1 //===-- llvm/CodeGen/GlobalISel/LegalizerHelper.cpp -----------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 /// \file This file implements the LegalizerHelper class to legalize
10 /// individual instructions and the LegalizeMachineIR wrapper pass for the
11 /// primary legalization.
13 //===----------------------------------------------------------------------===//
15 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
16 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
17 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
18 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/CodeGen/TargetInstrInfo.h"
21 #include "llvm/CodeGen/TargetLowering.h"
22 #include "llvm/CodeGen/TargetSubtargetInfo.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/MathExtras.h"
25 #include "llvm/Support/raw_ostream.h"
27 #define DEBUG_TYPE "legalizer"
30 using namespace LegalizeActions;
32 /// Try to break down \p OrigTy into \p NarrowTy sized pieces.
34 /// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy,
35 /// with any leftover piece as type \p LeftoverTy
37 /// Returns -1 in the first element of the pair if the breakdown is not
39 static std::pair<int, int>
40 getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) {
41 assert(!LeftoverTy.isValid() && "this is an out argument");
43 unsigned Size = OrigTy.getSizeInBits();
44 unsigned NarrowSize = NarrowTy.getSizeInBits();
45 unsigned NumParts = Size / NarrowSize;
46 unsigned LeftoverSize = Size - NumParts * NarrowSize;
47 assert(Size > NarrowSize);
49 if (LeftoverSize == 0)
52 if (NarrowTy.isVector()) {
53 unsigned EltSize = OrigTy.getScalarSizeInBits();
54 if (LeftoverSize % EltSize != 0)
56 LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize);
58 LeftoverTy = LLT::scalar(LeftoverSize);
61 int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits();
62 return std::make_pair(NumParts, NumLeftover);
65 LegalizerHelper::LegalizerHelper(MachineFunction &MF,
66 GISelChangeObserver &Observer,
67 MachineIRBuilder &Builder)
68 : MIRBuilder(Builder), MRI(MF.getRegInfo()),
69 LI(*MF.getSubtarget().getLegalizerInfo()), Observer(Observer) {
71 MIRBuilder.setChangeObserver(Observer);
74 LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI,
75 GISelChangeObserver &Observer,
77 : MIRBuilder(B), MRI(MF.getRegInfo()), LI(LI), Observer(Observer) {
79 MIRBuilder.setChangeObserver(Observer);
81 LegalizerHelper::LegalizeResult
82 LegalizerHelper::legalizeInstrStep(MachineInstr &MI) {
83 LLVM_DEBUG(dbgs() << "Legalizing: "; MI.print(dbgs()));
85 if (MI.getOpcode() == TargetOpcode::G_INTRINSIC ||
86 MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS)
87 return LI.legalizeIntrinsic(MI, MRI, MIRBuilder) ? Legalized
89 auto Step = LI.getAction(MI, MRI);
90 switch (Step.Action) {
92 LLVM_DEBUG(dbgs() << ".. Already legal\n");
95 LLVM_DEBUG(dbgs() << ".. Convert to libcall\n");
98 LLVM_DEBUG(dbgs() << ".. Narrow scalar\n");
99 return narrowScalar(MI, Step.TypeIdx, Step.NewType);
101 LLVM_DEBUG(dbgs() << ".. Widen scalar\n");
102 return widenScalar(MI, Step.TypeIdx, Step.NewType);
104 LLVM_DEBUG(dbgs() << ".. Lower\n");
105 return lower(MI, Step.TypeIdx, Step.NewType);
107 LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n");
108 return fewerElementsVector(MI, Step.TypeIdx, Step.NewType);
110 LLVM_DEBUG(dbgs() << ".. Increase number of elements\n");
111 return moreElementsVector(MI, Step.TypeIdx, Step.NewType);
113 LLVM_DEBUG(dbgs() << ".. Custom legalization\n");
114 return LI.legalizeCustom(MI, MRI, MIRBuilder, Observer) ? Legalized
117 LLVM_DEBUG(dbgs() << ".. Unable to legalize\n");
118 return UnableToLegalize;
122 void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts,
123 SmallVectorImpl<Register> &VRegs) {
124 for (int i = 0; i < NumParts; ++i)
125 VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
126 MIRBuilder.buildUnmerge(VRegs, Reg);
129 bool LegalizerHelper::extractParts(Register Reg, LLT RegTy,
130 LLT MainTy, LLT &LeftoverTy,
131 SmallVectorImpl<Register> &VRegs,
132 SmallVectorImpl<Register> &LeftoverRegs) {
133 assert(!LeftoverTy.isValid() && "this is an out argument");
135 unsigned RegSize = RegTy.getSizeInBits();
136 unsigned MainSize = MainTy.getSizeInBits();
137 unsigned NumParts = RegSize / MainSize;
138 unsigned LeftoverSize = RegSize - NumParts * MainSize;
140 // Use an unmerge when possible.
141 if (LeftoverSize == 0) {
142 for (unsigned I = 0; I < NumParts; ++I)
143 VRegs.push_back(MRI.createGenericVirtualRegister(MainTy));
144 MIRBuilder.buildUnmerge(VRegs, Reg);
148 if (MainTy.isVector()) {
149 unsigned EltSize = MainTy.getScalarSizeInBits();
150 if (LeftoverSize % EltSize != 0)
152 LeftoverTy = LLT::scalarOrVector(LeftoverSize / EltSize, EltSize);
154 LeftoverTy = LLT::scalar(LeftoverSize);
157 // For irregular sizes, extract the individual parts.
158 for (unsigned I = 0; I != NumParts; ++I) {
159 Register NewReg = MRI.createGenericVirtualRegister(MainTy);
160 VRegs.push_back(NewReg);
161 MIRBuilder.buildExtract(NewReg, Reg, MainSize * I);
164 for (unsigned Offset = MainSize * NumParts; Offset < RegSize;
165 Offset += LeftoverSize) {
166 Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy);
167 LeftoverRegs.push_back(NewReg);
168 MIRBuilder.buildExtract(NewReg, Reg, Offset);
174 void LegalizerHelper::insertParts(Register DstReg,
175 LLT ResultTy, LLT PartTy,
176 ArrayRef<Register> PartRegs,
178 ArrayRef<Register> LeftoverRegs) {
179 if (!LeftoverTy.isValid()) {
180 assert(LeftoverRegs.empty());
182 if (!ResultTy.isVector()) {
183 MIRBuilder.buildMerge(DstReg, PartRegs);
187 if (PartTy.isVector())
188 MIRBuilder.buildConcatVectors(DstReg, PartRegs);
190 MIRBuilder.buildBuildVector(DstReg, PartRegs);
194 unsigned PartSize = PartTy.getSizeInBits();
195 unsigned LeftoverPartSize = LeftoverTy.getSizeInBits();
197 Register CurResultReg = MRI.createGenericVirtualRegister(ResultTy);
198 MIRBuilder.buildUndef(CurResultReg);
201 for (Register PartReg : PartRegs) {
202 Register NewResultReg = MRI.createGenericVirtualRegister(ResultTy);
203 MIRBuilder.buildInsert(NewResultReg, CurResultReg, PartReg, Offset);
204 CurResultReg = NewResultReg;
208 for (unsigned I = 0, E = LeftoverRegs.size(); I != E; ++I) {
209 // Use the original output register for the final insert to avoid a copy.
