpcode) {
14338 case ISD::SETULT: {
14339 // If the comparison is against a constant we can turn this into a
14340 // setule. With psubus, setule does not require a swap. This is
14341 // beneficial because the constant in the register is no longer
14342 // destructed as the destination so it can be hoisted out of a loop.
14343 // Only do this pre-AVX since vpcmp* is no longer destructive.
14344 if (Subtarget->hasAVX())
14346 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14347 if (ULEOp1.getNode()) {
14349 Subus = true; Invert = false; Swap = false;
14353 // Psubus is better than flip-sign because it requires no inversion.
14354 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14355 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14359 Opc = X86ISD::SUBUS;
14365 std::swap(Op0, Op1);
14367 // Check that the operation in question is available (most are plain SSE2,
14368 // but PCMPGTQ and PCMPEQQ have different requirements).
14369 if (VT == MVT::v2i64) {
14370 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14371 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14373 // First cast everything to the right type.
14374 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14375 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14377 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14378 // bits of the inputs before performing those operations. The lower
14379 // compare is always unsigned.
14382 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14384 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14385 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14386 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14387 Sign, Zero, Sign, Zero);
14389 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14390 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14392 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14393 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14394 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14396 // Create masks for only the low parts/high parts of the 64 bit integers.
14397 static const int MaskHi[] = { 1, 1, 3, 3 };
14398 static const int MaskLo[] = { 0, 0, 2, 2 };
14399 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14400 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14401 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14403 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14404 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14407 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14409 return DAG.getBitcast(VT, Result);
14412 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14413 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14414 // pcmpeqd + pshufd + pand.
14415 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14417 // First cast everything to the right type.
14418 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14419 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14422 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14424 // Make sure the lower and upper halves are both all-ones.
14425 static const int Mask[] = { 1, 0, 3, 2 };
14426 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14427 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14430 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14432 return DAG.getBitcast(VT, Result);
14436 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14437 // bits of the inputs before performing those operations.
14439 MVT EltVT = VT.getVectorElementType();
14440 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14442 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14443 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14446 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14448 // If the logical-not of the result is required, perform that now.
14450 Result = DAG.getNOT(dl, Result, VT);
14453 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14456 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14457 getZeroVector(VT, Subtarget, DAG, dl));
14462 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14464 MVT VT = Op.getSimpleValueType();
14466 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14468 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14469 && "SetCC type must be 8-bit or 1-bit integer");
14470 SDValue Op0 = Op.getOperand(0);
14471 SDValue Op1 = Op.getOperand(1);
14473 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14475 // Optimize to BT if possible.
14476 // Lower (X & (1 << N)) == 0 to BT(X, N).
14477 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14478 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14479 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14480 Op1.getOpcode() == ISD::Constant &&
14481 cast<ConstantSDNode>(Op1)->isNullValue() &&
14482 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14483 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14484 if (NewSetCC.getNode()) {
14486 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14491 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14493 if (Op1.getOpcode() == ISD::Constant &&
14494 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14495 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14496 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14498 // If the input is a setcc, then reuse the input setcc or use a new one with
14499 // the inverted condition.
14500 if (Op0.getOpcode() == X86ISD::SETCC) {
14501 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14502 bool Invert = (CC == ISD::SETNE) ^
14503 cast<ConstantSDNode>(Op1)->isNullValue();
14507 CCode = X86::GetOppositeBranchCondition(CCode);
14508 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14509 DAG.getConstant(CCode, dl, MVT::i8),
14510 Op0.getOperand(1));
14512 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14516 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14517 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14518 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14520 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14521 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14524 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14525 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14526 if (X86CC == X86::COND_INVALID)
14529 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14530 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14531 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14532 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14534 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14538 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14539 static bool isX86LogicalCmp(SDValue Op) {
14540 unsigned Opc = Op.getNode()->getOpcode();
14541 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14542 Opc == X86ISD::SAHF)
14544 if (Op.getResNo() == 1 &&
14545 (Opc == X86ISD::ADD ||
14546 Opc == X86ISD::SUB ||
14547 Opc == X86ISD::ADC ||
14548 Opc == X86ISD::SBB ||
14549 Opc == X86ISD::SMUL ||
14550 Opc == X86ISD::UMUL ||
14551 Opc == X86ISD::INC ||
14552 Opc == X86ISD::DEC ||
14553 Opc == X86ISD::OR ||
14554 Opc == X86ISD::XOR ||
14555 Opc == X86ISD::AND))
14558 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14564 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14565 if (V.getOpcode() != ISD::TRUNCATE)
14568 SDValue VOp0 = V.getOperand(0);
14569 unsigned InBits = VOp0.getValueSizeInBits();
14570 unsigned Bits = V.getValueSizeInBits();
14571 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14574 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14575 bool addTest = true;
14576 SDValue Cond = Op.getOperand(0);
14577 SDValue Op1 = Op.getOperand(1);
14578 SDValue Op2 = Op.getOperand(2);
14580 EVT VT = Op1.getValueType();
14583 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14584 // are available or VBLENDV if AVX is available.
14585 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14586 if (Cond.getOpcode() == ISD::SETCC &&
14587 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14588 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14589 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14590 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14591 int SSECC = translateX86FSETCC(
14592 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14595 if (Subtarget->hasAVX512()) {
14596 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14597 DAG.getConstant(SSECC, DL, MVT::i8));
14598 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14601 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14602 DAG.getConstant(SSECC, DL, MVT::i8));
14604 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14605 // of 3 logic instructions for size savings and potentially speed.
14606 // Unfortunately, there is no scalar form of VBLENDV.
14608 // If either operand is a constant, don't try this. We can expect to
14609 // optimize away at least one of the logic instructions later in that
14610 // case, so that sequence would be faster than a variable blend.
14612 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14613 // uses XMM0 as the selection register. That may need just as many
14614 // instructions as the AND/ANDN/OR sequence due to register moves, so
14617 if (Subtarget->hasAVX() &&
14618 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14620 // Convert to vectors, do a VSELECT, and convert back to scalar.
14621 // All of the conversions should be optimized away.
14623 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14624 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14625 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14626 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14628 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14629 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14631 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14633 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14634 VSel, DAG.getIntPtrConstant(0, DL));
14636 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14637 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14638 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14642 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14644 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14645 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14646 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14647 Op1Scalar = Op1.getOperand(0);
14649 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14650 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14651 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14652 Op2Scalar = Op2.getOperand(0);
14653 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14654 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14655 Op1Scalar.getValueType(),
14656 Cond, Op1Scalar, Op2Scalar);
14657 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14658 return DAG.getBitcast(VT, newSelect);
14659 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14660 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14661 DAG.getIntPtrConstant(0, DL));
14665 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14666 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14667 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14668 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14669 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14670 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14671 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14673 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14676 if (Cond.getOpcode() == ISD::SETCC) {
14677 SDValue NewCond = LowerSETCC(Cond, DAG);
14678 if (NewCond.getNode())
14682 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14683 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14684 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14685 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14686 if (Cond.getOpcode() == X86ISD::SETCC &&
14687 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14688 isZero(Cond.getOperand(1).getOperand(1))) {
14689 SDValue Cmp = Cond.getOperand(1);
14691 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14693 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14694 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14695 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14697 SDValue CmpOp0 = Cmp.getOperand(0);
14698 // Apply further optimizations for special cases
14699 // (select (x != 0), -1, 0) -> neg & sbb
14700 // (select (x == 0), 0, -1) -> neg & sbb
14701 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14702 if (YC->isNullValue() &&
14703 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14704 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14705 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14706 DAG.getConstant(0, DL,
14707 CmpOp0.getValueType()),
14709 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14710 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14711 SDValue(Neg.getNode(), 1));
14715 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14716 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14717 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14719 SDValue Res = // Res = 0 or -1.
14720 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14721 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14723 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14724 Res = DAG.getNOT(DL, Res, Res.getValueType());
14726 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14727 if (!N2C || !N2C->isNullValue())
14728 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14733 // Look past (and (setcc_carry (cmp ...)), 1).
14734 if (Cond.getOpcode() == ISD::AND &&
14735 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14736 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14737 if (C && C->getAPIntValue() == 1)
14738 Cond = Cond.getOperand(0);
14741 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14742 // setting operand in place of the X86ISD::SETCC.
14743 unsigned CondOpcode = Cond.getOpcode();
14744 if (CondOpcode == X86ISD::SETCC ||
14745 CondOpcode == X86ISD::SETCC_CARRY) {
14746 CC = Cond.getOperand(0);
14748 SDValue Cmp = Cond.getOperand(1);
14749 unsigned Opc = Cmp.getOpcode();
14750 MVT VT = Op.getSimpleValueType();
14752 bool IllegalFPCMov = false;
14753 if (VT.isFloatingPoint() && !VT.isVector() &&
14754 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14755 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14757 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14758 Opc == X86ISD::BT) { // FIXME
14762 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14763 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14764 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14765 Cond.getOperand(0).getValueType() != MVT::i8)) {
14766 SDValue LHS = Cond.getOperand(0);
14767 SDValue RHS = Cond.getOperand(1);
14768 unsigned X86Opcode;
14771 switch (CondOpcode) {
14772 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14773 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14774 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14775 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14776 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14777 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14778 default: llvm_unreachable("unexpected overflowing operator");
14780 if (CondOpcode == ISD::UMULO)
14781 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14784 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14786 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14788 if (CondOpcode == ISD::UMULO)
14789 Cond = X86Op.getValue(2);
14791 Cond = X86Op.getValue(1);
14793 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14798 // Look past the truncate if the high bits are known zero.
14799 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14800 Cond = Cond.getOperand(0);
14802 // We know the result of AND is compared against zero. Try to match
14804 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14805 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14806 if (NewSetCC.getNode()) {
14807 CC = NewSetCC.getOperand(0);
14808 Cond = NewSetCC.getOperand(1);
14815 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14816 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14819 // a < b ? -1 : 0 -> RES = ~setcc_carry
14820 // a < b ? 0 : -1 -> RES = setcc_carry
14821 // a >= b ? -1 : 0 -> RES = setcc_carry
14822 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14823 if (Cond.getOpcode() == X86ISD::SUB) {
14824 Cond = ConvertCmpIfNecessary(Cond, DAG);
14825 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14827 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14828 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14829 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14830 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14832 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14833 return DAG.getNOT(DL, Res, Res.getValueType());
14838 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14839 // widen the cmov and push the truncate through. This avoids introducing a new
14840 // branch during isel and doesn't add any extensions.
14841 if (Op.getValueType() == MVT::i8 &&
14842 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14843 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14844 if (T1.getValueType() == T2.getValueType() &&
14845 // Blacklist CopyFromReg to avoid partial register stalls.
14846 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14847 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14848 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14849 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14853 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14854 // condition is true.
14855 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14856 SDValue Ops[] = { Op2, Op1, CC, Cond };
14857 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14860 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14861 const X86Subtarget *Subtarget,
14862 SelectionDAG &DAG) {
14863 MVT VT = Op->getSimpleValueType(0);
14864 SDValue In = Op->getOperand(0);
14865 MVT InVT = In.getSimpleValueType();
14866 MVT VTElt = VT.getVectorElementType();
14867 MVT InVTElt = InVT.getVectorElementType();
14871 if ((InVTElt == MVT::i1) &&
14872 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14873 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14875 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14876 VTElt.getSizeInBits() <= 16)) ||
14878 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14879 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14881 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14882 VTElt.getSizeInBits() >= 32))))
14883 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14885 unsigned int NumElts = VT.getVectorNumElements();
14887 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14890 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14891 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14892 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14893 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14896 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14897 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14899 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14902 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14904 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14905 if (VT.is512BitVector())
14907 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14910 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14911 const X86Subtarget *Subtarget,
14912 SelectionDAG &DAG) {
14913 SDValue In = Op->getOperand(0);
14914 MVT VT = Op->getSimpleValueType(0);
14915 MVT InVT = In.getSimpleValueType();
14916 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14918 MVT InSVT = InVT.getScalarType();
14919 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14921 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14923 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14928 // SSE41 targets can use the pmovsx* instructions directly.
14929 if (Subtarget->hasSSE41())
14930 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14932 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14936 // As SRAI is only available on i16/i32 types, we expand only up to i32
14937 // and handle i64 separately.
14938 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14939 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14940 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14941 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14942 Curr = DAG.getBitcast(CurrVT, Curr);
14945 SDValue SignExt = Curr;
14946 if (CurrVT != InVT) {
14947 unsigned SignExtShift =
14948 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14949 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14950 DAG.getConstant(SignExtShift, dl, MVT::i8));
14956 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14957 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14958 DAG.getConstant(31, dl, MVT::i8));
14959 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14960 return DAG.getBitcast(VT, Ext);
14966 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14967 SelectionDAG &DAG) {
14968 MVT VT = Op->getSimpleValueType(0);
14969 SDValue In = Op->getOperand(0);
14970 MVT InVT = In.getSimpleValueType();
14973 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14974 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14976 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14977 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14978 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14981 if (Subtarget->hasInt256())
14982 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14984 // Optimize vectors in AVX mode
14985 // Sign extend v8i16 to v8i32 and
14988 // Divide input vector into two parts
14989 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14990 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14991 // concat the vectors to original VT
14993 unsigned NumElems = InVT.getVectorNumElements();
14994 SDValue Undef = DAG.getUNDEF(InVT);
14996 SmallVector<int,8> ShufMask1(NumElems, -1);
14997 for (unsigned i = 0; i != NumElems/2; ++i)
15000 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15002 SmallVector<int,8> ShufMask2(NumElems, -1);
15003 for (unsigned i = 0; i != NumElems/2; ++i)
15004 ShufMask2[i] = i + NumElems/2;
15006 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15008 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
15009 VT.getVectorNumElements()/2);
15011 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15012 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15014 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15017 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15018 // may emit an illegal shuffle but the expansion is still better than scalar
15019 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15020 // we'll emit a shuffle and a arithmetic shift.
15021 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15022 // TODO: It is possible to support ZExt by zeroing the undef values during
15023 // the shuffle phase or after the shuffle.
15024 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15025 SelectionDAG &DAG) {
15026 MVT RegVT = Op.getSimpleValueType();
15027 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15028 assert(RegVT.isInteger() &&
15029 "We only custom lower integer vector sext loads.");
15031 // Nothing useful we can do without SSE2 shuffles.
15032 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15034 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15036 EVT MemVT = Ld->getMemoryVT();
15037 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15038 unsigned RegSz = RegVT.getSizeInBits();
15040 ISD::LoadExtType Ext = Ld->getExtensionType();
15042 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15043 && "Only anyext and sext are currently implemented.");
15044 assert(MemVT != RegVT && "Cannot extend to the same type");
15045 assert(MemVT.isVector() && "Must load a vector from memory");
15047 unsigned NumElems = RegVT.getVectorNumElements();
15048 unsigned MemSz = MemVT.getSizeInBits();
15049 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15051 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15052 // The only way in which we have a legal 256-bit vector result but not the
15053 // integer 256-bit operations needed to directly lower a sextload is if we
15054 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15055 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15056 // correctly legalized. We do this late to allow the canonical form of
15057 // sextload to persist throughout the rest of the DAG combiner -- it wants
15058 // to fold together any extensions it can, and so will fuse a sign_extend
15059 // of an sextload into a sextload targeting a wider value.
15061 if (MemSz == 128) {
15062 // Just switch this to a normal load.
15063 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15064 "it must be a legal 128-bit vector "
15066 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15067 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15068 Ld->isInvariant(), Ld->getAlignment());
15070 assert(MemSz < 128 &&
15071 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15072 // Do an sext load to a 128-bit vector type. We want to use the same
15073 // number of elements, but elements half as wide. This will end up being
15074 // recursively lowered by this routine, but will succeed as we definitely
15075 // have all the necessary features if we're using AVX1.
15077 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15078 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15080 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15081 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15082 Ld->isNonTemporal(), Ld->isInvariant(),
15083 Ld->getAlignment());
15086 // Replace chain users with the new chain.
15087 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15088 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15090 // Finally, do a normal sign-extend to the desired register.
15091 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15094 // All sizes must be a power of two.
15095 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15096 "Non-power-of-two elements are not custom lowered!");
15098 // Attempt to load the original value using scalar loads.
15099 // Find the largest scalar type that divides the total loaded size.
15100 MVT SclrLoadTy = MVT::i8;
15101 for (MVT Tp : MVT::integer_valuetypes()) {
15102 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15107 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15108 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15110 SclrLoadTy = MVT::f64;
15112 // Calculate the number of scalar loads that we need to perform
15113 // in order to load our vector from memory.
15114 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15116 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15117 "Can only lower sext loads with a single scalar load!");
15119 unsigned loadRegZize = RegSz;
15120 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15123 // Represent our vector as a sequence of elements which are the
15124 // largest scalar that we can load.
15125 EVT LoadUnitVecVT = EVT::getVectorVT(
15126 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15128 // Represent the data using the same element type that is stored in
15129 // memory. In practice, we ''widen'' MemVT.
15131 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15132 loadRegZize / MemVT.getScalarSizeInBits());
15134 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15135 "Invalid vector type");
15137 // We can't shuffle using an illegal type.
15138 assert(TLI.isTypeLegal(WideVecVT) &&
15139 "We only lower types that form legal widened vector types");
15141 SmallVector<SDValue, 8> Chains;
15142 SDValue Ptr = Ld->getBasePtr();
15143 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15144 TLI.getPointerTy(DAG.getDataLayout()));
15145 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15147 for (unsigned i = 0; i < NumLoads; ++i) {
15148 // Perform a single load.
15149 SDValue ScalarLoad =
15150 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15151 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15152 Ld->getAlignment());
15153 Chains.push_back(ScalarLoad.getValue(1));
15154 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15155 // another round of DAGCombining.
15157 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15159 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15160 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15162 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15165 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15167 // Bitcast the loaded value to a vector of the original element type, in
15168 // the size of the target vector type.
15169 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15170 unsigned SizeRatio = RegSz / MemSz;
15172 if (Ext == ISD::SEXTLOAD) {
15173 // If we have SSE4.1, we can directly emit a VSEXT node.
15174 if (Subtarget->hasSSE41()) {
15175 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15176 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15180 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15182 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15183 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15185 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15186 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15190 // Redistribute the loaded elements into the different locations.
15191 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15192 for (unsigned i = 0; i != NumElems; ++i)
15193 ShuffleVec[i * SizeRatio] = i;
15195 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15196 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15198 // Bitcast to the requested type.
15199 Shuff = DAG.getBitcast(RegVT, Shuff);
15200 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15204 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15205 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15206 // from the AND / OR.
15207 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15208 Opc = Op.getOpcode();
15209 if (Opc != ISD::OR && Opc != ISD::AND)
15211 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15212 Op.getOperand(0).hasOneUse() &&
15213 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15214 Op.getOperand(1).hasOneUse());
15217 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15218 // 1 and that the SETCC node has a single use.
15219 static bool isXor1OfSetCC(SDValue Op) {
15220 if (Op.getOpcode() != ISD::XOR)
15222 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15223 if (N1C && N1C->getAPIntValue() == 1) {
15224 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15225 Op.getOperand(0).hasOneUse();
15230 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15231 bool addTest = true;
15232 SDValue Chain = Op.getOperand(0);
15233 SDValue Cond = Op.getOperand(1);
15234 SDValue Dest = Op.getOperand(2);
15237 bool Inverted = false;
15239 if (Cond.getOpcode() == ISD::SETCC) {
15240 // Check for setcc([su]{add,sub,mul}o == 0).
15241 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15242 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15243 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15244 Cond.getOperand(0).getResNo() == 1 &&
15245 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15246 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15247 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15248 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15249 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15250 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15252 Cond = Cond.getOperand(0);
15254 SDValue NewCond = LowerSETCC(Cond, DAG);
15255 if (NewCond.getNode())
15260 // FIXME: LowerXALUO doesn't handle these!!
15261 else if (Cond.getOpcode() == X86ISD::ADD ||
15262 Cond.getOpcode() == X86ISD::SUB ||
15263 Cond.getOpcode() == X86ISD::SMUL ||
15264 Cond.getOpcode() == X86ISD::UMUL)
15265 Cond = LowerXALUO(Cond, DAG);
15268 // Look pass (and (setcc_carry (cmp ...)), 1).
15269 if (Cond.getOpcode() == ISD::AND &&
15270 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15271 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15272 if (C && C->getAPIntValue() == 1)
15273 Cond = Cond.getOperand(0);
15276 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15277 // setting operand in place of the X86ISD::SETCC.
15278 unsigned CondOpcode = Cond.getOpcode();
15279 if (CondOpcode == X86ISD::SETCC ||
15280 CondOpcode == X86ISD::SETCC_CARRY) {
15281 CC = Cond.getOperand(0);
15283 SDValue Cmp = Cond.getOperand(1);
15284 unsigned Opc = Cmp.getOpcode();
15285 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15286 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15290 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15294 // These can only come from an arithmetic instruction with overflow,
15295 // e.g. SADDO, UADDO.
15296 Cond = Cond.getNode()->getOperand(1);
15302 CondOpcode = Cond.getOpcode();
15303 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15304 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15305 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15306 Cond.getOperand(0).getValueType() != MVT::i8)) {
15307 SDValue LHS = Cond.getOperand(0);
15308 SDValue RHS = Cond.getOperand(1);
15309 unsigned X86Opcode;
15312 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15313 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15315 switch (CondOpcode) {
15316 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15318 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15320 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15323 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15324 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15326 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15328 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15331 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15332 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15333 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15334 default: llvm_unreachable("unexpected overflowing operator");
15337 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15338 if (CondOpcode == ISD::UMULO)
15339 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15342 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15344 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15346 if (CondOpcode == ISD::UMULO)
15347 Cond = X86Op.getValue(2);
15349 Cond = X86Op.getValue(1);
15351 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15355 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15356 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15357 if (CondOpc == ISD::OR) {
15358 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15359 // two branches instead of an explicit OR instruction with a
15361 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15362 isX86LogicalCmp(Cmp)) {
15363 CC = Cond.getOperand(0).getOperand(0);
15364 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15365 Chain, Dest, CC, Cmp);
15366 CC = Cond.getOperand(1).getOperand(0);
15370 } else { // ISD::AND
15371 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15372 // two branches instead of an explicit AND instruction with a
15373 // separate test. However, we only do this if this block doesn't
15374 // have a fall-through edge, because this requires an explicit
15375 // jmp when the condition is false.
15376 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15377 isX86LogicalCmp(Cmp) &&
15378 Op.getNode()->hasOneUse()) {
15379 X86::CondCode CCode =
15380 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15381 CCode = X86::GetOppositeBranchCondition(CCode);
15382 CC = DAG.getConstant(CCode, dl, MVT::i8);
15383 SDNode *User = *Op.getNode()->use_begin();
15384 // Look for an unconditional branch following this conditional branch.
15385 // We need this because we need to reverse the successors in order
15386 // to implement FCMP_OEQ.
15387 if (User->getOpcode() == ISD::BR) {
15388 SDValue FalseBB = User->getOperand(1);
15390 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15391 assert(NewBR == User);
15395 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15396 Chain, Dest, CC, Cmp);
15397 X86::CondCode CCode =
15398 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15399 CCode = X86::GetOppositeBranchCondition(CCode);
15400 CC = DAG.getConstant(CCode, dl, MVT::i8);
15406 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15407 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15408 // It should be transformed during dag combiner except when the condition
15409 // is set by a arithmetics with overflow node.
15410 X86::CondCode CCode =
15411 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15412 CCode = X86::GetOppositeBranchCondition(CCode);
15413 CC = DAG.getConstant(CCode, dl, MVT::i8);
15414 Cond = Cond.getOperand(0).getOperand(1);
15416 } else if (Cond.getOpcode() == ISD::SETCC &&
15417 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15418 // For FCMP_OEQ, we can emit
15419 // two branches instead of an explicit AND instruction with a
15420 // separate test. However, we only do this if this block doesn't
15421 // have a fall-through edge, because this requires an explicit
15422 // jmp when the condition is false.
15423 if (Op.getNode()->hasOneUse()) {
15424 SDNode *User = *Op.getNode()->use_begin();
15425 // Look for an unconditional branch following this conditional branch.
15426 // We need this because we need to reverse the successors in order
15427 // to implement FCMP_OEQ.
15428 if (User->getOpcode() == ISD::BR) {
15429 SDValue FalseBB = User->getOperand(1);
15431 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15432 assert(NewBR == User);
15436 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15437 Cond.getOperand(0), Cond.getOperand(1));
15438 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15439 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15440 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15441 Chain, Dest, CC, Cmp);
15442 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15447 } else if (Cond.getOpcode() == ISD::SETCC &&
15448 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15449 // For FCMP_UNE, we can emit
15450 // two branches instead of an explicit AND instruction with a
15451 // separate test. However, we only do this if this block doesn't
15452 // have a fall-through edge, because this requires an explicit
15453 // jmp when the condition is false.
15454 if (Op.getNode()->hasOneUse()) {
15455 SDNode *User = *Op.getNode()->use_begin();
15456 // Look for an unconditional branch following this conditional branch.
15457 // We need this because we need to reverse the successors in order
15458 // to implement FCMP_UNE.
15459 if (User->getOpcode() == ISD::BR) {
15460 SDValue FalseBB = User->getOperand(1);
15462 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15463 assert(NewBR == User);
15466 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15467 Cond.getOperand(0), Cond.getOperand(1));
15468 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15469 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15470 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15471 Chain, Dest, CC, Cmp);
15472 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15482 // Look pass the truncate if the high bits are known zero.
15483 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15484 Cond = Cond.getOperand(0);
15486 // We know the result of AND is compared against zero. Try to match
15488 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15489 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15490 if (NewSetCC.getNode()) {
15491 CC = NewSetCC.getOperand(0);
15492 Cond = NewSetCC.getOperand(1);
15499 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15500 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15501 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15503 Cond = ConvertCmpIfNecessary(Cond, DAG);
15504 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15505 Chain, Dest, CC, Cond);
15508 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15509 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15510 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15511 // that the guard pages used by the OS virtual memory manager are allocated in
15512 // correct sequence.
15514 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15515 SelectionDAG &DAG) const {
15516 MachineFunction &MF = DAG.getMachineFunction();
15517 bool SplitStack = MF.shouldSplitStack();
15518 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15523 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15524 SDNode* Node = Op.getNode();
15526 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15527 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15528 " not tell us which reg is the stack pointer!");
15529 EVT VT = Node->getValueType(0);
15530 SDValue Tmp1 = SDValue(Node, 0);
15531 SDValue Tmp2 = SDValue(Node, 1);
15532 SDValue Tmp3 = Node->getOperand(2);
15533 SDValue Chain = Tmp1.getOperand(0);
15535 // Chain the dynamic stack allocation so that it doesn't modify the stack
15536 // pointer when other instructions are using the stack.
15537 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15540 SDValue Size = Tmp2.getOperand(1);
15541 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15542 Chain = SP.getValue(1);
15543 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15544 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15545 unsigned StackAlign = TFI.getStackAlignment();
15546 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15547 if (Align > StackAlign)
15548 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15549 DAG.getConstant(-(uint64_t)Align, dl, VT));
15550 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15552 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15553 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15556 SDValue Ops[2] = { Tmp1, Tmp2 };
15557 return DAG.getMergeValues(Ops, dl);
15561 SDValue Chain = Op.getOperand(0);
15562 SDValue Size = Op.getOperand(1);
15563 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15564 EVT VT = Op.getNode()->getValueType(0);
15566 bool Is64Bit = Subtarget->is64Bit();
15567 MVT SPTy = getPointerTy(DAG.getDataLayout());
15570 MachineRegisterInfo &MRI = MF.getRegInfo();
15573 // The 64 bit implementation of segmented stacks needs to clobber both r10
15574 // r11. This makes it impossible to use it along with nested parameters.
15575 const Function *F = MF.getFunction();
15577 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15579 if (I->hasNestAttr())
15580 report_fatal_error("Cannot use segmented stacks with functions that "
15581 "have nested arguments.");
15584 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15585 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15586 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15587 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15588 DAG.getRegister(Vreg, SPTy));
15589 SDValue Ops1[2] = { Value, Chain };
15590 return DAG.getMergeValues(Ops1, dl);
15593 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15595 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15596 Flag = Chain.getValue(1);
15597 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15599 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15601 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15602 unsigned SPReg = RegInfo->getStackRegister();
15603 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15604 Chain = SP.getValue(1);
15607 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15608 DAG.getConstant(-(uint64_t)Align, dl, VT));
15609 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15612 SDValue Ops1[2] = { SP, Chain };
15613 return DAG.getMergeValues(Ops1, dl);
15617 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15618 MachineFunction &MF = DAG.getMachineFunction();
15619 auto PtrVT = getPointerTy(MF.getDataLayout());
15620 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15622 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15625 if (!Subtarget->is64Bit() ||
15626 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15627 // vastart just stores the address of the VarArgsFrameIndex slot into the
15628 // memory location argument.
15629 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15630 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15631 MachinePointerInfo(SV), false, false, 0);
15635 // gp_offset (0 - 6 * 8)
15636 // fp_offset (48 - 48 + 8 * 16)
15637 // overflow_arg_area (point to parameters coming in memory).
15639 SmallVector<SDValue, 8> MemOps;
15640 SDValue FIN = Op.getOperand(1);
15642 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15643 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15645 FIN, MachinePointerInfo(SV), false, false, 0);
15646 MemOps.push_back(Store);
15649 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15650 Store = DAG.getStore(Op.getOperand(0), DL,
15651 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15653 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15654 MemOps.push_back(Store);
15656 // Store ptr to overflow_arg_area
15657 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15658 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15659 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15660 MachinePointerInfo(SV, 8),
15662 MemOps.push_back(Store);
15664 // Store ptr to reg_save_area.
15665 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15666 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15667 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15668 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15669 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15670 MemOps.push_back(Store);
15671 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15674 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15675 assert(Subtarget->is64Bit() &&
15676 "LowerVAARG only handles 64-bit va_arg!");
15677 assert(Op.getNode()->getNumOperands() == 4);
15679 MachineFunction &MF = DAG.getMachineFunction();
15680 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15681 // The Win64 ABI uses char* instead of a structure.
15682 return DAG.expandVAArg(Op.getNode());
15684 SDValue Chain = Op.getOperand(0);
15685 SDValue SrcPtr = Op.getOperand(1);
15686 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15687 unsigned Align = Op.getConstantOperandVal(3);
15690 EVT ArgVT = Op.getNode()->getValueType(0);
15691 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15692 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15695 // Decide which area this value should be read from.
15696 // TODO: Implement the AMD64 ABI in its entirety. This simple
15697 // selection mechanism works only for the basic types.
15698 if (ArgVT == MVT::f80) {
15699 llvm_unreachable("va_arg for f80 not yet implemented");
15700 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15701 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15702 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15703 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15705 llvm_unreachable("Unhandled argument type in LowerVAARG");
15708 if (ArgMode == 2) {
15709 // Sanity Check: Make sure using fp_offset makes sense.
15710 assert(!Subtarget->useSoftFloat() &&
15711 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15712 Subtarget->hasSSE1());
15715 // Insert VAARG_64 node into the DAG
15716 // VAARG_64 returns two values: Variable Argument Address, Chain
15717 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15718 DAG.getConstant(ArgMode, dl, MVT::i8),
15719 DAG.getConstant(Align, dl, MVT::i32)};
15720 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15721 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15722 VTs, InstOps, MVT::i64,
15723 MachinePointerInfo(SV),
15725 /*Volatile=*/false,
15727 /*WriteMem=*/true);
15728 Chain = VAARG.getValue(1);
15730 // Load the next argument and return it
15731 return DAG.getLoad(ArgVT, dl,
15734 MachinePointerInfo(),
15735 false, false, false, 0);
15738 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15739 SelectionDAG &DAG) {
15740 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15741 // where a va_list is still an i8*.
15742 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15743 if (Subtarget->isCallingConvWin64(
15744 DAG.getMachineFunction().getFunction()->getCallingConv()))
15745 // Probably a Win64 va_copy.
15746 return DAG.expandVACopy(Op.getNode());
15748 SDValue Chain = Op.getOperand(0);
15749 SDValue DstPtr = Op.getOperand(1);
15750 SDValue SrcPtr = Op.getOperand(2);
15751 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15752 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15755 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15756 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15758 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15761 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15762 // amount is a constant. Takes immediate version of shift as input.
15763 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15764 SDValue SrcOp, uint64_t ShiftAmt,
15765 SelectionDAG &DAG) {
15766 MVT ElementType = VT.getVectorElementType();
15768 // Fold this packed shift into its first operand if ShiftAmt is 0.
15772 // Check for ShiftAmt >= element width
15773 if (ShiftAmt >= ElementType.getSizeInBits()) {
15774 if (Opc == X86ISD::VSRAI)
15775 ShiftAmt = ElementType.getSizeInBits() - 1;
15777 return DAG.getConstant(0, dl, VT);
15780 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15781 && "Unknown target vector shift-by-constant node");
15783 // Fold this packed vector shift into a build vector if SrcOp is a
15784 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15785 if (VT == SrcOp.getSimpleValueType() &&
15786 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15787 SmallVector<SDValue, 8> Elts;
15788 unsigned NumElts = SrcOp->getNumOperands();
15789 ConstantSDNode *ND;
15792 default: llvm_unreachable(nullptr);
15793 case X86ISD::VSHLI:
15794 for (unsigned i=0; i!=NumElts; ++i) {
15795 SDValue CurrentOp = SrcOp->getOperand(i);
15796 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15797 Elts.push_back(CurrentOp);
15800 ND = cast<ConstantSDNode>(CurrentOp);
15801 const APInt &C = ND->getAPIntValue();
15802 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15805 case X86ISD::VSRLI:
15806 for (unsigned i=0; i!=NumElts; ++i) {
15807 SDValue CurrentOp = SrcOp->getOperand(i);
15808 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15809 Elts.push_back(CurrentOp);
15812 ND = cast<ConstantSDNode>(CurrentOp);
15813 const APInt &C = ND->getAPIntValue();
15814 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15817 case X86ISD::VSRAI:
15818 for (unsigned i=0; i!=NumElts; ++i) {
15819 SDValue CurrentOp = SrcOp->getOperand(i);
15820 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15821 Elts.push_back(CurrentOp);
15824 ND = cast<ConstantSDNode>(CurrentOp);
15825 const APInt &C = ND->getAPIntValue();
15826 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15831 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15834 return DAG.getNode(Opc, dl, VT, SrcOp,
15835 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15838 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15839 // may or may not be a constant. Takes immediate version of shift as input.
15840 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15841 SDValue SrcOp, SDValue ShAmt,
15842 SelectionDAG &DAG) {
15843 MVT SVT = ShAmt.getSimpleValueType();
15844 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15846 // Catch shift-by-constant.
15847 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15848 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15849 CShAmt->getZExtValue(), DAG);
15851 // Change opcode to non-immediate version
15853 default: llvm_unreachable("Unknown target vector shift node");
15854 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15855 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15856 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15859 const X86Subtarget &Subtarget =
15860 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15861 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15862 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15863 // Let the shuffle legalizer expand this shift amount node.
15864 SDValue Op0 = ShAmt.getOperand(0);
15865 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15866 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15868 // Need to build a vector containing shift amount.
15869 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15870 SmallVector<SDValue, 4> ShOps;
15871 ShOps.push_back(ShAmt);
15872 if (SVT == MVT::i32) {
15873 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15874 ShOps.push_back(DAG.getUNDEF(SVT));
15876 ShOps.push_back(DAG.getUNDEF(SVT));
15878 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15879 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15882 // The return type has to be a 128-bit type with the same element
15883 // type as the input type.
15884 MVT EltVT = VT.getVectorElementType();
15885 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15887 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15888 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15891 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15892 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15893 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15894 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15895 SDValue PreservedSrc,
15896 const X86Subtarget *Subtarget,
15897 SelectionDAG &DAG) {
15898 EVT VT = Op.getValueType();
15899 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15900 MVT::i1, VT.getVectorNumElements());
15901 SDValue VMask = SDValue();
15902 unsigned OpcodeSelect = ISD::VSELECT;
15905 assert(MaskVT.isSimple() && "invalid mask type");
15907 if (isAllOnes(Mask))
15910 if (MaskVT.bitsGT(Mask.getValueType())) {
15911 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15912 MaskVT.getSizeInBits());
15913 VMask = DAG.getBitcast(MaskVT,
15914 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15916 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15917 Mask.getValueType().getSizeInBits());
15918 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15919 // are extracted by EXTRACT_SUBVECTOR.
15920 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15921 DAG.getBitcast(BitcastVT, Mask),
15922 DAG.getIntPtrConstant(0, dl));
15925 switch (Op.getOpcode()) {
15927 case X86ISD::PCMPEQM:
15928 case X86ISD::PCMPGTM:
15930 case X86ISD::CMPMU:
15931 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15932 case X86ISD::VFPCLASS:
15933 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15934 case X86ISD::VTRUNC:
15935 case X86ISD::VTRUNCS:
15936 case X86ISD::VTRUNCUS:
15937 // We can't use ISD::VSELECT here because it is not always "Legal"
15938 // for the destination type. For example vpmovqb require only AVX512
15939 // and vselect that can operate on byte element type require BWI
15940 OpcodeSelect = X86ISD::SELECT;
15943 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15944 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15945 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15948 /// \brief Creates an SDNode for a predicated scalar operation.
15949 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15950 /// The mask is coming as MVT::i8 and it should be truncated
15951 /// to MVT::i1 while lowering masking intrinsics.
15952 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15953 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15954 /// for a scalar instruction.
15955 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15956 SDValue PreservedSrc,
15957 const X86Subtarget *Subtarget,
15958 SelectionDAG &DAG) {
15959 if (isAllOnes(Mask))
15962 EVT VT = Op.getValueType();
15964 // The mask should be of type MVT::i1
15965 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15967 if (Op.getOpcode() == X86ISD::FSETCC)
15968 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
15969 if (Op.getOpcode() == X86ISD::VFPCLASS)
15970 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
15972 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15973 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15974 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15977 static int getSEHRegistrationNodeSize(const Function *Fn) {
15978 if (!Fn->hasPersonalityFn())
15979 report_fatal_error(
15980 "querying registration node size for function without personality");
15981 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15982 // WinEHStatePass for the full struct definition.
15983 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15984 case EHPersonality::MSVC_X86SEH: return 24;
15985 case EHPersonality::MSVC_CXX: return 16;
15988 report_fatal_error("can only recover FP for MSVC EH personality functions");
15991 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15992 /// function or when returning to a parent frame after catching an exception, we
15993 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15994 /// Here's the math:
15995 /// RegNodeBase = EntryEBP - RegNodeSize
15996 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15997 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15998 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15999 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16000 SDValue EntryEBP) {
16001 MachineFunction &MF = DAG.getMachineFunction();
16004 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16005 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16007 // It's possible that the parent function no longer has a personality function
16008 // if the exceptional code was optimized away, in which case we just return
16009 // the incoming EBP.
16010 if (!Fn->hasPersonalityFn())
16013 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16015 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16017 MCSymbol *OffsetSym =
16018 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16019 GlobalValue::getRealLinkageName(Fn->getName()));
16020 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16021 SDValue RegNodeFrameOffset =
16022 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16024 // RegNodeBase = EntryEBP - RegNodeSize
16025 // ParentFP = RegNodeBase - RegNodeFrameOffset
16026 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16027 DAG.getConstant(RegNodeSize, dl, PtrVT));
16028 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16031 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16032 SelectionDAG &DAG) {
16034 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16035 EVT VT = Op.getValueType();
16036 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16038 switch(IntrData->Type) {
16039 case INTR_TYPE_1OP:
16040 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16041 case INTR_TYPE_2OP:
16042 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16044 case INTR_TYPE_2OP_IMM8:
16045 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16046 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16047 case INTR_TYPE_3OP:
16048 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16049 Op.getOperand(2), Op.getOperand(3));
16050 case INTR_TYPE_4OP:
16051 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16052 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16053 case INTR_TYPE_1OP_MASK_RM: {
16054 SDValue Src = Op.getOperand(1);
16055 SDValue PassThru = Op.getOperand(2);
16056 SDValue Mask = Op.getOperand(3);
16057 SDValue RoundingMode;
16058 // We allways add rounding mode to the Node.
16059 // If the rounding mode is not specified, we add the
16060 // "current direction" mode.
16061 if (Op.getNumOperands() == 4)
16063 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16065 RoundingMode = Op.getOperand(4);
16066 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16067 if (IntrWithRoundingModeOpcode != 0)
16068 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16069 X86::STATIC_ROUNDING::CUR_DIRECTION)
16070 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16071 dl, Op.getValueType(), Src, RoundingMode),
16072 Mask, PassThru, Subtarget, DAG);
16073 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16075 Mask, PassThru, Subtarget, DAG);
16077 case INTR_TYPE_1OP_MASK: {
16078 SDValue Src = Op.getOperand(1);
16079 SDValue PassThru = Op.getOperand(2);
16080 SDValue Mask = Op.getOperand(3);
16081 // We add rounding mode to the Node when
16082 // - RM Opcode is specified and
16083 // - RM is not "current direction".