210 Register NewResultReg = (I + 1 == E) ?
211 DstReg : MRI.createGenericVirtualRegister(ResultTy);
213 MIRBuilder.buildInsert(NewResultReg, CurResultReg, LeftoverRegs[I], Offset);
214 CurResultReg = NewResultReg;
215 Offset += LeftoverPartSize;
219 static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
221 case TargetOpcode::G_SDIV:
222 assert((Size == 32 || Size == 64) && "Unsupported size");
223 return Size == 64 ? RTLIB::SDIV_I64 : RTLIB::SDIV_I32;
224 case TargetOpcode::G_UDIV:
225 assert((Size == 32 || Size == 64) && "Unsupported size");
226 return Size == 64 ? RTLIB::UDIV_I64 : RTLIB::UDIV_I32;
227 case TargetOpcode::G_SREM:
228 assert((Size == 32 || Size == 64) && "Unsupported size");
229 return Size == 64 ? RTLIB::SREM_I64 : RTLIB::SREM_I32;
230 case TargetOpcode::G_UREM:
231 assert((Size == 32 || Size == 64) && "Unsupported size");
232 return Size == 64 ? RTLIB::UREM_I64 : RTLIB::UREM_I32;
233 case TargetOpcode::G_CTLZ_ZERO_UNDEF:
234 assert(Size == 32 && "Unsupported size");
235 return RTLIB::CTLZ_I32;
236 case TargetOpcode::G_FADD:
237 assert((Size == 32 || Size == 64) && "Unsupported size");
238 return Size == 64 ? RTLIB::ADD_F64 : RTLIB::ADD_F32;
239 case TargetOpcode::G_FSUB:
240 assert((Size == 32 || Size == 64) && "Unsupported size");
241 return Size == 64 ? RTLIB::SUB_F64 : RTLIB::SUB_F32;
242 case TargetOpcode::G_FMUL:
243 assert((Size == 32 || Size == 64) && "Unsupported size");
244 return Size == 64 ? RTLIB::MUL_F64 : RTLIB::MUL_F32;
245 case TargetOpcode::G_FDIV:
246 assert((Size == 32 || Size == 64) && "Unsupported size");
247 return Size == 64 ? RTLIB::DIV_F64 : RTLIB::DIV_F32;
248 case TargetOpcode::G_FEXP:
249 assert((Size == 32 || Size == 64) && "Unsupported size");
250 return Size == 64 ? RTLIB::EXP_F64 : RTLIB::EXP_F32;
251 case TargetOpcode::G_FEXP2:
252 assert((Size == 32 || Size == 64) && "Unsupported size");
253 return Size == 64 ? RTLIB::EXP2_F64 : RTLIB::EXP2_F32;
254 case TargetOpcode::G_FREM:
255 return Size == 64 ? RTLIB::REM_F64 : RTLIB::REM_F32;
256 case TargetOpcode::G_FPOW:
257 return Size == 64 ? RTLIB::POW_F64 : RTLIB::POW_F32;
258 case TargetOpcode::G_FMA:
259 assert((Size == 32 || Size == 64) && "Unsupported size");
260 return Size == 64 ? RTLIB::FMA_F64 : RTLIB::FMA_F32;
261 case TargetOpcode::G_FSIN:
262 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
263 return Size == 128 ? RTLIB::SIN_F128
264 : Size == 64 ? RTLIB::SIN_F64 : RTLIB::SIN_F32;
265 case TargetOpcode::G_FCOS:
266 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
267 return Size == 128 ? RTLIB::COS_F128
268 : Size == 64 ? RTLIB::COS_F64 : RTLIB::COS_F32;
269 case TargetOpcode::G_FLOG10:
270 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
271 return Size == 128 ? RTLIB::LOG10_F128
272 : Size == 64 ? RTLIB::LOG10_F64 : RTLIB::LOG10_F32;
273 case TargetOpcode::G_FLOG:
274 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
275 return Size == 128 ? RTLIB::LOG_F128
276 : Size == 64 ? RTLIB::LOG_F64 : RTLIB::LOG_F32;
277 case TargetOpcode::G_FLOG2:
278 assert((Size == 32 || Size == 64 || Size == 128) && "Unsupported size");
279 return Size == 128 ? RTLIB::LOG2_F128
280 : Size == 64 ? RTLIB::LOG2_F64 : RTLIB::LOG2_F32;
281 case TargetOpcode::G_FCEIL:
282 assert((Size == 32 || Size == 64) && "Unsupported size");
283 return Size == 64 ? RTLIB::CEIL_F64 : RTLIB::CEIL_F32;
284 case TargetOpcode::G_FFLOOR:
285 assert((Size == 32 || Size == 64) && "Unsupported size");
286 return Size == 64 ? RTLIB::FLOOR_F64 : RTLIB::FLOOR_F32;
288 llvm_unreachable("Unknown libcall function");
291 LegalizerHelper::LegalizeResult
292 llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall,
293 const CallLowering::ArgInfo &Result,
294 ArrayRef<CallLowering::ArgInfo> Args) {
295 auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
296 auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
297 const char *Name = TLI.getLibcallName(Libcall);
299 MIRBuilder.getMF().getFrameInfo().setHasCalls(true);
300 if (!CLI.lowerCall(MIRBuilder, TLI.getLibcallCallingConv(Libcall),
301 MachineOperand::CreateES(Name), Result, Args))
302 return LegalizerHelper::UnableToLegalize;
304 return LegalizerHelper::Legalized;
307 // Useful for libcalls where all operands have the same type.
308 static LegalizerHelper::LegalizeResult
309 simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size,
311 auto Libcall = getRTLibDesc(MI.getOpcode(), Size);
313 SmallVector<CallLowering::ArgInfo, 3> Args;
314 for (unsigned i = 1; i < MI.getNumOperands(); i++)
315 Args.push_back({MI.getOperand(i).getReg(), OpType});
316 return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), OpType},
320 static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType,
322 auto ToMVT = MVT::getVT(ToType);
323 auto FromMVT = MVT::getVT(FromType);
326 case TargetOpcode::G_FPEXT:
327 return RTLIB::getFPEXT(FromMVT, ToMVT);
328 case TargetOpcode::G_FPTRUNC:
329 return RTLIB::getFPROUND(FromMVT, ToMVT);
330 case TargetOpcode::G_FPTOSI:
331 return RTLIB::getFPTOSINT(FromMVT, ToMVT);
332 case TargetOpcode::G_FPTOUI:
333 return RTLIB::getFPTOUINT(FromMVT, ToMVT);
334 case TargetOpcode::G_SITOFP:
335 return RTLIB::getSINTTOFP(FromMVT, ToMVT);
336 case TargetOpcode::G_UITOFP:
337 return RTLIB::getUINTTOFP(FromMVT, ToMVT);
339 llvm_unreachable("Unsupported libcall function");
342 static LegalizerHelper::LegalizeResult
343 conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType,
345 RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType);
346 return createLibcall(MIRBuilder, Libcall, {MI.getOperand(0).getReg(), ToType},
347 {{MI.getOperand(1).getReg(), FromType}});
350 LegalizerHelper::LegalizeResult
351 LegalizerHelper::libcall(MachineInstr &MI) {
352 LLT LLTy = MRI.getType(MI.getOperand(0).getReg());
353 unsigned Size = LLTy.getSizeInBits();
354 auto &Ctx = MIRBuilder.getMF().getFunction().getContext();
356 MIRBuilder.setInstr(MI);
358 switch (MI.getOpcode()) {
360 return UnableToLegalize;
361 case TargetOpcode::G_SDIV:
362 case TargetOpcode::G_UDIV:
363 case TargetOpcode::G_SREM:
364 case TargetOpcode::G_UREM:
365 case TargetOpcode::G_CTLZ_ZERO_UNDEF: {
366 Type *HLTy = IntegerType::get(Ctx, Size);
367 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
368 if (Status != Legalized)
372 case TargetOpcode::G_FADD:
373 case TargetOpcode::G_FSUB:
374 case TargetOpcode::G_FMUL:
375 case TargetOpcode::G_FDIV:
376 case TargetOpcode::G_FMA:
377 case TargetOpcode::G_FPOW:
378 case TargetOpcode::G_FREM:
379 case TargetOpcode::G_FCOS:
380 case TargetOpcode::G_FSIN:
381 case TargetOpcode::G_FLOG10:
382 case TargetOpcode::G_FLOG:
383 case TargetOpcode::G_FLOG2:
384 case TargetOpcode::G_FEXP:
385 case TargetOpcode::G_FEXP2:
386 case TargetOpcode::G_FCEIL:
387 case TargetOpcode::G_FFLOOR: {
389 LLVM_DEBUG(dbgs() << "Size " << Size << " too large to legalize.\n");
390 return UnableToLegalize;
392 Type *HLTy = Size == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx);
393 auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
394 if (Status != Legalized)
398 case TargetOpcode::G_FPEXT: {
399 // FIXME: Support other floating point types (half, fp128 etc)
400 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
401 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
402 if (ToSize != 64 || FromSize != 32)
403 return UnableToLegalize;
404 LegalizeResult Status = conversionLibcall(
405 MI, MIRBuilder, Type::getDoubleTy(Ctx), Type::getFloatTy(Ctx));
406 if (Status != Legalized)
410 case TargetOpcode::G_FPTRUNC: {
411 // FIXME: Support other floating point types (half, fp128 etc)
412 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
413 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
414 if (ToSize != 32 || FromSize != 64)
415 return UnableToLegalize;
416 LegalizeResult Status = conversionLibcall(
417 MI, MIRBuilder, Type::getFloatTy(Ctx), Type::getDoubleTy(Ctx));
418 if (Status != Legalized)
422 case TargetOpcode::G_FPTOSI:
423 case TargetOpcode::G_FPTOUI: {
424 // FIXME: Support other types
425 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
426 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
427 if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64))
428 return UnableToLegalize;
429 LegalizeResult Status = conversionLibcall(
431 ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx),
432 FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx));
433 if (Status != Legalized)
437 case TargetOpcode::G_SITOFP:
438 case TargetOpcode::G_UITOFP: {
439 // FIXME: Support other types
440 unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
441 unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
442 if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64))
443 return UnableToLegalize;
444 LegalizeResult Status = conversionLibcall(
446 ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx),
447 FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx));
448 if (Status != Legalized)
454 MI.eraseFromParent();
458 LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
461 MIRBuilder.setInstr(MI);
463 uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
464 uint64_t NarrowSize = NarrowTy.getSizeInBits();
466 switch (MI.getOpcode()) {
468 return UnableToLegalize;
469 case TargetOpcode::G_IMPLICIT_DEF: {
470 // FIXME: add support for when SizeOp0 isn't an exact multiple of
472 if (SizeOp0 % NarrowSize != 0)
473 return UnableToLegalize;
474 int NumParts = SizeOp0 / NarrowSize;
476 SmallVector<Register, 2> DstRegs;
477 for (int i = 0; i < NumParts; ++i)
479 MIRBuilder.buildUndef(NarrowTy)->getOperand(0).getReg());
481 Register DstReg = MI.getOperand(0).getReg();
482 if(MRI.getType(DstReg).isVector())
483 MIRBuilder.buildBuildVector(DstReg, DstRegs);
485 MIRBuilder.buildMerge(DstReg, DstRegs);
486 MI.eraseFromParent();
489 case TargetOpcode::G_CONSTANT: {
490 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
491 const APInt &Val = MI.getOperand(1).getCImm()->getValue();
492 unsigned TotalSize = Ty.getSizeInBits();
493 unsigned NarrowSize = NarrowTy.getSizeInBits();
494 int NumParts = TotalSize / NarrowSize;
496 SmallVector<Register, 4> PartRegs;
497 for (int I = 0; I != NumParts; ++I) {
498 unsigned Offset = I * NarrowSize;
499 auto K = MIRBuilder.buildConstant(NarrowTy,
500 Val.lshr(Offset).trunc(NarrowSize));
501 PartRegs.push_back(K.getReg(0));
505 unsigned LeftoverBits = TotalSize - NumParts * NarrowSize;
506 SmallVector<Register, 1> LeftoverRegs;
507 if (LeftoverBits != 0) {
508 LeftoverTy = LLT::scalar(LeftoverBits);
509 auto K = MIRBuilder.buildConstant(
511 Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits));
512 LeftoverRegs.push_back(K.getReg(0));
515 insertParts(MI.getOperand(0).getReg(),
516 Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs);
518 MI.eraseFromParent();
521 case TargetOpcode::G_ADD: {
522 // FIXME: add support for when SizeOp0 isn't an exact multiple of
524 if (SizeOp0 % NarrowSize != 0)
525 return UnableToLegalize;
526 // Expand in terms of carry-setting/consuming G_ADDE instructions.
527 int NumParts = SizeOp0 / NarrowTy.getSizeInBits();
529 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs;
530 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
531 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
533 Register CarryIn = MRI.createGenericVirtualRegister(LLT::scalar(1));
534 MIRBuilder.buildConstant(CarryIn, 0);
536 for (int i = 0; i < NumParts; ++i) {
537 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
538 Register CarryOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
540 MIRBuilder.buildUAdde(DstReg, CarryOut, Src1Regs[i],
541 Src2Regs[i], CarryIn);
543 DstRegs.push_back(DstReg);
546 Register DstReg = MI.getOperand(0).getReg();
547 if(MRI.getType(DstReg).isVector())
548 MIRBuilder.buildBuildVector(DstReg, DstRegs);
550 MIRBuilder.buildMerge(DstReg, DstRegs);
551 MI.eraseFromParent();
554 case TargetOpcode::G_SUB: {
555 // FIXME: add support for when SizeOp0 isn't an exact multiple of
557 if (SizeOp0 % NarrowSize != 0)
558 return UnableToLegalize;
560 int NumParts = SizeOp0 / NarrowTy.getSizeInBits();
562 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs;
563 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src1Regs);
564 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src2Regs);
566 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
567 Register BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
568 MIRBuilder.buildInstr(TargetOpcode::G_USUBO, {DstReg, BorrowOut},
569 {Src1Regs[0], Src2Regs[0]});
570 DstRegs.push_back(DstReg);
571 Register BorrowIn = BorrowOut;
572 for (int i = 1; i < NumParts; ++i) {
573 DstReg = MRI.createGenericVirtualRegister(NarrowTy);
574 BorrowOut = MRI.createGenericVirtualRegister(LLT::scalar(1));
576 MIRBuilder.buildInstr(TargetOpcode::G_USUBE, {DstReg, BorrowOut},
577 {Src1Regs[i], Src2Regs[i], BorrowIn});
579 DstRegs.push_back(DstReg);
580 BorrowIn = BorrowOut;
582 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs);
583 MI.eraseFromParent();
586 case TargetOpcode::G_MUL:
587 case TargetOpcode::G_UMULH:
588 return narrowScalarMul(MI, NarrowTy);
589 case TargetOpcode::G_EXTRACT:
590 return narrowScalarExtract(MI, TypeIdx, NarrowTy);
591 case TargetOpcode::G_INSERT:
592 return narrowScalarInsert(MI, TypeIdx, NarrowTy);
593 case TargetOpcode::G_LOAD: {
594 const auto &MMO = **MI.memoperands_begin();
595 Register DstReg = MI.getOperand(0).getReg();
596 LLT DstTy = MRI.getType(DstReg);
597 if (DstTy.isVector())
598 return UnableToLegalize;
600 if (8 * MMO.getSize() != DstTy.getSizeInBits()) {
601 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
602 auto &MMO = **MI.memoperands_begin();
603 MIRBuilder.buildLoad(TmpReg, MI.getOperand(1).getReg(), MMO);
604 MIRBuilder.buildAnyExt(DstReg, TmpReg);
605 MI.eraseFromParent();
609 return reduceLoadStoreWidth(MI, TypeIdx, NarrowTy);
611 case TargetOpcode::G_ZEXTLOAD:
612 case TargetOpcode::G_SEXTLOAD: {
613 bool ZExt = MI.getOpcode() == TargetOpcode::G_ZEXTLOAD;
614 Register DstReg = MI.getOperand(0).getReg();
615 Register PtrReg = MI.getOperand(1).getReg();
617 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
618 auto &MMO = **MI.memoperands_begin();
619 if (MMO.getSizeInBits() == NarrowSize) {
620 MIRBuilder.buildLoad(TmpReg, PtrReg, MMO);
622 unsigned ExtLoad = ZExt ? TargetOpcode::G_ZEXTLOAD