16084 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16085 if (IntrWithRoundingModeOpcode != 0) {
16086 SDValue Rnd = Op.getOperand(4);
16087 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16088 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16089 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16090 dl, Op.getValueType(),
16092 Mask, PassThru, Subtarget, DAG);
16095 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16096 Mask, PassThru, Subtarget, DAG);
16098 case INTR_TYPE_SCALAR_MASK: {
16099 SDValue Src1 = Op.getOperand(1);
16100 SDValue Src2 = Op.getOperand(2);
16101 SDValue passThru = Op.getOperand(3);
16102 SDValue Mask = Op.getOperand(4);
16103 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16104 Mask, passThru, Subtarget, DAG);
16106 case INTR_TYPE_SCALAR_MASK_RM: {
16107 SDValue Src1 = Op.getOperand(1);
16108 SDValue Src2 = Op.getOperand(2);
16109 SDValue Src0 = Op.getOperand(3);
16110 SDValue Mask = Op.getOperand(4);
16111 // There are 2 kinds of intrinsics in this group:
16112 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16113 // (2) With rounding mode and sae - 7 operands.
16114 if (Op.getNumOperands() == 6) {
16115 SDValue Sae = Op.getOperand(5);
16116 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16117 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16119 Mask, Src0, Subtarget, DAG);
16121 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16122 SDValue RoundingMode = Op.getOperand(5);
16123 SDValue Sae = Op.getOperand(6);
16124 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16125 RoundingMode, Sae),
16126 Mask, Src0, Subtarget, DAG);
16128 case INTR_TYPE_2OP_MASK:
16129 case INTR_TYPE_2OP_IMM8_MASK: {
16130 SDValue Src1 = Op.getOperand(1);
16131 SDValue Src2 = Op.getOperand(2);
16132 SDValue PassThru = Op.getOperand(3);
16133 SDValue Mask = Op.getOperand(4);
16135 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16136 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16138 // We specify 2 possible opcodes for intrinsics with rounding modes.
16139 // First, we check if the intrinsic may have non-default rounding mode,
16140 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16141 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16142 if (IntrWithRoundingModeOpcode != 0) {
16143 SDValue Rnd = Op.getOperand(5);
16144 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16145 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16146 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16147 dl, Op.getValueType(),
16149 Mask, PassThru, Subtarget, DAG);
16152 // TODO: Intrinsics should have fast-math-flags to propagate.
16153 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16154 Mask, PassThru, Subtarget, DAG);
16156 case INTR_TYPE_2OP_MASK_RM: {
16157 SDValue Src1 = Op.getOperand(1);
16158 SDValue Src2 = Op.getOperand(2);
16159 SDValue PassThru = Op.getOperand(3);
16160 SDValue Mask = Op.getOperand(4);
16161 // We specify 2 possible modes for intrinsics, with/without rounding
16163 // First, we check if the intrinsic have rounding mode (6 operands),
16164 // if not, we set rounding mode to "current".
16166 if (Op.getNumOperands() == 6)
16167 Rnd = Op.getOperand(5);
16169 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16170 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16172 Mask, PassThru, Subtarget, DAG);
16174 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16175 SDValue Src1 = Op.getOperand(1);
16176 SDValue Src2 = Op.getOperand(2);
16177 SDValue Src3 = Op.getOperand(3);
16178 SDValue PassThru = Op.getOperand(4);
16179 SDValue Mask = Op.getOperand(5);
16180 SDValue Sae = Op.getOperand(6);
16182 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16184 Mask, PassThru, Subtarget, DAG);
16186 case INTR_TYPE_3OP_MASK_RM: {
16187 SDValue Src1 = Op.getOperand(1);
16188 SDValue Src2 = Op.getOperand(2);
16189 SDValue Imm = Op.getOperand(3);
16190 SDValue PassThru = Op.getOperand(4);
16191 SDValue Mask = Op.getOperand(5);
16192 // We specify 2 possible modes for intrinsics, with/without rounding
16194 // First, we check if the intrinsic have rounding mode (7 operands),
16195 // if not, we set rounding mode to "current".
16197 if (Op.getNumOperands() == 7)
16198 Rnd = Op.getOperand(6);
16200 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16201 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16202 Src1, Src2, Imm, Rnd),
16203 Mask, PassThru, Subtarget, DAG);
16205 case INTR_TYPE_3OP_IMM8_MASK:
16206 case INTR_TYPE_3OP_MASK:
16207 case INSERT_SUBVEC: {
16208 SDValue Src1 = Op.getOperand(1);
16209 SDValue Src2 = Op.getOperand(2);
16210 SDValue Src3 = Op.getOperand(3);
16211 SDValue PassThru = Op.getOperand(4);
16212 SDValue Mask = Op.getOperand(5);
16214 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16215 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16216 else if (IntrData->Type == INSERT_SUBVEC) {
16217 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16218 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16219 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16220 Imm *= Src2.getValueType().getVectorNumElements();
16221 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16224 // We specify 2 possible opcodes for intrinsics with rounding modes.
16225 // First, we check if the intrinsic may have non-default rounding mode,
16226 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16227 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16228 if (IntrWithRoundingModeOpcode != 0) {
16229 SDValue Rnd = Op.getOperand(6);
16230 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16231 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16232 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16233 dl, Op.getValueType(),
16234 Src1, Src2, Src3, Rnd),
16235 Mask, PassThru, Subtarget, DAG);
16238 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16240 Mask, PassThru, Subtarget, DAG);
16242 case VPERM_3OP_MASKZ:
16243 case VPERM_3OP_MASK:
16246 case FMA_OP_MASK: {
16247 SDValue Src1 = Op.getOperand(1);
16248 SDValue Src2 = Op.getOperand(2);
16249 SDValue Src3 = Op.getOperand(3);
16250 SDValue Mask = Op.getOperand(4);
16251 EVT VT = Op.getValueType();
16252 SDValue PassThru = SDValue();
16254 // set PassThru element
16255 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16256 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16257 else if (IntrData->Type == FMA_OP_MASK3)
16262 // We specify 2 possible opcodes for intrinsics with rounding modes.
16263 // First, we check if the intrinsic may have non-default rounding mode,
16264 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16265 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16266 if (IntrWithRoundingModeOpcode != 0) {
16267 SDValue Rnd = Op.getOperand(5);
16268 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16269 X86::STATIC_ROUNDING::CUR_DIRECTION)
16270 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16271 dl, Op.getValueType(),
16272 Src1, Src2, Src3, Rnd),
16273 Mask, PassThru, Subtarget, DAG);
16275 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16276 dl, Op.getValueType(),
16278 Mask, PassThru, Subtarget, DAG);
16280 case TERLOG_OP_MASK:
16281 case TERLOG_OP_MASKZ: {
16282 SDValue Src1 = Op.getOperand(1);
16283 SDValue Src2 = Op.getOperand(2);
16284 SDValue Src3 = Op.getOperand(3);
16285 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16286 SDValue Mask = Op.getOperand(5);
16287 EVT VT = Op.getValueType();
16288 SDValue PassThru = Src1;
16289 // Set PassThru element.
16290 if (IntrData->Type == TERLOG_OP_MASKZ)
16291 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16293 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16294 Src1, Src2, Src3, Src4),
16295 Mask, PassThru, Subtarget, DAG);
16298 // FPclass intrinsics with mask
16299 SDValue Src1 = Op.getOperand(1);
16300 EVT VT = Src1.getValueType();
16301 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16302 VT.getVectorNumElements());
16303 SDValue Imm = Op.getOperand(2);
16304 SDValue Mask = Op.getOperand(3);
16305 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16306 Mask.getValueType().getSizeInBits());
16307 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16308 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16309 DAG.getTargetConstant(0, dl, MaskVT),
16311 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16312 DAG.getUNDEF(BitcastVT), FPclassMask,
16313 DAG.getIntPtrConstant(0, dl));
16314 return DAG.getBitcast(Op.getValueType(), Res);
16317 SDValue Src1 = Op.getOperand(1);
16318 SDValue Imm = Op.getOperand(2);
16319 SDValue Mask = Op.getOperand(3);
16320 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16321 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16322 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16323 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16326 case CMP_MASK_CC: {
16327 // Comparison intrinsics with masks.
16328 // Example of transformation:
16329 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16330 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16332 // (v8i1 (insert_subvector undef,
16333 // (v2i1 (and (PCMPEQM %a, %b),
16334 // (extract_subvector
16335 // (v8i1 (bitcast %mask)), 0))), 0))))
16336 EVT VT = Op.getOperand(1).getValueType();
16337 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16338 VT.getVectorNumElements());
16339 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16340 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16341 Mask.getValueType().getSizeInBits());
16343 if (IntrData->Type == CMP_MASK_CC) {
16344 SDValue CC = Op.getOperand(3);
16345 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16346 // We specify 2 possible opcodes for intrinsics with rounding modes.
16347 // First, we check if the intrinsic may have non-default rounding mode,
16348 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16349 if (IntrData->Opc1 != 0) {
16350 SDValue Rnd = Op.getOperand(5);
16351 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16352 X86::STATIC_ROUNDING::CUR_DIRECTION)
16353 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16354 Op.getOperand(2), CC, Rnd);
16356 //default rounding mode
16358 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16359 Op.getOperand(2), CC);
16362 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16363 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16366 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16367 DAG.getTargetConstant(0, dl,
16370 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16371 DAG.getUNDEF(BitcastVT), CmpMask,
16372 DAG.getIntPtrConstant(0, dl));
16373 return DAG.getBitcast(Op.getValueType(), Res);
16375 case CMP_MASK_SCALAR_CC: {
16376 SDValue Src1 = Op.getOperand(1);
16377 SDValue Src2 = Op.getOperand(2);
16378 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16379 SDValue Mask = Op.getOperand(4);
16382 if (IntrData->Opc1 != 0) {
16383 SDValue Rnd = Op.getOperand(5);
16384 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16385 X86::STATIC_ROUNDING::CUR_DIRECTION)
16386 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16388 //default rounding mode
16390 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16392 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16393 DAG.getTargetConstant(0, dl,
16397 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16398 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16399 DAG.getValueType(MVT::i1));
16401 case COMI: { // Comparison intrinsics
16402 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16403 SDValue LHS = Op.getOperand(1);
16404 SDValue RHS = Op.getOperand(2);
16405 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16406 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16407 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16408 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16409 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16410 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16413 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16414 Op.getOperand(1), Op.getOperand(2), DAG);
16416 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16417 Op.getSimpleValueType(),
16419 Op.getOperand(2), DAG),
16420 Op.getOperand(4), Op.getOperand(3), Subtarget,
16422 case COMPRESS_EXPAND_IN_REG: {
16423 SDValue Mask = Op.getOperand(3);
16424 SDValue DataToCompress = Op.getOperand(1);
16425 SDValue PassThru = Op.getOperand(2);
16426 if (isAllOnes(Mask)) // return data as is
16427 return Op.getOperand(1);
16429 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16431 Mask, PassThru, Subtarget, DAG);
16434 SDValue Mask = Op.getOperand(3);
16435 EVT VT = Op.getValueType();
16436 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16437 VT.getVectorNumElements());
16438 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16439 Mask.getValueType().getSizeInBits());
16441 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16442 DAG.getBitcast(BitcastVT, Mask),
16443 DAG.getIntPtrConstant(0, dl));
16444 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16453 default: return SDValue(); // Don't custom lower most intrinsics.
16455 case Intrinsic::x86_avx2_permd:
16456 case Intrinsic::x86_avx2_permps:
16457 // Operands intentionally swapped. Mask is last operand to intrinsic,
16458 // but second operand for node/instruction.
16459 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16460 Op.getOperand(2), Op.getOperand(1));
16462 // ptest and testp intrinsics. The intrinsic these come from are designed to
16463 // return an integer value, not just an instruction so lower it to the ptest
16464 // or testp pattern and a setcc for the result.
16465 case Intrinsic::x86_sse41_ptestz:
16466 case Intrinsic::x86_sse41_ptestc:
16467 case Intrinsic::x86_sse41_ptestnzc:
16468 case Intrinsic::x86_avx_ptestz_256:
16469 case Intrinsic::x86_avx_ptestc_256:
16470 case Intrinsic::x86_avx_ptestnzc_256:
16471 case Intrinsic::x86_avx_vtestz_ps:
16472 case Intrinsic::x86_avx_vtestc_ps:
16473 case Intrinsic::x86_avx_vtestnzc_ps:
16474 case Intrinsic::x86_avx_vtestz_pd:
16475 case Intrinsic::x86_avx_vtestc_pd:
16476 case Intrinsic::x86_avx_vtestnzc_pd:
16477 case Intrinsic::x86_avx_vtestz_ps_256:
16478 case Intrinsic::x86_avx_vtestc_ps_256:
16479 case Intrinsic::x86_avx_vtestnzc_ps_256:
16480 case Intrinsic::x86_avx_vtestz_pd_256:
16481 case Intrinsic::x86_avx_vtestc_pd_256:
16482 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16483 bool IsTestPacked = false;
16486 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16487 case Intrinsic::x86_avx_vtestz_ps:
16488 case Intrinsic::x86_avx_vtestz_pd:
16489 case Intrinsic::x86_avx_vtestz_ps_256:
16490 case Intrinsic::x86_avx_vtestz_pd_256:
16491 IsTestPacked = true; // Fallthrough
16492 case Intrinsic::x86_sse41_ptestz:
16493 case Intrinsic::x86_avx_ptestz_256:
16495 X86CC = X86::COND_E;
16497 case Intrinsic::x86_avx_vtestc_ps:
16498 case Intrinsic::x86_avx_vtestc_pd:
16499 case Intrinsic::x86_avx_vtestc_ps_256:
16500 case Intrinsic::x86_avx_vtestc_pd_256:
16501 IsTestPacked = true; // Fallthrough
16502 case Intrinsic::x86_sse41_ptestc:
16503 case Intrinsic::x86_avx_ptestc_256:
16505 X86CC = X86::COND_B;
16507 case Intrinsic::x86_avx_vtestnzc_ps:
16508 case Intrinsic::x86_avx_vtestnzc_pd:
16509 case Intrinsic::x86_avx_vtestnzc_ps_256:
16510 case Intrinsic::x86_avx_vtestnzc_pd_256:
16511 IsTestPacked = true; // Fallthrough
16512 case Intrinsic::x86_sse41_ptestnzc:
16513 case Intrinsic::x86_avx_ptestnzc_256:
16515 X86CC = X86::COND_A;
16519 SDValue LHS = Op.getOperand(1);
16520 SDValue RHS = Op.getOperand(2);
16521 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16522 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16523 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16524 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16525 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16527 case Intrinsic::x86_avx512_kortestz_w:
16528 case Intrinsic::x86_avx512_kortestc_w: {
16529 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16530 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16531 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16532 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16533 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16534 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16535 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16538 case Intrinsic::x86_sse42_pcmpistria128:
16539 case Intrinsic::x86_sse42_pcmpestria128:
16540 case Intrinsic::x86_sse42_pcmpistric128:
16541 case Intrinsic::x86_sse42_pcmpestric128:
16542 case Intrinsic::x86_sse42_pcmpistrio128:
16543 case Intrinsic::x86_sse42_pcmpestrio128:
16544 case Intrinsic::x86_sse42_pcmpistris128:
16545 case Intrinsic::x86_sse42_pcmpestris128:
16546 case Intrinsic::x86_sse42_pcmpistriz128:
16547 case Intrinsic::x86_sse42_pcmpestriz128: {
16551 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16552 case Intrinsic::x86_sse42_pcmpistria128:
16553 Opcode = X86ISD::PCMPISTRI;
16554 X86CC = X86::COND_A;
16556 case Intrinsic::x86_sse42_pcmpestria128:
16557 Opcode = X86ISD::PCMPESTRI;
16558 X86CC = X86::COND_A;
16560 case Intrinsic::x86_sse42_pcmpistric128:
16561 Opcode = X86ISD::PCMPISTRI;
16562 X86CC = X86::COND_B;
16564 case Intrinsic::x86_sse42_pcmpestric128:
16565 Opcode = X86ISD::PCMPESTRI;
16566 X86CC = X86::COND_B;
16568 case Intrinsic::x86_sse42_pcmpistrio128:
16569 Opcode = X86ISD::PCMPISTRI;
16570 X86CC = X86::COND_O;
16572 case Intrinsic::x86_sse42_pcmpestrio128:
16573 Opcode = X86ISD::PCMPESTRI;
16574 X86CC = X86::COND_O;
16576 case Intrinsic::x86_sse42_pcmpistris128:
16577 Opcode = X86ISD::PCMPISTRI;
16578 X86CC = X86::COND_S;
16580 case Intrinsic::x86_sse42_pcmpestris128:
16581 Opcode = X86ISD::PCMPESTRI;
16582 X86CC = X86::COND_S;
16584 case Intrinsic::x86_sse42_pcmpistriz128:
16585 Opcode = X86ISD::PCMPISTRI;
16586 X86CC = X86::COND_E;
16588 case Intrinsic::x86_sse42_pcmpestriz128:
16589 Opcode = X86ISD::PCMPESTRI;
16590 X86CC = X86::COND_E;
16593 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16594 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16595 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16596 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16597 DAG.getConstant(X86CC, dl, MVT::i8),
16598 SDValue(PCMP.getNode(), 1));
16599 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16602 case Intrinsic::x86_sse42_pcmpistri128:
16603 case Intrinsic::x86_sse42_pcmpestri128: {
16605 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16606 Opcode = X86ISD::PCMPISTRI;
16608 Opcode = X86ISD::PCMPESTRI;
16610 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16611 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16612 return DAG.getNode(Opcode, dl, VTs, NewOps);
16615 case Intrinsic::x86_seh_lsda: {
16616 // Compute the symbol for the LSDA. We know it'll get emitted later.
16617 MachineFunction &MF = DAG.getMachineFunction();
16618 SDValue Op1 = Op.getOperand(1);
16619 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16620 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16621 GlobalValue::getRealLinkageName(Fn->getName()));
16623 // Generate a simple absolute symbol reference. This intrinsic is only
16624 // supported on 32-bit Windows, which isn't PIC.
16625 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16626 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16629 case Intrinsic::x86_seh_recoverfp: {
16630 SDValue FnOp = Op.getOperand(1);
16631 SDValue IncomingFPOp = Op.getOperand(2);
16632 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16633 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16635 report_fatal_error(
16636 "llvm.x86.seh.recoverfp must take a function as the first argument");
16637 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16640 case Intrinsic::localaddress: {
16641 // Returns one of the stack, base, or frame pointer registers, depending on
16642 // which is used to reference local variables.
16643 MachineFunction &MF = DAG.getMachineFunction();
16644 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16646 if (RegInfo->hasBasePointer(MF))
16647 Reg = RegInfo->getBaseRegister();
16648 else // This function handles the SP or FP case.
16649 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16650 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16655 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16656 SDValue Src, SDValue Mask, SDValue Base,
16657 SDValue Index, SDValue ScaleOp, SDValue Chain,
16658 const X86Subtarget * Subtarget) {
16660 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16662 llvm_unreachable("Invalid scale type");
16663 unsigned ScaleVal = C->getZExtValue();
16664 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16665 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16667 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16668 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16669 Index.getSimpleValueType().getVectorNumElements());
16671 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16673 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16675 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16676 Mask.getValueType().getSizeInBits());
16678 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16679 // are extracted by EXTRACT_SUBVECTOR.
16680 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16681 DAG.getBitcast(BitcastVT, Mask),
16682 DAG.getIntPtrConstant(0, dl));
16684 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16685 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16686 SDValue Segment = DAG.getRegister(0, MVT::i32);
16687 if (Src.getOpcode() == ISD::UNDEF)
16688 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16689 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16690 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16691 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16692 return DAG.getMergeValues(RetOps, dl);
16695 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16696 SDValue Src, SDValue Mask, SDValue Base,
16697 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16699 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16701 llvm_unreachable("Invalid scale type");
16702 unsigned ScaleVal = C->getZExtValue();
16703 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16704 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16706 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16707 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16708 SDValue Segment = DAG.getRegister(0, MVT::i32);
16709 MVT MaskVT = MVT::getVectorVT(MVT::i1,
16710 Index.getSimpleValueType().getVectorNumElements());
16712 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16714 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16716 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16717 Mask.getValueType().getSizeInBits());
16719 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16720 // are extracted by EXTRACT_SUBVECTOR.
16721 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16722 DAG.getBitcast(BitcastVT, Mask),
16723 DAG.getIntPtrConstant(0, dl));
16725 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16726 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16727 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16728 return SDValue(Res, 1);
16731 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16732 SDValue Mask, SDValue Base, SDValue Index,
16733 SDValue ScaleOp, SDValue Chain) {
16735 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16736 assert(C && "Invalid scale type");
16737 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16738 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16739 SDValue Segment = DAG.getRegister(0, MVT::i32);
16741 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16743 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16745 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16747 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16748 //SDVTList VTs = DAG.getVTList(MVT::Other);
16749 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16750 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16751 return SDValue(Res, 0);
16754 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16755 // read performance monitor counters (x86_rdpmc).
16756 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16757 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16758 SmallVectorImpl<SDValue> &Results) {
16759 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16760 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16763 // The ECX register is used to select the index of the performance counter
16765 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16767 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16769 // Reads the content of a 64-bit performance counter and returns it in the
16770 // registers EDX:EAX.
16771 if (Subtarget->is64Bit()) {
16772 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16773 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16776 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16777 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16780 Chain = HI.getValue(1);
16782 if (Subtarget->is64Bit()) {
16783 // The EAX register is loaded with the low-order 32 bits. The EDX register
16784 // is loaded with the supported high-order bits of the counter.
16785 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16786 DAG.getConstant(32, DL, MVT::i8));
16787 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16788 Results.push_back(Chain);
16792 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16793 SDValue Ops[] = { LO, HI };
16794 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16795 Results.push_back(Pair);
16796 Results.push_back(Chain);
16799 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16800 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16801 // also used to custom lower READCYCLECOUNTER nodes.
16802 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16803 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16804 SmallVectorImpl<SDValue> &Results) {
16805 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16806 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16809 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16810 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16811 // and the EAX register is loaded with the low-order 32 bits.
16812 if (Subtarget->is64Bit()) {
16813 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16814 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16817 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16818 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16821 SDValue Chain = HI.getValue(1);
16823 if (Opcode == X86ISD::RDTSCP_DAG) {
16824 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16826 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16827 // the ECX register. Add 'ecx' explicitly to the chain.
16828 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16830 // Explicitly store the content of ECX at the location passed in input
16831 // to the 'rdtscp' intrinsic.
16832 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16833 MachinePointerInfo(), false, false, 0);
16836 if (Subtarget->is64Bit()) {
16837 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16838 // the EAX register is loaded with the low-order 32 bits.
16839 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16840 DAG.getConstant(32, DL, MVT::i8));
16841 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16842 Results.push_back(Chain);
16846 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16847 SDValue Ops[] = { LO, HI };
16848 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16849 Results.push_back(Pair);
16850 Results.push_back(Chain);
16853 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16854 SelectionDAG &DAG) {
16855 SmallVector<SDValue, 2> Results;
16857 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16859 return DAG.getMergeValues(Results, DL);
16862 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16863 SelectionDAG &DAG) {
16864 MachineFunction &MF = DAG.getMachineFunction();
16865 const Function *Fn = MF.getFunction();
16867 SDValue Chain = Op.getOperand(0);
16869 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16870 "using llvm.x86.seh.restoreframe requires a frame pointer");
16872 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16873 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16875 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16876 unsigned FrameReg =
16877 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16878 unsigned SPReg = RegInfo->getStackRegister();
16879 unsigned SlotSize = RegInfo->getSlotSize();
16881 // Get incoming EBP.
16882 SDValue IncomingEBP =
16883 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16885 // SP is saved in the first field of every registration node, so load
16886 // [EBP-RegNodeSize] into SP.
16887 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16888 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16889 DAG.getConstant(-RegNodeSize, dl, VT));
16891 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16892 false, VT.getScalarSizeInBits() / 8);
16893 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16895 if (!RegInfo->needsStackRealignment(MF)) {
16896 // Adjust EBP to point back to the original frame position.
16897 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16898 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16900 assert(RegInfo->hasBasePointer(MF) &&
16901 "functions with Win32 EH must use frame or base pointer register");
16903 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16904 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16905 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16907 // Reload the spilled EBP value, now that the stack and base pointers are
16909 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16910 X86FI->setHasSEHFramePtrSave(true);
16911 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16912 X86FI->setSEHFramePtrSaveIndex(FI);
16913 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16914 MachinePointerInfo(), false, false, false,
16915 VT.getScalarSizeInBits() / 8);
16916 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16922 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16923 /// return truncate Store/MaskedStore Node
16924 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16928 SDValue Mask = Op.getOperand(4);
16929 SDValue DataToTruncate = Op.getOperand(3);
16930 SDValue Addr = Op.getOperand(2);
16931 SDValue Chain = Op.getOperand(0);
16933 EVT VT = DataToTruncate.getValueType();
16934 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16935 ElementType, VT.getVectorNumElements());
16937 if (isAllOnes(Mask)) // return just a truncate store
16938 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16939 MachinePointerInfo(), SVT, false, false,
16940 SVT.getScalarSizeInBits()/8);
16942 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16943 MVT::i1, VT.getVectorNumElements());
16944 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16945 Mask.getValueType().getSizeInBits());
16946 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16947 // are extracted by EXTRACT_SUBVECTOR.
16948 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16949 DAG.getBitcast(BitcastVT, Mask),
16950 DAG.getIntPtrConstant(0, dl));
16952 MachineMemOperand *MMO = DAG.getMachineFunction().
16953 getMachineMemOperand(MachinePointerInfo(),
16954 MachineMemOperand::MOStore, SVT.getStoreSize(),
16955 SVT.getScalarSizeInBits()/8);
16957 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16958 VMask, SVT, MMO, true);
16961 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16962 SelectionDAG &DAG) {
16963 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16965 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16967 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16968 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16973 switch(IntrData->Type) {
16975 llvm_unreachable("Unknown Intrinsic Type");
16979 // Emit the node with the right value type.
16980 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16981 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16983 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16984 // Otherwise return the value from Rand, which is always 0, casted to i32.
16985 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16986 DAG.getConstant(1, dl, Op->getValueType(1)),
16987 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16988 SDValue(Result.getNode(), 1) };
16989 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16990 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16993 // Return { result, isValid, chain }.
16994 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16995 SDValue(Result.getNode(), 2));
16998 //gather(v1, mask, index, base, scale);
16999 SDValue Chain = Op.getOperand(0);
17000 SDValue Src = Op.getOperand(2);
17001 SDValue Base = Op.getOperand(3);
17002 SDValue Index = Op.getOperand(4);
17003 SDValue Mask = Op.getOperand(5);
17004 SDValue Scale = Op.getOperand(6);
17005 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17009 //scatter(base, mask, index, v1, scale);
17010 SDValue Chain = Op.getOperand(0);
17011 SDValue Base = Op.getOperand(2);
17012 SDValue Mask = Op.getOperand(3);
17013 SDValue Index = Op.getOperand(4);
17014 SDValue Src = Op.getOperand(5);
17015 SDValue Scale = Op.getOperand(6);
17016 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17020 SDValue Hint = Op.getOperand(6);
17021 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17022 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17023 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17024 SDValue Chain = Op.getOperand(0);
17025 SDValue Mask = Op.getOperand(2);
17026 SDValue Index = Op.getOperand(3);
17027 SDValue Base = Op.getOperand(4);
17028 SDValue Scale = Op.getOperand(5);
17029 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17031 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17033 SmallVector<SDValue, 2> Results;
17034 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17036 return DAG.getMergeValues(Results, dl);
17038 // Read Performance Monitoring Counters.
17040 SmallVector<SDValue, 2> Results;
17041 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17042 return DAG.getMergeValues(Results, dl);
17044 // XTEST intrinsics.
17046 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17047 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17048 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17049 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17051 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17052 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17053 Ret, SDValue(InTrans.getNode(), 1));
17057 SmallVector<SDValue, 2> Results;
17058 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17059 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17060 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17061 DAG.getConstant(-1, dl, MVT::i8));
17062 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17063 Op.getOperand(4), GenCF.getValue(1));
17064 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17065 Op.getOperand(5), MachinePointerInfo(),
17067 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17068 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17070 Results.push_back(SetCC);
17071 Results.push_back(Store);
17072 return DAG.getMergeValues(Results, dl);
17074 case COMPRESS_TO_MEM: {
17076 SDValue Mask = Op.getOperand(4);
17077 SDValue DataToCompress = Op.getOperand(3);
17078 SDValue Addr = Op.getOperand(2);
17079 SDValue Chain = Op.getOperand(0);
17081 EVT VT = DataToCompress.getValueType();
17082 if (isAllOnes(Mask)) // return just a store
17083 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17084 MachinePointerInfo(), false, false,
17085 VT.getScalarSizeInBits()/8);
17087 SDValue Compressed =
17088 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17089 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17090 return DAG.getStore(Chain, dl, Compressed, Addr,
17091 MachinePointerInfo(), false, false,
17092 VT.getScalarSizeInBits()/8);
17094 case TRUNCATE_TO_MEM_VI8:
17095 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17096 case TRUNCATE_TO_MEM_VI16:
17097 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17098 case TRUNCATE_TO_MEM_VI32:
17099 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17100 case EXPAND_FROM_MEM: {
17102 SDValue Mask = Op.getOperand(4);
17103 SDValue PassThru = Op.getOperand(3);
17104 SDValue Addr = Op.getOperand(2);
17105 SDValue Chain = Op.getOperand(0);
17106 EVT VT = Op.getValueType();
17108 if (isAllOnes(Mask)) // return just a load
17109 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17110 false, VT.getScalarSizeInBits()/8);
17112 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17113 false, false, false,
17114 VT.getScalarSizeInBits()/8);
17116 SDValue Results[] = {
17117 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17118 Mask, PassThru, Subtarget, DAG), Chain};
17119 return DAG.getMergeValues(Results, dl);
17124 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17125 SelectionDAG &DAG) const {
17126 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17127 MFI->setReturnAddressIsTaken(true);
17129 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17132 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17134 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17137 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17138 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17139 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17140 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17141 DAG.getNode(ISD::ADD, dl, PtrVT,
17142 FrameAddr, Offset),
17143 MachinePointerInfo(), false, false, false, 0);
17146 // Just load the return address.
17147 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17148 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17149 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17152 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17153 MachineFunction &MF = DAG.getMachineFunction();
17154 MachineFrameInfo *MFI = MF.getFrameInfo();
17155 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17156 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17157 EVT VT = Op.getValueType();
17159 MFI->setFrameAddressIsTaken(true);
17161 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17162 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17163 // is not possible to crawl up the stack without looking at the unwind codes
17165 int FrameAddrIndex = FuncInfo->getFAIndex();
17166 if (!FrameAddrIndex) {
17167 // Set up a frame object for the return address.
17168 unsigned SlotSize = RegInfo->getSlotSize();
17169 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17170 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17171 FuncInfo->setFAIndex(FrameAddrIndex);
17173 return DAG.getFrameIndex(FrameAddrIndex, VT);
17176 unsigned FrameReg =
17177 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17178 SDLoc dl(Op); // FIXME probably not meaningful
17179 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17180 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17181 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17182 "Invalid Frame Register!");
17183 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17185 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17186 MachinePointerInfo(),
17187 false, false, false, 0);
17191 // FIXME? Maybe this could be a TableGen attribute on some registers and
17192 // this table could be generated automatically from RegInfo.
17193 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17194 SelectionDAG &DAG) const {
17195 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17196 const MachineFunction &MF = DAG.getMachineFunction();
17198 unsigned Reg = StringSwitch<unsigned>(RegName)
17199 .Case("esp", X86::ESP)
17200 .Case("rsp", X86::RSP)
17201 .Case("ebp", X86::EBP)
17202 .Case("rbp", X86::RBP)
17205 if (Reg == X86::EBP || Reg == X86::RBP) {
17206 if (!TFI.hasFP(MF))
17207 report_fatal_error("register " + StringRef(RegName) +
17208 " is allocatable: function has no frame pointer");
17211 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17212 unsigned FrameReg =
17213 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17214 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17215 "Invalid Frame Register!");
17223 report_fatal_error("Invalid register name global variable");
17226 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17227 SelectionDAG &DAG) const {
17228 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17229 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17232 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17233 SDValue Chain = Op.getOperand(0);
17234 SDValue Offset = Op.getOperand(1);
17235 SDValue Handler = Op.getOperand(2);
17238 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17239 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17240 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17241 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17242 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17243 "Invalid Frame Register!");
17244 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17245 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17247 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17248 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17250 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17251 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17253 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17255 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17256 DAG.getRegister(StoreAddrReg, PtrVT));
17259 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17260 SelectionDAG &DAG) const {
17262 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17263 DAG.getVTList(MVT::i32, MVT::Other),
17264 Op.getOperand(0), Op.getOperand(1));
17267 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17268 SelectionDAG &DAG) const {
17270 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17271 Op.getOperand(0), Op.getOperand(1));
17274 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17275 return Op.getOperand(0);
17278 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17279 SelectionDAG &DAG) const {
17280 SDValue Root = Op.getOperand(0);
17281 SDValue Trmp = Op.getOperand(1); // trampoline
17282 SDValue FPtr = Op.getOperand(2); // nested function
17283 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17286 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17287 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17289 if (Subtarget->is64Bit()) {
17290 SDValue OutChains[6];
17292 // Large code-model.
17293 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17294 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17296 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17297 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17299 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17301 // Load the pointer to the nested function into R11.
17302 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17303 SDValue Addr = Trmp;
17304 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17305 Addr, MachinePointerInfo(TrmpAddr),
17308 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17309 DAG.getConstant(2, dl, MVT::i64));
17310 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17311 MachinePointerInfo(TrmpAddr, 2),
17314 // Load the 'nest' parameter value into R10.
17315 // R10 is specified in X86CallingConv.td
17316 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17317 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17318 DAG.getConstant(10, dl, MVT::i64));
17319 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17320 Addr, MachinePointerInfo(TrmpAddr, 10),
17323 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17324 DAG.getConstant(12, dl, MVT::i64));
17325 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17326 MachinePointerInfo(TrmpAddr, 12),
17329 // Jump to the nested function.
17330 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17331 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17332 DAG.getConstant(20, dl, MVT::i64));
17333 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17334 Addr, MachinePointerInfo(TrmpAddr, 20),
17337 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17338 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17339 DAG.getConstant(22, dl, MVT::i64));
17340 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17341 Addr, MachinePointerInfo(TrmpAddr, 22),
17344 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17346 const Function *Func =
17347 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17348 CallingConv::ID CC = Func->getCallingConv();
17353 llvm_unreachable("Unsupported calling convention");
17354 case CallingConv::C:
17355 case CallingConv::X86_StdCall: {
17356 // Pass 'nest' parameter in ECX.
17357 // Must be kept in sync with X86CallingConv.td
17358 NestReg = X86::ECX;
17360 // Check that ECX wasn't needed by an 'inreg' parameter.
17361 FunctionType *FTy = Func->getFunctionType();
17362 const AttributeSet &Attrs = Func->getAttributes();
17364 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17365 unsigned InRegCount = 0;
17368 for (FunctionType::param_iterator I = FTy->param_begin(),
17369 E = FTy->param_end(); I != E; ++I, ++Idx)
17370 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17371 auto &DL = DAG.getDataLayout();
17372 // FIXME: should only count parameters that are lowered to integers.
17373 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17376 if (InRegCount > 2) {
17377 report_fatal_error("Nest register in use - reduce number of inreg"
17383 case CallingConv::X86_FastCall:
17384 case CallingConv::X86_ThisCall:
17385 case CallingConv::Fast:
17386 // Pass 'nest' parameter in EAX.
17387 // Must be kept in sync with X86CallingConv.td
17388 NestReg = X86::EAX;
17392 SDValue OutChains[4];
17393 SDValue Addr, Disp;
17395 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17396 DAG.getConstant(10, dl, MVT::i32));
17397 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17399 // This is storing the opcode for MOV32ri.
17400 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17401 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17402 OutChains[0] = DAG.getStore(Root, dl,
17403 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17404 Trmp, MachinePointerInfo(TrmpAddr),
17407 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17408 DAG.getConstant(1, dl, MVT::i32));
17409 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17410 MachinePointerInfo(TrmpAddr, 1),
17413 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17414 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17415 DAG.getConstant(5, dl, MVT::i32));
17416 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17417 Addr, MachinePointerInfo(TrmpAddr, 5),
17420 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17421 DAG.getConstant(6, dl, MVT::i32));
17422 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17423 MachinePointerInfo(TrmpAddr, 6),
17426 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17430 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17431 SelectionDAG &DAG) const {
17433 The rounding mode is in bits 11:10 of FPSR, and has the following
17435 00 Round to nearest
17440 FLT_ROUNDS, on the other hand, expects the following:
17447 To perform the conversion, we do:
17448 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17451 MachineFunction &MF = DAG.getMachineFunction();
17452 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17453 unsigned StackAlignment = TFI.getStackAlignment();
17454 MVT VT = Op.getSimpleValueType();
17457 // Save FP Control Word to stack slot
17458 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17459 SDValue StackSlot =
17460 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17462 MachineMemOperand *MMO =
17463 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17464 MachineMemOperand::MOStore, 2, 2);
17466 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17467 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17468 DAG.getVTList(MVT::Other),
17469 Ops, MVT::i16, MMO);
17471 // Load FP Control Word from stack slot
17472 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17473 MachinePointerInfo(), false, false, false, 0);
17475 // Transform as necessary
17477 DAG.getNode(ISD::SRL, DL, MVT::i16,
17478 DAG.getNode(ISD::AND, DL, MVT::i16,
17479 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17480 DAG.getConstant(11, DL, MVT::i8));
17482 DAG.getNode(ISD::SRL, DL, MVT::i16,
17483 DAG.getNode(ISD::AND, DL, MVT::i16,
17484 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17485 DAG.getConstant(9, DL, MVT::i8));
17488 DAG.getNode(ISD::AND, DL, MVT::i16,
17489 DAG.getNode(ISD::ADD, DL, MVT::i16,
17490 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17491 DAG.getConstant(1, DL, MVT::i16)),
17492 DAG.getConstant(3, DL, MVT::i16));
17494 return DAG.getNode((VT.getSizeInBits() < 16 ?
17495 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17498 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17500 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17501 // to 512-bit vector.
17502 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17503 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17504 // split the vector, perform operation on it's Lo a Hi part and
17505 // concatenate the results.
17506 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17508 MVT VT = Op.getSimpleValueType();
17509 MVT EltVT = VT.getVectorElementType();
17510 unsigned NumElems = VT.getVectorNumElements();
17512 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17513 // Extend to 512 bit vector.
17514 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17515 "Unsupported value type for operation");
17517 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17518 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17519 DAG.getUNDEF(NewVT),
17521 DAG.getIntPtrConstant(0, dl));
17522 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17524 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17525 DAG.getIntPtrConstant(0, dl));
17528 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17529 "Unsupported element type");
17531 if (16 < NumElems) {
17532 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17534 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17535 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17537 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17538 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17540 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17543 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17545 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17546 "Unsupported value type for operation");
17548 // Use native supported vector instruction vplzcntd.
17549 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17550 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17551 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17552 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17554 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17557 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17558 SelectionDAG &DAG) {
17559 MVT VT = Op.getSimpleValueType();
17561 unsigned NumBits = VT.getSizeInBits();
17564 if (VT.isVector() && Subtarget->hasAVX512())
17565 return LowerVectorCTLZ_AVX512(Op, DAG);
17567 Op = Op.getOperand(0);
17568 if (VT == MVT::i8) {
17569 // Zero extend to i32 since there is not an i8 bsr.
17571 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17574 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17575 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17576 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17578 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17581 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17582 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17585 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17587 // Finally xor with NumBits-1.
17588 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17589 DAG.getConstant(NumBits - 1, dl, OpVT));
17592 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17596 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17597 SelectionDAG &DAG) {
17598 MVT VT = Op.getSimpleValueType();
17600 unsigned NumBits = VT.getSizeInBits();
17603 if (VT.isVector() && Subtarget->hasAVX512())
17604 return LowerVectorCTLZ_AVX512(Op, DAG);
17606 Op = Op.getOperand(0);
17607 if (VT == MVT::i8) {
17608 // Zero extend to i32 since there is not an i8 bsr.
17610 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17613 // Issue a bsr (scan bits in reverse).
17614 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17615 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17617 // And xor with NumBits-1.