623 : TargetOpcode::G_SEXTLOAD;
624 MIRBuilder.buildInstr(ExtLoad)
627 .addMemOperand(&MMO);
631 MIRBuilder.buildZExt(DstReg, TmpReg);
633 MIRBuilder.buildSExt(DstReg, TmpReg);
635 MI.eraseFromParent();
638 case TargetOpcode::G_STORE: {
639 const auto &MMO = **MI.memoperands_begin();
641 Register SrcReg = MI.getOperand(0).getReg();
642 LLT SrcTy = MRI.getType(SrcReg);
643 if (SrcTy.isVector())
644 return UnableToLegalize;
646 int NumParts = SizeOp0 / NarrowSize;
647 unsigned HandledSize = NumParts * NarrowTy.getSizeInBits();
648 unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize;
649 if (SrcTy.isVector() && LeftoverBits != 0)
650 return UnableToLegalize;
652 if (8 * MMO.getSize() != SrcTy.getSizeInBits()) {
653 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
654 auto &MMO = **MI.memoperands_begin();
655 MIRBuilder.buildTrunc(TmpReg, SrcReg);
656 MIRBuilder.buildStore(TmpReg, MI.getOperand(1).getReg(), MMO);
657 MI.eraseFromParent();
661 return reduceLoadStoreWidth(MI, 0, NarrowTy);
663 case TargetOpcode::G_SELECT:
664 return narrowScalarSelect(MI, TypeIdx, NarrowTy);
665 case TargetOpcode::G_AND:
666 case TargetOpcode::G_OR:
667 case TargetOpcode::G_XOR: {
668 // Legalize bitwise operation:
669 // A = BinOp<Ty> B, C
671 // B1, ..., BN = G_UNMERGE_VALUES B
672 // C1, ..., CN = G_UNMERGE_VALUES C
673 // A1 = BinOp<Ty/N> B1, C2
675 // AN = BinOp<Ty/N> BN, CN
676 // A = G_MERGE_VALUES A1, ..., AN
677 return narrowScalarBasic(MI, TypeIdx, NarrowTy);
679 case TargetOpcode::G_SHL:
680 case TargetOpcode::G_LSHR:
681 case TargetOpcode::G_ASHR:
682 return narrowScalarShift(MI, TypeIdx, NarrowTy);
683 case TargetOpcode::G_CTLZ:
684 case TargetOpcode::G_CTLZ_ZERO_UNDEF:
685 case TargetOpcode::G_CTTZ:
686 case TargetOpcode::G_CTTZ_ZERO_UNDEF:
687 case TargetOpcode::G_CTPOP:
689 return UnableToLegalize; // TODO
691 Observer.changingInstr(MI);
692 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
693 Observer.changedInstr(MI);
695 case TargetOpcode::G_INTTOPTR:
697 return UnableToLegalize;
699 Observer.changingInstr(MI);
700 narrowScalarSrc(MI, NarrowTy, 1);
701 Observer.changedInstr(MI);
703 case TargetOpcode::G_PTRTOINT:
705 return UnableToLegalize;
707 Observer.changingInstr(MI);
708 narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
709 Observer.changedInstr(MI);
711 case TargetOpcode::G_PHI: {
712 unsigned NumParts = SizeOp0 / NarrowSize;
713 SmallVector<Register, 2> DstRegs;
714 SmallVector<SmallVector<Register, 2>, 2> SrcRegs;
715 DstRegs.resize(NumParts);
716 SrcRegs.resize(MI.getNumOperands() / 2);
717 Observer.changingInstr(MI);
718 for (unsigned i = 1; i < MI.getNumOperands(); i += 2) {
719 MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB();
720 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
721 extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts,
724 MachineBasicBlock &MBB = *MI.getParent();
725 MIRBuilder.setInsertPt(MBB, MI);
726 for (unsigned i = 0; i < NumParts; ++i) {
727 DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy);
728 MachineInstrBuilder MIB =
729 MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]);
730 for (unsigned j = 1; j < MI.getNumOperands(); j += 2)
731 MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1));
733 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI());
734 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), DstRegs);
735 Observer.changedInstr(MI);
736 MI.eraseFromParent();
739 case TargetOpcode::G_EXTRACT_VECTOR_ELT:
740 case TargetOpcode::G_INSERT_VECTOR_ELT: {
742 return UnableToLegalize;
744 int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3;
745 Observer.changingInstr(MI);
746 narrowScalarSrc(MI, NarrowTy, OpIdx);
747 Observer.changedInstr(MI);
750 case TargetOpcode::G_ICMP: {
751 uint64_t SrcSize = MRI.getType(MI.getOperand(2).getReg()).getSizeInBits();
752 if (NarrowSize * 2 != SrcSize)
753 return UnableToLegalize;
755 Observer.changingInstr(MI);
756 Register LHSL = MRI.createGenericVirtualRegister(NarrowTy);
757 Register LHSH = MRI.createGenericVirtualRegister(NarrowTy);
758 MIRBuilder.buildUnmerge({LHSL, LHSH}, MI.getOperand(2).getReg());
760 Register RHSL = MRI.createGenericVirtualRegister(NarrowTy);
761 Register RHSH = MRI.createGenericVirtualRegister(NarrowTy);
762 MIRBuilder.buildUnmerge({RHSL, RHSH}, MI.getOperand(3).getReg());
764 CmpInst::Predicate Pred =
765 static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
767 if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) {
768 MachineInstrBuilder XorL = MIRBuilder.buildXor(NarrowTy, LHSL, RHSL);
769 MachineInstrBuilder XorH = MIRBuilder.buildXor(NarrowTy, LHSH, RHSH);
770 MachineInstrBuilder Or = MIRBuilder.buildOr(NarrowTy, XorL, XorH);
771 MachineInstrBuilder Zero = MIRBuilder.buildConstant(NarrowTy, 0);
772 MIRBuilder.buildICmp(Pred, MI.getOperand(0).getReg(), Or, Zero);
774 const LLT s1 = LLT::scalar(1);
775 MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, s1, LHSH, RHSH);
776 MachineInstrBuilder CmpHEQ =
777 MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, s1, LHSH, RHSH);
778 MachineInstrBuilder CmpLU = MIRBuilder.buildICmp(
779 ICmpInst::getUnsignedPredicate(Pred), s1, LHSL, RHSL);
780 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), CmpHEQ, CmpLU, CmpH);
782 Observer.changedInstr(MI);
783 MI.eraseFromParent();
789 void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy,
790 unsigned OpIdx, unsigned ExtOpcode) {
791 MachineOperand &MO = MI.getOperand(OpIdx);
792 auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO.getReg()});
793 MO.setReg(ExtB->getOperand(0).getReg());
796 void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy,
798 MachineOperand &MO = MI.getOperand(OpIdx);
799 auto ExtB = MIRBuilder.buildInstr(TargetOpcode::G_TRUNC, {NarrowTy},
801 MO.setReg(ExtB->getOperand(0).getReg());
804 void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy,
805 unsigned OpIdx, unsigned TruncOpcode) {
806 MachineOperand &MO = MI.getOperand(OpIdx);
807 Register DstExt = MRI.createGenericVirtualRegister(WideTy);
808 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
809 MIRBuilder.buildInstr(TruncOpcode, {MO.getReg()}, {DstExt});
813 void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy,
814 unsigned OpIdx, unsigned ExtOpcode) {
815 MachineOperand &MO = MI.getOperand(OpIdx);
816 Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy);
817 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
818 MIRBuilder.buildInstr(ExtOpcode, {MO.getReg()}, {DstTrunc});
822 void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy,
824 MachineOperand &MO = MI.getOperand(OpIdx);
825 Register DstExt = MRI.createGenericVirtualRegister(WideTy);
826 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
827 MIRBuilder.buildExtract(MO.getReg(), DstExt, 0);
831 void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy,
833 MachineOperand &MO = MI.getOperand(OpIdx);
835 LLT OldTy = MRI.getType(MO.getReg());
836 unsigned OldElts = OldTy.getNumElements();
837 unsigned NewElts = MoreTy.getNumElements();
839 unsigned NumParts = NewElts / OldElts;
841 // Use concat_vectors if the result is a multiple of the number of elements.
842 if (NumParts * OldElts == NewElts) {
843 SmallVector<Register, 8> Parts;
844 Parts.push_back(MO.getReg());
846 Register ImpDef = MIRBuilder.buildUndef(OldTy).getReg(0);
847 for (unsigned I = 1; I != NumParts; ++I)
848 Parts.push_back(ImpDef);
850 auto Concat = MIRBuilder.buildConcatVectors(MoreTy, Parts);
851 MO.setReg(Concat.getReg(0));
855 Register MoreReg = MRI.createGenericVirtualRegister(MoreTy);
856 Register ImpDef = MIRBuilder.buildUndef(MoreTy).getReg(0);
857 MIRBuilder.buildInsert(MoreReg, ImpDef, MO.getReg(), 0);
861 LegalizerHelper::LegalizeResult
862 LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx,
865 return UnableToLegalize;
867 Register DstReg = MI.getOperand(0).getReg();
868 LLT DstTy = MRI.getType(DstReg);
869 if (DstTy.isVector())
870 return UnableToLegalize;
872 Register Src1 = MI.getOperand(1).getReg();
873 LLT SrcTy = MRI.getType(Src1);
874 const int DstSize = DstTy.getSizeInBits();
875 const int SrcSize = SrcTy.getSizeInBits();
876 const int WideSize = WideTy.getSizeInBits();
877 const int NumMerge = (DstSize + WideSize - 1) / WideSize;
879 unsigned NumOps = MI.getNumOperands();
880 unsigned NumSrc = MI.getNumOperands() - 1;
881 unsigned PartSize = DstTy.getSizeInBits() / NumSrc;
883 if (WideSize >= DstSize) {
884 // Directly pack the bits in the target type.
885 Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1).getReg(0);
887 for (unsigned I = 2; I != NumOps; ++I) {
888 const unsigned Offset = (I - 1) * PartSize;
890 Register SrcReg = MI.getOperand(I).getReg();
891 assert(MRI.getType(SrcReg) == LLT::scalar(PartSize));
893 auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg);
895 Register NextResult = I + 1 == NumOps && WideSize == DstSize ? DstReg :
896 MRI.createGenericVirtualRegister(WideTy);
898 auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset);
899 auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt);
900 MIRBuilder.buildOr(NextResult, ResultReg, Shl);
901 ResultReg = NextResult;
904 if (WideSize > DstSize)
905 MIRBuilder.buildTrunc(DstReg, ResultReg);
907 MI.eraseFromParent();
911 // Unmerge the original values to the GCD type, and recombine to the next
912 // multiple greater than the original type.
914 // %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6
915 // %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0
916 // %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1
917 // %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2
918 // %10:_(s6) = G_MERGE_VALUES %4, %5, %6
919 // %11:_(s6) = G_MERGE_VALUES %7, %8, %9
920 // %12:_(s12) = G_MERGE_VALUES %10, %11
922 // Padding with undef if necessary:
924 // %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6
925 // %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0
926 // %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1
927 // %7:_(s2) = G_IMPLICIT_DEF
928 // %8:_(s6) = G_MERGE_VALUES %3, %4, %5
929 // %9:_(s6) = G_MERGE_VALUES %6, %7, %7
930 // %10:_(s12) = G_MERGE_VALUES %8, %9
932 const int GCD = greatestCommonDivisor(SrcSize, WideSize);
933 LLT GCDTy = LLT::scalar(GCD);
935 SmallVector<Register, 8> Parts;
936 SmallVector<Register, 8> NewMergeRegs;
937 SmallVector<Register, 8> Unmerges;
938 LLT WideDstTy = LLT::scalar(NumMerge * WideSize);
940 // Decompose the original operands if they don't evenly divide.
941 for (int I = 1, E = MI.getNumOperands(); I != E; ++I) {
942 Register SrcReg = MI.getOperand(I).getReg();
943 if (GCD == SrcSize) {
944 Unmerges.push_back(SrcReg);
946 auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg);
947 for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J)
948 Unmerges.push_back(Unmerge.getReg(J));
952 // Pad with undef to the next size that is a multiple of the requested size.
953 if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) {
954 Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0);
955 for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I)
956 Unmerges.push_back(UndefReg);
959 const int PartsPerGCD = WideSize / GCD;
961 // Build merges of each piece.
962 ArrayRef<Register> Slicer(Unmerges);
963 for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) {
964 auto Merge = MIRBuilder.buildMerge(WideTy, Slicer.take_front(PartsPerGCD));
965 NewMergeRegs.push_back(Merge.getReg(0));
968 // A truncate may be necessary if the requested type doesn't evenly divide the
969 // original result type.
970 if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) {
971 MIRBuilder.buildMerge(DstReg, NewMergeRegs);
973 auto FinalMerge = MIRBuilder.buildMerge(WideDstTy, NewMergeRegs);
974 MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0));
977 MI.eraseFromParent();
981 LegalizerHelper::LegalizeResult
982 LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx,
985 return UnableToLegalize;
987 unsigned NumDst = MI.getNumOperands() - 1;
988 Register SrcReg = MI.getOperand(NumDst).getReg();
989 LLT SrcTy = MRI.getType(SrcReg);
990 if (!SrcTy.isScalar())
991 return UnableToLegalize;
993 Register Dst0Reg = MI.getOperand(0).getReg();
994 LLT DstTy = MRI.getType(Dst0Reg);
995 if (!DstTy.isScalar())
996 return UnableToLegalize;
998 unsigned NewSrcSize = NumDst * WideTy.getSizeInBits();
999 LLT NewSrcTy = LLT::scalar(NewSrcSize);
1000 unsigned SizeDiff = WideTy.getSizeInBits() - DstTy.getSizeInBits();
1002 auto WideSrc = MIRBuilder.buildZExt(NewSrcTy, SrcReg);
1004 for (unsigned I = 1; I != NumDst; ++I) {
1005 auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, SizeDiff * I);
1006 auto Shl = MIRBuilder.buildShl(NewSrcTy, WideSrc, ShiftAmt);
1007 WideSrc = MIRBuilder.buildOr(NewSrcTy, WideSrc, Shl);
1010 Observer.changingInstr(MI);
1012 MI.getOperand(NumDst).setReg(WideSrc->getOperand(0).getReg());
1013 for (unsigned I = 0; I != NumDst; ++I)
1014 widenScalarDst(MI, WideTy, I);
1016 Observer.changedInstr(MI);
1021 LegalizerHelper::LegalizeResult
1022 LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx,
1024 Register DstReg = MI.getOperand(0).getReg();
1025 Register SrcReg = MI.getOperand(1).getReg();
1026 LLT SrcTy = MRI.getType(SrcReg);
1028 LLT DstTy = MRI.getType(DstReg);
1029 unsigned Offset = MI.getOperand(2).getImm();
1032 if (SrcTy.isVector() || DstTy.isVector())
1033 return UnableToLegalize;
1036 if (SrcTy.isPointer()) {
1037 // Extracts from pointers can be handled only if they are really just
1039 const DataLayout &DL = MIRBuilder.getDataLayout();
1040 if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace()))
1041 return UnableToLegalize;
1043 LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits());
1044 Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src);
1048 if (DstTy.isPointer())
1049 return UnableToLegalize;
1052 // Avoid a shift in the degenerate case.