17618 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17619 DAG.getConstant(NumBits - 1, dl, OpVT));
17622 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17626 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17627 MVT VT = Op.getSimpleValueType();
17628 unsigned NumBits = VT.getScalarSizeInBits();
17631 if (VT.isVector()) {
17632 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17634 SDValue N0 = Op.getOperand(0);
17635 SDValue Zero = DAG.getConstant(0, dl, VT);
17637 // lsb(x) = (x & -x)
17638 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17639 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17641 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17642 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17643 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17644 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17645 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17646 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17649 // cttz(x) = ctpop(lsb - 1)
17650 SDValue One = DAG.getConstant(1, dl, VT);
17651 return DAG.getNode(ISD::CTPOP, dl, VT,
17652 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17655 assert(Op.getOpcode() == ISD::CTTZ &&
17656 "Only scalar CTTZ requires custom lowering");
17658 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17659 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17660 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17662 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17665 DAG.getConstant(NumBits, dl, VT),
17666 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17669 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17672 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17673 // ones, and then concatenate the result back.
17674 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17675 MVT VT = Op.getSimpleValueType();
17677 assert(VT.is256BitVector() && VT.isInteger() &&
17678 "Unsupported value type for operation");
17680 unsigned NumElems = VT.getVectorNumElements();
17683 // Extract the LHS vectors
17684 SDValue LHS = Op.getOperand(0);
17685 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17686 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17688 // Extract the RHS vectors
17689 SDValue RHS = Op.getOperand(1);
17690 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17691 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17693 MVT EltVT = VT.getVectorElementType();
17694 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17696 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17697 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17698 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17701 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17702 if (Op.getValueType() == MVT::i1)
17703 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17704 Op.getOperand(0), Op.getOperand(1));
17705 assert(Op.getSimpleValueType().is256BitVector() &&
17706 Op.getSimpleValueType().isInteger() &&
17707 "Only handle AVX 256-bit vector integer operation");
17708 return Lower256IntArith(Op, DAG);
17711 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17712 if (Op.getValueType() == MVT::i1)
17713 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17714 Op.getOperand(0), Op.getOperand(1));
17715 assert(Op.getSimpleValueType().is256BitVector() &&
17716 Op.getSimpleValueType().isInteger() &&
17717 "Only handle AVX 256-bit vector integer operation");
17718 return Lower256IntArith(Op, DAG);
17721 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17722 assert(Op.getSimpleValueType().is256BitVector() &&
17723 Op.getSimpleValueType().isInteger() &&
17724 "Only handle AVX 256-bit vector integer operation");
17725 return Lower256IntArith(Op, DAG);
17728 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17729 SelectionDAG &DAG) {
17731 MVT VT = Op.getSimpleValueType();
17734 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17736 // Decompose 256-bit ops into smaller 128-bit ops.
17737 if (VT.is256BitVector() && !Subtarget->hasInt256())
17738 return Lower256IntArith(Op, DAG);
17740 SDValue A = Op.getOperand(0);
17741 SDValue B = Op.getOperand(1);
17743 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17744 // pairs, multiply and truncate.
17745 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17746 if (Subtarget->hasInt256()) {
17747 if (VT == MVT::v32i8) {
17748 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17749 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17750 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17751 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17752 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17753 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17754 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17755 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17756 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17757 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17760 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17761 return DAG.getNode(
17762 ISD::TRUNCATE, dl, VT,
17763 DAG.getNode(ISD::MUL, dl, ExVT,
17764 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17765 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17768 assert(VT == MVT::v16i8 &&
17769 "Pre-AVX2 support only supports v16i8 multiplication");
17770 MVT ExVT = MVT::v8i16;
17772 // Extract the lo parts and sign extend to i16
17774 if (Subtarget->hasSSE41()) {
17775 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17776 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17778 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17779 -1, 4, -1, 5, -1, 6, -1, 7};
17780 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17781 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17782 ALo = DAG.getBitcast(ExVT, ALo);
17783 BLo = DAG.getBitcast(ExVT, BLo);
17784 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17785 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17788 // Extract the hi parts and sign extend to i16
17790 if (Subtarget->hasSSE41()) {
17791 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17792 -1, -1, -1, -1, -1, -1, -1, -1};
17793 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17794 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17795 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17796 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17798 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17799 -1, 12, -1, 13, -1, 14, -1, 15};
17800 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17801 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17802 AHi = DAG.getBitcast(ExVT, AHi);
17803 BHi = DAG.getBitcast(ExVT, BHi);
17804 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17805 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17808 // Multiply, mask the lower 8bits of the lo/hi results and pack
17809 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17810 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17811 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17812 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17813 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17816 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17817 if (VT == MVT::v4i32) {
17818 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17819 "Should not custom lower when pmuldq is available!");
17821 // Extract the odd parts.
17822 static const int UnpackMask[] = { 1, -1, 3, -1 };
17823 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17824 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17826 // Multiply the even parts.
17827 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17828 // Now multiply odd parts.
17829 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17831 Evens = DAG.getBitcast(VT, Evens);
17832 Odds = DAG.getBitcast(VT, Odds);
17834 // Merge the two vectors back together with a shuffle. This expands into 2
17836 static const int ShufMask[] = { 0, 4, 2, 6 };
17837 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17840 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17841 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17843 // Ahi = psrlqi(a, 32);
17844 // Bhi = psrlqi(b, 32);
17846 // AloBlo = pmuludq(a, b);
17847 // AloBhi = pmuludq(a, Bhi);
17848 // AhiBlo = pmuludq(Ahi, b);
17850 // AloBhi = psllqi(AloBhi, 32);
17851 // AhiBlo = psllqi(AhiBlo, 32);
17852 // return AloBlo + AloBhi + AhiBlo;
17854 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17855 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17857 SDValue AhiBlo = Ahi;
17858 SDValue AloBhi = Bhi;
17859 // Bit cast to 32-bit vectors for MULUDQ
17860 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17861 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17862 A = DAG.getBitcast(MulVT, A);
17863 B = DAG.getBitcast(MulVT, B);
17864 Ahi = DAG.getBitcast(MulVT, Ahi);
17865 Bhi = DAG.getBitcast(MulVT, Bhi);
17867 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17868 // After shifting right const values the result may be all-zero.
17869 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17870 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17871 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17873 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17874 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17875 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17878 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17879 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17882 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17883 assert(Subtarget->isTargetWin64() && "Unexpected target");
17884 EVT VT = Op.getValueType();
17885 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17886 "Unexpected return type for lowering");
17890 switch (Op->getOpcode()) {
17891 default: llvm_unreachable("Unexpected request for libcall!");
17892 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17893 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17894 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17895 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17896 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17897 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17901 SDValue InChain = DAG.getEntryNode();
17903 TargetLowering::ArgListTy Args;
17904 TargetLowering::ArgListEntry Entry;
17905 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17906 EVT ArgVT = Op->getOperand(i).getValueType();
17907 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17908 "Unexpected argument type for lowering");
17909 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17910 Entry.Node = StackPtr;
17911 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17913 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17914 Entry.Ty = PointerType::get(ArgTy,0);
17915 Entry.isSExt = false;
17916 Entry.isZExt = false;
17917 Args.push_back(Entry);
17920 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17921 getPointerTy(DAG.getDataLayout()));
17923 TargetLowering::CallLoweringInfo CLI(DAG);
17924 CLI.setDebugLoc(dl).setChain(InChain)
17925 .setCallee(getLibcallCallingConv(LC),
17926 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17927 Callee, std::move(Args), 0)
17928 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17930 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17931 return DAG.getBitcast(VT, CallInfo.first);
17934 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17935 SelectionDAG &DAG) {
17936 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17937 EVT VT = Op0.getValueType();
17940 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17941 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17943 // PMULxD operations multiply each even value (starting at 0) of LHS with
17944 // the related value of RHS and produce a widen result.
17945 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17946 // => <2 x i64> <ae|cg>
17948 // In other word, to have all the results, we need to perform two PMULxD:
17949 // 1. one with the even values.
17950 // 2. one with the odd values.
17951 // To achieve #2, with need to place the odd values at an even position.
17953 // Place the odd value at an even position (basically, shift all values 1
17954 // step to the left):
17955 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17956 // <a|b|c|d> => <b|undef|d|undef>
17957 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17958 // <e|f|g|h> => <f|undef|h|undef>
17959 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17961 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17963 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17964 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17966 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17967 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17968 // => <2 x i64> <ae|cg>
17969 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17970 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17971 // => <2 x i64> <bf|dh>
17972 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17974 // Shuffle it back into the right order.
17975 SDValue Highs, Lows;
17976 if (VT == MVT::v8i32) {
17977 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17978 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17979 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17980 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17982 const int HighMask[] = {1, 5, 3, 7};
17983 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17984 const int LowMask[] = {0, 4, 2, 6};
17985 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17988 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17989 // unsigned multiply.
17990 if (IsSigned && !Subtarget->hasSSE41()) {
17991 SDValue ShAmt = DAG.getConstant(
17993 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17994 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17995 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17996 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17997 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17999 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18000 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18003 // The first result of MUL_LOHI is actually the low value, followed by the
18005 SDValue Ops[] = {Lows, Highs};
18006 return DAG.getMergeValues(Ops, dl);
18009 // Return true if the required (according to Opcode) shift-imm form is natively
18010 // supported by the Subtarget
18011 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18013 if (VT.getScalarSizeInBits() < 16)
18016 if (VT.is512BitVector() &&
18017 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18020 bool LShift = VT.is128BitVector() ||
18021 (VT.is256BitVector() && Subtarget->hasInt256());
18023 bool AShift = LShift && (Subtarget->hasVLX() ||
18024 (VT != MVT::v2i64 && VT != MVT::v4i64));
18025 return (Opcode == ISD::SRA) ? AShift : LShift;
18028 // The shift amount is a variable, but it is the same for all vector lanes.
18029 // These instructions are defined together with shift-immediate.
18031 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18033 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18036 // Return true if the required (according to Opcode) variable-shift form is
18037 // natively supported by the Subtarget
18038 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18041 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18044 // vXi16 supported only on AVX-512, BWI
18045 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18048 if (VT.is512BitVector() || Subtarget->hasVLX())
18051 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18052 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18053 return (Opcode == ISD::SRA) ? AShift : LShift;
18056 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18057 const X86Subtarget *Subtarget) {
18058 MVT VT = Op.getSimpleValueType();
18060 SDValue R = Op.getOperand(0);
18061 SDValue Amt = Op.getOperand(1);
18063 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18064 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18066 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18067 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18068 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18069 SDValue Ex = DAG.getBitcast(ExVT, R);
18071 if (ShiftAmt >= 32) {
18072 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18074 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18075 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18076 ShiftAmt - 32, DAG);
18077 if (VT == MVT::v2i64)
18078 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18079 if (VT == MVT::v4i64)
18080 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18081 {9, 1, 11, 3, 13, 5, 15, 7});
18083 // SRA upper i32, SHL whole i64 and select lower i32.
18084 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18087 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18088 Lower = DAG.getBitcast(ExVT, Lower);
18089 if (VT == MVT::v2i64)
18090 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18091 if (VT == MVT::v4i64)
18092 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18093 {8, 1, 10, 3, 12, 5, 14, 7});
18095 return DAG.getBitcast(VT, Ex);
18098 // Optimize shl/srl/sra with constant shift amount.
18099 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18100 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18101 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18103 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18104 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18106 // i64 SRA needs to be performed as partial shifts.
18107 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18108 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18109 return ArithmeticShiftRight64(ShiftAmt);
18111 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18112 unsigned NumElts = VT.getVectorNumElements();
18113 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18115 // Simple i8 add case
18116 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18117 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18119 // ashr(R, 7) === cmp_slt(R, 0)
18120 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18121 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18122 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18125 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18126 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18129 if (Op.getOpcode() == ISD::SHL) {
18130 // Make a large shift.
18131 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18133 SHL = DAG.getBitcast(VT, SHL);
18134 // Zero out the rightmost bits.
18135 SmallVector<SDValue, 32> V(
18136 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18137 return DAG.getNode(ISD::AND, dl, VT, SHL,
18138 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18140 if (Op.getOpcode() == ISD::SRL) {
18141 // Make a large shift.
18142 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18144 SRL = DAG.getBitcast(VT, SRL);
18145 // Zero out the leftmost bits.
18146 SmallVector<SDValue, 32> V(
18147 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18148 return DAG.getNode(ISD::AND, dl, VT, SRL,
18149 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18151 if (Op.getOpcode() == ISD::SRA) {
18152 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18153 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18154 SmallVector<SDValue, 32> V(NumElts,
18155 DAG.getConstant(128 >> ShiftAmt, dl,
18157 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18158 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18159 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18162 llvm_unreachable("Unknown shift opcode.");
18167 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18168 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18169 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18171 // Peek through any splat that was introduced for i64 shift vectorization.
18172 int SplatIndex = -1;
18173 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18174 if (SVN->isSplat()) {
18175 SplatIndex = SVN->getSplatIndex();
18176 Amt = Amt.getOperand(0);
18177 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18178 "Splat shuffle referencing second operand");
18181 if (Amt.getOpcode() != ISD::BITCAST ||
18182 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18185 Amt = Amt.getOperand(0);
18186 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18187 VT.getVectorNumElements();
18188 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18189 uint64_t ShiftAmt = 0;
18190 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18191 for (unsigned i = 0; i != Ratio; ++i) {
18192 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18196 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18199 // Check remaining shift amounts (if not a splat).
18200 if (SplatIndex < 0) {
18201 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18202 uint64_t ShAmt = 0;
18203 for (unsigned j = 0; j != Ratio; ++j) {
18204 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18208 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18210 if (ShAmt != ShiftAmt)
18215 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18216 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18218 if (Op.getOpcode() == ISD::SRA)
18219 return ArithmeticShiftRight64(ShiftAmt);
18225 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18226 const X86Subtarget* Subtarget) {
18227 MVT VT = Op.getSimpleValueType();
18229 SDValue R = Op.getOperand(0);
18230 SDValue Amt = Op.getOperand(1);
18232 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18233 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18235 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18236 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18238 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18240 MVT EltVT = VT.getVectorElementType();
18242 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18243 // Check if this build_vector node is doing a splat.
18244 // If so, then set BaseShAmt equal to the splat value.
18245 BaseShAmt = BV->getSplatValue();
18246 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18247 BaseShAmt = SDValue();
18249 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18250 Amt = Amt.getOperand(0);
18252 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18253 if (SVN && SVN->isSplat()) {
18254 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18255 SDValue InVec = Amt.getOperand(0);
18256 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18257 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18258 "Unexpected shuffle index found!");
18259 BaseShAmt = InVec.getOperand(SplatIdx);
18260 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18261 if (ConstantSDNode *C =
18262 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18263 if (C->getZExtValue() == SplatIdx)
18264 BaseShAmt = InVec.getOperand(1);
18269 // Avoid introducing an extract element from a shuffle.
18270 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18271 DAG.getIntPtrConstant(SplatIdx, dl));
18275 if (BaseShAmt.getNode()) {
18276 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18277 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18278 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18279 else if (EltVT.bitsLT(MVT::i32))
18280 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18282 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18286 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18287 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18288 Amt.getOpcode() == ISD::BITCAST &&
18289 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18290 Amt = Amt.getOperand(0);
18291 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18292 VT.getVectorNumElements();
18293 std::vector<SDValue> Vals(Ratio);
18294 for (unsigned i = 0; i != Ratio; ++i)
18295 Vals[i] = Amt.getOperand(i);
18296 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18297 for (unsigned j = 0; j != Ratio; ++j)
18298 if (Vals[j] != Amt.getOperand(i + j))
18302 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18303 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18308 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18309 SelectionDAG &DAG) {
18310 MVT VT = Op.getSimpleValueType();
18312 SDValue R = Op.getOperand(0);
18313 SDValue Amt = Op.getOperand(1);
18315 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18316 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18318 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18321 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18324 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18327 // XOP has 128-bit variable logical/arithmetic shifts.
18328 // +ve/-ve Amt = shift left/right.
18329 if (Subtarget->hasXOP() &&
18330 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18331 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18332 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18333 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18334 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18336 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18337 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18338 if (Op.getOpcode() == ISD::SRA)
18339 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18342 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18343 // shifts per-lane and then shuffle the partial results back together.
18344 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18345 // Splat the shift amounts so the scalar shifts above will catch it.
18346 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18347 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18348 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18349 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18350 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18353 // i64 vector arithmetic shift can be emulated with the transform:
18354 // M = lshr(SIGN_BIT, Amt)
18355 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18356 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18357 Op.getOpcode() == ISD::SRA) {
18358 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18359 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18360 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18361 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18362 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18366 // If possible, lower this packed shift into a vector multiply instead of
18367 // expanding it into a sequence of scalar shifts.
18368 // Do this only if the vector shift count is a constant build_vector.
18369 if (Op.getOpcode() == ISD::SHL &&
18370 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18371 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18372 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18373 SmallVector<SDValue, 8> Elts;
18374 EVT SVT = VT.getScalarType();
18375 unsigned SVTBits = SVT.getSizeInBits();
18376 APInt One(SVTBits, 1);
18377 unsigned NumElems = VT.getVectorNumElements();
18379 for (unsigned i=0; i !=NumElems; ++i) {
18380 SDValue Op = Amt->getOperand(i);
18381 if (Op->getOpcode() == ISD::UNDEF) {
18382 Elts.push_back(Op);
18386 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18387 APInt C(SVTBits, ND->getAPIntValue().getZExtValue());
18388 uint64_t ShAmt = C.getZExtValue();
18389 if (ShAmt >= SVTBits) {
18390 Elts.push_back(DAG.getUNDEF(SVT));
18393 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18395 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18396 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18399 // Lower SHL with variable shift amount.
18400 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18401 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18403 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18404 DAG.getConstant(0x3f800000U, dl, VT));
18405 Op = DAG.getBitcast(MVT::v4f32, Op);
18406 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18407 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18410 // If possible, lower this shift as a sequence of two shifts by
18411 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18413 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18415 // Could be rewritten as:
18416 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18418 // The advantage is that the two shifts from the example would be
18419 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18420 // the vector shift into four scalar shifts plus four pairs of vector
18422 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18423 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18424 unsigned TargetOpcode = X86ISD::MOVSS;
18425 bool CanBeSimplified;
18426 // The splat value for the first packed shift (the 'X' from the example).
18427 SDValue Amt1 = Amt->getOperand(0);
18428 // The splat value for the second packed shift (the 'Y' from the example).
18429 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18430 Amt->getOperand(2);
18432 // See if it is possible to replace this node with a sequence of
18433 // two shifts followed by a MOVSS/MOVSD
18434 if (VT == MVT::v4i32) {
18435 // Check if it is legal to use a MOVSS.
18436 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18437 Amt2 == Amt->getOperand(3);
18438 if (!CanBeSimplified) {
18439 // Otherwise, check if we can still simplify this node using a MOVSD.
18440 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18441 Amt->getOperand(2) == Amt->getOperand(3);
18442 TargetOpcode = X86ISD::MOVSD;
18443 Amt2 = Amt->getOperand(2);
18446 // Do similar checks for the case where the machine value type
18448 CanBeSimplified = Amt1 == Amt->getOperand(1);
18449 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18450 CanBeSimplified = Amt2 == Amt->getOperand(i);
18452 if (!CanBeSimplified) {
18453 TargetOpcode = X86ISD::MOVSD;
18454 CanBeSimplified = true;
18455 Amt2 = Amt->getOperand(4);
18456 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18457 CanBeSimplified = Amt1 == Amt->getOperand(i);
18458 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18459 CanBeSimplified = Amt2 == Amt->getOperand(j);
18463 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18464 isa<ConstantSDNode>(Amt2)) {
18465 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18466 MVT CastVT = MVT::v4i32;
18468 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18469 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18471 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18472 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18473 if (TargetOpcode == X86ISD::MOVSD)
18474 CastVT = MVT::v2i64;
18475 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18476 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18477 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18479 return DAG.getBitcast(VT, Result);
18483 // v4i32 Non Uniform Shifts.
18484 // If the shift amount is constant we can shift each lane using the SSE2
18485 // immediate shifts, else we need to zero-extend each lane to the lower i64
18486 // and shift using the SSE2 variable shifts.
18487 // The separate results can then be blended together.
18488 if (VT == MVT::v4i32) {
18489 unsigned Opc = Op.getOpcode();
18490 SDValue Amt0, Amt1, Amt2, Amt3;
18491 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18492 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18493 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18494 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18495 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18497 // ISD::SHL is handled above but we include it here for completeness.
18500 llvm_unreachable("Unknown target vector shift node");
18502 Opc = X86ISD::VSHL;
18505 Opc = X86ISD::VSRL;
18508 Opc = X86ISD::VSRA;
18511 // The SSE2 shifts use the lower i64 as the same shift amount for
18512 // all lanes and the upper i64 is ignored. These shuffle masks
18513 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18514 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18515 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18516 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18517 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18518 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18521 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18522 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18523 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18524 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18525 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18526 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18527 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18530 if (VT == MVT::v16i8 ||
18531 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18532 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18533 unsigned ShiftOpcode = Op->getOpcode();
18535 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18536 // On SSE41 targets we make use of the fact that VSELECT lowers
18537 // to PBLENDVB which selects bytes based just on the sign bit.
18538 if (Subtarget->hasSSE41()) {
18539 V0 = DAG.getBitcast(VT, V0);
18540 V1 = DAG.getBitcast(VT, V1);
18541 Sel = DAG.getBitcast(VT, Sel);
18542 return DAG.getBitcast(SelVT,
18543 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18545 // On pre-SSE41 targets we test for the sign bit by comparing to
18546 // zero - a negative value will set all bits of the lanes to true
18547 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18548 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18549 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18550 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18553 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18554 // We can safely do this using i16 shifts as we're only interested in
18555 // the 3 lower bits of each byte.
18556 Amt = DAG.getBitcast(ExtVT, Amt);
18557 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18558 Amt = DAG.getBitcast(VT, Amt);
18560 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18561 // r = VSELECT(r, shift(r, 4), a);
18563 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18564 R = SignBitSelect(VT, Amt, M, R);
18567 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18569 // r = VSELECT(r, shift(r, 2), a);
18570 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18571 R = SignBitSelect(VT, Amt, M, R);
18574 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18576 // return VSELECT(r, shift(r, 1), a);
18577 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18578 R = SignBitSelect(VT, Amt, M, R);
18582 if (Op->getOpcode() == ISD::SRA) {
18583 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18584 // so we can correctly sign extend. We don't care what happens to the
18586 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18587 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18588 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18589 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18590 ALo = DAG.getBitcast(ExtVT, ALo);
18591 AHi = DAG.getBitcast(ExtVT, AHi);
18592 RLo = DAG.getBitcast(ExtVT, RLo);
18593 RHi = DAG.getBitcast(ExtVT, RHi);
18595 // r = VSELECT(r, shift(r, 4), a);
18596 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18597 DAG.getConstant(4, dl, ExtVT));
18598 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18599 DAG.getConstant(4, dl, ExtVT));
18600 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18601 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18604 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18605 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18607 // r = VSELECT(r, shift(r, 2), a);
18608 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18609 DAG.getConstant(2, dl, ExtVT));
18610 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18611 DAG.getConstant(2, dl, ExtVT));
18612 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18613 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18616 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18617 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18619 // r = VSELECT(r, shift(r, 1), a);
18620 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18621 DAG.getConstant(1, dl, ExtVT));
18622 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18623 DAG.getConstant(1, dl, ExtVT));
18624 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18625 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18627 // Logical shift the result back to the lower byte, leaving a zero upper
18629 // meaning that we can safely pack with PACKUSWB.
18631 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18633 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18634 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18638 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18639 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18640 // solution better.
18641 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18642 MVT ExtVT = MVT::v8i32;
18644 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18645 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18646 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18647 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18648 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18651 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18652 MVT ExtVT = MVT::v8i32;
18653 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18654 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18655 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18656 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18657 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18658 ALo = DAG.getBitcast(ExtVT, ALo);
18659 AHi = DAG.getBitcast(ExtVT, AHi);
18660 RLo = DAG.getBitcast(ExtVT, RLo);
18661 RHi = DAG.getBitcast(ExtVT, RHi);
18662 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18663 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18664 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18665 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18666 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18669 if (VT == MVT::v8i16) {
18670 unsigned ShiftOpcode = Op->getOpcode();
18672 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18673 // On SSE41 targets we make use of the fact that VSELECT lowers
18674 // to PBLENDVB which selects bytes based just on the sign bit.
18675 if (Subtarget->hasSSE41()) {
18676 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18677 V0 = DAG.getBitcast(ExtVT, V0);
18678 V1 = DAG.getBitcast(ExtVT, V1);
18679 Sel = DAG.getBitcast(ExtVT, Sel);
18680 return DAG.getBitcast(
18681 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18683 // On pre-SSE41 targets we splat the sign bit - a negative value will
18684 // set all bits of the lanes to true and VSELECT uses that in
18685 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18687 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18688 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18691 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18692 if (Subtarget->hasSSE41()) {
18693 // On SSE41 targets we need to replicate the shift mask in both
18694 // bytes for PBLENDVB.
18697 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18698 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18700 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18703 // r = VSELECT(r, shift(r, 8), a);
18704 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18705 R = SignBitSelect(Amt, M, R);
18708 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18710 // r = VSELECT(r, shift(r, 4), a);
18711 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18712 R = SignBitSelect(Amt, M, R);
18715 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18717 // r = VSELECT(r, shift(r, 2), a);
18718 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18719 R = SignBitSelect(Amt, M, R);
18722 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18724 // return VSELECT(r, shift(r, 1), a);
18725 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18726 R = SignBitSelect(Amt, M, R);
18730 // Decompose 256-bit shifts into smaller 128-bit shifts.
18731 if (VT.is256BitVector()) {
18732 unsigned NumElems = VT.getVectorNumElements();
18733 MVT EltVT = VT.getVectorElementType();
18734 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18736 // Extract the two vectors
18737 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18738 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18740 // Recreate the shift amount vectors
18741 SDValue Amt1, Amt2;
18742 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18743 // Constant shift amount
18744 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18745 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18746 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18748 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18749 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18751 // Variable shift amount
18752 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18753 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18756 // Issue new vector shifts for the smaller types
18757 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18758 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18760 // Concatenate the result back
18761 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18767 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18768 SelectionDAG &DAG) {
18769 MVT VT = Op.getSimpleValueType();
18771 SDValue R = Op.getOperand(0);
18772 SDValue Amt = Op.getOperand(1);
18774 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18775 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18776 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18778 // XOP has 128-bit vector variable + immediate rotates.
18779 // +ve/-ve Amt = rotate left/right.
18781 // Split 256-bit integers.
18782 if (VT.is256BitVector())
18783 return Lower256IntArith(Op, DAG);
18785 assert(VT.is128BitVector() && "Only rotate 128-bit vectors!");
18787 // Attempt to rotate by immediate.
18788 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18789 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18790 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18791 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18792 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18793 DAG.getConstant(RotateAmt, DL, MVT::i8));
18797 // Use general rotate by variable (per-element).
18798 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18801 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18802 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18803 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18804 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18805 // has only one use.
18806 SDNode *N = Op.getNode();
18807 SDValue LHS = N->getOperand(0);
18808 SDValue RHS = N->getOperand(1);
18809 unsigned BaseOp = 0;
18812 switch (Op.getOpcode()) {
18813 default: llvm_unreachable("Unknown ovf instruction!");
18815 // A subtract of one will be selected as a INC. Note that INC doesn't
18816 // set CF, so we can't do this for UADDO.
18817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18819 BaseOp = X86ISD::INC;
18820 Cond = X86::COND_O;
18823 BaseOp = X86ISD::ADD;
18824 Cond = X86::COND_O;
18827 BaseOp = X86ISD::ADD;
18828 Cond = X86::COND_B;
18831 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18832 // set CF, so we can't do this for USUBO.
18833 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18835 BaseOp = X86ISD::DEC;
18836 Cond = X86::COND_O;
18839 BaseOp = X86ISD::SUB;
18840 Cond = X86::COND_O;
18843 BaseOp = X86ISD::SUB;
18844 Cond = X86::COND_B;
18847 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18848 Cond = X86::COND_O;
18850 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18851 if (N->getValueType(0) == MVT::i8) {
18852 BaseOp = X86ISD::UMUL8;
18853 Cond = X86::COND_O;
18856 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18858 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18861 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18862 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18863 SDValue(Sum.getNode(), 2));
18865 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18869 // Also sets EFLAGS.
18870 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18871 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18874 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18875 DAG.getConstant(Cond, DL, MVT::i32),
18876 SDValue(Sum.getNode(), 1));
18878 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18881 /// Returns true if the operand type is exactly twice the native width, and
18882 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18883 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18884 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18885 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18886 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18889 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18890 else if (OpWidth == 128)
18891 return Subtarget->hasCmpxchg16b();
18896 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18897 return needsCmpXchgNb(SI->getValueOperand()->getType());
18900 // Note: this turns large loads into lock cmpxchg8b/16b.
18901 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18902 TargetLowering::AtomicExpansionKind
18903 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18904 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18905 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18906 : AtomicExpansionKind::None;
18909 TargetLowering::AtomicExpansionKind
18910 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18911 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18912 Type *MemType = AI->getType();
18914 // If the operand is too big, we must see if cmpxchg8/16b is available
18915 // and default to library calls otherwise.
18916 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18917 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18918 : AtomicExpansionKind::None;
18921 AtomicRMWInst::BinOp Op = AI->getOperation();
18924 llvm_unreachable("Unknown atomic operation");
18925 case AtomicRMWInst::Xchg:
18926 case AtomicRMWInst::Add:
18927 case AtomicRMWInst::Sub:
18928 // It's better to use xadd, xsub or xchg for these in all cases.
18929 return AtomicExpansionKind::None;
18930 case AtomicRMWInst::Or:
18931 case AtomicRMWInst::And:
18932 case AtomicRMWInst::Xor:
18933 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18934 // prefix to a normal instruction for these operations.
18935 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18936 : AtomicExpansionKind::None;
18937 case AtomicRMWInst::Nand:
18938 case AtomicRMWInst::Max:
18939 case AtomicRMWInst::Min:
18940 case AtomicRMWInst::UMax:
18941 case AtomicRMWInst::UMin:
18942 // These always require a non-trivial set of data operations on x86. We must
18943 // use a cmpxchg loop.
18944 return AtomicExpansionKind::CmpXChg;
18948 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18949 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18950 // no-sse2). There isn't any reason to disable it if the target processor
18952 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18956 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18957 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18958 Type *MemType = AI->getType();
18959 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18960 // there is no benefit in turning such RMWs into loads, and it is actually
18961 // harmful as it introduces a mfence.
18962 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18965 auto Builder = IRBuilder<>(AI);
18966 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18967 auto SynchScope = AI->getSynchScope();
18968 // We must restrict the ordering to avoid generating loads with Release or
18969 // ReleaseAcquire orderings.
18970 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18971 auto Ptr = AI->getPointerOperand();
18973 // Before the load we need a fence. Here is an example lifted from
18974 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18977 // x.store(1, relaxed);
18978 // r1 = y.fetch_add(0, release);
18980 // y.fetch_add(42, acquire);
18981 // r2 = x.load(relaxed);
18982 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18983 // lowered to just a load without a fence. A mfence flushes the store buffer,
18984 // making the optimization clearly correct.
18985 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18986 // otherwise, we might be able to be more aggressive on relaxed idempotent
18987 // rmw. In practice, they do not look useful, so we don't try to be
18988 // especially clever.
18989 if (SynchScope == SingleThread)
18990 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18991 // the IR level, so we must wrap it in an intrinsic.
18994 if (!hasMFENCE(*Subtarget))
18995 // FIXME: it might make sense to use a locked operation here but on a
18996 // different cache-line to prevent cache-line bouncing. In practice it
18997 // is probably a small win, and x86 processors without mfence are rare
18998 // enough that we do not bother.
19002 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19003 Builder.CreateCall(MFence, {});
19005 // Finally we can emit the atomic load.
19006 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19007 AI->getType()->getPrimitiveSizeInBits());
19008 Loaded->setAtomic(Order, SynchScope);
19009 AI->replaceAllUsesWith(Loaded);
19010 AI->eraseFromParent();
19014 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19015 SelectionDAG &DAG) {
19017 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19018 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19019 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19020 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19022 // The only fence that needs an instruction is a sequentially-consistent
19023 // cross-thread fence.
19024 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19025 if (hasMFENCE(*Subtarget))
19026 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19028 SDValue Chain = Op.getOperand(0);
19029 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19031 DAG.getRegister(X86::ESP, MVT::i32), // Base
19032 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19033 DAG.getRegister(0, MVT::i32), // Index
19034 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19035 DAG.getRegister(0, MVT::i32), // Segment.
19039 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19040 return SDValue(Res, 0);
19043 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19044 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19047 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19048 SelectionDAG &DAG) {
19049 MVT T = Op.getSimpleValueType();
19053 switch(T.SimpleTy) {
19054 default: llvm_unreachable("Invalid value type!");
19055 case MVT::i8: Reg = X86::AL; size = 1; break;
19056 case MVT::i16: Reg = X86::AX; size = 2; break;
19057 case MVT::i32: Reg = X86::EAX; size = 4; break;
19059 assert(Subtarget->is64Bit() && "Node not type legal!");
19060 Reg = X86::RAX; size = 8;
19063 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19064 Op.getOperand(2), SDValue());
19065 SDValue Ops[] = { cpIn.getValue(0),
19068 DAG.getTargetConstant(size, DL, MVT::i8),
19069 cpIn.getValue(1) };
19070 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19071 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19072 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19076 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19077 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19078 MVT::i32, cpOut.getValue(2));
19079 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19080 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19083 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19084 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19085 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19089 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19090 SelectionDAG &DAG) {
19091 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19092 MVT DstVT = Op.getSimpleValueType();
19094 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19095 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19096 if (DstVT != MVT::f64)
19097 // This conversion needs to be expanded.
19100 SDValue InVec = Op->getOperand(0);
19102 unsigned NumElts = SrcVT.getVectorNumElements();
19103 MVT SVT = SrcVT.getVectorElementType();
19105 // Widen the vector in input in the case of MVT::v2i32.
19106 // Example: from MVT::v2i32 to MVT::v4i32.
19107 SmallVector<SDValue, 16> Elts;
19108 for (unsigned i = 0, e = NumElts; i != e; ++i)
19109 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19110 DAG.getIntPtrConstant(i, dl)));
19112 // Explicitly mark the extra elements as Undef.
19113 Elts.append(NumElts, DAG.getUNDEF(SVT));
19115 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19116 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19117 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19118 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19119 DAG.getIntPtrConstant(0, dl));
19122 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19123 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19124 assert((DstVT == MVT::i64 ||
19125 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19126 "Unexpected custom BITCAST");
19127 // i64 <=> MMX conversions are Legal.
19128 if (SrcVT==MVT::i64 && DstVT.isVector())
19130 if (DstVT==MVT::i64 && SrcVT.isVector())
19132 // MMX <=> MMX conversions are Legal.
19133 if (SrcVT.isVector() && DstVT.isVector())
19135 // All other conversions need to be expanded.
19139 /// Compute the horizontal sum of bytes in V for the elements of VT.
19141 /// Requires V to be a byte vector and VT to be an integer vector type with
19142 /// wider elements than V's type. The width of the elements of VT determines
19143 /// how many bytes of V are summed horizontally to produce each element of the
19145 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19146 const X86Subtarget *Subtarget,
19147 SelectionDAG &DAG) {
19149 MVT ByteVecVT = V.getSimpleValueType();
19150 MVT EltVT = VT.getVectorElementType();
19151 int NumElts = VT.getVectorNumElements();
19152 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19153 "Expected value to have byte element type.");
19154 assert(EltVT != MVT::i8 &&
19155 "Horizontal byte sum only makes sense for wider elements!");
19156 unsigned VecSize = VT.getSizeInBits();
19157 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19159 // PSADBW instruction horizontally add all bytes and leave the result in i64
19160 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19161 if (EltVT == MVT::i64) {
19162 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19163 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19164 return DAG.getBitcast(VT, V);
19167 if (EltVT == MVT::i32) {
19168 // We unpack the low half and high half into i32s interleaved with zeros so
19169 // that we can use PSADBW to horizontally sum them. The most useful part of
19170 // this is that it lines up the results of two PSADBW instructions to be
19171 // two v2i64 vectors which concatenated are the 4 population counts. We can
19172 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19173 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19174 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19175 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19177 // Do the horizontal sums into two v2i64s.
19178 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19179 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19180 DAG.getBitcast(ByteVecVT, Low), Zeros);
19181 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19182 DAG.getBitcast(ByteVecVT, High), Zeros);
19184 // Merge them together.
19185 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19186 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19187 DAG.getBitcast(ShortVecVT, Low),
19188 DAG.getBitcast(ShortVecVT, High));
19190 return DAG.getBitcast(VT, V);
19193 // The only element type left is i16.
19194 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19196 // To obtain pop count for each i16 element starting from the pop count for
19197 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19198 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19199 // directly supported.
19200 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19201 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19202 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19203 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19204 DAG.getBitcast(ByteVecVT, V));
19205 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19208 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19209 const X86Subtarget *Subtarget,
19210 SelectionDAG &DAG) {
19211 MVT VT = Op.getSimpleValueType();
19212 MVT EltVT = VT.getVectorElementType();
19213 unsigned VecSize = VT.getSizeInBits();
19215 // Implement a lookup table in register by using an algorithm based on:
19216 // http://wm.ite.pl/articles/sse-popcount.html
19218 // The general idea is that every lower byte nibble in the input vector is an
19219 // index into a in-register pre-computed pop count table. We then split up the
19220 // input vector in two new ones: (1) a vector with only the shifted-right
19221 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19222 // masked out higher ones) for each byte. PSHUB is used separately with both
19223 // to index the in-register table. Next, both are added and the result is a
19224 // i8 vector where each element contains the pop count for input byte.
19226 // To obtain the pop count for elements != i8, we follow up with the same
19227 // approach and use additional tricks as described below.
19229 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19230 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19231 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19232 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19234 int NumByteElts = VecSize / 8;
19235 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19236 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19237 SmallVector<SDValue, 16> LUTVec;
19238 for (int i = 0; i < NumByteElts; ++i)
19239 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19240 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19241 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19242 DAG.getConstant(0x0F, DL, MVT::i8));
19243 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19246 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19247 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19248 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19251 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19253 // The input vector is used as the shuffle mask that index elements into the
19254 // LUT. After counting low and high nibbles, add the vector to obtain the
19255 // final pop count per i8 element.
19256 SDValue HighPopCnt =
19257 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19258 SDValue LowPopCnt =
19259 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19260 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19262 if (EltVT == MVT::i8)
19265 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19268 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19269 const X86Subtarget *Subtarget,
19270 SelectionDAG &DAG) {
19271 MVT VT = Op.getSimpleValueType();
19272 assert(VT.is128BitVector() &&
19273 "Only 128-bit vector bitmath lowering supported.");
19275 int VecSize = VT.getSizeInBits();
19276 MVT EltVT = VT.getVectorElementType();
19277 int Len = EltVT.getSizeInBits();
19279 // This is the vectorized version of the "best" algorithm from
19280 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19281 // with a minor tweak to use a series of adds + shifts instead of vector
19282 // multiplications. Implemented for all integer vector types. We only use
19283 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19284 // much faster, even faster than using native popcnt instructions.
19286 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19287 MVT VT = V.getSimpleValueType();
19288 SmallVector<SDValue, 32> Shifters(
19289 VT.getVectorNumElements(),
19290 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19291 return DAG.getNode(OpCode, DL, VT, V,
19292 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19294 auto GetMask = [&](SDValue V, APInt Mask) {
19295 MVT VT = V.getSimpleValueType();
19296 SmallVector<SDValue, 32> Masks(
19297 VT.getVectorNumElements(),
19298 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19299 return DAG.getNode(ISD::AND, DL, VT, V,
19300 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19303 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19304 // x86, so set the SRL type to have elements at least i16 wide. This is
19305 // correct because all of our SRLs are followed immediately by a mask anyways
19306 // that handles any bits that sneak into the high bits of the byte elements.
19307 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19311 // v = v - ((v >> 1) & 0x55555555...)
19313 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19314 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19315 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19317 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19318 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19319 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19320 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19321 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19323 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19324 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19325 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19326 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19328 // At this point, V contains the byte-wise population count, and we are
19329 // merely doing a horizontal sum if necessary to get the wider element
19331 if (EltVT == MVT::i8)
19334 return LowerHorizontalByteSum(
19335 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19339 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19340 SelectionDAG &DAG) {
19341 MVT VT = Op.getSimpleValueType();
19342 // FIXME: Need to add AVX-512 support here!
19343 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19344 "Unknown CTPOP type to handle");
19345 SDLoc DL(Op.getNode());
19346 SDValue Op0 = Op.getOperand(0);
19348 if (!Subtarget->hasSSSE3()) {
19349 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19350 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19351 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19354 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19355 unsigned NumElems = VT.getVectorNumElements();
19357 // Extract each 128-bit vector, compute pop count and concat the result.
19358 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19359 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19361 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19362 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19363 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19366 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19369 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19370 SelectionDAG &DAG) {
19371 assert(Op.getValueType().isVector() &&
19372 "We only do custom lowering for vector population count.");
19373 return LowerVectorCTPOP(Op, Subtarget, DAG);
19376 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19377 SDNode *Node = Op.getNode();
19379 EVT T = Node->getValueType(0);
19380 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19381 DAG.getConstant(0, dl, T), Node->getOperand(2));
19382 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19383 cast<AtomicSDNode>(Node)->getMemoryVT(),
19384 Node->getOperand(0),
19385 Node->getOperand(1), negOp,
19386 cast<AtomicSDNode>(Node)->getMemOperand(),
19387 cast<AtomicSDNode>(Node)->getOrdering(),
19388 cast<AtomicSDNode>(Node)->getSynchScope());
19391 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19392 SDNode *Node = Op.getNode();
19394 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19396 // Convert seq_cst store -> xchg
19397 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19398 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19399 // (The only way to get a 16-byte store is cmpxchg16b)
19400 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19401 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19402 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19403 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19404 cast<AtomicSDNode>(Node)->getMemoryVT(),
19405 Node->getOperand(0),
19406 Node->getOperand(1), Node->getOperand(2),
19407 cast<AtomicSDNode>(Node)->getMemOperand(),
19408 cast<AtomicSDNode>(Node)->getOrdering(),
19409 cast<AtomicSDNode>(Node)->getSynchScope());
19410 return Swap.getValue(1);
19412 // Other atomic stores have a simple pattern.
19416 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19417 EVT VT = Op.getNode()->getSimpleValueType(0);
19419 // Let legalize expand this if it isn't a legal type yet.
19420 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19423 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19426 bool ExtraOp = false;
19427 switch (Op.getOpcode()) {
19428 default: llvm_unreachable("Invalid code");
19429 case ISD::ADDC: Opc = X86ISD::ADD; break;
19430 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19431 case ISD::SUBC: Opc = X86ISD::SUB; break;
19432 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19436 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19438 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19439 Op.getOperand(1), Op.getOperand(2));
19442 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19443 SelectionDAG &DAG) {
19444 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19446 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19447 // which returns the values as { float, float } (in XMM0) or
19448 // { double, double } (which is returned in XMM0, XMM1).
19450 SDValue Arg = Op.getOperand(0);
19451 EVT ArgVT = Arg.getValueType();
19452 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19454 TargetLowering::ArgListTy Args;
19455 TargetLowering::ArgListEntry Entry;
19459 Entry.isSExt = false;
19460 Entry.isZExt = false;
19461 Args.push_back(Entry);
19463 bool isF64 = ArgVT == MVT::f64;
19464 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19465 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19466 // the results are returned via SRet in memory.
19467 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19468 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19470 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19472 Type *RetTy = isF64
19473 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19474 : (Type*)VectorType::get(ArgTy, 4);
19476 TargetLowering::CallLoweringInfo CLI(DAG);
19477 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19478 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19480 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19483 // Returned in xmm0 and xmm1.
19484 return CallResult.first;
19486 // Returned in bits 0:31 and 32:64 xmm0.
19487 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19488 CallResult.first, DAG.getIntPtrConstant(0, dl));
19489 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19490 CallResult.first, DAG.getIntPtrConstant(1, dl));
19491 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19492 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19495 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19496 SelectionDAG &DAG) {
19497 assert(Subtarget->hasAVX512() &&
19498 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19500 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19501 EVT VT = N->getValue().getValueType();
19502 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19505 // X86 scatter kills mask register, so its type should be added to
19506 // the list of return values
19507 if (N->getNumValues() == 1) {
19508 SDValue Index = N->getIndex();
19509 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19510 !Index.getValueType().is512BitVector())
19511 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19513 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19514 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19515 N->getOperand(3), Index };
19517 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19518 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19519 return SDValue(NewScatter.getNode(), 0);
19524 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19525 SelectionDAG &DAG) {
19526 assert(Subtarget->hasAVX512() &&
19527 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19529 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19530 EVT VT = Op.getValueType();
19531 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19534 SDValue Index = N->getIndex();
19535 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19536 !Index.getValueType().is512BitVector()) {
19537 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19538 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19539 N->getOperand(3), Index };
19540 DAG.UpdateNodeOperands(N, Ops);
19545 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19546 SelectionDAG &DAG) const {
19547 // TODO: Eventually, the lowering of these nodes should be informed by or
19548 // deferred to the GC strategy for the function in which they appear. For
19549 // now, however, they must be lowered to something. Since they are logically
19550 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19551 // require special handling for these nodes), lower them as literal NOOPs for
19553 SmallVector<SDValue, 2> Ops;
19555 Ops.push_back(Op.getOperand(0));
19556 if (Op->getGluedNode())
19557 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19560 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19561 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19566 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19567 SelectionDAG &DAG) const {
19568 // TODO: Eventually, the lowering of these nodes should be informed by or
19569 // deferred to the GC strategy for the function in which they appear. For
19570 // now, however, they must be lowered to something. Since they are logically
19571 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19572 // require special handling for these nodes), lower them as literal NOOPs for
19574 SmallVector<SDValue, 2> Ops;
19576 Ops.push_back(Op.getOperand(0));
19577 if (Op->getGluedNode())
19578 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19581 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19582 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19587 /// LowerOperation - Provide custom lowering hooks for some operations.
19589 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19590 switch (Op.getOpcode()) {
19591 default: llvm_unreachable("Should not custom lower this!");
19592 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19593 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19594 return LowerCMP_SWAP(Op, Subtarget, DAG);
19595 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19596 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19597 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19598 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19599 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19600 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19601 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19602 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19603 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19604 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19605 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19606 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19607 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19608 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19609 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19610 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19611 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19612 case ISD::SHL_PARTS:
19613 case ISD::SRA_PARTS:
19614 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19615 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19616 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19617 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19618 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19619 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19620 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19621 case ISD::SIGN_EXTEND_VECTOR_INREG:
19622 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19623 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19624 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19625 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19626 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19628 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19629 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19630 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19631 case ISD::SETCC: return LowerSETCC(Op, DAG);
19632 case ISD::SELECT: return LowerSELECT(Op, DAG);
19633 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19634 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19635 case ISD::VASTART: return LowerVASTART(Op, DAG);
19636 case ISD::VAARG: return LowerVAARG(Op, DAG);
19637 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19638 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19639 case ISD::INTRINSIC_VOID:
19640 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19641 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19642 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19643 case ISD::FRAME_TO_ARGS_OFFSET:
19644 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19645 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19646 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19647 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19648 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19649 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19650 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19651 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19652 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19653 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19655 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19656 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19657 case ISD::UMUL_LOHI:
19658 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19659 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19662 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19668 case ISD::UMULO: return LowerXALUO(Op, DAG);
19669 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19670 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19674 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19675 case ISD::ADD: return LowerADD(Op, DAG);
19676 case ISD::SUB: return LowerSUB(Op, DAG);
19680 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19681 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19682 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19683 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19684 case ISD::GC_TRANSITION_START:
19685 return LowerGC_TRANSITION_START(Op, DAG);
19686 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19690 /// ReplaceNodeResults - Replace a node with an illegal result type
19691 /// with a new node built out of custom code.
19692 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19693 SmallVectorImpl<SDValue>&Results,
19694 SelectionDAG &DAG) const {
19696 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19697 switch (N->getOpcode()) {
19699 llvm_unreachable("Do not know how to custom type legalize this operation!");
19700 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19701 case X86ISD::FMINC:
19703 case X86ISD::FMAXC:
19704 case X86ISD::FMAX: {
19705 EVT VT = N->getValueType(0);
19706 if (VT != MVT::v2f32)
19707 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19708 SDValue UNDEF = DAG.getUNDEF(VT);
19709 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19710 N->getOperand(0), UNDEF);
19711 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19712 N->getOperand(1), UNDEF);
19713 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19716 case ISD::SIGN_EXTEND_INREG:
19721 // We don't want to expand or promote these.
19728 case ISD::UDIVREM: {
19729 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19730 Results.push_back(V);
19733 case ISD::FP_TO_SINT:
19734 case ISD::FP_TO_UINT: {
19735 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19737 std::pair<SDValue,SDValue> Vals =
19738 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19739 SDValue FIST = Vals.first, StackSlot = Vals.second;
19740 if (FIST.getNode()) {
19741 EVT VT = N->getValueType(0);
19742 // Return a load from the stack slot.
19743 if (StackSlot.getNode())
19744 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19745 MachinePointerInfo(),
19746 false, false, false, 0));
19748 Results.push_back(FIST);
19752 case ISD::UINT_TO_FP: {
19753 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19754 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19755 N->getValueType(0) != MVT::v2f32)
19757 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19759 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19761 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19762 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19763 DAG.getBitcast(MVT::v2i64, VBias));
19764 Or = DAG.getBitcast(MVT::v2f64, Or);
19765 // TODO: Are there any fast-math-flags to propagate here?
19766 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19767 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19770 case ISD::FP_ROUND: {
19771 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19773 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19774 Results.push_back(V);
19777 case ISD::FP_EXTEND: {
19778 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19779 // No other ValueType for FP_EXTEND should reach this point.
19780 assert(N->getValueType(0) == MVT::v2f32 &&
19781 "Do not know how to legalize this Node");
19784 case ISD::INTRINSIC_W_CHAIN: {
19785 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19787 default : llvm_unreachable("Do not know how to custom type "
19788 "legalize this intrinsic operation!");
19789 case Intrinsic::x86_rdtsc:
19790 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19792 case Intrinsic::x86_rdtscp:
19793 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19795 case Intrinsic::x86_rdpmc:
19796 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19799 case ISD::READCYCLECOUNTER: {
19800 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19803 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19804 EVT T = N->getValueType(0);
19805 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19806 bool Regs64bit = T == MVT::i128;
19807 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19808 SDValue cpInL, cpInH;
19809 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19810 DAG.getConstant(0, dl, HalfT));
19811 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19812 DAG.getConstant(1, dl, HalfT));
19813 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19814 Regs64bit ? X86::RAX : X86::EAX,
19816 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19817 Regs64bit ? X86::RDX : X86::EDX,
19818 cpInH, cpInL.getValue(1));
19819 SDValue swapInL, swapInH;
19820 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19821 DAG.getConstant(0, dl, HalfT));
19822 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19823 DAG.getConstant(1, dl, HalfT));
19824 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19825 Regs64bit ? X86::RBX : X86::EBX,
19826 swapInL, cpInH.getValue(1));
19827 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19828 Regs64bit ? X86::RCX : X86::ECX,
19829 swapInH, swapInL.getValue(1));
19830 SDValue Ops[] = { swapInH.getValue(0),
19832 swapInH.getValue(1) };
19833 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19834 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19835 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19836 X86ISD::LCMPXCHG8_DAG;
19837 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19838 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19839 Regs64bit ? X86::RAX : X86::EAX,
19840 HalfT, Result.getValue(1));
19841 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19842 Regs64bit ? X86::RDX : X86::EDX,
19843 HalfT, cpOutL.getValue(2));
19844 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19846 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19847 MVT::i32, cpOutH.getValue(2));
19849 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19850 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19851 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19853 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19854 Results.push_back(Success);
19855 Results.push_back(EFLAGS.getValue(1));
19858 case ISD::ATOMIC_SWAP:
19859 case ISD::ATOMIC_LOAD_ADD:
19860 case ISD::ATOMIC_LOAD_SUB:
19861 case ISD::ATOMIC_LOAD_AND:
19862 case ISD::ATOMIC_LOAD_OR:
19863 case ISD::ATOMIC_LOAD_XOR:
19864 case ISD::ATOMIC_LOAD_NAND:
19865 case ISD::ATOMIC_LOAD_MIN:
19866 case ISD::ATOMIC_LOAD_MAX:
19867 case ISD::ATOMIC_LOAD_UMIN:
19868 case ISD::ATOMIC_LOAD_UMAX:
19869 case ISD::ATOMIC_LOAD: {
19870 // Delegate to generic TypeLegalization. Situations we can really handle
19871 // should have already been dealt with by AtomicExpandPass.cpp.
19874 case ISD::BITCAST: {
19875 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19876 EVT DstVT = N->getValueType(0);
19877 EVT SrcVT = N->getOperand(0)->getValueType(0);
19879 if (SrcVT != MVT::f64 ||
19880 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19883 unsigned NumElts = DstVT.getVectorNumElements();
19884 EVT SVT = DstVT.getVectorElementType();
19885 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19886 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19887 MVT::v2f64, N->getOperand(0));
19888 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19890 if (ExperimentalVectorWideningLegalization) {
19891 // If we are legalizing vectors by widening, we already have the desired
19892 // legal vector type, just return it.
19893 Results.push_back(ToVecInt);
19897 SmallVector<SDValue, 8> Elts;
19898 for (unsigned i = 0, e = NumElts; i != e; ++i)
19899 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19900 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19902 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19907 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19908 switch ((X86ISD::NodeType)Opcode) {
19909 case X86ISD::FIRST_NUMBER: break;
19910 case X86ISD::BSF: return "X86ISD::BSF";
19911 case X86ISD::BSR: return "X86ISD::BSR";
19912 case X86ISD::SHLD: return "X86ISD::SHLD";
19913 case X86ISD::SHRD: return "X86ISD::SHRD";
19914 case X86ISD::FAND: return "X86ISD::FAND";
19915 case X86ISD::FANDN: return "X86ISD::FANDN";
19916 case X86ISD::FOR: return "X86ISD::FOR";
19917 case X86ISD::FXOR: return "X86ISD::FXOR";
19918 case X86ISD::FILD: return "X86ISD::FILD";
19919 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19920 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19921 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19922 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19923 case X86ISD::FLD: return "X86ISD::FLD";
19924 case X86ISD::FST: return "X86ISD::FST";
19925 case X86ISD::CALL: return "X86ISD::CALL";
19926 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19927 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19928 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19929 case X86ISD::BT: return "X86ISD::BT";
19930 case X86ISD::CMP: return "X86ISD::CMP";
19931 case X86ISD::COMI: return "X86ISD::COMI";
19932 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19933 case X86ISD::CMPM: return "X86ISD::CMPM";
19934 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19935 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19936 case X86ISD::SETCC: return "X86ISD::SETCC";
19937 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19938 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19939 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19940 case X86ISD::CMOV: return "X86ISD::CMOV";
19941 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19942 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19943 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19944 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19945 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19946 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19947 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19948 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19949 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19950 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19951 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19952 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19953 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19954 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19955 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19956 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19957 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19958 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19959 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19960 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19961 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19962 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19963 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19964 case X86ISD::HADD: return "X86ISD::HADD";
19965 case X86ISD::HSUB: return "X86ISD::HSUB";
19966 case X86ISD::FHADD: return "X86ISD::FHADD";
19967 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19968 case X86ISD::ABS: return "X86ISD::ABS";
19969 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19970 case X86ISD::FMAX: return "X86ISD::FMAX";
19971 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19972 case X86ISD::FMIN: return "X86ISD::FMIN";
19973 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19974 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19975 case X86ISD::FMINC: return "X86ISD::FMINC";
19976 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19977 case X86ISD::FRCP: return "X86ISD::FRCP";
19978 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19979 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19980 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19981 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19982 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19983 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19984 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19985 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19986 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19987 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19988 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19989 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19990 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19991 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19992 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19993 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19994 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19995 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19996 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19997 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19998 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19999 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20000 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20001 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20002 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20003 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20004 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20005 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20006 case X86ISD::VSHL: return "X86ISD::VSHL";
20007 case X86ISD::VSRL: return "X86ISD::VSRL";
20008 case X86ISD::VSRA: return "X86ISD::VSRA";
20009 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20010 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20011 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20012 case X86ISD::CMPP: return "X86ISD::CMPP";
20013 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20014 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20015 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20016 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20017 case X86ISD::ADD: return "X86ISD::ADD";
20018 case X86ISD::SUB: return "X86ISD::SUB";
20019 case X86ISD::ADC: return "X86ISD::ADC";
20020 case X86ISD::SBB: return "X86ISD::SBB";
20021 case X86ISD::SMUL: return "X86ISD::SMUL";
20022 case X86ISD::UMUL: return "X86ISD::UMUL";
20023 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20024 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20025 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20026 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20027 case X86ISD::INC: return "X86ISD::INC";
20028 case X86ISD::DEC: return "X86ISD::DEC";
20029 case X86ISD::OR: return "X86ISD::OR";
20030 case X86ISD::XOR: return "X86ISD::XOR";
20031 case X86ISD::AND: return "X86ISD::AND";
20032 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20033 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20034 case X86ISD::PTEST: return "X86ISD::PTEST";
20035 case X86ISD::TESTP: return "X86ISD::TESTP";
20036 case X86ISD::TESTM: return "X86ISD::TESTM";
20037 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20038 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20039 case X86ISD::KTEST: return "X86ISD::KTEST";
20040 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20041 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20042 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20043 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20044 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20045 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20046 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20047 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20048 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20049 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20050 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20051 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20052 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20053 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20054 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20055 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20056 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20057 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20058 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20059 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20060 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20061 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20062 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20063 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20064 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20065 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20066 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20067 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20068 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20069 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20070 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20071 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20072 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20073 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20074 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20075 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20076 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20077 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20078 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20079 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20080 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20081 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20082 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20083 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20084 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20085 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20086 case X86ISD::SAHF: return "X86ISD::SAHF";
20087 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20088 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20089 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20090 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20091 case X86ISD::VPROT: return "X86ISD::VPROT";
20092 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20093 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20094 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20095 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20096 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20097 case X86ISD::FMADD: return "X86ISD::FMADD";
20098 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20099 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20100 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20101 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20102 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20103 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20104 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20105 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20106 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20107 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20108 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20109 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20110 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20111 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20112 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20113 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20114 case X86ISD::XTEST: return "X86ISD::XTEST";
20115 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20116 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20117 case X86ISD::SELECT: return "X86ISD::SELECT";
20118 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20119 case X86ISD::RCP28: return "X86ISD::RCP28";
20120 case X86ISD::EXP2: return "X86ISD::EXP2";
20121 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20122 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20123 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20124 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20125 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20126 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20127 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20128 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20129 case X86ISD::ADDS: return "X86ISD::ADDS";
20130 case X86ISD::SUBS: return "X86ISD::SUBS";
20131 case X86ISD::AVG: return "X86ISD::AVG";
20132 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20133 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20134 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20135 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20136 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20137 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20142 // isLegalAddressingMode - Return true if the addressing mode represented
20143 // by AM is legal for this target, for a load/store of the specified type.
20144 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20145 const AddrMode &AM, Type *Ty,
20146 unsigned AS) const {
20147 // X86 supports extremely general addressing modes.
20148 CodeModel::Model M = getTargetMachine().getCodeModel();
20149 Reloc::Model R = getTargetMachine().getRelocationModel();
20151 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20152 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20157 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20159 // If a reference to this global requires an extra load, we can't fold it.
20160 if (isGlobalStubReference(GVFlags))
20163 // If BaseGV requires a register for the PIC base, we cannot also have a
20164 // BaseReg specified.
20165 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20168 // If lower 4G is not available, then we must use rip-relative addressing.
20169 if ((M != CodeModel::Small || R != Reloc::Static) &&
20170 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20174 switch (AM.Scale) {
20180 // These scales always work.
20185 // These scales are formed with basereg+scalereg. Only accept if there is
20190 default: // Other stuff never works.
20197 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20198 unsigned Bits = Ty->getScalarSizeInBits();
20200 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20201 // particularly cheaper than those without.
20205 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20206 // variable shifts just as cheap as scalar ones.
20207 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20210 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20211 // fully general vector.
20215 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20216 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20218 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20219 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20220 return NumBits1 > NumBits2;
20223 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20224 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20227 if (!isTypeLegal(EVT::getEVT(Ty1)))
20230 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20232 // Assuming the caller doesn't have a zeroext or signext return parameter,
20233 // truncation all the way down to i1 is valid.
20237 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20238 return isInt<32>(Imm);
20241 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20242 // Can also use sub to handle negated immediates.
20243 return isInt<32>(Imm);
20246 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20247 if (!VT1.isInteger() || !VT2.isInteger())
20249 unsigned NumBits1 = VT1.getSizeInBits();
20250 unsigned NumBits2 = VT2.getSizeInBits();
20251 return NumBits1 > NumBits2;
20254 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20255 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20256 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20259 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20260 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20261 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20264 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20265 EVT VT1 = Val.getValueType();
20266 if (isZExtFree(VT1, VT2))
20269 if (Val.getOpcode() != ISD::LOAD)
20272 if (!VT1.isSimple() || !VT1.isInteger() ||
20273 !VT2.isSimple() || !VT2.isInteger())
20276 switch (VT1.getSimpleVT().SimpleTy) {
20281 // X86 has 8, 16, and 32-bit zero-extending loads.
20288 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20291 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20292 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20295 VT = VT.getScalarType();
20297 if (!VT.isSimple())
20300 switch (VT.getSimpleVT().SimpleTy) {
20311 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20312 // i16 instructions are longer (0x66 prefix) and potentially slower.
20313 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20316 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20317 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20318 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20319 /// are assumed to be legal.
20321 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20323 if (!VT.isSimple())
20326 // Not for i1 vectors
20327 if (VT.getScalarType() == MVT::i1)
20330 // Very little shuffling can be done for 64-bit vectors right now.
20331 if (VT.getSizeInBits() == 64)
20334 // We only care that the types being shuffled are legal. The lowering can
20335 // handle any possible shuffle mask that results.
20336 return isTypeLegal(VT.getSimpleVT());
20340 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20342 // Just delegate to the generic legality, clear masks aren't special.
20343 return isShuffleMaskLegal(Mask, VT);
20346 //===----------------------------------------------------------------------===//
20347 // X86 Scheduler Hooks
20348 //===----------------------------------------------------------------------===//
20350 /// Utility function to emit xbegin specifying the start of an RTM region.
20351 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20352 const TargetInstrInfo *TII) {
20353 DebugLoc DL = MI->getDebugLoc();
20355 const BasicBlock *BB = MBB->getBasicBlock();
20356 MachineFunction::iterator I = ++MBB->getIterator();
20358 // For the v = xbegin(), we generate
20369 MachineBasicBlock *thisMBB = MBB;
20370 MachineFunction *MF = MBB->getParent();
20371 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20372 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20373 MF->insert(I, mainMBB);
20374 MF->insert(I, sinkMBB);
20376 // Transfer the remainder of BB and its successor edges to sinkMBB.
20377 sinkMBB->splice(sinkMBB->begin(), MBB,
20378 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20379 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20383 // # fallthrough to mainMBB
20384 // # abortion to sinkMBB
20385 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20386 thisMBB->addSuccessor(mainMBB);
20387 thisMBB->addSuccessor(sinkMBB);
20391 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20392 mainMBB->addSuccessor(sinkMBB);
20395 // EAX is live into the sinkMBB
20396 sinkMBB->addLiveIn(X86::EAX);
20397 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20398 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20401 MI->eraseFromParent();
20405 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20406 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20407 // in the .td file.
20408 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20409 const TargetInstrInfo *TII) {
20411 switch (MI->getOpcode()) {
20412 default: llvm_unreachable("illegal opcode!");
20413 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20414 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20415 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20416 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20417 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20418 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20419 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20420 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20423 DebugLoc dl = MI->getDebugLoc();
20424 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20426 unsigned NumArgs = MI->getNumOperands();
20427 for (unsigned i = 1; i < NumArgs; ++i) {
20428 MachineOperand &Op = MI->getOperand(i);
20429 if (!(Op.isReg() && Op.isImplicit()))
20430 MIB.addOperand(Op);
20432 if (MI->hasOneMemOperand())
20433 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20435 BuildMI(*BB, MI, dl,
20436 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20437 .addReg(X86::XMM0);
20439 MI->eraseFromParent();
20443 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20444 // defs in an instruction pattern
20445 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20446 const TargetInstrInfo *TII) {
20448 switch (MI->getOpcode()) {
20449 default: llvm_unreachable("illegal opcode!");
20450 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20451 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20452 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20453 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20454 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20455 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20456 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20457 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20460 DebugLoc dl = MI->getDebugLoc();
20461 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20463 unsigned NumArgs = MI->getNumOperands(); // remove the results
20464 for (unsigned i = 1; i < NumArgs; ++i) {
20465 MachineOperand &Op = MI->getOperand(i);
20466 if (!(Op.isReg() && Op.isImplicit()))
20467 MIB.addOperand(Op);
20469 if (MI->hasOneMemOperand())
20470 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20472 BuildMI(*BB, MI, dl,
20473 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20476 MI->eraseFromParent();
20480 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20481 const X86Subtarget *Subtarget) {
20482 DebugLoc dl = MI->getDebugLoc();
20483 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20484 // Address into RAX/EAX, other two args into ECX, EDX.
20485 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20486 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20487 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20488 for (int i = 0; i < X86::AddrNumOperands; ++i)
20489 MIB.addOperand(MI->getOperand(i));
20491 unsigned ValOps = X86::AddrNumOperands;
20492 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20493 .addReg(MI->getOperand(ValOps).getReg());
20494 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20495 .addReg(MI->getOperand(ValOps+1).getReg());
20497 // The instruction doesn't actually take any operands though.
20498 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20500 MI->eraseFromParent(); // The pseudo is gone now.
20504 MachineBasicBlock *
20505 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20506 MachineBasicBlock *MBB) const {
20507 // Emit va_arg instruction on X86-64.
20509 // Operands to this pseudo-instruction:
20510 // 0 ) Output : destination address (reg)
20511 // 1-5) Input : va_list address (addr, i64mem)
20512 // 6 ) ArgSize : Size (in bytes) of vararg type
20513 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20514 // 8 ) Align : Alignment of type
20515 // 9 ) EFLAGS (implicit-def)
20517 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20518 static_assert(X86::AddrNumOperands == 5,
20519 "VAARG_64 assumes 5 address operands");
20521 unsigned DestReg = MI->getOperand(0).getReg();
20522 MachineOperand &Base = MI->getOperand(1);
20523 MachineOperand &Scale = MI->getOperand(2);
20524 MachineOperand &Index = MI->getOperand(3);
20525 MachineOperand &Disp = MI->getOperand(4);
20526 MachineOperand &Segment = MI->getOperand(5);
20527 unsigned ArgSize = MI->getOperand(6).getImm();
20528 unsigned ArgMode = MI->getOperand(7).getImm();
20529 unsigned Align = MI->getOperand(8).getImm();
20531 // Memory Reference
20532 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20533 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20534 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20536 // Machine Information
20537 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20538 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20539 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20540 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20541 DebugLoc DL = MI->getDebugLoc();
20543 // struct va_list {
20546 // i64 overflow_area (address)
20547 // i64 reg_save_area (address)
20549 // sizeof(va_list) = 24
20550 // alignment(va_list) = 8
20552 unsigned TotalNumIntRegs = 6;
20553 unsigned TotalNumXMMRegs = 8;
20554 bool UseGPOffset = (ArgMode == 1);
20555 bool UseFPOffset = (ArgMode == 2);
20556 unsigned MaxOffset = TotalNumIntRegs * 8 +
20557 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20559 /* Align ArgSize to a multiple of 8 */
20560 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20561 bool NeedsAlign = (Align > 8);
20563 MachineBasicBlock *thisMBB = MBB;
20564 MachineBasicBlock *overflowMBB;
20565 MachineBasicBlock *offsetMBB;
20566 MachineBasicBlock *endMBB;
20568 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20569 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20570 unsigned OffsetReg = 0;
20572 if (!UseGPOffset && !UseFPOffset) {
20573 // If we only pull from the overflow region, we don't create a branch.
20574 // We don't need to alter control flow.
20575 OffsetDestReg = 0; // unused
20576 OverflowDestReg = DestReg;
20578 offsetMBB = nullptr;
20579 overflowMBB = thisMBB;
20582 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20583 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20584 // If not, pull from overflow_area. (branch to overflowMBB)
20589 // offsetMBB overflowMBB
20594 // Registers for the PHI in endMBB
20595 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20596 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20598 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20599 MachineFunction *MF = MBB->getParent();
20600 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20601 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20602 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20604 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20606 // Insert the new basic blocks
20607 MF->insert(MBBIter, offsetMBB);
20608 MF->insert(MBBIter, overflowMBB);
20609 MF->insert(MBBIter, endMBB);
20611 // Transfer the remainder of MBB and its successor edges to endMBB.
20612 endMBB->splice(endMBB->begin(), thisMBB,
20613 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20614 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20616 // Make offsetMBB and overflowMBB successors of thisMBB
20617 thisMBB->addSuccessor(offsetMBB);
20618 thisMBB->addSuccessor(overflowMBB);
20620 // endMBB is a successor of both offsetMBB and overflowMBB
20621 offsetMBB->addSuccessor(endMBB);
20622 overflowMBB->addSuccessor(endMBB);
20624 // Load the offset value into a register
20625 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20626 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20630 .addDisp(Disp, UseFPOffset ? 4 : 0)
20631 .addOperand(Segment)
20632 .setMemRefs(MMOBegin, MMOEnd);
20634 // Check if there is enough room left to pull this argument.
20635 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20637 .addImm(MaxOffset + 8 - ArgSizeA8);
20639 // Branch to "overflowMBB" if offset >= max
20640 // Fall through to "offsetMBB" otherwise
20641 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20642 .addMBB(overflowMBB);
20645 // In offsetMBB, emit code to use the reg_save_area.
20647 assert(OffsetReg != 0);
20649 // Read the reg_save_area address.
20650 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20651 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20656 .addOperand(Segment)
20657 .setMemRefs(MMOBegin, MMOEnd);
20659 // Zero-extend the offset
20660 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20661 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20664 .addImm(X86::sub_32bit);
20666 // Add the offset to the reg_save_area to get the final address.
20667 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20668 .addReg(OffsetReg64)
20669 .addReg(RegSaveReg);
20671 // Compute the offset for the next argument
20672 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20673 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20675 .addImm(UseFPOffset ? 16 : 8);
20677 // Store it back into the va_list.
20678 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20682 .addDisp(Disp, UseFPOffset ? 4 : 0)
20683 .addOperand(Segment)
20684 .addReg(NextOffsetReg)
20685 .setMemRefs(MMOBegin, MMOEnd);
20688 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20693 // Emit code to use overflow area
20696 // Load the overflow_area address into a register.
20697 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20698 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20703 .addOperand(Segment)
20704 .setMemRefs(MMOBegin, MMOEnd);
20706 // If we need to align it, do so. Otherwise, just copy the address
20707 // to OverflowDestReg.
20709 // Align the overflow address
20710 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20711 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20713 // aligned_addr = (addr + (align-1)) & ~(align-1)
20714 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20715 .addReg(OverflowAddrReg)
20718 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20720 .addImm(~(uint64_t)(Align-1));
20722 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20723 .addReg(OverflowAddrReg);
20726 // Compute the next overflow address after this argument.
20727 // (the overflow address should be kept 8-byte aligned)
20728 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20729 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20730 .addReg(OverflowDestReg)
20731 .addImm(ArgSizeA8);
20733 // Store the new overflow address.
20734 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20739 .addOperand(Segment)
20740 .addReg(NextAddrReg)
20741 .setMemRefs(MMOBegin, MMOEnd);
20743 // If we branched, emit the PHI to the front of endMBB.
20745 BuildMI(*endMBB, endMBB->begin(), DL,
20746 TII->get(X86::PHI), DestReg)
20747 .addReg(OffsetDestReg).addMBB(offsetMBB)
20748 .addReg(OverflowDestReg).addMBB(overflowMBB);
20751 // Erase the pseudo instruction
20752 MI->eraseFromParent();
20757 MachineBasicBlock *
20758 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20760 MachineBasicBlock *MBB) const {
20761 // Emit code to save XMM registers to the stack. The ABI says that the
20762 // number of registers to save is given in %al, so it's theoretically
20763 // possible to do an indirect jump trick to avoid saving all of them,
20764 // however this code takes a simpler approach and just executes all
20765 // of the stores if %al is non-zero. It's less code, and it's probably
20766 // easier on the hardware branch predictor, and stores aren't all that
20767 // expensive anyway.
20769 // Create the new basic blocks. One block contains all the XMM stores,
20770 // and one block is the final destination regardless of whether any
20771 // stores were performed.
20772 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20773 MachineFunction *F = MBB->getParent();
20774 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20775 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20776 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20777 F->insert(MBBIter, XMMSaveMBB);
20778 F->insert(MBBIter, EndMBB);
20780 // Transfer the remainder of MBB and its successor edges to EndMBB.
20781 EndMBB->splice(EndMBB->begin(), MBB,
20782 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20783 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20785 // The original block will now fall through to the XMM save block.
20786 MBB->addSuccessor(XMMSaveMBB);
20787 // The XMMSaveMBB will fall through to the end block.
20788 XMMSaveMBB->addSuccessor(EndMBB);
20790 // Now add the instructions.
20791 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20792 DebugLoc DL = MI->getDebugLoc();
20794 unsigned CountReg = MI->getOperand(0).getReg();
20795 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20796 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20798 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20799 // If %al is 0, branch around the XMM save block.
20800 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20801 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20802 MBB->addSuccessor(EndMBB);
20805 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20806 // that was just emitted, but clearly shouldn't be "saved".
20807 assert((MI->getNumOperands() <= 3 ||
20808 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20809 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20810 && "Expected last argument to be EFLAGS");
20811 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20812 // In the XMM save block, save all the XMM argument registers.
20813 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20814 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20815 MachineMemOperand *MMO = F->getMachineMemOperand(
20816 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20817 MachineMemOperand::MOStore,
20818 /*Size=*/16, /*Align=*/16);
20819 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20820 .addFrameIndex(RegSaveFrameIndex)
20821 .addImm(/*Scale=*/1)
20822 .addReg(/*IndexReg=*/0)
20823 .addImm(/*Disp=*/Offset)
20824 .addReg(/*Segment=*/0)
20825 .addReg(MI->getOperand(i).getReg())
20826 .addMemOperand(MMO);
20829 MI->eraseFromParent(); // The pseudo instruction is gone now.
20834 // The EFLAGS operand of SelectItr might be missing a kill marker
20835 // because there were multiple uses of EFLAGS, and ISel didn't know
20836 // which to mark. Figure out whether SelectItr should have had a
20837 // kill marker, and set it if it should. Returns the correct kill
20839 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20840 MachineBasicBlock* BB,
20841 const TargetRegisterInfo* TRI) {
20842 // Scan forward through BB for a use/def of EFLAGS.
20843 MachineBasicBlock::iterator miI(std::next(SelectItr));
20844 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20845 const MachineInstr& mi = *miI;
20846 if (mi.readsRegister(X86::EFLAGS))
20848 if (mi.definesRegister(X86::EFLAGS))
20849 break; // Should have kill-flag - update below.
20852 // If we hit the end of the block, check whether EFLAGS is live into a
20854 if (miI == BB->end()) {
20855 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20856 sEnd = BB->succ_end();
20857 sItr != sEnd; ++sItr) {
20858 MachineBasicBlock* succ = *sItr;
20859 if (succ->isLiveIn(X86::EFLAGS))
20864 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20865 // out. SelectMI should have a kill flag on EFLAGS.
20866 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20870 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20871 // together with other CMOV pseudo-opcodes into a single basic-block with
20872 // conditional jump around it.
20873 static bool isCMOVPseudo(MachineInstr *MI) {
20874 switch (MI->getOpcode()) {
20875 case X86::CMOV_FR32:
20876 case X86::CMOV_FR64:
20877 case X86::CMOV_GR8:
20878 case X86::CMOV_GR16:
20879 case X86::CMOV_GR32:
20880 case X86::CMOV_RFP32:
20881 case X86::CMOV_RFP64:
20882 case X86::CMOV_RFP80:
20883 case X86::CMOV_V2F64:
20884 case X86::CMOV_V2I64:
20885 case X86::CMOV_V4F32:
20886 case X86::CMOV_V4F64:
20887 case X86::CMOV_V4I64:
20888 case X86::CMOV_V16F32:
20889 case X86::CMOV_V8F32:
20890 case X86::CMOV_V8F64:
20891 case X86::CMOV_V8I64:
20892 case X86::CMOV_V8I1:
20893 case X86::CMOV_V16I1:
20894 case X86::CMOV_V32I1:
20895 case X86::CMOV_V64I1:
20903 MachineBasicBlock *
20904 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20905 MachineBasicBlock *BB) const {
20906 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20907 DebugLoc DL = MI->getDebugLoc();
20909 // To "insert" a SELECT_CC instruction, we actually have to insert the
20910 // diamond control-flow pattern. The incoming instruction knows the
20911 // destination vreg to set, the condition code register to branch on, the
20912 // true/false values to select between, and a branch opcode to use.
20913 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20914 MachineFunction::iterator It = ++BB->getIterator();
20919 // cmpTY ccX, r1, r2
20921 // fallthrough --> copy0MBB
20922 MachineBasicBlock *thisMBB = BB;
20923 MachineFunction *F = BB->getParent();
20925 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20926 // as described above, by inserting a BB, and then making a PHI at the join
20927 // point to select the true and false operands of the CMOV in the PHI.
20929 // The code also handles two different cases of multiple CMOV opcodes
20933 // In this case, there are multiple CMOVs in a row, all which are based on
20934 // the same condition setting (or the exact opposite condition setting).
20935 // In this case we can lower all the CMOVs using a single inserted BB, and
20936 // then make a number of PHIs at the join point to model the CMOVs. The only
20937 // trickiness here, is that in a case like:
20939 // t2 = CMOV cond1 t1, f1
20940 // t3 = CMOV cond1 t2, f2
20942 // when rewriting this into PHIs, we have to perform some renaming on the
20943 // temps since you cannot have a PHI operand refer to a PHI result earlier
20944 // in the same block. The "simple" but wrong lowering would be:
20946 // t2 = PHI t1(BB1), f1(BB2)
20947 // t3 = PHI t2(BB1), f2(BB2)
20949 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20950 // renaming is to note that on the path through BB1, t2 is really just a
20951 // copy of t1, and do that renaming, properly generating:
20953 // t2 = PHI t1(BB1), f1(BB2)
20954 // t3 = PHI t1(BB1), f2(BB2)
20956 // Case 2, we lower cascaded CMOVs such as
20958 // (CMOV (CMOV F, T, cc1), T, cc2)
20960 // to two successives branches. For that, we look for another CMOV as the
20961 // following instruction.
20963 // Without this, we would add a PHI between the two jumps, which ends up
20964 // creating a few copies all around. For instance, for
20966 // (sitofp (zext (fcmp une)))
20968 // we would generate:
20970 // ucomiss %xmm1, %xmm0
20971 // movss <1.0f>, %xmm0
20972 // movaps %xmm0, %xmm1
20974 // xorps %xmm1, %xmm1
20977 // movaps %xmm1, %xmm0
20981 // because this custom-inserter would have generated:
20993 // A: X = ...; Y = ...
20995 // C: Z = PHI [X, A], [Y, B]
20997 // E: PHI [X, C], [Z, D]
20999 // If we lower both CMOVs in a single step, we can instead generate:
21011 // A: X = ...; Y = ...
21013 // E: PHI [X, A], [X, C], [Y, D]
21015 // Which, in our sitofp/fcmp example, gives us something like:
21017 // ucomiss %xmm1, %xmm0
21018 // movss <1.0f>, %xmm0
21021 // xorps %xmm0, %xmm0
21025 MachineInstr *CascadedCMOV = nullptr;
21026 MachineInstr *LastCMOV = MI;
21027 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21028 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21029 MachineBasicBlock::iterator NextMIIt =
21030 std::next(MachineBasicBlock::iterator(MI));
21032 // Check for case 1, where there are multiple CMOVs with the same condition
21033 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21034 // number of jumps the most.
21036 if (isCMOVPseudo(MI)) {
21037 // See if we have a string of CMOVS with the same condition.
21038 while (NextMIIt != BB->end() &&
21039 isCMOVPseudo(NextMIIt) &&
21040 (NextMIIt->getOperand(3).getImm() == CC ||
21041 NextMIIt->getOperand(3).getImm() == OppCC)) {
21042 LastCMOV = &*NextMIIt;
21047 // This checks for case 2, but only do this if we didn't already find
21048 // case 1, as indicated by LastCMOV == MI.
21049 if (LastCMOV == MI &&
21050 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21051 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21052 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21053 CascadedCMOV = &*NextMIIt;
21056 MachineBasicBlock *jcc1MBB = nullptr;
21058 // If we have a cascaded CMOV, we lower it to two successive branches to
21059 // the same block. EFLAGS is used by both, so mark it as live in the second.