1053 MIRBuilder.buildTrunc(DstReg,
1054 MIRBuilder.buildAnyExtOrTrunc(WideTy, Src));
1055 MI.eraseFromParent();
1059 // Do a shift in the source type.
1060 LLT ShiftTy = SrcTy;
1061 if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) {
1062 Src = MIRBuilder.buildAnyExt(WideTy, Src);
1064 } else if (WideTy.getSizeInBits() > SrcTy.getSizeInBits())
1065 return UnableToLegalize;
1067 auto LShr = MIRBuilder.buildLShr(
1068 ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset));
1069 MIRBuilder.buildTrunc(DstReg, LShr);
1070 MI.eraseFromParent();
1074 if (SrcTy.isScalar()) {
1075 Observer.changingInstr(MI);
1076 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1077 Observer.changedInstr(MI);
1081 if (!SrcTy.isVector())
1082 return UnableToLegalize;
1084 if (DstTy != SrcTy.getElementType())
1085 return UnableToLegalize;
1087 if (Offset % SrcTy.getScalarSizeInBits() != 0)
1088 return UnableToLegalize;
1090 Observer.changingInstr(MI);
1091 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1093 MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) *
1095 widenScalarDst(MI, WideTy.getScalarType(), 0);
1096 Observer.changedInstr(MI);
1100 LegalizerHelper::LegalizeResult
1101 LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx,
1104 return UnableToLegalize;
1105 Observer.changingInstr(MI);
1106 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1107 widenScalarDst(MI, WideTy);
1108 Observer.changedInstr(MI);
1112 LegalizerHelper::LegalizeResult
1113 LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
1114 MIRBuilder.setInstr(MI);
1116 switch (MI.getOpcode()) {
1118 return UnableToLegalize;
1119 case TargetOpcode::G_EXTRACT:
1120 return widenScalarExtract(MI, TypeIdx, WideTy);
1121 case TargetOpcode::G_INSERT:
1122 return widenScalarInsert(MI, TypeIdx, WideTy);
1123 case TargetOpcode::G_MERGE_VALUES:
1124 return widenScalarMergeValues(MI, TypeIdx, WideTy);
1125 case TargetOpcode::G_UNMERGE_VALUES:
1126 return widenScalarUnmergeValues(MI, TypeIdx, WideTy);
1127 case TargetOpcode::G_UADDO:
1128 case TargetOpcode::G_USUBO: {
1130 return UnableToLegalize; // TODO
1131 auto LHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy},
1132 {MI.getOperand(2).getReg()});
1133 auto RHSZext = MIRBuilder.buildInstr(TargetOpcode::G_ZEXT, {WideTy},
1134 {MI.getOperand(3).getReg()});
1135 unsigned Opcode = MI.getOpcode() == TargetOpcode::G_UADDO
1136 ? TargetOpcode::G_ADD
1137 : TargetOpcode::G_SUB;
1138 // Do the arithmetic in the larger type.
1139 auto NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSZext, RHSZext});
1140 LLT OrigTy = MRI.getType(MI.getOperand(0).getReg());
1141 APInt Mask = APInt::getAllOnesValue(OrigTy.getSizeInBits());
1142 auto AndOp = MIRBuilder.buildInstr(
1143 TargetOpcode::G_AND, {WideTy},
1144 {NewOp, MIRBuilder.buildConstant(WideTy, Mask.getZExtValue())});
1145 // There is no overflow if the AndOp is the same as NewOp.
1146 MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1).getReg(), NewOp,
1148 // Now trunc the NewOp to the original result.
1149 MIRBuilder.buildTrunc(MI.getOperand(0).getReg(), NewOp);
1150 MI.eraseFromParent();
1153 case TargetOpcode::G_CTTZ:
1154 case TargetOpcode::G_CTTZ_ZERO_UNDEF:
1155 case TargetOpcode::G_CTLZ:
1156 case TargetOpcode::G_CTLZ_ZERO_UNDEF:
1157 case TargetOpcode::G_CTPOP: {
1159 Observer.changingInstr(MI);
1160 widenScalarDst(MI, WideTy, 0);
1161 Observer.changedInstr(MI);
1165 Register SrcReg = MI.getOperand(1).getReg();
1167 // First ZEXT the input.
1168 auto MIBSrc = MIRBuilder.buildZExt(WideTy, SrcReg);
1169 LLT CurTy = MRI.getType(SrcReg);
1170 if (MI.getOpcode() == TargetOpcode::G_CTTZ) {
1171 // The count is the same in the larger type except if the original
1172 // value was zero. This can be handled by setting the bit just off
1173 // the top of the original type.
1175 APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits());
1176 MIBSrc = MIRBuilder.buildOr(
1177 WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit));
1180 // Perform the operation at the larger size.
1181 auto MIBNewOp = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy}, {MIBSrc});
1182 // This is already the correct result for CTPOP and CTTZs
1183 if (MI.getOpcode() == TargetOpcode::G_CTLZ ||
1184 MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) {
1185 // The correct result is NewOp - (Difference in widety and current ty).
1186 unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits();
1187 MIBNewOp = MIRBuilder.buildInstr(
1188 TargetOpcode::G_SUB, {WideTy},
1189 {MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff)});
1192 MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp);
1193 MI.eraseFromParent();
1196 case TargetOpcode::G_BSWAP: {
1197 Observer.changingInstr(MI);
1198 Register DstReg = MI.getOperand(0).getReg();
1200 Register ShrReg = MRI.createGenericVirtualRegister(WideTy);
1201 Register DstExt = MRI.createGenericVirtualRegister(WideTy);
1202 Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy);
1203 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1205 MI.getOperand(0).setReg(DstExt);
1207 MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
1209 LLT Ty = MRI.getType(DstReg);
1210 unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits();
1211 MIRBuilder.buildConstant(ShiftAmtReg, DiffBits);
1212 MIRBuilder.buildInstr(TargetOpcode::G_LSHR)
1215 .addUse(ShiftAmtReg);
1217 MIRBuilder.buildTrunc(DstReg, ShrReg);
1218 Observer.changedInstr(MI);
1221 case TargetOpcode::G_ADD:
1222 case TargetOpcode::G_AND:
1223 case TargetOpcode::G_MUL:
1224 case TargetOpcode::G_OR:
1225 case TargetOpcode::G_XOR:
1226 case TargetOpcode::G_SUB:
1227 // Perform operation at larger width (any extension is fines here, high bits
1228 // don't affect the result) and then truncate the result back to the
1230 Observer.changingInstr(MI);
1231 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1232 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
1233 widenScalarDst(MI, WideTy);
1234 Observer.changedInstr(MI);
1237 case TargetOpcode::G_SHL:
1238 Observer.changingInstr(MI);
1241 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
1242 widenScalarDst(MI, WideTy);
1244 assert(TypeIdx == 1);
1245 // The "number of bits to shift" operand must preserve its value as an
1246 // unsigned integer:
1247 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
1250 Observer.changedInstr(MI);
1253 case TargetOpcode::G_SDIV:
1254 case TargetOpcode::G_SREM:
1255 case TargetOpcode::G_SMIN:
1256 case TargetOpcode::G_SMAX:
1257 Observer.changingInstr(MI);
1258 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
1259 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
1260 widenScalarDst(MI, WideTy);
1261 Observer.changedInstr(MI);
1264 case TargetOpcode::G_ASHR:
1265 case TargetOpcode::G_LSHR:
1266 Observer.changingInstr(MI);
1269 unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ?