21060 if (CascadedCMOV) {
21061 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21062 F->insert(It, jcc1MBB);
21063 jcc1MBB->addLiveIn(X86::EFLAGS);
21066 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21067 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21068 F->insert(It, copy0MBB);
21069 F->insert(It, sinkMBB);
21071 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21072 // live into the sink and copy blocks.
21073 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21075 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21076 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21077 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21078 copy0MBB->addLiveIn(X86::EFLAGS);
21079 sinkMBB->addLiveIn(X86::EFLAGS);
21082 // Transfer the remainder of BB and its successor edges to sinkMBB.
21083 sinkMBB->splice(sinkMBB->begin(), BB,
21084 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21085 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21087 // Add the true and fallthrough blocks as its successors.
21088 if (CascadedCMOV) {
21089 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21090 BB->addSuccessor(jcc1MBB);
21092 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21093 // jump to the sinkMBB.
21094 jcc1MBB->addSuccessor(copy0MBB);
21095 jcc1MBB->addSuccessor(sinkMBB);
21097 BB->addSuccessor(copy0MBB);
21100 // The true block target of the first (or only) branch is always sinkMBB.
21101 BB->addSuccessor(sinkMBB);
21103 // Create the conditional branch instruction.
21104 unsigned Opc = X86::GetCondBranchFromCond(CC);
21105 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21107 if (CascadedCMOV) {
21108 unsigned Opc2 = X86::GetCondBranchFromCond(
21109 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21110 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21114 // %FalseValue = ...
21115 // # fallthrough to sinkMBB
21116 copy0MBB->addSuccessor(sinkMBB);
21119 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21121 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21122 MachineBasicBlock::iterator MIItEnd =
21123 std::next(MachineBasicBlock::iterator(LastCMOV));
21124 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21125 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21126 MachineInstrBuilder MIB;
21128 // As we are creating the PHIs, we have to be careful if there is more than
21129 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21130 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21131 // That also means that PHI construction must work forward from earlier to
21132 // later, and that the code must maintain a mapping from earlier PHI's
21133 // destination registers, and the registers that went into the PHI.
21135 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21136 unsigned DestReg = MIIt->getOperand(0).getReg();
21137 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21138 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21140 // If this CMOV we are generating is the opposite condition from
21141 // the jump we generated, then we have to swap the operands for the
21142 // PHI that is going to be generated.
21143 if (MIIt->getOperand(3).getImm() == OppCC)
21144 std::swap(Op1Reg, Op2Reg);
21146 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21147 Op1Reg = RegRewriteTable[Op1Reg].first;
21149 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21150 Op2Reg = RegRewriteTable[Op2Reg].second;
21152 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21153 TII->get(X86::PHI), DestReg)
21154 .addReg(Op1Reg).addMBB(copy0MBB)
21155 .addReg(Op2Reg).addMBB(thisMBB);
21157 // Add this PHI to the rewrite table.
21158 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21161 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21162 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21163 if (CascadedCMOV) {
21164 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21165 // Copy the PHI result to the register defined by the second CMOV.
21166 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21167 DL, TII->get(TargetOpcode::COPY),
21168 CascadedCMOV->getOperand(0).getReg())
21169 .addReg(MI->getOperand(0).getReg());
21170 CascadedCMOV->eraseFromParent();
21173 // Now remove the CMOV(s).
21174 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21175 (MIIt++)->eraseFromParent();
21180 MachineBasicBlock *
21181 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21182 MachineBasicBlock *BB) const {
21183 // Combine the following atomic floating-point modification pattern:
21184 // a.store(reg OP a.load(acquire), release)
21185 // Transform them into:
21186 // OPss (%gpr), %xmm
21187 // movss %xmm, (%gpr)
21188 // Or sd equivalent for 64-bit operations.
21190 switch (MI->getOpcode()) {
21191 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21192 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21193 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21195 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21196 DebugLoc DL = MI->getDebugLoc();
21197 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21198 MachineOperand MSrc = MI->getOperand(0);
21199 unsigned VSrc = MI->getOperand(5).getReg();
21200 const MachineOperand &Disp = MI->getOperand(3);
21201 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21202 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21203 if (hasDisp && MSrc.isReg())
21204 MSrc.setIsKill(false);
21205 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21206 .addOperand(/*Base=*/MSrc)
21207 .addImm(/*Scale=*/1)
21208 .addReg(/*Index=*/0)
21209 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21211 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21212 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21214 .addOperand(/*Base=*/MSrc)
21215 .addImm(/*Scale=*/1)
21216 .addReg(/*Index=*/0)
21217 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21218 .addReg(/*Segment=*/0);
21219 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21220 MI->eraseFromParent(); // The pseudo instruction is gone now.
21224 MachineBasicBlock *
21225 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21226 MachineBasicBlock *BB) const {
21227 MachineFunction *MF = BB->getParent();
21228 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21229 DebugLoc DL = MI->getDebugLoc();
21230 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21232 assert(MF->shouldSplitStack());
21234 const bool Is64Bit = Subtarget->is64Bit();
21235 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21237 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21238 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21241 // ... [Till the alloca]
21242 // If stacklet is not large enough, jump to mallocMBB
21245 // Allocate by subtracting from RSP
21246 // Jump to continueMBB
21249 // Allocate by call to runtime
21253 // [rest of original BB]
21256 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21257 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21258 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21260 MachineRegisterInfo &MRI = MF->getRegInfo();
21261 const TargetRegisterClass *AddrRegClass =
21262 getRegClassFor(getPointerTy(MF->getDataLayout()));
21264 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21265 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21266 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21267 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21268 sizeVReg = MI->getOperand(1).getReg(),
21269 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21271 MachineFunction::iterator MBBIter = ++BB->getIterator();
21273 MF->insert(MBBIter, bumpMBB);
21274 MF->insert(MBBIter, mallocMBB);
21275 MF->insert(MBBIter, continueMBB);
21277 continueMBB->splice(continueMBB->begin(), BB,
21278 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21279 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21281 // Add code to the main basic block to check if the stack limit has been hit,
21282 // and if so, jump to mallocMBB otherwise to bumpMBB.
21283 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21284 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21285 .addReg(tmpSPVReg).addReg(sizeVReg);
21286 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21287 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21288 .addReg(SPLimitVReg);
21289 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21291 // bumpMBB simply decreases the stack pointer, since we know the current
21292 // stacklet has enough space.
21293 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21294 .addReg(SPLimitVReg);
21295 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21296 .addReg(SPLimitVReg);
21297 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21299 // Calls into a routine in libgcc to allocate more space from the heap.
21300 const uint32_t *RegMask =
21301 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21303 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21305 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21306 .addExternalSymbol("__morestack_allocate_stack_space")
21307 .addRegMask(RegMask)
21308 .addReg(X86::RDI, RegState::Implicit)
21309 .addReg(X86::RAX, RegState::ImplicitDefine);
21310 } else if (Is64Bit) {
21311 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21313 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21314 .addExternalSymbol("__morestack_allocate_stack_space")
21315 .addRegMask(RegMask)
21316 .addReg(X86::EDI, RegState::Implicit)
21317 .addReg(X86::EAX, RegState::ImplicitDefine);
21319 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21321 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21322 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21323 .addExternalSymbol("__morestack_allocate_stack_space")
21324 .addRegMask(RegMask)
21325 .addReg(X86::EAX, RegState::ImplicitDefine);
21329 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21332 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21333 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21334 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21336 // Set up the CFG correctly.
21337 BB->addSuccessor(bumpMBB);
21338 BB->addSuccessor(mallocMBB);
21339 mallocMBB->addSuccessor(continueMBB);
21340 bumpMBB->addSuccessor(continueMBB);
21342 // Take care of the PHI nodes.
21343 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21344 MI->getOperand(0).getReg())
21345 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21346 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21348 // Delete the original pseudo instruction.
21349 MI->eraseFromParent();
21352 return continueMBB;
21355 MachineBasicBlock *
21356 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21357 MachineBasicBlock *BB) const {
21358 DebugLoc DL = MI->getDebugLoc();
21360 assert(!Subtarget->isTargetMachO());
21362 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
21365 MI->eraseFromParent(); // The pseudo instruction is gone now.
21369 MachineBasicBlock *
21370 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21371 MachineBasicBlock *BB) const {
21372 // This is pretty easy. We're taking the value that we received from
21373 // our load from the relocation, sticking it in either RDI (x86-64)
21374 // or EAX and doing an indirect call. The return value will then
21375 // be in the normal return register.
21376 MachineFunction *F = BB->getParent();
21377 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21378 DebugLoc DL = MI->getDebugLoc();
21380 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21381 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21383 // Get a register mask for the lowered call.
21384 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21385 // proper register mask.
21386 const uint32_t *RegMask =
21387 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21388 if (Subtarget->is64Bit()) {
21389 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21390 TII->get(X86::MOV64rm), X86::RDI)
21392 .addImm(0).addReg(0)
21393 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21394 MI->getOperand(3).getTargetFlags())
21396 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21397 addDirectMem(MIB, X86::RDI);
21398 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21399 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21400 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21401 TII->get(X86::MOV32rm), X86::EAX)
21403 .addImm(0).addReg(0)
21404 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21405 MI->getOperand(3).getTargetFlags())
21407 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21408 addDirectMem(MIB, X86::EAX);
21409 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21411 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21412 TII->get(X86::MOV32rm), X86::EAX)
21413 .addReg(TII->getGlobalBaseReg(F))
21414 .addImm(0).addReg(0)
21415 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21416 MI->getOperand(3).getTargetFlags())
21418 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21419 addDirectMem(MIB, X86::EAX);
21420 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21423 MI->eraseFromParent(); // The pseudo instruction is gone now.
21427 MachineBasicBlock *
21428 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21429 MachineBasicBlock *MBB) const {
21430 DebugLoc DL = MI->getDebugLoc();
21431 MachineFunction *MF = MBB->getParent();
21432 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21433 MachineRegisterInfo &MRI = MF->getRegInfo();
21435 const BasicBlock *BB = MBB->getBasicBlock();
21436 MachineFunction::iterator I = ++MBB->getIterator();
21438 // Memory Reference
21439 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21440 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21443 unsigned MemOpndSlot = 0;
21445 unsigned CurOp = 0;
21447 DstReg = MI->getOperand(CurOp++).getReg();
21448 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21449 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21450 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21451 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21453 MemOpndSlot = CurOp;
21455 MVT PVT = getPointerTy(MF->getDataLayout());
21456 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21457 "Invalid Pointer Size!");
21459 // For v = setjmp(buf), we generate
21462 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21463 // SjLjSetup restoreMBB
21469 // v = phi(main, restore)
21472 // if base pointer being used, load it from frame
21475 MachineBasicBlock *thisMBB = MBB;
21476 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21477 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21478 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21479 MF->insert(I, mainMBB);
21480 MF->insert(I, sinkMBB);
21481 MF->push_back(restoreMBB);
21482 restoreMBB->setHasAddressTaken();
21484 MachineInstrBuilder MIB;
21486 // Transfer the remainder of BB and its successor edges to sinkMBB.
21487 sinkMBB->splice(sinkMBB->begin(), MBB,
21488 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21489 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21492 unsigned PtrStoreOpc = 0;
21493 unsigned LabelReg = 0;
21494 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21495 Reloc::Model RM = MF->getTarget().getRelocationModel();
21496 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21497 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21499 // Prepare IP either in reg or imm.
21500 if (!UseImmLabel) {
21501 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21502 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21503 LabelReg = MRI.createVirtualRegister(PtrRC);
21504 if (Subtarget->is64Bit()) {
21505 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21509 .addMBB(restoreMBB)
21512 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21513 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21514 .addReg(XII->getGlobalBaseReg(MF))
21517 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21521 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21523 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21524 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21525 if (i == X86::AddrDisp)
21526 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21528 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21531 MIB.addReg(LabelReg);
21533 MIB.addMBB(restoreMBB);
21534 MIB.setMemRefs(MMOBegin, MMOEnd);
21536 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21537 .addMBB(restoreMBB);
21539 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21540 MIB.addRegMask(RegInfo->getNoPreservedMask());
21541 thisMBB->addSuccessor(mainMBB);
21542 thisMBB->addSuccessor(restoreMBB);
21546 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21547 mainMBB->addSuccessor(sinkMBB);
21550 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21551 TII->get(X86::PHI), DstReg)
21552 .addReg(mainDstReg).addMBB(mainMBB)
21553 .addReg(restoreDstReg).addMBB(restoreMBB);
21556 if (RegInfo->hasBasePointer(*MF)) {
21557 const bool Uses64BitFramePtr =
21558 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21559 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21560 X86FI->setRestoreBasePointer(MF);
21561 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21562 unsigned BasePtr = RegInfo->getBaseRegister();
21563 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21564 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21565 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21566 .setMIFlag(MachineInstr::FrameSetup);
21568 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21569 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21570 restoreMBB->addSuccessor(sinkMBB);
21572 MI->eraseFromParent();
21576 MachineBasicBlock *
21577 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21578 MachineBasicBlock *MBB) const {
21579 DebugLoc DL = MI->getDebugLoc();
21580 MachineFunction *MF = MBB->getParent();
21581 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21582 MachineRegisterInfo &MRI = MF->getRegInfo();
21584 // Memory Reference
21585 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21586 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21588 MVT PVT = getPointerTy(MF->getDataLayout());
21589 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21590 "Invalid Pointer Size!");
21592 const TargetRegisterClass *RC =
21593 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21594 unsigned Tmp = MRI.createVirtualRegister(RC);
21595 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21596 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21597 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21598 unsigned SP = RegInfo->getStackRegister();
21600 MachineInstrBuilder MIB;
21602 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21603 const int64_t SPOffset = 2 * PVT.getStoreSize();
21605 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21606 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21609 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21610 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21611 MIB.addOperand(MI->getOperand(i));
21612 MIB.setMemRefs(MMOBegin, MMOEnd);
21614 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21615 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21616 if (i == X86::AddrDisp)
21617 MIB.addDisp(MI->getOperand(i), LabelOffset);
21619 MIB.addOperand(MI->getOperand(i));
21621 MIB.setMemRefs(MMOBegin, MMOEnd);
21623 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21624 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21625 if (i == X86::AddrDisp)
21626 MIB.addDisp(MI->getOperand(i), SPOffset);
21628 MIB.addOperand(MI->getOperand(i));
21630 MIB.setMemRefs(MMOBegin, MMOEnd);
21632 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21634 MI->eraseFromParent();
21638 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21639 // accumulator loops. Writing back to the accumulator allows the coalescer
21640 // to remove extra copies in the loop.
21641 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21642 MachineBasicBlock *
21643 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21644 MachineBasicBlock *MBB) const {
21645 MachineOperand &AddendOp = MI->getOperand(3);
21647 // Bail out early if the addend isn't a register - we can't switch these.
21648 if (!AddendOp.isReg())
21651 MachineFunction &MF = *MBB->getParent();
21652 MachineRegisterInfo &MRI = MF.getRegInfo();
21654 // Check whether the addend is defined by a PHI:
21655 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21656 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21657 if (!AddendDef.isPHI())
21660 // Look for the following pattern:
21662 // %addend = phi [%entry, 0], [%loop, %result]
21664 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21668 // %addend = phi [%entry, 0], [%loop, %result]
21670 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21672 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21673 assert(AddendDef.getOperand(i).isReg());
21674 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21675 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21676 if (&PHISrcInst == MI) {
21677 // Found a matching instruction.
21678 unsigned NewFMAOpc = 0;
21679 switch (MI->getOpcode()) {
21680 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21681 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21682 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21683 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21684 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21685 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21686 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21687 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21688 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21689 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21690 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21691 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21692 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21693 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21694 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21695 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21696 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21697 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21698 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21699 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21701 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21702 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21703 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21704 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21705 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21706 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21707 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21708 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21709 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21710 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21711 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21712 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21713 default: llvm_unreachable("Unrecognized FMA variant.");
21716 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21717 MachineInstrBuilder MIB =
21718 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21719 .addOperand(MI->getOperand(0))
21720 .addOperand(MI->getOperand(3))
21721 .addOperand(MI->getOperand(2))
21722 .addOperand(MI->getOperand(1));
21723 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21724 MI->eraseFromParent();
21731 MachineBasicBlock *
21732 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21733 MachineBasicBlock *BB) const {
21734 switch (MI->getOpcode()) {
21735 default: llvm_unreachable("Unexpected instr type to insert");
21736 case X86::TAILJMPd64:
21737 case X86::TAILJMPr64:
21738 case X86::TAILJMPm64:
21739 case X86::TAILJMPd64_REX:
21740 case X86::TAILJMPr64_REX:
21741 case X86::TAILJMPm64_REX:
21742 llvm_unreachable("TAILJMP64 would not be touched here.");
21743 case X86::TCRETURNdi64:
21744 case X86::TCRETURNri64:
21745 case X86::TCRETURNmi64:
21747 case X86::WIN_ALLOCA:
21748 return EmitLoweredWinAlloca(MI, BB);
21749 case X86::SEG_ALLOCA_32:
21750 case X86::SEG_ALLOCA_64:
21751 return EmitLoweredSegAlloca(MI, BB);
21752 case X86::TLSCall_32:
21753 case X86::TLSCall_64:
21754 return EmitLoweredTLSCall(MI, BB);
21755 case X86::CMOV_FR32:
21756 case X86::CMOV_FR64:
21757 case X86::CMOV_GR8:
21758 case X86::CMOV_GR16:
21759 case X86::CMOV_GR32:
21760 case X86::CMOV_RFP32:
21761 case X86::CMOV_RFP64:
21762 case X86::CMOV_RFP80:
21763 case X86::CMOV_V2F64:
21764 case X86::CMOV_V2I64:
21765 case X86::CMOV_V4F32:
21766 case X86::CMOV_V4F64:
21767 case X86::CMOV_V4I64:
21768 case X86::CMOV_V16F32:
21769 case X86::CMOV_V8F32:
21770 case X86::CMOV_V8F64:
21771 case X86::CMOV_V8I64:
21772 case X86::CMOV_V8I1:
21773 case X86::CMOV_V16I1:
21774 case X86::CMOV_V32I1:
21775 case X86::CMOV_V64I1:
21776 return EmitLoweredSelect(MI, BB);
21778 case X86::RELEASE_FADD32mr:
21779 case X86::RELEASE_FADD64mr:
21780 return EmitLoweredAtomicFP(MI, BB);
21782 case X86::FP32_TO_INT16_IN_MEM:
21783 case X86::FP32_TO_INT32_IN_MEM:
21784 case X86::FP32_TO_INT64_IN_MEM:
21785 case X86::FP64_TO_INT16_IN_MEM:
21786 case X86::FP64_TO_INT32_IN_MEM:
21787 case X86::FP64_TO_INT64_IN_MEM:
21788 case X86::FP80_TO_INT16_IN_MEM:
21789 case X86::FP80_TO_INT32_IN_MEM:
21790 case X86::FP80_TO_INT64_IN_MEM: {
21791 MachineFunction *F = BB->getParent();
21792 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21793 DebugLoc DL = MI->getDebugLoc();
21795 // Change the floating point control register to use "round towards zero"
21796 // mode when truncating to an integer value.
21797 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21798 addFrameReference(BuildMI(*BB, MI, DL,
21799 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21801 // Load the old value of the high byte of the control word...
21803 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21804 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21807 // Set the high part to be round to zero...
21808 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21811 // Reload the modified control word now...
21812 addFrameReference(BuildMI(*BB, MI, DL,
21813 TII->get(X86::FLDCW16m)), CWFrameIdx);
21815 // Restore the memory image of control word to original value
21816 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21819 // Get the X86 opcode to use.
21821 switch (MI->getOpcode()) {
21822 default: llvm_unreachable("illegal opcode!");
21823 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21824 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21825 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21826 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21827 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21828 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21829 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21830 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21831 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21835 MachineOperand &Op = MI->getOperand(0);
21837 AM.BaseType = X86AddressMode::RegBase;
21838 AM.Base.Reg = Op.getReg();
21840 AM.BaseType = X86AddressMode::FrameIndexBase;
21841 AM.Base.FrameIndex = Op.getIndex();
21843 Op = MI->getOperand(1);
21845 AM.Scale = Op.getImm();
21846 Op = MI->getOperand(2);
21848 AM.IndexReg = Op.getImm();
21849 Op = MI->getOperand(3);
21850 if (Op.isGlobal()) {
21851 AM.GV = Op.getGlobal();
21853 AM.Disp = Op.getImm();
21855 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21856 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21858 // Reload the original control word now.
21859 addFrameReference(BuildMI(*BB, MI, DL,
21860 TII->get(X86::FLDCW16m)), CWFrameIdx);
21862 MI->eraseFromParent(); // The pseudo instruction is gone now.
21865 // String/text processing lowering.
21866 case X86::PCMPISTRM128REG:
21867 case X86::VPCMPISTRM128REG:
21868 case X86::PCMPISTRM128MEM:
21869 case X86::VPCMPISTRM128MEM:
21870 case X86::PCMPESTRM128REG:
21871 case X86::VPCMPESTRM128REG:
21872 case X86::PCMPESTRM128MEM:
21873 case X86::VPCMPESTRM128MEM:
21874 assert(Subtarget->hasSSE42() &&
21875 "Target must have SSE4.2 or AVX features enabled");
21876 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21878 // String/text processing lowering.
21879 case X86::PCMPISTRIREG:
21880 case X86::VPCMPISTRIREG:
21881 case X86::PCMPISTRIMEM:
21882 case X86::VPCMPISTRIMEM:
21883 case X86::PCMPESTRIREG:
21884 case X86::VPCMPESTRIREG:
21885 case X86::PCMPESTRIMEM:
21886 case X86::VPCMPESTRIMEM:
21887 assert(Subtarget->hasSSE42() &&
21888 "Target must have SSE4.2 or AVX features enabled");
21889 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21891 // Thread synchronization.
21893 return EmitMonitor(MI, BB, Subtarget);
21897 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21899 case X86::VASTART_SAVE_XMM_REGS:
21900 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21902 case X86::VAARG_64:
21903 return EmitVAARG64WithCustomInserter(MI, BB);
21905 case X86::EH_SjLj_SetJmp32:
21906 case X86::EH_SjLj_SetJmp64:
21907 return emitEHSjLjSetJmp(MI, BB);
21909 case X86::EH_SjLj_LongJmp32:
21910 case X86::EH_SjLj_LongJmp64:
21911 return emitEHSjLjLongJmp(MI, BB);
21913 case TargetOpcode::STATEPOINT:
21914 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21915 // this point in the process. We diverge later.
21916 return emitPatchPoint(MI, BB);
21918 case TargetOpcode::STACKMAP:
21919 case TargetOpcode::PATCHPOINT:
21920 return emitPatchPoint(MI, BB);
21922 case X86::VFMADDPDr213r:
21923 case X86::VFMADDPSr213r:
21924 case X86::VFMADDSDr213r:
21925 case X86::VFMADDSSr213r:
21926 case X86::VFMSUBPDr213r:
21927 case X86::VFMSUBPSr213r:
21928 case X86::VFMSUBSDr213r:
21929 case X86::VFMSUBSSr213r:
21930 case X86::VFNMADDPDr213r:
21931 case X86::VFNMADDPSr213r:
21932 case X86::VFNMADDSDr213r:
21933 case X86::VFNMADDSSr213r:
21934 case X86::VFNMSUBPDr213r:
21935 case X86::VFNMSUBPSr213r:
21936 case X86::VFNMSUBSDr213r:
21937 case X86::VFNMSUBSSr213r:
21938 case X86::VFMADDSUBPDr213r:
21939 case X86::VFMADDSUBPSr213r:
21940 case X86::VFMSUBADDPDr213r:
21941 case X86::VFMSUBADDPSr213r:
21942 case X86::VFMADDPDr213rY:
21943 case X86::VFMADDPSr213rY:
21944 case X86::VFMSUBPDr213rY:
21945 case X86::VFMSUBPSr213rY:
21946 case X86::VFNMADDPDr213rY:
21947 case X86::VFNMADDPSr213rY:
21948 case X86::VFNMSUBPDr213rY:
21949 case X86::VFNMSUBPSr213rY:
21950 case X86::VFMADDSUBPDr213rY:
21951 case X86::VFMADDSUBPSr213rY:
21952 case X86::VFMSUBADDPDr213rY:
21953 case X86::VFMSUBADDPSr213rY:
21954 return emitFMA3Instr(MI, BB);
21958 //===----------------------------------------------------------------------===//
21959 // X86 Optimization Hooks
21960 //===----------------------------------------------------------------------===//
21962 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21965 const SelectionDAG &DAG,
21966 unsigned Depth) const {
21967 unsigned BitWidth = KnownZero.getBitWidth();
21968 unsigned Opc = Op.getOpcode();
21969 assert((Opc >= ISD::BUILTIN_OP_END ||
21970 Opc == ISD::INTRINSIC_WO_CHAIN ||
21971 Opc == ISD::INTRINSIC_W_CHAIN ||
21972 Opc == ISD::INTRINSIC_VOID) &&
21973 "Should use MaskedValueIsZero if you don't know whether Op"
21974 " is a target node!");
21976 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21990 // These nodes' second result is a boolean.
21991 if (Op.getResNo() == 0)
21994 case X86ISD::SETCC:
21995 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21997 case ISD::INTRINSIC_WO_CHAIN: {
21998 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21999 unsigned NumLoBits = 0;
22002 case Intrinsic::x86_sse_movmsk_ps:
22003 case Intrinsic::x86_avx_movmsk_ps_256:
22004 case Intrinsic::x86_sse2_movmsk_pd:
22005 case Intrinsic::x86_avx_movmsk_pd_256:
22006 case Intrinsic::x86_mmx_pmovmskb:
22007 case Intrinsic::x86_sse2_pmovmskb_128:
22008 case Intrinsic::x86_avx2_pmovmskb: {
22009 // High bits of movmskp{s|d}, pmovmskb are known zero.
22011 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22012 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22013 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22014 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22015 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22016 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22017 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22018 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22020 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22029 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22031 const SelectionDAG &,
22032 unsigned Depth) const {
22033 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22034 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22035 return Op.getValueType().getScalarSizeInBits();
22041 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22042 /// node is a GlobalAddress + offset.
22043 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22044 const GlobalValue* &GA,
22045 int64_t &Offset) const {
22046 if (N->getOpcode() == X86ISD::Wrapper) {
22047 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22048 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22049 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22053 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22056 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22057 /// same as extracting the high 128-bit part of 256-bit vector and then
22058 /// inserting the result into the low part of a new 256-bit vector
22059 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22060 EVT VT = SVOp->getValueType(0);
22061 unsigned NumElems = VT.getVectorNumElements();
22063 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22064 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22065 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22066 SVOp->getMaskElt(j) >= 0)
22072 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22073 /// same as extracting the low 128-bit part of 256-bit vector and then
22074 /// inserting the result into the high part of a new 256-bit vector
22075 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22076 EVT VT = SVOp->getValueType(0);
22077 unsigned NumElems = VT.getVectorNumElements();
22079 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22080 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22081 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22082 SVOp->getMaskElt(j) >= 0)
22088 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22089 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22090 TargetLowering::DAGCombinerInfo &DCI,
22091 const X86Subtarget* Subtarget) {
22093 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22094 SDValue V1 = SVOp->getOperand(0);
22095 SDValue V2 = SVOp->getOperand(1);
22096 EVT VT = SVOp->getValueType(0);
22097 unsigned NumElems = VT.getVectorNumElements();
22099 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22100 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22104 // V UNDEF BUILD_VECTOR UNDEF
22106 // CONCAT_VECTOR CONCAT_VECTOR
22109 // RESULT: V + zero extended
22111 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22112 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22113 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22116 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22119 // To match the shuffle mask, the first half of the mask should
22120 // be exactly the first vector, and all the rest a splat with the
22121 // first element of the second one.
22122 for (unsigned i = 0; i != NumElems/2; ++i)
22123 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22124 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22127 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22128 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22129 if (Ld->hasNUsesOfValue(1, 0)) {
22130 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22131 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22133 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22135 Ld->getPointerInfo(),
22136 Ld->getAlignment(),
22137 false/*isVolatile*/, true/*ReadMem*/,
22138 false/*WriteMem*/);
22140 // Make sure the newly-created LOAD is in the same position as Ld in
22141 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22142 // and update uses of Ld's output chain to use the TokenFactor.
22143 if (Ld->hasAnyUseOfValue(1)) {
22144 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22145 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22146 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22147 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22148 SDValue(ResNode.getNode(), 1));
22151 return DAG.getBitcast(VT, ResNode);
22155 // Emit a zeroed vector and insert the desired subvector on its
22157 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22158 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22159 return DCI.CombineTo(N, InsV);
22162 //===--------------------------------------------------------------------===//
22163 // Combine some shuffles into subvector extracts and inserts:
22166 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22167 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22168 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22169 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22170 return DCI.CombineTo(N, InsV);
22173 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22174 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22175 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22176 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22177 return DCI.CombineTo(N, InsV);
22183 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22186 /// This is the leaf of the recursive combinine below. When we have found some
22187 /// chain of single-use x86 shuffle instructions and accumulated the combined
22188 /// shuffle mask represented by them, this will try to pattern match that mask
22189 /// into either a single instruction if there is a special purpose instruction
22190 /// for this operation, or into a PSHUFB instruction which is a fully general
22191 /// instruction but should only be used to replace chains over a certain depth.
22192 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22193 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22194 TargetLowering::DAGCombinerInfo &DCI,
22195 const X86Subtarget *Subtarget) {
22196 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22198 // Find the operand that enters the chain. Note that multiple uses are OK
22199 // here, we're not going to remove the operand we find.
22200 SDValue Input = Op.getOperand(0);
22201 while (Input.getOpcode() == ISD::BITCAST)
22202 Input = Input.getOperand(0);
22204 MVT VT = Input.getSimpleValueType();
22205 MVT RootVT = Root.getSimpleValueType();
22208 if (Mask.size() == 1) {
22209 int Index = Mask[0];
22210 assert((Index >= 0 || Index == SM_SentinelUndef ||
22211 Index == SM_SentinelZero) &&
22212 "Invalid shuffle index found!");
22214 // We may end up with an accumulated mask of size 1 as a result of
22215 // widening of shuffle operands (see function canWidenShuffleElements).
22216 // If the only shuffle index is equal to SM_SentinelZero then propagate
22217 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22218 // mask, and therefore the entire chain of shuffles can be folded away.
22219 if (Index == SM_SentinelZero)
22220 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22222 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22227 // Use the float domain if the operand type is a floating point type.
22228 bool FloatDomain = VT.isFloatingPoint();
22230 // For floating point shuffles, we don't have free copies in the shuffle
22231 // instructions or the ability to load as part of the instruction, so
22232 // canonicalize their shuffles to UNPCK or MOV variants.
22234 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22235 // vectors because it can have a load folded into it that UNPCK cannot. This
22236 // doesn't preclude something switching to the shorter encoding post-RA.
22238 // FIXME: Should teach these routines about AVX vector widths.
22239 if (FloatDomain && VT.is128BitVector()) {
22240 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22241 bool Lo = Mask.equals({0, 0});
22244 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22245 // is no slower than UNPCKLPD but has the option to fold the input operand
22246 // into even an unaligned memory load.
22247 if (Lo && Subtarget->hasSSE3()) {
22248 Shuffle = X86ISD::MOVDDUP;
22249 ShuffleVT = MVT::v2f64;
22251 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22252 // than the UNPCK variants.
22253 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22254 ShuffleVT = MVT::v4f32;
22256 if (Depth == 1 && Root->getOpcode() == Shuffle)
22257 return false; // Nothing to do!
22258 Op = DAG.getBitcast(ShuffleVT, Input);
22259 DCI.AddToWorklist(Op.getNode());
22260 if (Shuffle == X86ISD::MOVDDUP)
22261 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22263 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22264 DCI.AddToWorklist(Op.getNode());
22265 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22269 if (Subtarget->hasSSE3() &&
22270 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22271 bool Lo = Mask.equals({0, 0, 2, 2});
22272 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22273 MVT ShuffleVT = MVT::v4f32;
22274 if (Depth == 1 && Root->getOpcode() == Shuffle)
22275 return false; // Nothing to do!
22276 Op = DAG.getBitcast(ShuffleVT, Input);
22277 DCI.AddToWorklist(Op.getNode());
22278 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22279 DCI.AddToWorklist(Op.getNode());
22280 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22284 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22285 bool Lo = Mask.equals({0, 0, 1, 1});
22286 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22287 MVT ShuffleVT = MVT::v4f32;
22288 if (Depth == 1 && Root->getOpcode() == Shuffle)
22289 return false; // Nothing to do!
22290 Op = DAG.getBitcast(ShuffleVT, Input);
22291 DCI.AddToWorklist(Op.getNode());
22292 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22293 DCI.AddToWorklist(Op.getNode());
22294 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22300 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22301 // variants as none of these have single-instruction variants that are
22302 // superior to the UNPCK formulation.
22303 if (!FloatDomain && VT.is128BitVector() &&
22304 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22305 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22306 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22308 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22309 bool Lo = Mask[0] == 0;
22310 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22311 if (Depth == 1 && Root->getOpcode() == Shuffle)
22312 return false; // Nothing to do!
22314 switch (Mask.size()) {
22316 ShuffleVT = MVT::v8i16;
22319 ShuffleVT = MVT::v16i8;
22322 llvm_unreachable("Impossible mask size!");
22324 Op = DAG.getBitcast(ShuffleVT, Input);
22325 DCI.AddToWorklist(Op.getNode());
22326 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22327 DCI.AddToWorklist(Op.getNode());
22328 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22333 // Don't try to re-form single instruction chains under any circumstances now
22334 // that we've done encoding canonicalization for them.
22338 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22339 // can replace them with a single PSHUFB instruction profitably. Intel's
22340 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22341 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22342 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22343 SmallVector<SDValue, 16> PSHUFBMask;
22344 int NumBytes = VT.getSizeInBits() / 8;
22345 int Ratio = NumBytes / Mask.size();
22346 for (int i = 0; i < NumBytes; ++i) {
22347 if (Mask[i / Ratio] == SM_SentinelUndef) {
22348 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22351 int M = Mask[i / Ratio] != SM_SentinelZero
22352 ? Ratio * Mask[i / Ratio] + i % Ratio
22354 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22356 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22357 Op = DAG.getBitcast(ByteVT, Input);
22358 DCI.AddToWorklist(Op.getNode());
22359 SDValue PSHUFBMaskOp =
22360 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22361 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22362 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22363 DCI.AddToWorklist(Op.getNode());
22364 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22369 // Failed to find any combines.
22373 /// \brief Fully generic combining of x86 shuffle instructions.
22375 /// This should be the last combine run over the x86 shuffle instructions. Once
22376 /// they have been fully optimized, this will recursively consider all chains
22377 /// of single-use shuffle instructions, build a generic model of the cumulative
22378 /// shuffle operation, and check for simpler instructions which implement this
22379 /// operation. We use this primarily for two purposes:
22381 /// 1) Collapse generic shuffles to specialized single instructions when
22382 /// equivalent. In most cases, this is just an encoding size win, but
22383 /// sometimes we will collapse multiple generic shuffles into a single
22384 /// special-purpose shuffle.
22385 /// 2) Look for sequences of shuffle instructions with 3 or more total
22386 /// instructions, and replace them with the slightly more expensive SSSE3
22387 /// PSHUFB instruction if available. We do this as the last combining step
22388 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22389 /// a suitable short sequence of other instructions. The PHUFB will either
22390 /// use a register or have to read from memory and so is slightly (but only
22391 /// slightly) more expensive than the other shuffle instructions.
22393 /// Because this is inherently a quadratic operation (for each shuffle in
22394 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22395 /// This should never be an issue in practice as the shuffle lowering doesn't
22396 /// produce sequences of more than 8 instructions.
22398 /// FIXME: We will currently miss some cases where the redundant shuffling
22399 /// would simplify under the threshold for PSHUFB formation because of
22400 /// combine-ordering. To fix this, we should do the redundant instruction
22401 /// combining in this recursive walk.
22402 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22403 ArrayRef<int> RootMask,
22404 int Depth, bool HasPSHUFB,
22406 TargetLowering::DAGCombinerInfo &DCI,
22407 const X86Subtarget *Subtarget) {
22408 // Bound the depth of our recursive combine because this is ultimately
22409 // quadratic in nature.
22413 // Directly rip through bitcasts to find the underlying operand.
22414 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22415 Op = Op.getOperand(0);
22417 MVT VT = Op.getSimpleValueType();
22418 if (!VT.isVector())
22419 return false; // Bail if we hit a non-vector.
22421 assert(Root.getSimpleValueType().isVector() &&
22422 "Shuffles operate on vector types!");
22423 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22424 "Can only combine shuffles of the same vector register size.");
22426 if (!isTargetShuffle(Op.getOpcode()))
22428 SmallVector<int, 16> OpMask;
22430 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22431 // We only can combine unary shuffles which we can decode the mask for.
22432 if (!HaveMask || !IsUnary)
22435 assert(VT.getVectorNumElements() == OpMask.size() &&
22436 "Different mask size from vector size!");
22437 assert(((RootMask.size() > OpMask.size() &&
22438 RootMask.size() % OpMask.size() == 0) ||
22439 (OpMask.size() > RootMask.size() &&
22440 OpMask.size() % RootMask.size() == 0) ||
22441 OpMask.size() == RootMask.size()) &&
22442 "The smaller number of elements must divide the larger.");
22443 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22444 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22445 assert(((RootRatio == 1 && OpRatio == 1) ||
22446 (RootRatio == 1) != (OpRatio == 1)) &&
22447 "Must not have a ratio for both incoming and op masks!");
22449 SmallVector<int, 16> Mask;
22450 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22452 // Merge this shuffle operation's mask into our accumulated mask. Note that
22453 // this shuffle's mask will be the first applied to the input, followed by the
22454 // root mask to get us all the way to the root value arrangement. The reason
22455 // for this order is that we are recursing up the operation chain.
22456 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22457 int RootIdx = i / RootRatio;
22458 if (RootMask[RootIdx] < 0) {
22459 // This is a zero or undef lane, we're done.
22460 Mask.push_back(RootMask[RootIdx]);
22464 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22465 int OpIdx = RootMaskedIdx / OpRatio;
22466 if (OpMask[OpIdx] < 0) {
22467 // The incoming lanes are zero or undef, it doesn't matter which ones we
22469 Mask.push_back(OpMask[OpIdx]);
22473 // Ok, we have non-zero lanes, map them through.
22474 Mask.push_back(OpMask[OpIdx] * OpRatio +
22475 RootMaskedIdx % OpRatio);
22478 // See if we can recurse into the operand to combine more things.
22479 switch (Op.getOpcode()) {
22480 case X86ISD::PSHUFB:
22482 case X86ISD::PSHUFD:
22483 case X86ISD::PSHUFHW:
22484 case X86ISD::PSHUFLW:
22485 if (Op.getOperand(0).hasOneUse() &&
22486 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22487 HasPSHUFB, DAG, DCI, Subtarget))
22491 case X86ISD::UNPCKL:
22492 case X86ISD::UNPCKH:
22493 assert(Op.getOperand(0) == Op.getOperand(1) &&
22494 "We only combine unary shuffles!");
22495 // We can't check for single use, we have to check that this shuffle is the
22497 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22498 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22499 HasPSHUFB, DAG, DCI, Subtarget))
22504 // Minor canonicalization of the accumulated shuffle mask to make it easier
22505 // to match below. All this does is detect masks with squential pairs of
22506 // elements, and shrink them to the half-width mask. It does this in a loop
22507 // so it will reduce the size of the mask to the minimal width mask which
22508 // performs an equivalent shuffle.
22509 SmallVector<int, 16> WidenedMask;
22510 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22511 Mask = std::move(WidenedMask);
22512 WidenedMask.clear();
22515 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22519 /// \brief Get the PSHUF-style mask from PSHUF node.
22521 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22522 /// PSHUF-style masks that can be reused with such instructions.
22523 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22524 MVT VT = N.getSimpleValueType();
22525 SmallVector<int, 4> Mask;
22527 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22531 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22532 // matter. Check that the upper masks are repeats and remove them.
22533 if (VT.getSizeInBits() > 128) {
22534 int LaneElts = 128 / VT.getScalarSizeInBits();
22536 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22537 for (int j = 0; j < LaneElts; ++j)
22538 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22539 "Mask doesn't repeat in high 128-bit lanes!");
22541 Mask.resize(LaneElts);
22544 switch (N.getOpcode()) {
22545 case X86ISD::PSHUFD:
22547 case X86ISD::PSHUFLW:
22550 case X86ISD::PSHUFHW:
22551 Mask.erase(Mask.begin(), Mask.begin() + 4);
22552 for (int &M : Mask)
22556 llvm_unreachable("No valid shuffle instruction found!");
22560 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22562 /// We walk up the chain and look for a combinable shuffle, skipping over
22563 /// shuffles that we could hoist this shuffle's transformation past without
22564 /// altering anything.
22566 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22568 TargetLowering::DAGCombinerInfo &DCI) {
22569 assert(N.getOpcode() == X86ISD::PSHUFD &&
22570 "Called with something other than an x86 128-bit half shuffle!");
22573 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22574 // of the shuffles in the chain so that we can form a fresh chain to replace
22576 SmallVector<SDValue, 8> Chain;
22577 SDValue V = N.getOperand(0);
22578 for (; V.hasOneUse(); V = V.getOperand(0)) {
22579 switch (V.getOpcode()) {
22581 return SDValue(); // Nothing combined!
22584 // Skip bitcasts as we always know the type for the target specific
22588 case X86ISD::PSHUFD:
22589 // Found another dword shuffle.
22592 case X86ISD::PSHUFLW:
22593 // Check that the low words (being shuffled) are the identity in the
22594 // dword shuffle, and the high words are self-contained.
22595 if (Mask[0] != 0 || Mask[1] != 1 ||
22596 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22599 Chain.push_back(V);
22602 case X86ISD::PSHUFHW:
22603 // Check that the high words (being shuffled) are the identity in the
22604 // dword shuffle, and the low words are self-contained.
22605 if (Mask[2] != 2 || Mask[3] != 3 ||
22606 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22609 Chain.push_back(V);
22612 case X86ISD::UNPCKL:
22613 case X86ISD::UNPCKH:
22614 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22615 // shuffle into a preceding word shuffle.
22616 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22617 V.getSimpleValueType().getScalarType() != MVT::i16)
22620 // Search for a half-shuffle which we can combine with.
22621 unsigned CombineOp =
22622 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22623 if (V.getOperand(0) != V.getOperand(1) ||
22624 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22626 Chain.push_back(V);
22627 V = V.getOperand(0);
22629 switch (V.getOpcode()) {
22631 return SDValue(); // Nothing to combine.
22633 case X86ISD::PSHUFLW:
22634 case X86ISD::PSHUFHW:
22635 if (V.getOpcode() == CombineOp)
22638 Chain.push_back(V);
22642 V = V.getOperand(0);
22646 } while (V.hasOneUse());
22649 // Break out of the loop if we break out of the switch.
22653 if (!V.hasOneUse())
22654 // We fell out of the loop without finding a viable combining instruction.
22657 // Merge this node's mask and our incoming mask.
22658 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22659 for (int &M : Mask)
22661 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22662 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22664 // Rebuild the chain around this new shuffle.
22665 while (!Chain.empty()) {
22666 SDValue W = Chain.pop_back_val();
22668 if (V.getValueType() != W.getOperand(0).getValueType())
22669 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22671 switch (W.getOpcode()) {
22673 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22675 case X86ISD::UNPCKL:
22676 case X86ISD::UNPCKH:
22677 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22680 case X86ISD::PSHUFD:
22681 case X86ISD::PSHUFLW:
22682 case X86ISD::PSHUFHW:
22683 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22687 if (V.getValueType() != N.getValueType())
22688 V = DAG.getBitcast(N.getValueType(), V);
22690 // Return the new chain to replace N.
22694 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22697 /// We walk up the chain, skipping shuffles of the other half and looking
22698 /// through shuffles which switch halves trying to find a shuffle of the same
22699 /// pair of dwords.
22700 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22702 TargetLowering::DAGCombinerInfo &DCI) {
22704 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22705 "Called with something other than an x86 128-bit half shuffle!");
22707 unsigned CombineOpcode = N.getOpcode();
22709 // Walk up a single-use chain looking for a combinable shuffle.
22710 SDValue V = N.getOperand(0);
22711 for (; V.hasOneUse(); V = V.getOperand(0)) {
22712 switch (V.getOpcode()) {
22714 return false; // Nothing combined!
22717 // Skip bitcasts as we always know the type for the target specific
22721 case X86ISD::PSHUFLW:
22722 case X86ISD::PSHUFHW:
22723 if (V.getOpcode() == CombineOpcode)
22726 // Other-half shuffles are no-ops.
22729 // Break out of the loop if we break out of the switch.
22733 if (!V.hasOneUse())
22734 // We fell out of the loop without finding a viable combining instruction.
22737 // Combine away the bottom node as its shuffle will be accumulated into
22738 // a preceding shuffle.
22739 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22741 // Record the old value.
22744 // Merge this node's mask and our incoming mask (adjusted to account for all
22745 // the pshufd instructions encountered).
22746 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22747 for (int &M : Mask)
22749 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22750 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22752 // Check that the shuffles didn't cancel each other out. If not, we need to
22753 // combine to the new one.
22755 // Replace the combinable shuffle with the combined one, updating all users
22756 // so that we re-evaluate the chain here.
22757 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22762 /// \brief Try to combine x86 target specific shuffles.
22763 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22764 TargetLowering::DAGCombinerInfo &DCI,
22765 const X86Subtarget *Subtarget) {
22767 MVT VT = N.getSimpleValueType();
22768 SmallVector<int, 4> Mask;
22770 switch (N.getOpcode()) {
22771 case X86ISD::PSHUFD:
22772 case X86ISD::PSHUFLW:
22773 case X86ISD::PSHUFHW:
22774 Mask = getPSHUFShuffleMask(N);
22775 assert(Mask.size() == 4);
22781 // Nuke no-op shuffles that show up after combining.
22782 if (isNoopShuffleMask(Mask))
22783 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22785 // Look for simplifications involving one or two shuffle instructions.
22786 SDValue V = N.getOperand(0);
22787 switch (N.getOpcode()) {
22790 case X86ISD::PSHUFLW:
22791 case X86ISD::PSHUFHW:
22792 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22794 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22795 return SDValue(); // We combined away this shuffle, so we're done.
22797 // See if this reduces to a PSHUFD which is no more expensive and can
22798 // combine with more operations. Note that it has to at least flip the
22799 // dwords as otherwise it would have been removed as a no-op.
22800 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22801 int DMask[] = {0, 1, 2, 3};
22802 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22803 DMask[DOffset + 0] = DOffset + 1;
22804 DMask[DOffset + 1] = DOffset + 0;
22805 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22806 V = DAG.getBitcast(DVT, V);
22807 DCI.AddToWorklist(V.getNode());
22808 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22809 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22810 DCI.AddToWorklist(V.getNode());
22811 return DAG.getBitcast(VT, V);
22814 // Look for shuffle patterns which can be implemented as a single unpack.
22815 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22816 // only works when we have a PSHUFD followed by two half-shuffles.
22817 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22818 (V.getOpcode() == X86ISD::PSHUFLW ||
22819 V.getOpcode() == X86ISD::PSHUFHW) &&
22820 V.getOpcode() != N.getOpcode() &&
22822 SDValue D = V.getOperand(0);
22823 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22824 D = D.getOperand(0);
22825 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22826 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22827 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22828 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22829 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22831 for (int i = 0; i < 4; ++i) {
22832 WordMask[i + NOffset] = Mask[i] + NOffset;
22833 WordMask[i + VOffset] = VMask[i] + VOffset;
22835 // Map the word mask through the DWord mask.
22837 for (int i = 0; i < 8; ++i)
22838 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22839 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22840 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22841 // We can replace all three shuffles with an unpack.
22842 V = DAG.getBitcast(VT, D.getOperand(0));
22843 DCI.AddToWorklist(V.getNode());
22844 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22853 case X86ISD::PSHUFD:
22854 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22863 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22865 /// We combine this directly on the abstract vector shuffle nodes so it is
22866 /// easier to generically match. We also insert dummy vector shuffle nodes for
22867 /// the operands which explicitly discard the lanes which are unused by this
22868 /// operation to try to flow through the rest of the combiner the fact that
22869 /// they're unused.
22870 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22872 EVT VT = N->getValueType(0);
22874 // We only handle target-independent shuffles.
22875 // FIXME: It would be easy and harmless to use the target shuffle mask
22876 // extraction tool to support more.
22877 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22880 auto *SVN = cast<ShuffleVectorSDNode>(N);
22881 ArrayRef<int> Mask = SVN->getMask();
22882 SDValue V1 = N->getOperand(0);
22883 SDValue V2 = N->getOperand(1);
22885 // We require the first shuffle operand to be the SUB node, and the second to
22886 // be the ADD node.
22887 // FIXME: We should support the commuted patterns.
22888 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22891 // If there are other uses of these operations we can't fold them.
22892 if (!V1->hasOneUse() || !V2->hasOneUse())
22895 // Ensure that both operations have the same operands. Note that we can
22896 // commute the FADD operands.
22897 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22898 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22899 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22902 // We're looking for blends between FADD and FSUB nodes. We insist on these
22903 // nodes being lined up in a specific expected pattern.
22904 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22905 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22906 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22909 // Only specific types are legal at this point, assert so we notice if and
22910 // when these change.
22911 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22912 VT == MVT::v4f64) &&
22913 "Unknown vector type encountered!");
22915 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22918 /// PerformShuffleCombine - Performs several different shuffle combines.
22919 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22920 TargetLowering::DAGCombinerInfo &DCI,
22921 const X86Subtarget *Subtarget) {
22923 SDValue N0 = N->getOperand(0);
22924 SDValue N1 = N->getOperand(1);
22925 EVT VT = N->getValueType(0);
22927 // Don't create instructions with illegal types after legalize types has run.
22928 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22929 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22932 // If we have legalized the vector types, look for blends of FADD and FSUB
22933 // nodes that we can fuse into an ADDSUB node.
22934 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22935 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22938 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22939 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22940 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22941 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22943 // During Type Legalization, when promoting illegal vector types,
22944 // the backend might introduce new shuffle dag nodes and bitcasts.
22946 // This code performs the following transformation:
22947 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22948 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22950 // We do this only if both the bitcast and the BINOP dag nodes have
22951 // one use. Also, perform this transformation only if the new binary
22952 // operation is legal. This is to avoid introducing dag nodes that
22953 // potentially need to be further expanded (or custom lowered) into a
22954 // less optimal sequence of dag nodes.
22955 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22956 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22957 N0.getOpcode() == ISD::BITCAST) {
22958 SDValue BC0 = N0.getOperand(0);
22959 EVT SVT = BC0.getValueType();
22960 unsigned Opcode = BC0.getOpcode();
22961 unsigned NumElts = VT.getVectorNumElements();
22963 if (BC0.hasOneUse() && SVT.isVector() &&
22964 SVT.getVectorNumElements() * 2 == NumElts &&
22965 TLI.isOperationLegal(Opcode, VT)) {
22966 bool CanFold = false;
22978 unsigned SVTNumElts = SVT.getVectorNumElements();
22979 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22980 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22981 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22982 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22983 CanFold = SVOp->getMaskElt(i) < 0;
22986 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22987 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22988 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22989 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22994 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22995 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22996 // consecutive, non-overlapping, and in the right order.
22997 SmallVector<SDValue, 16> Elts;
22998 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22999 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23001 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23004 if (isTargetShuffle(N->getOpcode())) {
23006 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23007 if (Shuffle.getNode())
23010 // Try recursively combining arbitrary sequences of x86 shuffle
23011 // instructions into higher-order shuffles. We do this after combining
23012 // specific PSHUF instruction sequences into their minimal form so that we
23013 // can evaluate how many specialized shuffle instructions are involved in
23014 // a particular chain.
23015 SmallVector<int, 1> NonceMask; // Just a placeholder.
23016 NonceMask.push_back(0);
23017 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23018 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23020 return SDValue(); // This routine will use CombineTo to replace N.
23026 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23027 /// specific shuffle of a load can be folded into a single element load.
23028 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23029 /// shuffles have been custom lowered so we need to handle those here.
23030 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23031 TargetLowering::DAGCombinerInfo &DCI) {
23032 if (DCI.isBeforeLegalizeOps())
23035 SDValue InVec = N->getOperand(0);
23036 SDValue EltNo = N->getOperand(1);
23038 if (!isa<ConstantSDNode>(EltNo))
23041 EVT OriginalVT = InVec.getValueType();
23043 if (InVec.getOpcode() == ISD::BITCAST) {
23044 // Don't duplicate a load with other uses.
23045 if (!InVec.hasOneUse())
23047 EVT BCVT = InVec.getOperand(0).getValueType();
23048 if (!BCVT.isVector() ||
23049 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23051 InVec = InVec.getOperand(0);
23054 EVT CurrentVT = InVec.getValueType();
23056 if (!isTargetShuffle(InVec.getOpcode()))
23059 // Don't duplicate a load with other uses.
23060 if (!InVec.hasOneUse())
23063 SmallVector<int, 16> ShuffleMask;
23065 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23066 ShuffleMask, UnaryShuffle))
23069 // Select the input vector, guarding against out of range extract vector.
23070 unsigned NumElems = CurrentVT.getVectorNumElements();
23071 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23072 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23073 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23074 : InVec.getOperand(1);
23076 // If inputs to shuffle are the same for both ops, then allow 2 uses
23077 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23078 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23080 if (LdNode.getOpcode() == ISD::BITCAST) {
23081 // Don't duplicate a load with other uses.
23082 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23085 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23086 LdNode = LdNode.getOperand(0);
23089 if (!ISD::isNormalLoad(LdNode.getNode()))
23092 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23094 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23097 EVT EltVT = N->getValueType(0);
23098 // If there's a bitcast before the shuffle, check if the load type and
23099 // alignment is valid.
23100 unsigned Align = LN0->getAlignment();
23101 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23102 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23103 EltVT.getTypeForEVT(*DAG.getContext()));
23105 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23108 // All checks match so transform back to vector_shuffle so that DAG combiner
23109 // can finish the job
23112 // Create shuffle node taking into account the case that its a unary shuffle
23113 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23114 : InVec.getOperand(1);
23115 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23116 InVec.getOperand(0), Shuffle,
23118 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23119 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23123 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23124 const X86Subtarget *Subtarget) {
23125 SDValue N0 = N->getOperand(0);
23126 EVT VT = N->getValueType(0);
23128 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23129 // special and don't usually play with other vector types, it's better to
23130 // handle them early to be sure we emit efficient code by avoiding
23131 // store-load conversions.
23132 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23133 N0.getValueType() == MVT::v2i32 &&
23134 isa<ConstantSDNode>(N0.getOperand(1))) {
23135 SDValue N00 = N0->getOperand(0);
23136 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23137 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23140 // Convert a bitcasted integer logic operation that has one bitcasted
23141 // floating-point operand and one constant operand into a floating-point
23142 // logic operation. This may create a load of the constant, but that is
23143 // cheaper than materializing the constant in an integer register and
23144 // transferring it to an SSE register or transferring the SSE operand to
23145 // integer register and back.
23147 switch (N0.getOpcode()) {
23148 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23149 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23150 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23151 default: return SDValue();
23153 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23154 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23155 isa<ConstantSDNode>(N0.getOperand(1)) &&
23156 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23157 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23158 SDValue N000 = N0.getOperand(0).getOperand(0);
23159 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23160 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23166 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23167 /// generation and convert it from being a bunch of shuffles and extracts
23168 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23169 /// storing the value and loading scalars back, while for x64 we should
23170 /// use 64-bit extracts and shifts.
23171 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23172 TargetLowering::DAGCombinerInfo &DCI) {
23173 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23176 SDValue InputVector = N->getOperand(0);
23177 SDLoc dl(InputVector);
23178 // Detect mmx to i32 conversion through a v2i32 elt extract.
23179 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23180 N->getValueType(0) == MVT::i32 &&
23181 InputVector.getValueType() == MVT::v2i32) {
23183 // The bitcast source is a direct mmx result.
23184 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23185 if (MMXSrc.getValueType() == MVT::x86mmx)
23186 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23187 N->getValueType(0),
23188 InputVector.getNode()->getOperand(0));
23190 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23191 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23192 MMXSrc.getValueType() == MVT::i64) {
23193 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23194 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23195 MMXSrcOp.getValueType() == MVT::v1i64 &&
23196 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23197 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23198 N->getValueType(0), MMXSrcOp.getOperand(0));
23202 EVT VT = N->getValueType(0);
23204 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
23205 InputVector.getOpcode() == ISD::BITCAST &&
23206 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
23207 uint64_t ExtractedElt =
23208 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23209 uint64_t InputValue =
23210 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23211 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23212 return DAG.getConstant(Res, dl, MVT::i1);
23214 // Only operate on vectors of 4 elements, where the alternative shuffling
23215 // gets to be more expensive.
23216 if (InputVector.getValueType() != MVT::v4i32)
23219 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23220 // single use which is a sign-extend or zero-extend, and all elements are
23222 SmallVector<SDNode *, 4> Uses;
23223 unsigned ExtractedElements = 0;
23224 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23225 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23226 if (UI.getUse().getResNo() != InputVector.getResNo())
23229 SDNode *Extract = *UI;
23230 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23233 if (Extract->getValueType(0) != MVT::i32)
23235 if (!Extract->hasOneUse())
23237 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23238 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23240 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23243 // Record which element was extracted.
23244 ExtractedElements |=
23245 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23247 Uses.push_back(Extract);
23250 // If not all the elements were used, this may not be worthwhile.
23251 if (ExtractedElements != 15)
23254 // Ok, we've now decided to do the transformation.
23255 // If 64-bit shifts are legal, use the extract-shift sequence,
23256 // otherwise bounce the vector off the cache.
23257 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23260 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23261 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23262 auto &DL = DAG.getDataLayout();
23263 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23264 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23265 DAG.getConstant(0, dl, VecIdxTy));
23266 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23267 DAG.getConstant(1, dl, VecIdxTy));
23269 SDValue ShAmt = DAG.getConstant(
23270 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23271 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23272 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23273 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23274 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23275 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23276 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23278 // Store the value to a temporary stack slot.
23279 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23280 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23281 MachinePointerInfo(), false, false, 0);
23283 EVT ElementType = InputVector.getValueType().getVectorElementType();
23284 unsigned EltSize = ElementType.getSizeInBits() / 8;
23286 // Replace each use (extract) with a load of the appropriate element.
23287 for (unsigned i = 0; i < 4; ++i) {
23288 uint64_t Offset = EltSize * i;
23289 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23290 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23292 SDValue ScalarAddr =
23293 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23295 // Load the scalar.
23296 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23297 ScalarAddr, MachinePointerInfo(),
23298 false, false, false, 0);
23303 // Replace the extracts
23304 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23305 UE = Uses.end(); UI != UE; ++UI) {
23306 SDNode *Extract = *UI;
23308 SDValue Idx = Extract->getOperand(1);
23309 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23310 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23313 // The replacement was made in place; don't return anything.
23318 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23319 const X86Subtarget *Subtarget) {
23321 SDValue Cond = N->getOperand(0);
23322 SDValue LHS = N->getOperand(1);
23323 SDValue RHS = N->getOperand(2);
23325 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23326 SDValue CondSrc = Cond->getOperand(0);
23327 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23328 Cond = CondSrc->getOperand(0);
23331 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23334 // A vselect where all conditions and data are constants can be optimized into
23335 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23336 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23337 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23340 unsigned MaskValue = 0;
23341 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23344 MVT VT = N->getSimpleValueType(0);
23345 unsigned NumElems = VT.getVectorNumElements();
23346 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23347 for (unsigned i = 0; i < NumElems; ++i) {
23348 // Be sure we emit undef where we can.
23349 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23350 ShuffleMask[i] = -1;
23352 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23355 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23356 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23358 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23361 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23363 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23364 TargetLowering::DAGCombinerInfo &DCI,
23365 const X86Subtarget *Subtarget) {
23367 SDValue Cond = N->getOperand(0);
23368 // Get the LHS/RHS of the select.
23369 SDValue LHS = N->getOperand(1);
23370 SDValue RHS = N->getOperand(2);
23371 EVT VT = LHS.getValueType();
23372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23374 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23375 // instructions match the semantics of the common C idiom x<y?x:y but not
23376 // x<=y?x:y, because of how they handle negative zero (which can be
23377 // ignored in unsafe-math mode).
23378 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23379 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23380 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23381 (Subtarget->hasSSE2() ||
23382 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23383 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23385 unsigned Opcode = 0;
23386 // Check for x CC y ? x : y.
23387 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23388 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23392 // Converting this to a min would handle NaNs incorrectly, and swapping
23393 // the operands would cause it to handle comparisons between positive
23394 // and negative zero incorrectly.
23395 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23396 if (!DAG.getTarget().Options.UnsafeFPMath &&
23397 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23399 std::swap(LHS, RHS);
23401 Opcode = X86ISD::FMIN;
23404 // Converting this to a min would handle comparisons between positive
23405 // and negative zero incorrectly.
23406 if (!DAG.getTarget().Options.UnsafeFPMath &&
23407 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23409 Opcode = X86ISD::FMIN;
23412 // Converting this to a min would handle both negative zeros and NaNs
23413 // incorrectly, but we can swap the operands to fix both.
23414 std::swap(LHS, RHS);
23418 Opcode = X86ISD::FMIN;
23422 // Converting this to a max would handle comparisons between positive
23423 // and negative zero incorrectly.
23424 if (!DAG.getTarget().Options.UnsafeFPMath &&
23425 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23427 Opcode = X86ISD::FMAX;
23430 // Converting this to a max would handle NaNs incorrectly, and swapping
23431 // the operands would cause it to handle comparisons between positive
23432 // and negative zero incorrectly.
23433 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23434 if (!DAG.getTarget().Options.UnsafeFPMath &&
23435 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23437 std::swap(LHS, RHS);
23439 Opcode = X86ISD::FMAX;
23442 // Converting this to a max would handle both negative zeros and NaNs
23443 // incorrectly, but we can swap the operands to fix both.
23444 std::swap(LHS, RHS);
23448 Opcode = X86ISD::FMAX;
23451 // Check for x CC y ? y : x -- a min/max with reversed arms.
23452 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23453 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23457 // Converting this to a min would handle comparisons between positive
23458 // and negative zero incorrectly, and swapping the operands would
23459 // cause it to handle NaNs incorrectly.
23460 if (!DAG.getTarget().Options.UnsafeFPMath &&
23461 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23462 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23464 std::swap(LHS, RHS);
23466 Opcode = X86ISD::FMIN;
23469 // Converting this to a min would handle NaNs incorrectly.
23470 if (!DAG.getTarget().Options.UnsafeFPMath &&
23471 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23473 Opcode = X86ISD::FMIN;
23476 // Converting this to a min would handle both negative zeros and NaNs
23477 // incorrectly, but we can swap the operands to fix both.
23478 std::swap(LHS, RHS);
23482 Opcode = X86ISD::FMIN;
23486 // Converting this to a max would handle NaNs incorrectly.
23487 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23489 Opcode = X86ISD::FMAX;
23492 // Converting this to a max would handle comparisons between positive
23493 // and negative zero incorrectly, and swapping the operands would
23494 // cause it to handle NaNs incorrectly.
23495 if (!DAG.getTarget().Options.UnsafeFPMath &&
23496 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23497 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23499 std::swap(LHS, RHS);
23501 Opcode = X86ISD::FMAX;
23504 // Converting this to a max would handle both negative zeros and NaNs
23505 // incorrectly, but we can swap the operands to fix both.
23506 std::swap(LHS, RHS);
23510 Opcode = X86ISD::FMAX;
23516 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23519 EVT CondVT = Cond.getValueType();
23520 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23521 CondVT.getVectorElementType() == MVT::i1) {
23522 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23523 // lowering on KNL. In this case we convert it to
23524 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23525 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23526 // Since SKX these selects have a proper lowering.
23527 EVT OpVT = LHS.getValueType();
23528 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23529 (OpVT.getVectorElementType() == MVT::i8 ||
23530 OpVT.getVectorElementType() == MVT::i16) &&
23531 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23532 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23533 DCI.AddToWorklist(Cond.getNode());
23534 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23537 // If this is a select between two integer constants, try to do some
23539 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23540 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23541 // Don't do this for crazy integer types.
23542 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23543 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23544 // so that TrueC (the true value) is larger than FalseC.
23545 bool NeedsCondInvert = false;
23547 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23548 // Efficiently invertible.
23549 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23550 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23551 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23552 NeedsCondInvert = true;
23553 std::swap(TrueC, FalseC);
23556 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23557 if (FalseC->getAPIntValue() == 0 &&
23558 TrueC->getAPIntValue().isPowerOf2()) {
23559 if (NeedsCondInvert) // Invert the condition if needed.
23560 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23561 DAG.getConstant(1, DL, Cond.getValueType()));
23563 // Zero extend the condition if needed.
23564 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23566 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23567 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23568 DAG.getConstant(ShAmt, DL, MVT::i8));
23571 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23572 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23573 if (NeedsCondInvert) // Invert the condition if needed.
23574 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23575 DAG.getConstant(1, DL, Cond.getValueType()));
23577 // Zero extend the condition if needed.
23578 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23579 FalseC->getValueType(0), Cond);
23580 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23581 SDValue(FalseC, 0));
23584 // Optimize cases that will turn into an LEA instruction. This requires
23585 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23586 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23587 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23588 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23590 bool isFastMultiplier = false;
23592 switch ((unsigned char)Diff) {
23594 case 1: // result = add base, cond
23595 case 2: // result = lea base( , cond*2)
23596 case 3: // result = lea base(cond, cond*2)
23597 case 4: // result = lea base( , cond*4)
23598 case 5: // result = lea base(cond, cond*4)
23599 case 8: // result = lea base( , cond*8)
23600 case 9: // result = lea base(cond, cond*8)
23601 isFastMultiplier = true;
23606 if (isFastMultiplier) {
23607 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23608 if (NeedsCondInvert) // Invert the condition if needed.
23609 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23610 DAG.getConstant(1, DL, Cond.getValueType()));
23612 // Zero extend the condition if needed.
23613 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23615 // Scale the condition by the difference.
23617 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23618 DAG.getConstant(Diff, DL,
23619 Cond.getValueType()));
23621 // Add the base if non-zero.
23622 if (FalseC->getAPIntValue() != 0)
23623 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23624 SDValue(FalseC, 0));
23631 // Canonicalize max and min:
23632 // (x > y) ? x : y -> (x >= y) ? x : y
23633 // (x < y) ? x : y -> (x <= y) ? x : y
23634 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23635 // the need for an extra compare
23636 // against zero. e.g.
23637 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23639 // testl %edi, %edi
23641 // cmovgl %edi, %eax
23645 // cmovsl %eax, %edi
23646 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23647 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23648 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23649 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23654 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23655 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23656 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23657 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23662 // Early exit check
23663 if (!TLI.isTypeLegal(VT))
23666 // Match VSELECTs into subs with unsigned saturation.
23667 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23668 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23669 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23670 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23671 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23673 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23674 // left side invert the predicate to simplify logic below.
23676 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23678 CC = ISD::getSetCCInverse(CC, true);
23679 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23683 if (Other.getNode() && Other->getNumOperands() == 2 &&
23684 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23685 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23686 SDValue CondRHS = Cond->getOperand(1);
23688 // Look for a general sub with unsigned saturation first.
23689 // x >= y ? x-y : 0 --> subus x, y
23690 // x > y ? x-y : 0 --> subus x, y
23691 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23692 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23693 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23695 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23696 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23697 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23698 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23699 // If the RHS is a constant we have to reverse the const
23700 // canonicalization.
23701 // x > C-1 ? x+-C : 0 --> subus x, C
23702 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23703 CondRHSConst->getAPIntValue() ==
23704 (-OpRHSConst->getAPIntValue() - 1))
23705 return DAG.getNode(
23706 X86ISD::SUBUS, DL, VT, OpLHS,
23707 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23709 // Another special case: If C was a sign bit, the sub has been
23710 // canonicalized into a xor.
23711 // FIXME: Would it be better to use computeKnownBits to determine
23712 // whether it's safe to decanonicalize the xor?
23713 // x s< 0 ? x^C : 0 --> subus x, C
23714 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23715 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23716 OpRHSConst->getAPIntValue().isSignBit())
23717 // Note that we have to rebuild the RHS constant here to ensure we
23718 // don't rely on particular values of undef lanes.
23719 return DAG.getNode(
23720 X86ISD::SUBUS, DL, VT, OpLHS,
23721 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23726 // Simplify vector selection if condition value type matches vselect
23728 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23729 assert(Cond.getValueType().isVector() &&
23730 "vector select expects a vector selector!");
23732 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23733 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23735 // Try invert the condition if true value is not all 1s and false value
23737 if (!TValIsAllOnes && !FValIsAllZeros &&
23738 // Check if the selector will be produced by CMPP*/PCMP*
23739 Cond.getOpcode() == ISD::SETCC &&
23740 // Check if SETCC has already been promoted
23741 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23743 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23744 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23746 if (TValIsAllZeros || FValIsAllOnes) {
23747 SDValue CC = Cond.getOperand(2);
23748 ISD::CondCode NewCC =
23749 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23750 Cond.getOperand(0).getValueType().isInteger());
23751 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23752 std::swap(LHS, RHS);
23753 TValIsAllOnes = FValIsAllOnes;
23754 FValIsAllZeros = TValIsAllZeros;
23758 if (TValIsAllOnes || FValIsAllZeros) {
23761 if (TValIsAllOnes && FValIsAllZeros)
23763 else if (TValIsAllOnes)
23765 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23766 else if (FValIsAllZeros)
23767 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23768 DAG.getBitcast(CondVT, LHS));
23770 return DAG.getBitcast(VT, Ret);
23774 // We should generate an X86ISD::BLENDI from a vselect if its argument
23775 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23776 // constants. This specific pattern gets generated when we split a
23777 // selector for a 512 bit vector in a machine without AVX512 (but with
23778 // 256-bit vectors), during legalization:
23780 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23782 // Iff we find this pattern and the build_vectors are built from
23783 // constants, we translate the vselect into a shuffle_vector that we
23784 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23785 if ((N->getOpcode() == ISD::VSELECT ||
23786 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23787 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23788 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23789 if (Shuffle.getNode())
23793 // If this is a *dynamic* select (non-constant condition) and we can match
23794 // this node with one of the variable blend instructions, restructure the
23795 // condition so that the blends can use the high bit of each element and use
23796 // SimplifyDemandedBits to simplify the condition operand.
23797 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23798 !DCI.isBeforeLegalize() &&
23799 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23800 unsigned BitWidth = Cond.getValueType().getScalarSizeInBits();
23802 // Don't optimize vector selects that map to mask-registers.
23806 // We can only handle the cases where VSELECT is directly legal on the
23807 // subtarget. We custom lower VSELECT nodes with constant conditions and
23808 // this makes it hard to see whether a dynamic VSELECT will correctly
23809 // lower, so we both check the operation's status and explicitly handle the
23810 // cases where a *dynamic* blend will fail even though a constant-condition
23811 // blend could be custom lowered.
23812 // FIXME: We should find a better way to handle this class of problems.
23813 // Potentially, we should combine constant-condition vselect nodes
23814 // pre-legalization into shuffles and not mark as many types as custom
23816 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23818 // FIXME: We don't support i16-element blends currently. We could and
23819 // should support them by making *all* the bits in the condition be set
23820 // rather than just the high bit and using an i8-element blend.
23821 if (VT.getScalarType() == MVT::i16)
23823 // Dynamic blending was only available from SSE4.1 onward.
23824 if (VT.is128BitVector() && !Subtarget->hasSSE41())
23826 // Byte blends are only available in AVX2
23827 if (VT == MVT::v32i8 && !Subtarget->hasAVX2())
23830 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23831 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23833 APInt KnownZero, KnownOne;
23834 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23835 DCI.isBeforeLegalizeOps());
23836 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23837 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23839 // If we changed the computation somewhere in the DAG, this change
23840 // will affect all users of Cond.
23841 // Make sure it is fine and update all the nodes so that we do not
23842 // use the generic VSELECT anymore. Otherwise, we may perform
23843 // wrong optimizations as we messed up with the actual expectation
23844 // for the vector boolean values.
23845 if (Cond != TLO.Old) {
23846 // Check all uses of that condition operand to check whether it will be
23847 // consumed by non-BLEND instructions, which may depend on all bits are
23849 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23851 if (I->getOpcode() != ISD::VSELECT)
23852 // TODO: Add other opcodes eventually lowered into BLEND.
23855 // Update all the users of the condition, before committing the change,
23856 // so that the VSELECT optimizations that expect the correct vector
23857 // boolean value will not be triggered.
23858 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23860 DAG.ReplaceAllUsesOfValueWith(
23862 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23863 Cond, I->getOperand(1), I->getOperand(2)));
23864 DCI.CommitTargetLoweringOpt(TLO);
23867 // At this point, only Cond is changed. Change the condition
23868 // just for N to keep the opportunity to optimize all other
23869 // users their own way.
23870 DAG.ReplaceAllUsesOfValueWith(
23872 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23873 TLO.New, N->getOperand(1), N->getOperand(2)));
23881 // Check whether a boolean test is testing a boolean value generated by
23882 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23885 // Simplify the following patterns:
23886 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23887 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23888 // to (Op EFLAGS Cond)
23890 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23891 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23892 // to (Op EFLAGS !Cond)
23894 // where Op could be BRCOND or CMOV.
23896 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23897 // Quit if not CMP and SUB with its value result used.
23898 if (Cmp.getOpcode() != X86ISD::CMP &&
23899 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23902 // Quit if not used as a boolean value.
23903 if (CC != X86::COND_E && CC != X86::COND_NE)
23906 // Check CMP operands. One of them should be 0 or 1 and the other should be
23907 // an SetCC or extended from it.
23908 SDValue Op1 = Cmp.getOperand(0);
23909 SDValue Op2 = Cmp.getOperand(1);
23912 const ConstantSDNode* C = nullptr;
23913 bool needOppositeCond = (CC == X86::COND_E);
23914 bool checkAgainstTrue = false; // Is it a comparison against 1?
23916 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23918 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23920 else // Quit if all operands are not constants.
23923 if (C->getZExtValue() == 1) {
23924 needOppositeCond = !needOppositeCond;
23925 checkAgainstTrue = true;
23926 } else if (C->getZExtValue() != 0)
23927 // Quit if the constant is neither 0 or 1.
23930 bool truncatedToBoolWithAnd = false;
23931 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23932 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23933 SetCC.getOpcode() == ISD::TRUNCATE ||
23934 SetCC.getOpcode() == ISD::AND) {
23935 if (SetCC.getOpcode() == ISD::AND) {
23937 ConstantSDNode *CS;
23938 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23939 CS->getZExtValue() == 1)
23941 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23942 CS->getZExtValue() == 1)
23946 SetCC = SetCC.getOperand(OpIdx);
23947 truncatedToBoolWithAnd = true;
23949 SetCC = SetCC.getOperand(0);
23952 switch (SetCC.getOpcode()) {
23953 case X86ISD::SETCC_CARRY:
23954 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23955 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23956 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23957 // truncated to i1 using 'and'.
23958 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23960 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23961 "Invalid use of SETCC_CARRY!");
23963 case X86ISD::SETCC:
23964 // Set the condition code or opposite one if necessary.
23965 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23966 if (needOppositeCond)
23967 CC = X86::GetOppositeBranchCondition(CC);
23968 return SetCC.getOperand(1);
23969 case X86ISD::CMOV: {
23970 // Check whether false/true value has canonical one, i.e. 0 or 1.
23971 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23972 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23973 // Quit if true value is not a constant.
23976 // Quit if false value is not a constant.
23978 SDValue Op = SetCC.getOperand(0);
23979 // Skip 'zext' or 'trunc' node.
23980 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23981 Op.getOpcode() == ISD::TRUNCATE)
23982 Op = Op.getOperand(0);
23983 // A special case for rdrand/rdseed, where 0 is set if false cond is
23985 if ((Op.getOpcode() != X86ISD::RDRAND &&
23986 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23989 // Quit if false value is not the constant 0 or 1.
23990 bool FValIsFalse = true;
23991 if (FVal && FVal->getZExtValue() != 0) {
23992 if (FVal->getZExtValue() != 1)
23994 // If FVal is 1, opposite cond is needed.
23995 needOppositeCond = !needOppositeCond;
23996 FValIsFalse = false;
23998 // Quit if TVal is not the constant opposite of FVal.
23999 if (FValIsFalse && TVal->getZExtValue() != 1)
24001 if (!FValIsFalse && TVal->getZExtValue() != 0)
24003 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24004 if (needOppositeCond)
24005 CC = X86::GetOppositeBranchCondition(CC);
24006 return SetCC.getOperand(3);
24013 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24015 /// (X86or (X86setcc) (X86setcc))
24016 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24017 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24018 X86::CondCode &CC1, SDValue &Flags,
24020 if (Cond->getOpcode() == X86ISD::CMP) {
24021 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24022 if (!CondOp1C || !CondOp1C->isNullValue())
24025 Cond = Cond->getOperand(0);
24030 SDValue SetCC0, SetCC1;
24031 switch (Cond->getOpcode()) {
24032 default: return false;
24039 SetCC0 = Cond->getOperand(0);
24040 SetCC1 = Cond->getOperand(1);
24044 // Make sure we have SETCC nodes, using the same flags value.
24045 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24046 SetCC1.getOpcode() != X86ISD::SETCC ||
24047 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24050 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24051 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24052 Flags = SetCC0->getOperand(1);
24056 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24057 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24058 TargetLowering::DAGCombinerInfo &DCI,
24059 const X86Subtarget *Subtarget) {
24062 // If the flag operand isn't dead, don't touch this CMOV.
24063 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24066 SDValue FalseOp = N->getOperand(0);
24067 SDValue TrueOp = N->getOperand(1);
24068 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24069 SDValue Cond = N->getOperand(3);
24071 if (CC == X86::COND_E || CC == X86::COND_NE) {
24072 switch (Cond.getOpcode()) {
24076 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24077 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24078 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24084 Flags = checkBoolTestSetCCCombine(Cond, CC);
24085 if (Flags.getNode() &&
24086 // Extra check as FCMOV only supports a subset of X86 cond.
24087 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24088 SDValue Ops[] = { FalseOp, TrueOp,
24089 DAG.getConstant(CC, DL, MVT::i8), Flags };
24090 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24093 // If this is a select between two integer constants, try to do some
24094 // optimizations. Note that the operands are ordered the opposite of SELECT
24096 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24097 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24098 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24099 // larger than FalseC (the false value).
24100 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24101 CC = X86::GetOppositeBranchCondition(CC);
24102 std::swap(TrueC, FalseC);
24103 std::swap(TrueOp, FalseOp);
24106 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24107 // This is efficient for any integer data type (including i8/i16) and
24109 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24110 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24111 DAG.getConstant(CC, DL, MVT::i8), Cond);
24113 // Zero extend the condition if needed.
24114 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24116 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24117 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24118 DAG.getConstant(ShAmt, DL, MVT::i8));
24119 if (N->getNumValues() == 2) // Dead flag value?
24120 return DCI.CombineTo(N, Cond, SDValue());
24124 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24125 // for any integer data type, including i8/i16.
24126 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24127 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24128 DAG.getConstant(CC, DL, MVT::i8), Cond);
24130 // Zero extend the condition if needed.
24131 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24132 FalseC->getValueType(0), Cond);
24133 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24134 SDValue(FalseC, 0));
24136 if (N->getNumValues() == 2) // Dead flag value?
24137 return DCI.CombineTo(N, Cond, SDValue());
24141 // Optimize cases that will turn into an LEA instruction. This requires
24142 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24143 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24144 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24145 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24147 bool isFastMultiplier = false;
24149 switch ((unsigned char)Diff) {
24151 case 1: // result = add base, cond
24152 case 2: // result = lea base( , cond*2)
24153 case 3: // result = lea base(cond, cond*2)
24154 case 4: // result = lea base( , cond*4)
24155 case 5: // result = lea base(cond, cond*4)
24156 case 8: // result = lea base( , cond*8)
24157 case 9: // result = lea base(cond, cond*8)
24158 isFastMultiplier = true;
24163 if (isFastMultiplier) {
24164 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24165 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24166 DAG.getConstant(CC, DL, MVT::i8), Cond);
24167 // Zero extend the condition if needed.
24168 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24170 // Scale the condition by the difference.
24172 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24173 DAG.getConstant(Diff, DL, Cond.getValueType()));
24175 // Add the base if non-zero.
24176 if (FalseC->getAPIntValue() != 0)
24177 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24178 SDValue(FalseC, 0));
24179 if (N->getNumValues() == 2) // Dead flag value?
24180 return DCI.CombineTo(N, Cond, SDValue());
24187 // Handle these cases:
24188 // (select (x != c), e, c) -> select (x != c), e, x),
24189 // (select (x == c), c, e) -> select (x == c), x, e)
24190 // where the c is an integer constant, and the "select" is the combination
24191 // of CMOV and CMP.
24193 // The rationale for this change is that the conditional-move from a constant
24194 // needs two instructions, however, conditional-move from a register needs
24195 // only one instruction.
24197 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24198 // some instruction-combining opportunities. This opt needs to be
24199 // postponed as late as possible.
24201 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24202 // the DCI.xxxx conditions are provided to postpone the optimization as
24203 // late as possible.
24205 ConstantSDNode *CmpAgainst = nullptr;
24206 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24207 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24208 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24210 if (CC == X86::COND_NE &&
24211 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24212 CC = X86::GetOppositeBranchCondition(CC);
24213 std::swap(TrueOp, FalseOp);
24216 if (CC == X86::COND_E &&
24217 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24218 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24219 DAG.getConstant(CC, DL, MVT::i8), Cond };
24220 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24225 // Fold and/or of setcc's to double CMOV:
24226 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24227 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24229 // This combine lets us generate:
24230 // cmovcc1 (jcc1 if we don't have CMOV)
24236 // cmovne (jne if we don't have CMOV)
24237 // When we can't use the CMOV instruction, it might increase branch
24239 // When we can use CMOV, or when there is no mispredict, this improves
24240 // throughput and reduces register pressure.
24242 if (CC == X86::COND_NE) {
24244 X86::CondCode CC0, CC1;
24246 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24248 std::swap(FalseOp, TrueOp);
24249 CC0 = X86::GetOppositeBranchCondition(CC0);
24250 CC1 = X86::GetOppositeBranchCondition(CC1);
24253 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24255 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24256 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24257 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24258 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24266 /// PerformMulCombine - Optimize a single multiply with constant into two
24267 /// in order to implement it with two cheaper instructions, e.g.
24268 /// LEA + SHL, LEA + LEA.
24269 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24270 TargetLowering::DAGCombinerInfo &DCI) {
24271 // An imul is usually smaller than the alternative sequence.
24272 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24275 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24278 EVT VT = N->getValueType(0);
24279 if (VT != MVT::i64 && VT != MVT::i32)
24282 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24285 uint64_t MulAmt = C->getZExtValue();
24286 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24289 uint64_t MulAmt1 = 0;
24290 uint64_t MulAmt2 = 0;
24291 if ((MulAmt % 9) == 0) {
24293 MulAmt2 = MulAmt / 9;
24294 } else if ((MulAmt % 5) == 0) {
24296 MulAmt2 = MulAmt / 5;
24297 } else if ((MulAmt % 3) == 0) {
24299 MulAmt2 = MulAmt / 3;
24302 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24305 if (isPowerOf2_64(MulAmt2) &&
24306 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24307 // If second multiplifer is pow2, issue it first. We want the multiply by
24308 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24310 std::swap(MulAmt1, MulAmt2);
24313 if (isPowerOf2_64(MulAmt1))
24314 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24315 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24317 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24318 DAG.getConstant(MulAmt1, DL, VT));
24320 if (isPowerOf2_64(MulAmt2))
24321 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24322 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24324 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24325 DAG.getConstant(MulAmt2, DL, VT));
24327 // Do not add new nodes to DAG combiner worklist.
24328 DCI.CombineTo(N, NewMul, false);
24333 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24334 SDValue N0 = N->getOperand(0);
24335 SDValue N1 = N->getOperand(1);
24336 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24337 EVT VT = N0.getValueType();
24339 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24340 // since the result of setcc_c is all zero's or all ones.
24341 if (VT.isInteger() && !VT.isVector() &&
24342 N1C && N0.getOpcode() == ISD::AND &&
24343 N0.getOperand(1).getOpcode() == ISD::Constant) {
24344 SDValue N00 = N0.getOperand(0);
24345 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24346 APInt ShAmt = N1C->getAPIntValue();
24347 Mask = Mask.shl(ShAmt);
24348 bool MaskOK = false;
24349 // We can handle cases concerning bit-widening nodes containing setcc_c if
24350 // we carefully interrogate the mask to make sure we are semantics
24352 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24353 // of the underlying setcc_c operation if the setcc_c was zero extended.
24354 // Consider the following example:
24355 // zext(setcc_c) -> i32 0x0000FFFF
24356 // c1 -> i32 0x0000FFFF
24357 // c2 -> i32 0x00000001
24358 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24359 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24360 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24362 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24363 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24365 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24366 N00.getOpcode() == ISD::ANY_EXTEND) &&
24367 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24368 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24370 if (MaskOK && Mask != 0) {
24372 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24376 // Hardware support for vector shifts is sparse which makes us scalarize the
24377 // vector operations in many cases. Also, on sandybridge ADD is faster than
24379 // (shl V, 1) -> add V,V
24380 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24381 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24382 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24383 // We shift all of the values by one. In many cases we do not have
24384 // hardware support for this operation. This is better expressed as an ADD
24386 if (N1SplatC->getAPIntValue() == 1)
24387 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24393 /// \brief Returns a vector of 0s if the node in input is a vector logical
24394 /// shift by a constant amount which is known to be bigger than or equal
24395 /// to the vector element size in bits.
24396 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24397 const X86Subtarget *Subtarget) {
24398 EVT VT = N->getValueType(0);
24400 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24401 (!Subtarget->hasInt256() ||
24402 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24405 SDValue Amt = N->getOperand(1);
24407 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24408 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24409 APInt ShiftAmt = AmtSplat->getAPIntValue();
24410 unsigned MaxAmount =
24411 VT.getSimpleVT().getVectorElementType().getSizeInBits();
24413 // SSE2/AVX2 logical shifts always return a vector of 0s
24414 // if the shift amount is bigger than or equal to
24415 // the element size. The constant shift amount will be
24416 // encoded as a 8-bit immediate.
24417 if (ShiftAmt.trunc(8).uge(MaxAmount))
24418 return getZeroVector(VT, Subtarget, DAG, DL);
24424 /// PerformShiftCombine - Combine shifts.
24425 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24426 TargetLowering::DAGCombinerInfo &DCI,
24427 const X86Subtarget *Subtarget) {
24428 if (N->getOpcode() == ISD::SHL)
24429 if (SDValue V = PerformSHLCombine(N, DAG))
24432 // Try to fold this logical shift into a zero vector.
24433 if (N->getOpcode() != ISD::SRA)
24434 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24440 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24441 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24442 // and friends. Likewise for OR -> CMPNEQSS.
24443 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24444 TargetLowering::DAGCombinerInfo &DCI,
24445 const X86Subtarget *Subtarget) {
24448 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24449 // we're requiring SSE2 for both.
24450 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24451 SDValue N0 = N->getOperand(0);
24452 SDValue N1 = N->getOperand(1);
24453 SDValue CMP0 = N0->getOperand(1);
24454 SDValue CMP1 = N1->getOperand(1);
24457 // The SETCCs should both refer to the same CMP.
24458 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24461 SDValue CMP00 = CMP0->getOperand(0);
24462 SDValue CMP01 = CMP0->getOperand(1);
24463 EVT VT = CMP00.getValueType();
24465 if (VT == MVT::f32 || VT == MVT::f64) {
24466 bool ExpectingFlags = false;
24467 // Check for any users that want flags:
24468 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24469 !ExpectingFlags && UI != UE; ++UI)
24470 switch (UI->getOpcode()) {
24475 ExpectingFlags = true;
24477 case ISD::CopyToReg:
24478 case ISD::SIGN_EXTEND:
24479 case ISD::ZERO_EXTEND:
24480 case ISD::ANY_EXTEND:
24484 if (!ExpectingFlags) {
24485 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24486 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24488 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24489 X86::CondCode tmp = cc0;
24494 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24495 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24496 // FIXME: need symbolic constants for these magic numbers.
24497 // See X86ATTInstPrinter.cpp:printSSECC().
24498 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24499 if (Subtarget->hasAVX512()) {
24500 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24502 DAG.getConstant(x86cc, DL, MVT::i8));
24503 if (N->getValueType(0) != MVT::i1)
24504 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24508 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24509 CMP00.getValueType(), CMP00, CMP01,
24510 DAG.getConstant(x86cc, DL,
24513 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24514 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24516 if (is64BitFP && !Subtarget->is64Bit()) {
24517 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24518 // 64-bit integer, since that's not a legal type. Since
24519 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24520 // bits, but can do this little dance to extract the lowest 32 bits
24521 // and work with those going forward.
24522 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24524 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24525 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24526 Vector32, DAG.getIntPtrConstant(0, DL));
24530 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24531 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24532 DAG.getConstant(1, DL, IntVT));
24533 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24535 return OneBitOfTruth;
24543 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24544 /// so it can be folded inside ANDNP.
24545 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24546 EVT VT = N->getValueType(0);
24548 // Match direct AllOnes for 128 and 256-bit vectors
24549 if (ISD::isBuildVectorAllOnes(N))
24552 // Look through a bit convert.
24553 if (N->getOpcode() == ISD::BITCAST)
24554 N = N->getOperand(0).getNode();
24556 // Sometimes the operand may come from a insert_subvector building a 256-bit
24558 if (VT.is256BitVector() &&
24559 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24560 SDValue V1 = N->getOperand(0);
24561 SDValue V2 = N->getOperand(1);
24563 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24564 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24565 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24566 ISD::isBuildVectorAllOnes(V2.getNode()))
24573 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24574 // register. In most cases we actually compare or select YMM-sized registers
24575 // and mixing the two types creates horrible code. This method optimizes
24576 // some of the transition sequences.
24577 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24578 TargetLowering::DAGCombinerInfo &DCI,
24579 const X86Subtarget *Subtarget) {
24580 EVT VT = N->getValueType(0);
24581 if (!VT.is256BitVector())
24584 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24585 N->getOpcode() == ISD::ZERO_EXTEND ||
24586 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24588 SDValue Narrow = N->getOperand(0);
24589 EVT NarrowVT = Narrow->getValueType(0);
24590 if (!NarrowVT.is128BitVector())
24593 if (Narrow->getOpcode() != ISD::XOR &&
24594 Narrow->getOpcode() != ISD::AND &&
24595 Narrow->getOpcode() != ISD::OR)
24598 SDValue N0 = Narrow->getOperand(0);
24599 SDValue N1 = Narrow->getOperand(1);
24602 // The Left side has to be a trunc.
24603 if (N0.getOpcode() != ISD::TRUNCATE)
24606 // The type of the truncated inputs.
24607 EVT WideVT = N0->getOperand(0)->getValueType(0);
24611 // The right side has to be a 'trunc' or a constant vector.
24612 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24613 ConstantSDNode *RHSConstSplat = nullptr;
24614 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24615 RHSConstSplat = RHSBV->getConstantSplatNode();
24616 if (!RHSTrunc && !RHSConstSplat)
24619 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24621 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24624 // Set N0 and N1 to hold the inputs to the new wide operation.
24625 N0 = N0->getOperand(0);
24626 if (RHSConstSplat) {
24627 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24628 SDValue(RHSConstSplat, 0));
24629 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24630 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24631 } else if (RHSTrunc) {
24632 N1 = N1->getOperand(0);
24635 // Generate the wide operation.
24636 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24637 unsigned Opcode = N->getOpcode();
24639 case ISD::ANY_EXTEND:
24641 case ISD::ZERO_EXTEND: {
24642 unsigned InBits = NarrowVT.getScalarSizeInBits();
24643 APInt Mask = APInt::getAllOnesValue(InBits);
24644 Mask = Mask.zext(VT.getScalarSizeInBits());
24645 return DAG.getNode(ISD::AND, DL, VT,
24646 Op, DAG.getConstant(Mask, DL, VT));
24648 case ISD::SIGN_EXTEND:
24649 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24650 Op, DAG.getValueType(NarrowVT));
24652 llvm_unreachable("Unexpected opcode");
24656 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24657 TargetLowering::DAGCombinerInfo &DCI,
24658 const X86Subtarget *Subtarget) {
24659 SDValue N0 = N->getOperand(0);
24660 SDValue N1 = N->getOperand(1);
24663 // A vector zext_in_reg may be represented as a shuffle,
24664 // feeding into a bitcast (this represents anyext) feeding into
24665 // an and with a mask.
24666 // We'd like to try to combine that into a shuffle with zero
24667 // plus a bitcast, removing the and.
24668 if (N0.getOpcode() != ISD::BITCAST ||
24669 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24672 // The other side of the AND should be a splat of 2^C, where C
24673 // is the number of bits in the source type.
24674 if (N1.getOpcode() == ISD::BITCAST)
24675 N1 = N1.getOperand(0);
24676 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24678 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24680 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24681 EVT SrcType = Shuffle->getValueType(0);
24683 // We expect a single-source shuffle
24684 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24687 unsigned SrcSize = SrcType.getScalarSizeInBits();
24689 APInt SplatValue, SplatUndef;
24690 unsigned SplatBitSize;
24692 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24693 SplatBitSize, HasAnyUndefs))
24696 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24697 // Make sure the splat matches the mask we expect
24698 if (SplatBitSize > ResSize ||
24699 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24702 // Make sure the input and output size make sense
24703 if (SrcSize >= ResSize || ResSize % SrcSize)
24706 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24707 // The number of u's between each two values depends on the ratio between
24708 // the source and dest type.
24709 unsigned ZextRatio = ResSize / SrcSize;
24710 bool IsZext = true;
24711 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24712 if (i % ZextRatio) {
24713 if (Shuffle->getMaskElt(i) > 0) {
24719 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24720 // Expected element number
24730 // Ok, perform the transformation - replace the shuffle with
24731 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24732 // (instead of undef) where the k elements come from the zero vector.
24733 SmallVector<int, 8> Mask;
24734 unsigned NumElems = SrcType.getVectorNumElements();
24735 for (unsigned i = 0; i < NumElems; ++i)
24737 Mask.push_back(NumElems);
24739 Mask.push_back(i / ZextRatio);
24741 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24742 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24743 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24746 /// If both input operands of a logic op are being cast from floating point
24747 /// types, try to convert this into a floating point logic node to avoid
24748 /// unnecessary moves from SSE to integer registers.
24749 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24750 const X86Subtarget *Subtarget) {
24751 unsigned FPOpcode = ISD::DELETED_NODE;
24752 if (N->getOpcode() == ISD::AND)
24753 FPOpcode = X86ISD::FAND;
24754 else if (N->getOpcode() == ISD::OR)
24755 FPOpcode = X86ISD::FOR;
24756 else if (N->getOpcode() == ISD::XOR)
24757 FPOpcode = X86ISD::FXOR;
24759 assert(FPOpcode != ISD::DELETED_NODE &&
24760 "Unexpected input node for FP logic conversion");
24762 EVT VT = N->getValueType(0);
24763 SDValue N0 = N->getOperand(0);
24764 SDValue N1 = N->getOperand(1);
24766 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24767 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24768 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24769 SDValue N00 = N0.getOperand(0);
24770 SDValue N10 = N1.getOperand(0);
24771 EVT N00Type = N00.getValueType();
24772 EVT N10Type = N10.getValueType();
24773 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24774 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24775 return DAG.getBitcast(VT, FPLogic);
24781 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24782 TargetLowering::DAGCombinerInfo &DCI,
24783 const X86Subtarget *Subtarget) {
24784 if (DCI.isBeforeLegalizeOps())
24787 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24790 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24793 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24796 EVT VT = N->getValueType(0);
24797 SDValue N0 = N->getOperand(0);
24798 SDValue N1 = N->getOperand(1);
24801 // Create BEXTR instructions
24802 // BEXTR is ((X >> imm) & (2**size-1))
24803 if (VT == MVT::i32 || VT == MVT::i64) {
24804 // Check for BEXTR.
24805 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24806 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24807 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24808 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24809 if (MaskNode && ShiftNode) {
24810 uint64_t Mask = MaskNode->getZExtValue();
24811 uint64_t Shift = ShiftNode->getZExtValue();
24812 if (isMask_64(Mask)) {
24813 uint64_t MaskSize = countPopulation(Mask);
24814 if (Shift + MaskSize <= VT.getSizeInBits())
24815 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24816 DAG.getConstant(Shift | (MaskSize << 8), DL,
24825 // Want to form ANDNP nodes:
24826 // 1) In the hopes of then easily combining them with OR and AND nodes
24827 // to form PBLEND/PSIGN.
24828 // 2) To match ANDN packed intrinsics
24829 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24832 // Check LHS for vnot
24833 if (N0.getOpcode() == ISD::XOR &&
24834 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24835 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24836 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24838 // Check RHS for vnot
24839 if (N1.getOpcode() == ISD::XOR &&
24840 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24841 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24842 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24847 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24848 TargetLowering::DAGCombinerInfo &DCI,
24849 const X86Subtarget *Subtarget) {
24850 if (DCI.isBeforeLegalizeOps())
24853 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24856 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24859 SDValue N0 = N->getOperand(0);
24860 SDValue N1 = N->getOperand(1);
24861 EVT VT = N->getValueType(0);
24863 // look for psign/blend
24864 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24865 if (!Subtarget->hasSSSE3() ||
24866 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24869 // Canonicalize pandn to RHS
24870 if (N0.getOpcode() == X86ISD::ANDNP)
24872 // or (and (m, y), (pandn m, x))
24873 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24874 SDValue Mask = N1.getOperand(0);
24875 SDValue X = N1.getOperand(1);
24877 if (N0.getOperand(0) == Mask)
24878 Y = N0.getOperand(1);
24879 if (N0.getOperand(1) == Mask)
24880 Y = N0.getOperand(0);
24882 // Check to see if the mask appeared in both the AND and ANDNP and
24886 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24887 // Look through mask bitcast.
24888 if (Mask.getOpcode() == ISD::BITCAST)
24889 Mask = Mask.getOperand(0);
24890 if (X.getOpcode() == ISD::BITCAST)
24891 X = X.getOperand(0);
24892 if (Y.getOpcode() == ISD::BITCAST)
24893 Y = Y.getOperand(0);
24895 EVT MaskVT = Mask.getValueType();
24897 // Validate that the Mask operand is a vector sra node.
24898 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24899 // there is no psrai.b
24900 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24901 unsigned SraAmt = ~0;
24902 if (Mask.getOpcode() == ISD::SRA) {
24903 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24904 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24905 SraAmt = AmtConst->getZExtValue();
24906 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24907 SDValue SraC = Mask.getOperand(1);
24908 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24910 if ((SraAmt + 1) != EltBits)
24915 // Now we know we at least have a plendvb with the mask val. See if
24916 // we can form a psignb/w/d.
24917 // psign = x.type == y.type == mask.type && y = sub(0, x);
24918 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24919 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24920 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24921 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24922 "Unsupported VT for PSIGN");
24923 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24924 return DAG.getBitcast(VT, Mask);
24926 // PBLENDVB only available on SSE 4.1
24927 if (!Subtarget->hasSSE41())
24930 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24932 X = DAG.getBitcast(BlendVT, X);
24933 Y = DAG.getBitcast(BlendVT, Y);
24934 Mask = DAG.getBitcast(BlendVT, Mask);
24935 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24936 return DAG.getBitcast(VT, Mask);
24940 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24943 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24944 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24946 // SHLD/SHRD instructions have lower register pressure, but on some
24947 // platforms they have higher latency than the equivalent
24948 // series of shifts/or that would otherwise be generated.
24949 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24950 // have higher latencies and we are not optimizing for size.
24951 if (!OptForSize && Subtarget->isSHLDSlow())
24954 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24956 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24958 if (!N0.hasOneUse() || !N1.hasOneUse())
24961 SDValue ShAmt0 = N0.getOperand(1);
24962 if (ShAmt0.getValueType() != MVT::i8)
24964 SDValue ShAmt1 = N1.getOperand(1);
24965 if (ShAmt1.getValueType() != MVT::i8)
24967 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24968 ShAmt0 = ShAmt0.getOperand(0);
24969 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24970 ShAmt1 = ShAmt1.getOperand(0);
24973 unsigned Opc = X86ISD::SHLD;
24974 SDValue Op0 = N0.getOperand(0);
24975 SDValue Op1 = N1.getOperand(0);
24976 if (ShAmt0.getOpcode() == ISD::SUB) {
24977 Opc = X86ISD::SHRD;
24978 std::swap(Op0, Op1);
24979 std::swap(ShAmt0, ShAmt1);
24982 unsigned Bits = VT.getSizeInBits();
24983 if (ShAmt1.getOpcode() == ISD::SUB) {
24984 SDValue Sum = ShAmt1.getOperand(0);
24985 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24986 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24987 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24988 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24989 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24990 return DAG.getNode(Opc, DL, VT,
24992 DAG.getNode(ISD::TRUNCATE, DL,
24995 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24996 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24998 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24999 return DAG.getNode(Opc, DL, VT,
25000 N0.getOperand(0), N1.getOperand(0),
25001 DAG.getNode(ISD::TRUNCATE, DL,
25008 // Generate NEG and CMOV for integer abs.
25009 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25010 EVT VT = N->getValueType(0);
25012 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25013 // 8-bit integer abs to NEG and CMOV.
25014 if (VT.isInteger() && VT.getSizeInBits() == 8)
25017 SDValue N0 = N->getOperand(0);
25018 SDValue N1 = N->getOperand(1);
25021 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25022 // and change it to SUB and CMOV.
25023 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25024 N0.getOpcode() == ISD::ADD &&
25025 N0.getOperand(1) == N1 &&
25026 N1.getOpcode() == ISD::SRA &&
25027 N1.getOperand(0) == N0.getOperand(0))
25028 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25029 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25030 // Generate SUB & CMOV.
25031 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25032 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25034 SDValue Ops[] = { N0.getOperand(0), Neg,
25035 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25036 SDValue(Neg.getNode(), 1) };
25037 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25042 // Try to turn tests against the signbit in the form of:
25043 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25046 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25047 // This is only worth doing if the output type is i8.
25048 if (N->getValueType(0) != MVT::i8)
25051 SDValue N0 = N->getOperand(0);
25052 SDValue N1 = N->getOperand(1);
25054 // We should be performing an xor against a truncated shift.
25055 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25058 // Make sure we are performing an xor against one.
25059 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25062 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25063 SDValue Shift = N0.getOperand(0);
25064 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25067 // Make sure we are truncating from one of i16, i32 or i64.
25068 EVT ShiftTy = Shift.getValueType();
25069 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25072 // Make sure the shift amount extracts the sign bit.
25073 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25074 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25077 // Create a greater-than comparison against -1.
25078 // N.B. Using SETGE against 0 works but we want a canonical looking
25079 // comparison, using SETGT matches up with what TranslateX86CC.
25081 SDValue ShiftOp = Shift.getOperand(0);
25082 EVT ShiftOpTy = ShiftOp.getValueType();
25083 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25084 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25088 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25089 TargetLowering::DAGCombinerInfo &DCI,
25090 const X86Subtarget *Subtarget) {
25091 if (DCI.isBeforeLegalizeOps())
25094 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25097 if (Subtarget->hasCMov())
25098 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25101 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25107 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25108 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25109 TargetLowering::DAGCombinerInfo &DCI,
25110 const X86Subtarget *Subtarget) {
25111 LoadSDNode *Ld = cast<LoadSDNode>(N);
25112 EVT RegVT = Ld->getValueType(0);
25113 EVT MemVT = Ld->getMemoryVT();
25115 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25117 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25118 // into two 16-byte operations.
25119 ISD::LoadExtType Ext = Ld->getExtensionType();
25121 unsigned AddressSpace = Ld->getAddressSpace();
25122 unsigned Alignment = Ld->getAlignment();
25123 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25124 Ext == ISD::NON_EXTLOAD &&
25125 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25126 AddressSpace, Alignment, &Fast) && !Fast) {
25127 unsigned NumElems = RegVT.getVectorNumElements();
25131 SDValue Ptr = Ld->getBasePtr();
25132 SDValue Increment =
25133 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25135 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25137 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25138 Ld->getPointerInfo(), Ld->isVolatile(),
25139 Ld->isNonTemporal(), Ld->isInvariant(),
25141 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25142 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25143 Ld->getPointerInfo(), Ld->isVolatile(),
25144 Ld->isNonTemporal(), Ld->isInvariant(),
25145 std::min(16U, Alignment));
25146 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25148 Load2.getValue(1));
25150 SDValue NewVec = DAG.getUNDEF(RegVT);
25151 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25152 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25153 return DCI.CombineTo(N, NewVec, TF, true);
25159 /// PerformMLOADCombine - Resolve extending loads
25160 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25161 TargetLowering::DAGCombinerInfo &DCI,
25162 const X86Subtarget *Subtarget) {
25163 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25164 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25167 EVT VT = Mld->getValueType(0);
25168 unsigned NumElems = VT.getVectorNumElements();
25169 EVT LdVT = Mld->getMemoryVT();
25172 assert(LdVT != VT && "Cannot extend to the same type");
25173 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25174 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25175 // From, To sizes and ElemCount must be pow of two
25176 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25177 "Unexpected size for extending masked load");
25179 unsigned SizeRatio = ToSz / FromSz;
25180 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25182 // Create a type on which we perform the shuffle
25183 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25184 LdVT.getScalarType(), NumElems*SizeRatio);
25185 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25187 // Convert Src0 value
25188 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25189 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25190 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25191 for (unsigned i = 0; i != NumElems; ++i)
25192 ShuffleVec[i] = i * SizeRatio;
25194 // Can't shuffle using an illegal type.
25195 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25196 "WideVecVT should be legal");
25197 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25198 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25200 // Prepare the new mask
25202 SDValue Mask = Mld->getMask();
25203 if (Mask.getValueType() == VT) {
25204 // Mask and original value have the same type
25205 NewMask = DAG.getBitcast(WideVecVT, Mask);
25206 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25207 for (unsigned i = 0; i != NumElems; ++i)
25208 ShuffleVec[i] = i * SizeRatio;
25209 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25210 ShuffleVec[i] = NumElems*SizeRatio;
25211 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25212 DAG.getConstant(0, dl, WideVecVT),
25216 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25217 unsigned WidenNumElts = NumElems*SizeRatio;
25218 unsigned MaskNumElts = VT.getVectorNumElements();
25219 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25222 unsigned NumConcat = WidenNumElts / MaskNumElts;
25223 SmallVector<SDValue, 16> Ops(NumConcat);
25224 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25226 for (unsigned i = 1; i != NumConcat; ++i)
25229 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25232 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25233 Mld->getBasePtr(), NewMask, WideSrc0,
25234 Mld->getMemoryVT(), Mld->getMemOperand(),
25236 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25237 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25239 /// PerformMSTORECombine - Resolve truncating stores
25240 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25241 const X86Subtarget *Subtarget) {
25242 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25243 if (!Mst->isTruncatingStore())
25246 EVT VT = Mst->getValue().getValueType();
25247 unsigned NumElems = VT.getVectorNumElements();
25248 EVT StVT = Mst->getMemoryVT();
25251 assert(StVT != VT && "Cannot truncate to the same type");
25252 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25253 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25257 // The truncating store is legal in some cases. For example
25258 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25259 // are designated for truncate store.
25260 // In this case we don't need any further transformations.
25261 if (TLI.isTruncStoreLegal(VT, StVT))
25264 // From, To sizes and ElemCount must be pow of two
25265 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25266 "Unexpected size for truncating masked store");
25267 // We are going to use the original vector elt for storing.
25268 // Accumulated smaller vector elements must be a multiple of the store size.
25269 assert (((NumElems * FromSz) % ToSz) == 0 &&
25270 "Unexpected ratio for truncating masked store");
25272 unsigned SizeRatio = FromSz / ToSz;
25273 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25275 // Create a type on which we perform the shuffle
25276 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25277 StVT.getScalarType(), NumElems*SizeRatio);
25279 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25281 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25282 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25283 for (unsigned i = 0; i != NumElems; ++i)
25284 ShuffleVec[i] = i * SizeRatio;
25286 // Can't shuffle using an illegal type.
25287 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25288 "WideVecVT should be legal");
25290 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25291 DAG.getUNDEF(WideVecVT),
25295 SDValue Mask = Mst->getMask();
25296 if (Mask.getValueType() == VT) {
25297 // Mask and original value have the same type
25298 NewMask = DAG.getBitcast(WideVecVT, Mask);
25299 for (unsigned i = 0; i != NumElems; ++i)
25300 ShuffleVec[i] = i * SizeRatio;
25301 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25302 ShuffleVec[i] = NumElems*SizeRatio;
25303 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25304 DAG.getConstant(0, dl, WideVecVT),
25308 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25309 unsigned WidenNumElts = NumElems*SizeRatio;
25310 unsigned MaskNumElts = VT.getVectorNumElements();
25311 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25314 unsigned NumConcat = WidenNumElts / MaskNumElts;
25315 SmallVector<SDValue, 16> Ops(NumConcat);
25316 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25318 for (unsigned i = 1; i != NumConcat; ++i)
25321 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25324 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25325 NewMask, StVT, Mst->getMemOperand(), false);
25327 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25328 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25329 const X86Subtarget *Subtarget) {
25330 StoreSDNode *St = cast<StoreSDNode>(N);
25331 EVT VT = St->getValue().getValueType();
25332 EVT StVT = St->getMemoryVT();
25334 SDValue StoredVal = St->getOperand(1);
25335 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25337 // If we are saving a concatenation of two XMM registers and 32-byte stores
25338 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25340 unsigned AddressSpace = St->getAddressSpace();
25341 unsigned Alignment = St->getAlignment();
25342 if (VT.is256BitVector() && StVT == VT &&
25343 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25344 AddressSpace, Alignment, &Fast) && !Fast) {
25345 unsigned NumElems = VT.getVectorNumElements();
25349 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25350 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25353 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25354 SDValue Ptr0 = St->getBasePtr();
25355 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25357 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25358 St->getPointerInfo(), St->isVolatile(),
25359 St->isNonTemporal(), Alignment);
25360 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25361 St->getPointerInfo(), St->isVolatile(),
25362 St->isNonTemporal(),
25363 std::min(16U, Alignment));
25364 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25367 // Optimize trunc store (of multiple scalars) to shuffle and store.
25368 // First, pack all of the elements in one place. Next, store to memory
25369 // in fewer chunks.
25370 if (St->isTruncatingStore() && VT.isVector()) {
25371 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25372 unsigned NumElems = VT.getVectorNumElements();
25373 assert(StVT != VT && "Cannot truncate to the same type");
25374 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25375 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25377 // The truncating store is legal in some cases. For example
25378 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25379 // are designated for truncate store.
25380 // In this case we don't need any further transformations.
25381 if (TLI.isTruncStoreLegal(VT, StVT))
25384 // From, To sizes and ElemCount must be pow of two
25385 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25386 // We are going to use the original vector elt for storing.
25387 // Accumulated smaller vector elements must be a multiple of the store size.
25388 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25390 unsigned SizeRatio = FromSz / ToSz;
25392 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25394 // Create a type on which we perform the shuffle
25395 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25396 StVT.getScalarType(), NumElems*SizeRatio);
25398 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25400 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25401 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25402 for (unsigned i = 0; i != NumElems; ++i)
25403 ShuffleVec[i] = i * SizeRatio;
25405 // Can't shuffle using an illegal type.
25406 if (!TLI.isTypeLegal(WideVecVT))
25409 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25410 DAG.getUNDEF(WideVecVT),
25412 // At this point all of the data is stored at the bottom of the
25413 // register. We now need to save it to mem.
25415 // Find the largest store unit
25416 MVT StoreType = MVT::i8;
25417 for (MVT Tp : MVT::integer_valuetypes()) {
25418 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25422 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25423 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25424 (64 <= NumElems * ToSz))
25425 StoreType = MVT::f64;
25427 // Bitcast the original vector into a vector of store-size units
25428 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25429 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25430 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25431 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25432 SmallVector<SDValue, 8> Chains;
25433 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25434 TLI.getPointerTy(DAG.getDataLayout()));
25435 SDValue Ptr = St->getBasePtr();
25437 // Perform one or more big stores into memory.
25438 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25439 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25440 StoreType, ShuffWide,
25441 DAG.getIntPtrConstant(i, dl));
25442 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25443 St->getPointerInfo(), St->isVolatile(),
25444 St->isNonTemporal(), St->getAlignment());
25445 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25446 Chains.push_back(Ch);
25449 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25452 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25453 // the FP state in cases where an emms may be missing.
25454 // A preferable solution to the general problem is to figure out the right
25455 // places to insert EMMS. This qualifies as a quick hack.
25457 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25458 if (VT.getSizeInBits() != 64)
25461 const Function *F = DAG.getMachineFunction().getFunction();
25462 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25464 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25465 if ((VT.isVector() ||
25466 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25467 isa<LoadSDNode>(St->getValue()) &&
25468 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25469 St->getChain().hasOneUse() && !St->isVolatile()) {
25470 SDNode* LdVal = St->getValue().getNode();
25471 LoadSDNode *Ld = nullptr;
25472 int TokenFactorIndex = -1;
25473 SmallVector<SDValue, 8> Ops;
25474 SDNode* ChainVal = St->getChain().getNode();
25475 // Must be a store of a load. We currently handle two cases: the load
25476 // is a direct child, and it's under an intervening TokenFactor. It is
25477 // possible to dig deeper under nested TokenFactors.
25478 if (ChainVal == LdVal)
25479 Ld = cast<LoadSDNode>(St->getChain());
25480 else if (St->getValue().hasOneUse() &&
25481 ChainVal->getOpcode() == ISD::TokenFactor) {
25482 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25483 if (ChainVal->getOperand(i).getNode() == LdVal) {
25484 TokenFactorIndex = i;
25485 Ld = cast<LoadSDNode>(St->getValue());
25487 Ops.push_back(ChainVal->getOperand(i));
25491 if (!Ld || !ISD::isNormalLoad(Ld))
25494 // If this is not the MMX case, i.e. we are just turning i64 load/store
25495 // into f64 load/store, avoid the transformation if there are multiple
25496 // uses of the loaded value.
25497 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25502 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25503 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25505 if (Subtarget->is64Bit() || F64IsLegal) {
25506 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25507 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25508 Ld->getPointerInfo(), Ld->isVolatile(),
25509 Ld->isNonTemporal(), Ld->isInvariant(),
25510 Ld->getAlignment());
25511 SDValue NewChain = NewLd.getValue(1);
25512 if (TokenFactorIndex != -1) {
25513 Ops.push_back(NewChain);
25514 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25516 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25517 St->getPointerInfo(),
25518 St->isVolatile(), St->isNonTemporal(),
25519 St->getAlignment());
25522 // Otherwise, lower to two pairs of 32-bit loads / stores.
25523 SDValue LoAddr = Ld->getBasePtr();
25524 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25525 DAG.getConstant(4, LdDL, MVT::i32));
25527 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25528 Ld->getPointerInfo(),
25529 Ld->isVolatile(), Ld->isNonTemporal(),
25530 Ld->isInvariant(), Ld->getAlignment());
25531 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25532 Ld->getPointerInfo().getWithOffset(4),
25533 Ld->isVolatile(), Ld->isNonTemporal(),
25535 MinAlign(Ld->getAlignment(), 4));
25537 SDValue NewChain = LoLd.getValue(1);
25538 if (TokenFactorIndex != -1) {
25539 Ops.push_back(LoLd);
25540 Ops.push_back(HiLd);
25541 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25544 LoAddr = St->getBasePtr();
25545 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25546 DAG.getConstant(4, StDL, MVT::i32));
25548 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25549 St->getPointerInfo(),
25550 St->isVolatile(), St->isNonTemporal(),
25551 St->getAlignment());
25552 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25553 St->getPointerInfo().getWithOffset(4),
25555 St->isNonTemporal(),
25556 MinAlign(St->getAlignment(), 4));
25557 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25560 // This is similar to the above case, but here we handle a scalar 64-bit
25561 // integer store that is extracted from a vector on a 32-bit target.
25562 // If we have SSE2, then we can treat it like a floating-point double
25563 // to get past legalization. The execution dependencies fixup pass will
25564 // choose the optimal machine instruction for the store if this really is
25565 // an integer or v2f32 rather than an f64.
25566 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25567 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25568 SDValue OldExtract = St->getOperand(1);
25569 SDValue ExtOp0 = OldExtract.getOperand(0);
25570 unsigned VecSize = ExtOp0.getValueSizeInBits();
25571 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25572 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25573 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25574 BitCast, OldExtract.getOperand(1));
25575 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25576 St->getPointerInfo(), St->isVolatile(),
25577 St->isNonTemporal(), St->getAlignment());
25583 /// Return 'true' if this vector operation is "horizontal"
25584 /// and return the operands for the horizontal operation in LHS and RHS. A
25585 /// horizontal operation performs the binary operation on successive elements
25586 /// of its first operand, then on successive elements of its second operand,
25587 /// returning the resulting values in a vector. For example, if
25588 /// A = < float a0, float a1, float a2, float a3 >
25590 /// B = < float b0, float b1, float b2, float b3 >
25591 /// then the result of doing a horizontal operation on A and B is
25592 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25593 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25594 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25595 /// set to A, RHS to B, and the routine returns 'true'.
25596 /// Note that the binary operation should have the property that if one of the
25597 /// operands is UNDEF then the result is UNDEF.
25598 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25599 // Look for the following pattern: if
25600 // A = < float a0, float a1, float a2, float a3 >
25601 // B = < float b0, float b1, float b2, float b3 >
25603 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25604 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25605 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25606 // which is A horizontal-op B.
25608 // At least one of the operands should be a vector shuffle.
25609 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25610 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25613 MVT VT = LHS.getSimpleValueType();
25615 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25616 "Unsupported vector type for horizontal add/sub");
25618 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25619 // operate independently on 128-bit lanes.
25620 unsigned NumElts = VT.getVectorNumElements();
25621 unsigned NumLanes = VT.getSizeInBits()/128;
25622 unsigned NumLaneElts = NumElts / NumLanes;
25623 assert((NumLaneElts % 2 == 0) &&
25624 "Vector type should have an even number of elements in each lane");
25625 unsigned HalfLaneElts = NumLaneElts/2;
25627 // View LHS in the form
25628 // LHS = VECTOR_SHUFFLE A, B, LMask
25629 // If LHS is not a shuffle then pretend it is the shuffle
25630 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25631 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25634 SmallVector<int, 16> LMask(NumElts);
25635 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25636 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25637 A = LHS.getOperand(0);
25638 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25639 B = LHS.getOperand(1);
25640 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25641 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25643 if (LHS.getOpcode() != ISD::UNDEF)
25645 for (unsigned i = 0; i != NumElts; ++i)
25649 // Likewise, view RHS in the form
25650 // RHS = VECTOR_SHUFFLE C, D, RMask
25652 SmallVector<int, 16> RMask(NumElts);
25653 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25654 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25655 C = RHS.getOperand(0);
25656 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25657 D = RHS.getOperand(1);
25658 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25659 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25661 if (RHS.getOpcode() != ISD::UNDEF)
25663 for (unsigned i = 0; i != NumElts; ++i)
25667 // Check that the shuffles are both shuffling the same vectors.
25668 if (!(A == C && B == D) && !(A == D && B == C))
25671 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25672 if (!A.getNode() && !B.getNode())
25675 // If A and B occur in reverse order in RHS, then "swap" them (which means
25676 // rewriting the mask).
25678 ShuffleVectorSDNode::commuteMask(RMask);
25680 // At this point LHS and RHS are equivalent to
25681 // LHS = VECTOR_SHUFFLE A, B, LMask
25682 // RHS = VECTOR_SHUFFLE A, B, RMask
25683 // Check that the masks correspond to performing a horizontal operation.
25684 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25685 for (unsigned i = 0; i != NumLaneElts; ++i) {
25686 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25688 // Ignore any UNDEF components.
25689 if (LIdx < 0 || RIdx < 0 ||
25690 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25691 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25694 // Check that successive elements are being operated on. If not, this is
25695 // not a horizontal operation.
25696 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25697 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25698 if (!(LIdx == Index && RIdx == Index + 1) &&
25699 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25704 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25705 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25709 /// Do target-specific dag combines on floating point adds.
25710 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25711 const X86Subtarget *Subtarget) {
25712 EVT VT = N->getValueType(0);
25713 SDValue LHS = N->getOperand(0);
25714 SDValue RHS = N->getOperand(1);
25716 // Try to synthesize horizontal adds from adds of shuffles.
25717 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25718 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25719 isHorizontalBinOp(LHS, RHS, true))
25720 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25724 /// Do target-specific dag combines on floating point subs.
25725 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25726 const X86Subtarget *Subtarget) {
25727 EVT VT = N->getValueType(0);
25728 SDValue LHS = N->getOperand(0);
25729 SDValue RHS = N->getOperand(1);
25731 // Try to synthesize horizontal subs from subs of shuffles.
25732 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25733 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25734 isHorizontalBinOp(LHS, RHS, false))
25735 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25739 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25740 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25741 const X86Subtarget *Subtarget) {
25742 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25744 // F[X]OR(0.0, x) -> x
25745 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25746 if (C->getValueAPF().isPosZero())
25747 return N->getOperand(1);
25749 // F[X]OR(x, 0.0) -> x
25750 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25751 if (C->getValueAPF().isPosZero())
25752 return N->getOperand(0);
25754 EVT VT = N->getValueType(0);
25755 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25757 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25758 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25760 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25761 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25762 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25763 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25764 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25769 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25770 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25771 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25773 // Only perform optimizations if UnsafeMath is used.
25774 if (!DAG.getTarget().Options.UnsafeFPMath)
25777 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25778 // into FMINC and FMAXC, which are Commutative operations.
25779 unsigned NewOp = 0;
25780 switch (N->getOpcode()) {
25781 default: llvm_unreachable("unknown opcode");
25782 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25783 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25786 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25787 N->getOperand(0), N->getOperand(1));
25790 /// Do target-specific dag combines on X86ISD::FAND nodes.
25791 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25792 // FAND(0.0, x) -> 0.0
25793 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25794 if (C->getValueAPF().isPosZero())
25795 return N->getOperand(0);
25797 // FAND(x, 0.0) -> 0.0
25798 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25799 if (C->getValueAPF().isPosZero())
25800 return N->getOperand(1);
25805 /// Do target-specific dag combines on X86ISD::FANDN nodes
25806 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25807 // FANDN(0.0, x) -> x
25808 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25809 if (C->getValueAPF().isPosZero())
25810 return N->getOperand(1);
25812 // FANDN(x, 0.0) -> 0.0
25813 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25814 if (C->getValueAPF().isPosZero())
25815 return N->getOperand(1);
25820 static SDValue PerformBTCombine(SDNode *N,
25822 TargetLowering::DAGCombinerInfo &DCI) {
25823 // BT ignores high bits in the bit index operand.
25824 SDValue Op1 = N->getOperand(1);
25825 if (Op1.hasOneUse()) {
25826 unsigned BitWidth = Op1.getValueSizeInBits();
25827 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25828 APInt KnownZero, KnownOne;
25829 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25830 !DCI.isBeforeLegalizeOps());
25831 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25832 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25833 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25834 DCI.CommitTargetLoweringOpt(TLO);
25839 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25840 SDValue Op = N->getOperand(0);
25841 if (Op.getOpcode() == ISD::BITCAST)
25842 Op = Op.getOperand(0);
25843 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25844 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25845 VT.getVectorElementType().getSizeInBits() ==
25846 OpVT.getVectorElementType().getSizeInBits()) {
25847 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25852 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25853 const X86Subtarget *Subtarget) {
25854 EVT VT = N->getValueType(0);
25855 if (!VT.isVector())
25858 SDValue N0 = N->getOperand(0);
25859 SDValue N1 = N->getOperand(1);
25860 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25863 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25864 // both SSE and AVX2 since there is no sign-extended shift right
25865 // operation on a vector with 64-bit elements.
25866 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25867 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25868 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25869 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25870 SDValue N00 = N0.getOperand(0);
25872 // EXTLOAD has a better solution on AVX2,
25873 // it may be replaced with X86ISD::VSEXT node.
25874 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25875 if (!ISD::isNormalLoad(N00.getNode()))
25878 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25879 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25881 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25887 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25888 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25889 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25890 /// eliminate extend, add, and shift instructions.
25891 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25892 const X86Subtarget *Subtarget) {
25893 // TODO: This should be valid for other integer types.
25894 EVT VT = Sext->getValueType(0);
25895 if (VT != MVT::i64)
25898 // We need an 'add nsw' feeding into the 'sext'.
25899 SDValue Add = Sext->getOperand(0);
25900 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
25903 // Having a constant operand to the 'add' ensures that we are not increasing
25904 // the instruction count because the constant is extended for free below.
25905 // A constant operand can also become the displacement field of an LEA.
25906 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
25910 // Don't make the 'add' bigger if there's no hope of combining it with some
25911 // other 'add' or 'shl' instruction.
25912 // TODO: It may be profitable to generate simpler LEA instructions in place
25913 // of single 'add' instructions, but the cost model for selecting an LEA
25914 // currently has a high threshold.
25915 bool HasLEAPotential = false;
25916 for (auto *User : Sext->uses()) {
25917 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
25918 HasLEAPotential = true;
25922 if (!HasLEAPotential)
25925 // Everything looks good, so pull the 'sext' ahead of the 'add'.
25926 int64_t AddConstant = AddOp1->getSExtValue();
25927 SDValue AddOp0 = Add.getOperand(0);
25928 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
25929 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
25931 // The wider add is guaranteed to not wrap because both operands are
25934 Flags.setNoSignedWrap(true);
25935 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
25938 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25939 TargetLowering::DAGCombinerInfo &DCI,
25940 const X86Subtarget *Subtarget) {
25941 SDValue N0 = N->getOperand(0);
25942 EVT VT = N->getValueType(0);
25943 EVT SVT = VT.getScalarType();
25944 EVT InVT = N0.getValueType();
25945 EVT InSVT = InVT.getScalarType();
25948 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25949 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25950 // This exposes the sext to the sdivrem lowering, so that it directly extends
25951 // from AH (which we otherwise need to do contortions to access).
25952 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25953 InVT == MVT::i8 && VT == MVT::i32) {
25954 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25955 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25956 N0.getOperand(0), N0.getOperand(1));
25957 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25958 return R.getValue(1);
25961 if (!DCI.isBeforeLegalizeOps()) {
25962 if (InVT == MVT::i1) {
25963 SDValue Zero = DAG.getConstant(0, DL, VT);
25965 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25966 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25971 if (VT.isVector() && Subtarget->hasSSE2()) {
25972 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25973 EVT InVT = N.getValueType();
25974 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25975 Size / InVT.getScalarSizeInBits());
25976 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25977 DAG.getUNDEF(InVT));
25979 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25982 // If target-size is less than 128-bits, extend to a type that would extend
25983 // to 128 bits, extend that and extract the original target vector.
25984 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25985 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25986 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25987 unsigned Scale = 128 / VT.getSizeInBits();
25989 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25990 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25991 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25992 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25993 DAG.getIntPtrConstant(0, DL));
25996 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25997 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25998 if (VT.getSizeInBits() == 128 &&
25999 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26000 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26001 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26002 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26005 // On pre-AVX2 targets, split into 128-bit nodes of
26006 // ISD::SIGN_EXTEND_VECTOR_INREG.
26007 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26008 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26009 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26010 unsigned NumVecs = VT.getSizeInBits() / 128;
26011 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26012 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26013 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26015 SmallVector<SDValue, 8> Opnds;
26016 for (unsigned i = 0, Offset = 0; i != NumVecs;
26017 ++i, Offset += NumSubElts) {
26018 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26019 DAG.getIntPtrConstant(Offset, DL));
26020 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26021 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26022 Opnds.push_back(SrcVec);
26024 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26028 if (Subtarget->hasAVX() && VT.is256BitVector())
26029 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26032 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26038 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26039 const X86Subtarget* Subtarget) {
26041 EVT VT = N->getValueType(0);
26043 // Let legalize expand this if it isn't a legal type yet.
26044 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26047 EVT ScalarVT = VT.getScalarType();
26048 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26049 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26050 !Subtarget->hasAVX512()))
26053 SDValue A = N->getOperand(0);
26054 SDValue B = N->getOperand(1);
26055 SDValue C = N->getOperand(2);
26057 bool NegA = (A.getOpcode() == ISD::FNEG);
26058 bool NegB = (B.getOpcode() == ISD::FNEG);
26059 bool NegC = (C.getOpcode() == ISD::FNEG);
26061 // Negative multiplication when NegA xor NegB
26062 bool NegMul = (NegA != NegB);
26064 A = A.getOperand(0);
26066 B = B.getOperand(0);
26068 C = C.getOperand(0);
26072 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26074 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26076 return DAG.getNode(Opcode, dl, VT, A, B, C);
26079 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26080 TargetLowering::DAGCombinerInfo &DCI,
26081 const X86Subtarget *Subtarget) {
26082 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26083 // (and (i32 x86isd::setcc_carry), 1)
26084 // This eliminates the zext. This transformation is necessary because
26085 // ISD::SETCC is always legalized to i8.
26087 SDValue N0 = N->getOperand(0);
26088 EVT VT = N->getValueType(0);
26090 if (N0.getOpcode() == ISD::AND &&
26092 N0.getOperand(0).hasOneUse()) {
26093 SDValue N00 = N0.getOperand(0);
26094 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26095 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26096 if (!C || C->getZExtValue() != 1)
26098 return DAG.getNode(ISD::AND, dl, VT,
26099 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26100 N00.getOperand(0), N00.getOperand(1)),
26101 DAG.getConstant(1, dl, VT));
26105 if (N0.getOpcode() == ISD::TRUNCATE &&
26107 N0.getOperand(0).hasOneUse()) {
26108 SDValue N00 = N0.getOperand(0);
26109 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26110 return DAG.getNode(ISD::AND, dl, VT,
26111 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26112 N00.getOperand(0), N00.getOperand(1)),
26113 DAG.getConstant(1, dl, VT));
26117 if (VT.is256BitVector())
26118 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26121 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26122 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26123 // This exposes the zext to the udivrem lowering, so that it directly extends
26124 // from AH (which we otherwise need to do contortions to access).
26125 if (N0.getOpcode() == ISD::UDIVREM &&
26126 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26127 (VT == MVT::i32 || VT == MVT::i64)) {
26128 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26129 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26130 N0.getOperand(0), N0.getOperand(1));
26131 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26132 return R.getValue(1);
26138 // Optimize x == -y --> x+y == 0
26139 // x != -y --> x+y != 0
26140 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26141 const X86Subtarget* Subtarget) {
26142 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26143 SDValue LHS = N->getOperand(0);
26144 SDValue RHS = N->getOperand(1);
26145 EVT VT = N->getValueType(0);
26148 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26149 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26150 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26151 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26152 LHS.getOperand(1));
26153 return DAG.getSetCC(DL, N->getValueType(0), addV,
26154 DAG.getConstant(0, DL, addV.getValueType()), CC);
26156 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26157 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26158 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26159 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26160 RHS.getOperand(1));
26161 return DAG.getSetCC(DL, N->getValueType(0), addV,
26162 DAG.getConstant(0, DL, addV.getValueType()), CC);
26165 if (VT.getScalarType() == MVT::i1 &&
26166 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26168 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26169 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26170 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26172 if (!IsSEXT0 || !IsVZero1) {
26173 // Swap the operands and update the condition code.
26174 std::swap(LHS, RHS);
26175 CC = ISD::getSetCCSwappedOperands(CC);
26177 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26178 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26179 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26182 if (IsSEXT0 && IsVZero1) {
26183 assert(VT == LHS.getOperand(0).getValueType() &&
26184 "Uexpected operand type");
26185 if (CC == ISD::SETGT)
26186 return DAG.getConstant(0, DL, VT);
26187 if (CC == ISD::SETLE)
26188 return DAG.getConstant(1, DL, VT);
26189 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26190 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26192 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26193 "Unexpected condition code!");
26194 return LHS.getOperand(0);
26201 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
26202 SelectionDAG &DAG) {
26204 MVT VT = Load->getSimpleValueType(0);
26205 MVT EVT = VT.getVectorElementType();
26206 SDValue Addr = Load->getOperand(1);
26207 SDValue NewAddr = DAG.getNode(
26208 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
26209 DAG.getConstant(Index * EVT.getStoreSize(), dl,
26210 Addr.getSimpleValueType()));
26213 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
26214 DAG.getMachineFunction().getMachineMemOperand(
26215 Load->getMemOperand(), 0, EVT.getStoreSize()));
26219 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26220 const X86Subtarget *Subtarget) {
26222 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26223 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26224 "X86insertps is only defined for v4x32");
26226 SDValue Ld = N->getOperand(1);
26227 if (MayFoldLoad(Ld)) {
26228 // Extract the countS bits from the immediate so we can get the proper
26229 // address when narrowing the vector load to a specific element.
26230 // When the second source op is a memory address, insertps doesn't use
26231 // countS and just gets an f32 from that address.
26232 unsigned DestIndex =
26233 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26235 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26237 // Create this as a scalar to vector to match the instruction pattern.
26238 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26239 // countS bits are ignored when loading from memory on insertps, which
26240 // means we don't need to explicitly set them to 0.
26241 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26242 LoadScalarToVector, N->getOperand(2));
26247 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26248 SDValue V0 = N->getOperand(0);
26249 SDValue V1 = N->getOperand(1);
26251 EVT VT = N->getValueType(0);
26253 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26254 // operands and changing the mask to 1. This saves us a bunch of
26255 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26256 // x86InstrInfo knows how to commute this back after instruction selection
26257 // if it would help register allocation.
26259 // TODO: If optimizing for size or a processor that doesn't suffer from
26260 // partial register update stalls, this should be transformed into a MOVSD
26261 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26263 if (VT == MVT::v2f64)
26264 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26265 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26266 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26267 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26273 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26274 // as "sbb reg,reg", since it can be extended without zext and produces
26275 // an all-ones bit which is more useful than 0/1 in some cases.
26276 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26279 return DAG.getNode(ISD::AND, DL, VT,
26280 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26281 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26283 DAG.getConstant(1, DL, VT));
26284 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26285 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26286 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26287 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26291 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26292 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26293 TargetLowering::DAGCombinerInfo &DCI,
26294 const X86Subtarget *Subtarget) {
26296 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26297 SDValue EFLAGS = N->getOperand(1);
26299 if (CC == X86::COND_A) {
26300 // Try to convert COND_A into COND_B in an attempt to facilitate
26301 // materializing "setb reg".
26303 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26304 // cannot take an immediate as its first operand.
26306 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26307 EFLAGS.getValueType().isInteger() &&
26308 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26309 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26310 EFLAGS.getNode()->getVTList(),
26311 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26312 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26313 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26317 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26318 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26320 if (CC == X86::COND_B)
26321 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26323 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26324 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26325 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26331 // Optimize branch condition evaluation.
26333 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26334 TargetLowering::DAGCombinerInfo &DCI,
26335 const X86Subtarget *Subtarget) {
26337 SDValue Chain = N->getOperand(0);
26338 SDValue Dest = N->getOperand(1);
26339 SDValue EFLAGS = N->getOperand(3);
26340 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26342 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26343 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26344 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26351 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26352 SelectionDAG &DAG) {
26353 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26354 // optimize away operation when it's from a constant.
26356 // The general transformation is:
26357 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26358 // AND(VECTOR_CMP(x,y), constant2)
26359 // constant2 = UNARYOP(constant)
26361 // Early exit if this isn't a vector operation, the operand of the
26362 // unary operation isn't a bitwise AND, or if the sizes of the operations
26363 // aren't the same.
26364 EVT VT = N->getValueType(0);
26365 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26366 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26367 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26370 // Now check that the other operand of the AND is a constant. We could
26371 // make the transformation for non-constant splats as well, but it's unclear
26372 // that would be a benefit as it would not eliminate any operations, just
26373 // perform one more step in scalar code before moving to the vector unit.
26374 if (BuildVectorSDNode *BV =
26375 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26376 // Bail out if the vector isn't a constant.
26377 if (!BV->isConstant())
26380 // Everything checks out. Build up the new and improved node.
26382 EVT IntVT = BV->getValueType(0);
26383 // Create a new constant of the appropriate type for the transformed
26385 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26386 // The AND node needs bitcasts to/from an integer vector type around it.
26387 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26388 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26389 N->getOperand(0)->getOperand(0), MaskConst);
26390 SDValue Res = DAG.getBitcast(VT, NewAnd);
26397 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26398 const X86Subtarget *Subtarget) {
26399 SDValue Op0 = N->getOperand(0);
26400 EVT VT = N->getValueType(0);
26401 EVT InVT = Op0.getValueType();
26402 EVT InSVT = InVT.getScalarType();
26403 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26405 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26406 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26407 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26409 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26410 InVT.getVectorNumElements());
26411 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26413 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26414 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26416 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26422 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26423 const X86Subtarget *Subtarget) {
26424 // First try to optimize away the conversion entirely when it's
26425 // conditionally from a constant. Vectors only.
26426 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26429 // Now move on to more general possibilities.
26430 SDValue Op0 = N->getOperand(0);
26431 EVT VT = N->getValueType(0);
26432 EVT InVT = Op0.getValueType();
26433 EVT InSVT = InVT.getScalarType();
26435 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26436 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26437 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26439 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26440 InVT.getVectorNumElements());
26441 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26442 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26445 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26446 // a 32-bit target where SSE doesn't support i64->FP operations.
26447 if (Op0.getOpcode() == ISD::LOAD) {
26448 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26449 EVT LdVT = Ld->getValueType(0);
26451 // This transformation is not supported if the result type is f16
26452 if (VT == MVT::f16)
26455 if (!Ld->isVolatile() && !VT.isVector() &&
26456 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26457 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26458 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26459 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26460 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26467 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26468 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26469 X86TargetLowering::DAGCombinerInfo &DCI) {
26470 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26471 // the result is either zero or one (depending on the input carry bit).
26472 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26473 if (X86::isZeroNode(N->getOperand(0)) &&
26474 X86::isZeroNode(N->getOperand(1)) &&
26475 // We don't have a good way to replace an EFLAGS use, so only do this when
26477 SDValue(N, 1).use_empty()) {
26479 EVT VT = N->getValueType(0);
26480 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26481 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26482 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26483 DAG.getConstant(X86::COND_B, DL,
26486 DAG.getConstant(1, DL, VT));
26487 return DCI.CombineTo(N, Res1, CarryOut);
26493 // fold (add Y, (sete X, 0)) -> adc 0, Y
26494 // (add Y, (setne X, 0)) -> sbb -1, Y
26495 // (sub (sete X, 0), Y) -> sbb 0, Y
26496 // (sub (setne X, 0), Y) -> adc -1, Y
26497 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26500 // Look through ZExts.
26501 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26502 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26505 SDValue SetCC = Ext.getOperand(0);
26506 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26509 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26510 if (CC != X86::COND_E && CC != X86::COND_NE)
26513 SDValue Cmp = SetCC.getOperand(1);
26514 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26515 !X86::isZeroNode(Cmp.getOperand(1)) ||
26516 !Cmp.getOperand(0).getValueType().isInteger())
26519 SDValue CmpOp0 = Cmp.getOperand(0);
26520 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26521 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26523 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26524 if (CC == X86::COND_NE)
26525 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26526 DL, OtherVal.getValueType(), OtherVal,
26527 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26529 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26530 DL, OtherVal.getValueType(), OtherVal,
26531 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26534 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26535 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26536 const X86Subtarget *Subtarget) {
26537 EVT VT = N->getValueType(0);
26538 SDValue Op0 = N->getOperand(0);
26539 SDValue Op1 = N->getOperand(1);
26541 // Try to synthesize horizontal adds from adds of shuffles.
26542 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26543 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26544 isHorizontalBinOp(Op0, Op1, true))
26545 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26547 return OptimizeConditionalInDecrement(N, DAG);
26550 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26551 const X86Subtarget *Subtarget) {
26552 SDValue Op0 = N->getOperand(0);
26553 SDValue Op1 = N->getOperand(1);
26555 // X86 can't encode an immediate LHS of a sub. See if we can push the
26556 // negation into a preceding instruction.
26557 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26558 // If the RHS of the sub is a XOR with one use and a constant, invert the
26559 // immediate. Then add one to the LHS of the sub so we can turn
26560 // X-Y -> X+~Y+1, saving one register.
26561 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26562 isa<ConstantSDNode>(Op1.getOperand(1))) {
26563 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26564 EVT VT = Op0.getValueType();
26565 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26567 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26568 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26569 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26573 // Try to synthesize horizontal adds from adds of shuffles.
26574 EVT VT = N->getValueType(0);
26575 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26576 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26577 isHorizontalBinOp(Op0, Op1, true))
26578 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26580 return OptimizeConditionalInDecrement(N, DAG);
26583 /// performVZEXTCombine - Performs build vector combines
26584 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26585 TargetLowering::DAGCombinerInfo &DCI,
26586 const X86Subtarget *Subtarget) {
26588 MVT VT = N->getSimpleValueType(0);
26589 SDValue Op = N->getOperand(0);
26590 MVT OpVT = Op.getSimpleValueType();
26591 MVT OpEltVT = OpVT.getVectorElementType();
26592 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26594 // (vzext (bitcast (vzext (x)) -> (vzext x)
26596 while (V.getOpcode() == ISD::BITCAST)
26597 V = V.getOperand(0);
26599 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26600 MVT InnerVT = V.getSimpleValueType();
26601 MVT InnerEltVT = InnerVT.getVectorElementType();
26603 // If the element sizes match exactly, we can just do one larger vzext. This
26604 // is always an exact type match as vzext operates on integer types.
26605 if (OpEltVT == InnerEltVT) {
26606 assert(OpVT == InnerVT && "Types must match for vzext!");
26607 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26610 // The only other way we can combine them is if only a single element of the
26611 // inner vzext is used in the input to the outer vzext.
26612 if (InnerEltVT.getSizeInBits() < InputBits)
26615 // In this case, the inner vzext is completely dead because we're going to
26616 // only look at bits inside of the low element. Just do the outer vzext on
26617 // a bitcast of the input to the inner.
26618 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26621 // Check if we can bypass extracting and re-inserting an element of an input
26622 // vector. Essentially:
26623 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26624 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26625 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26626 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26627 SDValue ExtractedV = V.getOperand(0);
26628 SDValue OrigV = ExtractedV.getOperand(0);
26629 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26630 if (ExtractIdx->getZExtValue() == 0) {
26631 MVT OrigVT = OrigV.getSimpleValueType();
26632 // Extract a subvector if necessary...
26633 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26634 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26635 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26636 OrigVT.getVectorNumElements() / Ratio);
26637 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26638 DAG.getIntPtrConstant(0, DL));
26640 Op = DAG.getBitcast(OpVT, OrigV);
26641 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26648 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26649 DAGCombinerInfo &DCI) const {
26650 SelectionDAG &DAG = DCI.DAG;
26651 switch (N->getOpcode()) {
26653 case ISD::EXTRACT_VECTOR_ELT:
26654 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26657 case X86ISD::SHRUNKBLEND:
26658 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26659 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26660 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26661 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26662 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26663 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26664 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26667 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26668 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26669 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26670 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26671 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26672 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26673 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26674 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26675 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26676 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26677 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26678 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26680 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26682 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26683 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26684 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26685 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26686 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26687 case ISD::ANY_EXTEND:
26688 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26689 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26690 case ISD::SIGN_EXTEND_INREG:
26691 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26692 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26693 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26694 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26695 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26696 case X86ISD::SHUFP: // Handle all target specific shuffles
26697 case X86ISD::PALIGNR:
26698 case X86ISD::UNPCKH:
26699 case X86ISD::UNPCKL:
26700 case X86ISD::MOVHLPS:
26701 case X86ISD::MOVLHPS:
26702 case X86ISD::PSHUFB:
26703 case X86ISD::PSHUFD:
26704 case X86ISD::PSHUFHW:
26705 case X86ISD::PSHUFLW:
26706 case X86ISD::MOVSS:
26707 case X86ISD::MOVSD:
26708 case X86ISD::VPERMILPI:
26709 case X86ISD::VPERM2X128:
26710 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26711 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26712 case X86ISD::INSERTPS: {
26713 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26714 return PerformINSERTPSCombine(N, DAG, Subtarget);
26717 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26723 /// isTypeDesirableForOp - Return true if the target has native support for
26724 /// the specified value type and it is 'desirable' to use the type for the
26725 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26726 /// instruction encodings are longer and some i16 instructions are slow.
26727 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26728 if (!isTypeLegal(VT))
26730 if (VT != MVT::i16)
26737 case ISD::SIGN_EXTEND:
26738 case ISD::ZERO_EXTEND:
26739 case ISD::ANY_EXTEND:
26752 /// IsDesirableToPromoteOp - This method query the target whether it is
26753 /// beneficial for dag combiner to promote the specified node. If true, it
26754 /// should return the desired promotion type by reference.
26755 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26756 EVT VT = Op.getValueType();
26757 if (VT != MVT::i16)
26760 bool Promote = false;
26761 bool Commute = false;
26762 switch (Op.getOpcode()) {
26765 LoadSDNode *LD = cast<LoadSDNode>(Op);
26766 // If the non-extending load has a single use and it's not live out, then it
26767 // might be folded.
26768 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26769 Op.hasOneUse()*/) {
26770 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26771 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26772 // The only case where we'd want to promote LOAD (rather then it being
26773 // promoted as an operand is when it's only use is liveout.
26774 if (UI->getOpcode() != ISD::CopyToReg)
26781 case ISD::SIGN_EXTEND:
26782 case ISD::ZERO_EXTEND:
26783 case ISD::ANY_EXTEND:
26788 SDValue N0 = Op.getOperand(0);
26789 // Look out for (store (shl (load), x)).
26790 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26803 SDValue N0 = Op.getOperand(0);
26804 SDValue N1 = Op.getOperand(1);
26805 if (!Commute && MayFoldLoad(N1))
26807 // Avoid disabling potential load folding opportunities.
26808 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26810 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26820 //===----------------------------------------------------------------------===//
26821 // X86 Inline Assembly Support
26822 //===----------------------------------------------------------------------===//
26824 // Helper to match a string separated by whitespace.
26825 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26826 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26828 for (StringRef Piece : Pieces) {
26829 if (!S.startswith(Piece)) // Check if the piece matches.
26832 S = S.substr(Piece.size());
26833 StringRef::size_type Pos = S.find_first_not_of(" \t");
26834 if (Pos == 0) // We matched a prefix.
26843 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26845 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26846 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26847 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26848 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26850 if (AsmPieces.size() == 3)
26852 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26859 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26860 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26862 std::string AsmStr = IA->getAsmString();
26864 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26865 if (!Ty || Ty->getBitWidth() % 16 != 0)
26868 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26869 SmallVector<StringRef, 4> AsmPieces;
26870 SplitString(AsmStr, AsmPieces, ";\n");
26872 switch (AsmPieces.size()) {
26873 default: return false;
26875 // FIXME: this should verify that we are targeting a 486 or better. If not,
26876 // we will turn this bswap into something that will be lowered to logical
26877 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26878 // lower so don't worry about this.
26880 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26881 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26882 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26883 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26884 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26885 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26886 // No need to check constraints, nothing other than the equivalent of
26887 // "=r,0" would be valid here.
26888 return IntrinsicLowering::LowerToByteSwap(CI);
26891 // rorw $$8, ${0:w} --> llvm.bswap.i16
26892 if (CI->getType()->isIntegerTy(16) &&
26893 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26894 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26895 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26897 StringRef ConstraintsStr = IA->getConstraintString();
26898 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26899 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26900 if (clobbersFlagRegisters(AsmPieces))
26901 return IntrinsicLowering::LowerToByteSwap(CI);
26905 if (CI->getType()->isIntegerTy(32) &&
26906 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26907 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26908 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26909 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26911 StringRef ConstraintsStr = IA->getConstraintString();
26912 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26913 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26914 if (clobbersFlagRegisters(AsmPieces))
26915 return IntrinsicLowering::LowerToByteSwap(CI);
26918 if (CI->getType()->isIntegerTy(64)) {
26919 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26920 if (Constraints.size() >= 2 &&
26921 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26922 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26923 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26924 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26925 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26926 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26927 return IntrinsicLowering::LowerToByteSwap(CI);
26935 /// getConstraintType - Given a constraint letter, return the type of
26936 /// constraint it is for this target.
26937 X86TargetLowering::ConstraintType
26938 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26939 if (Constraint.size() == 1) {
26940 switch (Constraint[0]) {
26951 return C_RegisterClass;
26975 return TargetLowering::getConstraintType(Constraint);
26978 /// Examine constraint type and operand type and determine a weight value.
26979 /// This object must already have been set up with the operand type
26980 /// and the current alternative constraint selected.
26981 TargetLowering::ConstraintWeight
26982 X86TargetLowering::getSingleConstraintMatchWeight(
26983 AsmOperandInfo &info, const char *constraint) const {
26984 ConstraintWeight weight = CW_Invalid;
26985 Value *CallOperandVal = info.CallOperandVal;
26986 // If we don't have a value, we can't do a match,
26987 // but allow it at the lowest weight.
26988 if (!CallOperandVal)
26990 Type *type = CallOperandVal->getType();
26991 // Look at the constraint type.
26992 switch (*constraint) {
26994 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27005 if (CallOperandVal->getType()->isIntegerTy())
27006 weight = CW_SpecificReg;
27011 if (type->isFloatingPointTy())
27012 weight = CW_SpecificReg;
27015 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27016 weight = CW_SpecificReg;
27020 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27021 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27022 weight = CW_Register;
27025 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27026 if (C->getZExtValue() <= 31)
27027 weight = CW_Constant;
27031 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27032 if (C->getZExtValue() <= 63)
27033 weight = CW_Constant;
27037 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27038 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27039 weight = CW_Constant;
27043 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27044 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27045 weight = CW_Constant;
27049 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27050 if (C->getZExtValue() <= 3)
27051 weight = CW_Constant;
27055 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27056 if (C->getZExtValue() <= 0xff)
27057 weight = CW_Constant;
27062 if (isa<ConstantFP>(CallOperandVal)) {
27063 weight = CW_Constant;
27067 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27068 if ((C->getSExtValue() >= -0x80000000LL) &&
27069 (C->getSExtValue() <= 0x7fffffffLL))
27070 weight = CW_Constant;
27074 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27075 if (C->getZExtValue() <= 0xffffffff)
27076 weight = CW_Constant;
27083 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27084 /// with another that has more specific requirements based on the type of the
27085 /// corresponding operand.
27086 const char *X86TargetLowering::
27087 LowerXConstraint(EVT ConstraintVT) const {
27088 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27089 // 'f' like normal targets.
27090 if (ConstraintVT.isFloatingPoint()) {
27091 if (Subtarget->hasSSE2())
27093 if (Subtarget->hasSSE1())
27097 return TargetLowering::LowerXConstraint(ConstraintVT);
27100 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27101 /// vector. If it is invalid, don't add anything to Ops.
27102 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27103 std::string &Constraint,
27104 std::vector<SDValue>&Ops,
27105 SelectionDAG &DAG) const {
27108 // Only support length 1 constraints for now.
27109 if (Constraint.length() > 1) return;
27111 char ConstraintLetter = Constraint[0];
27112 switch (ConstraintLetter) {
27115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27116 if (C->getZExtValue() <= 31) {
27117 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27118 Op.getValueType());
27124 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27125 if (C->getZExtValue() <= 63) {
27126 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27127 Op.getValueType());
27133 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27134 if (isInt<8>(C->getSExtValue())) {
27135 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27136 Op.getValueType());
27142 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27143 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27144 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27145 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27146 Op.getValueType());
27152 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27153 if (C->getZExtValue() <= 3) {
27154 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27155 Op.getValueType());
27161 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27162 if (C->getZExtValue() <= 255) {
27163 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27164 Op.getValueType());
27170 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27171 if (C->getZExtValue() <= 127) {
27172 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27173 Op.getValueType());
27179 // 32-bit signed value
27180 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27181 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27182 C->getSExtValue())) {
27183 // Widen to 64 bits here to get it sign extended.
27184 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27187 // FIXME gcc accepts some relocatable values here too, but only in certain
27188 // memory models; it's complicated.
27193 // 32-bit unsigned value
27194 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27195 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27196 C->getZExtValue())) {
27197 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27198 Op.getValueType());
27202 // FIXME gcc accepts some relocatable values here too, but only in certain
27203 // memory models; it's complicated.
27207 // Literal immediates are always ok.
27208 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27209 // Widen to 64 bits here to get it sign extended.
27210 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27214 // In any sort of PIC mode addresses need to be computed at runtime by
27215 // adding in a register or some sort of table lookup. These can't
27216 // be used as immediates.
27217 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27220 // If we are in non-pic codegen mode, we allow the address of a global (with
27221 // an optional displacement) to be used with 'i'.
27222 GlobalAddressSDNode *GA = nullptr;
27223 int64_t Offset = 0;
27225 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27227 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27228 Offset += GA->getOffset();
27230 } else if (Op.getOpcode() == ISD::ADD) {
27231 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27232 Offset += C->getZExtValue();
27233 Op = Op.getOperand(0);
27236 } else if (Op.getOpcode() == ISD::SUB) {
27237 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27238 Offset += -C->getZExtValue();
27239 Op = Op.getOperand(0);
27244 // Otherwise, this isn't something we can handle, reject it.
27248 const GlobalValue *GV = GA->getGlobal();
27249 // If we require an extra load to get this address, as in PIC mode, we
27250 // can't accept it.
27251 if (isGlobalStubReference(
27252 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27255 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27256 GA->getValueType(0), Offset);
27261 if (Result.getNode()) {
27262 Ops.push_back(Result);
27265 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27268 std::pair<unsigned, const TargetRegisterClass *>
27269 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27270 StringRef Constraint,
27272 // First, see if this is a constraint that directly corresponds to an LLVM
27274 if (Constraint.size() == 1) {
27275 // GCC Constraint Letters
27276 switch (Constraint[0]) {
27278 // TODO: Slight differences here in allocation order and leaving
27279 // RIP in the class. Do they matter any more here than they do
27280 // in the normal allocation?
27281 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27282 if (Subtarget->is64Bit()) {
27283 if (VT == MVT::i32 || VT == MVT::f32)
27284 return std::make_pair(0U, &X86::GR32RegClass);
27285 if (VT == MVT::i16)
27286 return std::make_pair(0U, &X86::GR16RegClass);
27287 if (VT == MVT::i8 || VT == MVT::i1)
27288 return std::make_pair(0U, &X86::GR8RegClass);
27289 if (VT == MVT::i64 || VT == MVT::f64)
27290 return std::make_pair(0U, &X86::GR64RegClass);
27293 // 32-bit fallthrough
27294 case 'Q': // Q_REGS
27295 if (VT == MVT::i32 || VT == MVT::f32)
27296 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27297 if (VT == MVT::i16)
27298 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27299 if (VT == MVT::i8 || VT == MVT::i1)
27300 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27301 if (VT == MVT::i64)
27302 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27304 case 'r': // GENERAL_REGS
27305 case 'l': // INDEX_REGS
27306 if (VT == MVT::i8 || VT == MVT::i1)
27307 return std::make_pair(0U, &X86::GR8RegClass);
27308 if (VT == MVT::i16)
27309 return std::make_pair(0U, &X86::GR16RegClass);
27310 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27311 return std::make_pair(0U, &X86::GR32RegClass);
27312 return std::make_pair(0U, &X86::GR64RegClass);
27313 case 'R': // LEGACY_REGS
27314 if (VT == MVT::i8 || VT == MVT::i1)
27315 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27316 if (VT == MVT::i16)
27317 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27318 if (VT == MVT::i32 || !Subtarget->is64Bit())
27319 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27320 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27321 case 'f': // FP Stack registers.
27322 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27323 // value to the correct fpstack register class.
27324 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27325 return std::make_pair(0U, &X86::RFP32RegClass);
27326 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27327 return std::make_pair(0U, &X86::RFP64RegClass);
27328 return std::make_pair(0U, &X86::RFP80RegClass);
27329 case 'y': // MMX_REGS if MMX allowed.
27330 if (!Subtarget->hasMMX()) break;
27331 return std::make_pair(0U, &X86::VR64RegClass);
27332 case 'Y': // SSE_REGS if SSE2 allowed
27333 if (!Subtarget->hasSSE2()) break;
27335 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27336 if (!Subtarget->hasSSE1()) break;
27338 switch (VT.SimpleTy) {
27340 // Scalar SSE types.
27343 return std::make_pair(0U, &X86::FR32RegClass);
27346 return std::make_pair(0U, &X86::FR64RegClass);
27354 return std::make_pair(0U, &X86::VR128RegClass);
27362 return std::make_pair(0U, &X86::VR256RegClass);
27367 return std::make_pair(0U, &X86::VR512RegClass);
27373 // Use the default implementation in TargetLowering to convert the register
27374 // constraint into a member of a register class.
27375 std::pair<unsigned, const TargetRegisterClass*> Res;
27376 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27378 // Not found as a standard register?
27380 // Map st(0) -> st(7) -> ST0
27381 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27382 tolower(Constraint[1]) == 's' &&
27383 tolower(Constraint[2]) == 't' &&
27384 Constraint[3] == '(' &&
27385 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27386 Constraint[5] == ')' &&
27387 Constraint[6] == '}') {
27389 Res.first = X86::FP0+Constraint[4]-'0';
27390 Res.second = &X86::RFP80RegClass;
27394 // GCC allows "st(0)" to be called just plain "st".
27395 if (StringRef("{st}").equals_lower(Constraint)) {
27396 Res.first = X86::FP0;
27397 Res.second = &X86::RFP80RegClass;
27402 if (StringRef("{flags}").equals_lower(Constraint)) {
27403 Res.first = X86::EFLAGS;
27404 Res.second = &X86::CCRRegClass;
27408 // 'A' means EAX + EDX.
27409 if (Constraint == "A") {
27410 Res.first = X86::EAX;
27411 Res.second = &X86::GR32_ADRegClass;
27417 // Otherwise, check to see if this is a register class of the wrong value
27418 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27419 // turn into {ax},{dx}.
27420 // MVT::Other is used to specify clobber names.
27421 if (Res.second->hasType(VT) || VT == MVT::Other)
27422 return Res; // Correct type already, nothing to do.
27424 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27425 // return "eax". This should even work for things like getting 64bit integer
27426 // registers when given an f64 type.
27427 const TargetRegisterClass *Class = Res.second;
27428 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27429 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27430 unsigned Size = VT.getSizeInBits();
27431 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27432 : Size == 16 ? MVT::i16
27433 : Size == 32 ? MVT::i32
27434 : Size == 64 ? MVT::i64
27436 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27438 Res.first = DestReg;
27439 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27440 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27441 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27442 : &X86::GR64RegClass;
27443 assert(Res.second->contains(Res.first) && "Register in register class");
27445 // No register found/type mismatch.
27447 Res.second = nullptr;
27449 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27450 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27451 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27452 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27453 Class == &X86::VR512RegClass) {
27454 // Handle references to XMM physical registers that got mapped into the
27455 // wrong class. This can happen with constraints like {xmm0} where the
27456 // target independent register mapper will just pick the first match it can
27457 // find, ignoring the required type.
27459 if (VT == MVT::f32 || VT == MVT::i32)
27460 Res.second = &X86::FR32RegClass;
27461 else if (VT == MVT::f64 || VT == MVT::i64)
27462 Res.second = &X86::FR64RegClass;
27463 else if (X86::VR128RegClass.hasType(VT))
27464 Res.second = &X86::VR128RegClass;
27465 else if (X86::VR256RegClass.hasType(VT))
27466 Res.second = &X86::VR256RegClass;
27467 else if (X86::VR512RegClass.hasType(VT))
27468 Res.second = &X86::VR512RegClass;
27470 // Type mismatch and not a clobber: Return an error;
27472 Res.second = nullptr;
27479 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27480 const AddrMode &AM, Type *Ty,
27481 unsigned AS) const {
27482 // Scaling factors are not free at all.
27483 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27484 // will take 2 allocations in the out of order engine instead of 1
27485 // for plain addressing mode, i.e. inst (reg1).
27487 // vaddps (%rsi,%drx), %ymm0, %ymm1
27488 // Requires two allocations (one for the load, one for the computation)
27490 // vaddps (%rsi), %ymm0, %ymm1
27491 // Requires just 1 allocation, i.e., freeing allocations for other operations
27492 // and having less micro operations to execute.
27494 // For some X86 architectures, this is even worse because for instance for
27495 // stores, the complex addressing mode forces the instruction to use the
27496 // "load" ports instead of the dedicated "store" port.
27497 // E.g., on Haswell:
27498 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27499 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27500 if (isLegalAddressingMode(DL, AM, Ty, AS))
27501 // Scale represents reg2 * scale, thus account for 1
27502 // as soon as we use a second register.
27503 return AM.Scale != 0;
27507 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27508 // Integer division on x86 is expensive. However, when aggressively optimizing
27509 // for code size, we prefer to use a div instruction, as it is usually smaller
27510 // than the alternative sequence.
27511 // The exception to this is vector division. Since x86 doesn't have vector
27512 // integer division, leaving the division as-is is a loss even in terms of
27513 // size, because it will have to be scalarized, while the alternative code
27514 // sequence can be performed in vector form.
27515 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27516 Attribute::MinSize);
27517 return OptSize && !VT.isVector();
27520 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27521 TargetLowering::ArgListTy& Args) const {
27522 // The MCU psABI requires some arguments to be passed in-register.
27523 // For regular calls, the inreg arguments are marked by the front-end.
27524 // However, for compiler generated library calls, we have to patch this
27526 if (!Subtarget->isTargetMCU() || !Args.size())
27529 unsigned FreeRegs = 3;
27530 for (auto &Arg : Args) {
27531 // For library functions, we do not expect any fancy types.
27532 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27533 unsigned SizeInRegs = (Size + 31) / 32;
27534 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27537 Arg.isInReg = true;
27538 FreeRegs -= SizeInRegs;