1270 TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
1272 widenScalarSrc(MI, WideTy, 1, CvtOp);
1273 widenScalarDst(MI, WideTy);
1275 assert(TypeIdx == 1);
1276 // The "number of bits to shift" operand must preserve its value as an
1277 // unsigned integer:
1278 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
1281 Observer.changedInstr(MI);
1283 case TargetOpcode::G_UDIV:
1284 case TargetOpcode::G_UREM:
1285 case TargetOpcode::G_UMIN:
1286 case TargetOpcode::G_UMAX:
1287 Observer.changingInstr(MI);
1288 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
1289 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
1290 widenScalarDst(MI, WideTy);
1291 Observer.changedInstr(MI);
1294 case TargetOpcode::G_SELECT:
1295 Observer.changingInstr(MI);
1297 // Perform operation at larger width (any extension is fine here, high
1298 // bits don't affect the result) and then truncate the result back to the
1300 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
1301 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
1302 widenScalarDst(MI, WideTy);
1304 bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector();
1305 // Explicit extension is required here since high bits affect the result.
1306 widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false));
1308 Observer.changedInstr(MI);
1311 case TargetOpcode::G_FPTOSI:
1312 case TargetOpcode::G_FPTOUI:
1314 return UnableToLegalize;
1315 Observer.changingInstr(MI);
1316 widenScalarDst(MI, WideTy);
1317 Observer.changedInstr(MI);
1320 case TargetOpcode::G_SITOFP:
1322 return UnableToLegalize;
1323 Observer.changingInstr(MI);
1324 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
1325 Observer.changedInstr(MI);
1328 case TargetOpcode::G_UITOFP:
1330 return UnableToLegalize;
1331 Observer.changingInstr(MI);
1332 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
1333 Observer.changedInstr(MI);
1336 case TargetOpcode::G_LOAD:
1337 case TargetOpcode::G_SEXTLOAD:
1338 case TargetOpcode::G_ZEXTLOAD:
1339 Observer.changingInstr(MI);
1340 widenScalarDst(MI, WideTy);
1341 Observer.changedInstr(MI);
1344 case TargetOpcode::G_STORE: {
1346 return UnableToLegalize;
1348 LLT Ty = MRI.getType(MI.getOperand(0).getReg());
1349 if (!isPowerOf2_32(Ty.getSizeInBits()))
1350 return UnableToLegalize;
1352 Observer.changingInstr(MI);
1354 unsigned ExtType = Ty.getScalarSizeInBits() == 1 ?
1355 TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT;
1356 widenScalarSrc(MI, WideTy, 0, ExtType);
1358 Observer.changedInstr(MI);
1361 case TargetOpcode::G_CONSTANT: {
1362 MachineOperand &SrcMO = MI.getOperand(1);
1363 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
1364 const APInt &Val = SrcMO.getCImm()->getValue().sext(WideTy.getSizeInBits());
1365 Observer.changingInstr(MI);
1366 SrcMO.setCImm(ConstantInt::get(Ctx, Val));
1368 widenScalarDst(MI, WideTy);
1369 Observer.changedInstr(MI);
1372 case TargetOpcode::G_FCONSTANT: {
1373 MachineOperand &SrcMO = MI.getOperand(1);
1374 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
1375 APFloat Val = SrcMO.getFPImm()->getValueAPF();
1377 switch (WideTy.getSizeInBits()) {
1379 Val.convert(APFloat::IEEEsingle(), APFloat::rmNearestTiesToEven,
1383 Val.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
1387 return UnableToLegalize;
1390 assert(!LosesInfo && "extend should always be lossless");
1392 Observer.changingInstr(MI);
1393 SrcMO.setFPImm(ConstantFP::get(Ctx, Val));
1395 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
1396 Observer.changedInstr(MI);
1399 case TargetOpcode::G_IMPLICIT_DEF: {
1400 Observer.changingInstr(MI);
1401 widenScalarDst(MI, WideTy);
1402 Observer.changedInstr(MI);
1405 case TargetOpcode::G_BRCOND:
1406 Observer.changingInstr(MI);
1407 widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false));
1408 Observer.changedInstr(MI);
1411 case TargetOpcode::G_FCMP:
1412 Observer.changingInstr(MI);
1414 widenScalarDst(MI, WideTy);
1416 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT);
1417 widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT);
1419 Observer.changedInstr(MI);
1422 case TargetOpcode::G_ICMP:
1423 Observer.changingInstr(MI);
1425 widenScalarDst(MI, WideTy);
1427 unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>(
1428 MI.getOperand(1).getPredicate()))
1429 ? TargetOpcode::G_SEXT
1430 : TargetOpcode::G_ZEXT;
1431 widenScalarSrc(MI, WideTy, 2, ExtOpcode);
1432 widenScalarSrc(MI, WideTy, 3, ExtOpcode);
1434 Observer.changedInstr(MI);
1437 case TargetOpcode::G_GEP:
1438 assert(TypeIdx == 1 && "unable to legalize pointer of GEP");
1439 Observer.changingInstr(MI);
1440 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
1441 Observer.changedInstr(MI);
1444 case TargetOpcode::G_PHI: {
1445 assert(TypeIdx == 0 && "Expecting only Idx 0");
1447 Observer.changingInstr(MI);
1448 for (unsigned I = 1; I < MI.getNumOperands(); I += 2) {
1449 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB();
1450 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
1451 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT);
1454 MachineBasicBlock &MBB = *MI.getParent();
1455 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI());
1456 widenScalarDst(MI, WideTy);
1457 Observer.changedInstr(MI);
1460 case TargetOpcode::G_EXTRACT_VECTOR_ELT: {
1462 Register VecReg = MI.getOperand(1).getReg();
1463 LLT VecTy = MRI.getType(VecReg);
1464 Observer.changingInstr(MI);
1466 widenScalarSrc(MI, LLT::vector(VecTy.getNumElements(),
1467 WideTy.getSizeInBits()),
1468 1, TargetOpcode::G_SEXT);
1470 widenScalarDst(MI, WideTy, 0);
1471 Observer.changedInstr(MI);
1476 return UnableToLegalize;
1477 Observer.changingInstr(MI);
1478 widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
1479 Observer.changedInstr(MI);
1482 case TargetOpcode::G_FADD:
1483 case TargetOpcode::G_FMUL:
1484 case TargetOpcode::G_FSUB:
1485 case TargetOpcode::G_FMA:
1486 case TargetOpcode::G_FNEG:
1487 case TargetOpcode::G_FABS:
1488 case TargetOpcode::G_FCANONICALIZE:
1489 case TargetOpcode::G_FMINNUM:
1490 case TargetOpcode::G_FMAXNUM:
1491 case TargetOpcode::G_FMINNUM_IEEE:
1492 case TargetOpcode::G_FMAXNUM_IEEE:
1493 case TargetOpcode::G_FMINIMUM:
1494 case TargetOpcode::G_FMAXIMUM:
1495 case TargetOpcode::G_FDIV:
1496 case TargetOpcode::G_FREM:
1497 case TargetOpcode::G_FCEIL:
1498 case TargetOpcode::G_FFLOOR:
1499 case TargetOpcode::G_FCOS:
1500 case TargetOpcode::G_FSIN:
1501 case TargetOpcode::G_FLOG10:
1502 case TargetOpcode::G_FLOG:
1503 case TargetOpcode::G_FLOG2:
1504 case TargetOpcode::G_FRINT:
1505 case TargetOpcode::G_FNEARBYINT:
1506 case TargetOpcode::G_FSQRT:
1507 case TargetOpcode::G_FEXP:
1508 case TargetOpcode::G_FEXP2:
1509 case TargetOpcode::G_FPOW:
1510 case TargetOpcode::G_INTRINSIC_TRUNC:
1511 case TargetOpcode::G_INTRINSIC_ROUND:
1512 assert(TypeIdx == 0);
1513 Observer.changingInstr(MI);
1515 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I)
1516 widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT);
1518 widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
1519 Observer.changedInstr(MI);
1521 case TargetOpcode::G_INTTOPTR:
1523 return UnableToLegalize;
1525 Observer.changingInstr(MI);
1526 widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
1527 Observer.changedInstr(MI);
1529 case TargetOpcode::G_PTRTOINT:
1531 return UnableToLegalize;
1533 Observer.changingInstr(MI);
1534 widenScalarDst(MI, WideTy, 0);
1535 Observer.changedInstr(MI);
1537 case TargetOpcode::G_BUILD_VECTOR: {
1538 Observer.changingInstr(MI);
1540 const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType();
1541 for (int I = 1, E = MI.getNumOperands(); I != E; ++I)
1542 widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT);
1544 // Avoid changing the result vector type if the source element type was
1547 auto &TII = *MI.getMF()->getSubtarget().getInstrInfo();
1548 MI.setDesc(TII.get(TargetOpcode::G_BUILD_VECTOR_TRUNC));
1550 widenScalarDst(MI, WideTy, 0);
1553 Observer.changedInstr(MI);
1559 LegalizerHelper::LegalizeResult
1560 LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
1561 using namespace TargetOpcode;
1562 MIRBuilder.setInstr(MI);
1564 switch(MI.getOpcode()) {
1566 return UnableToLegalize;
1567 case TargetOpcode::G_SREM:
1568 case TargetOpcode::G_UREM: {
1569 Register QuotReg = MRI.createGenericVirtualRegister(Ty);
1570 MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV)
1572 .addUse(MI.getOperand(1).getReg())
1573 .addUse(MI.getOperand(2).getReg());
1575 Register ProdReg = MRI.createGenericVirtualRegister(Ty);
1576 MIRBuilder.buildMul(ProdReg, QuotReg, MI.getOperand(2).getReg());
1577 MIRBuilder.buildSub(MI.getOperand(0).getReg(), MI.getOperand(1).getReg(),
1579 MI.eraseFromParent();
1582 case TargetOpcode::G_SMULO:
1583 case TargetOpcode::G_UMULO: {
1584 // Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the
1586 Register Res = MI.getOperand(0).getReg();
1587 Register Overflow = MI.getOperand(1).getReg();
1588 Register LHS = MI.getOperand(2).getReg();
1589 Register RHS = MI.getOperand(3).getReg();
1591 MIRBuilder.buildMul(Res, LHS, RHS);
1593 unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO
1594 ? TargetOpcode::G_SMULH
1595 : TargetOpcode::G_UMULH;
1597 Register HiPart = MRI.createGenericVirtualRegister(Ty);
1598 MIRBuilder.buildInstr(Opcode)
1603 Register Zero = MRI.createGenericVirtualRegister(Ty);
1604 MIRBuilder.buildConstant(Zero, 0);
1606 // For *signed* multiply, overflow is detected by checking:
1607 // (hi != (lo >> bitwidth-1))
1608 if (Opcode == TargetOpcode::G_SMULH) {
1609 Register Shifted = MRI.createGenericVirtualRegister(Ty);
1610 Register ShiftAmt = MRI.createGenericVirtualRegister(Ty);
1611 MIRBuilder.buildConstant(ShiftAmt, Ty.getSizeInBits() - 1);
1612 MIRBuilder.buildInstr(TargetOpcode::G_ASHR)
1616 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted);
1618 MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero);
1620 MI.eraseFromParent();
1623 case TargetOpcode::G_FNEG: {
1624 // TODO: Handle vector types once we are able to
1627 return UnableToLegalize;
1628 Register Res = MI.getOperand(0).getReg();
1630 LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
1631 switch (Ty.getSizeInBits()) {
1633 ZeroTy = Type::getHalfTy(Ctx);
1636 ZeroTy = Type::getFloatTy(Ctx);
1639 ZeroTy = Type::getDoubleTy(Ctx);
1642 ZeroTy = Type::getFP128Ty(Ctx);
1645 llvm_unreachable("unexpected floating-point type");
1647 ConstantFP &ZeroForNegation =
1648 *cast<ConstantFP>(ConstantFP::getZeroValueForNegation(ZeroTy));
1649 auto Zero = MIRBuilder.buildFConstant(Ty, ZeroForNegation);
1650 Register SubByReg = MI.getOperand(1).getReg();
1651 Register ZeroReg = Zero->getOperand(0).getReg();
1652 MIRBuilder.buildInstr(TargetOpcode::G_FSUB, {Res}, {ZeroReg, SubByReg},
1654 MI.eraseFromParent();
1657 case TargetOpcode::G_FSUB: {
1658 // Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)).
1659 // First, check if G_FNEG is marked as Lower. If so, we may
1660 // end up with an infinite loop as G_FSUB is used to legalize G_FNEG.
1661 if (LI.getAction({G_FNEG, {Ty}}).Action == Lower)
1662 return UnableToLegalize;
1663 Register Res = MI.getOperand(0).getReg();
1664 Register LHS = MI.getOperand(1).getReg();
1665 Register RHS = MI.getOperand(2).getReg();
1666 Register Neg = MRI.createGenericVirtualRegister(Ty);
1667 MIRBuilder.buildInstr(TargetOpcode::G_FNEG).addDef(Neg).addUse(RHS);
1668 MIRBuilder.buildInstr(TargetOpcode::G_FADD, {Res}, {LHS, Neg}, MI.getFlags());
1669 MI.eraseFromParent();
1672 case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
1673 Register OldValRes = MI.getOperand(0).getReg();
1674 Register SuccessRes = MI.getOperand(1).getReg();
1675 Register Addr = MI.getOperand(2).getReg();
1676 Register CmpVal = MI.getOperand(3).getReg();
1677 Register NewVal = MI.getOperand(4).getReg();
1678 MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal,
1679 **MI.memoperands_begin());
1680 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal);
1681 MI.eraseFromParent();
1684 case TargetOpcode::G_LOAD:
1685 case TargetOpcode::G_SEXTLOAD:
1686 case TargetOpcode::G_ZEXTLOAD: {
1687 // Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT
1688 Register DstReg = MI.getOperand(0).getReg();
1689 Register PtrReg = MI.getOperand(1).getReg();
1690 LLT DstTy = MRI.getType(DstReg);
1691 auto &MMO = **MI.memoperands_begin();
1693 if (DstTy.getSizeInBits() == MMO.getSize() /* in bytes */ * 8) {
1694 // In the case of G_LOAD, this was a non-extending load already and we're
1695 // about to lower to the same instruction.
1696 if (MI.getOpcode() == TargetOpcode::G_LOAD)
1697 return UnableToLegalize;
1698 MIRBuilder.buildLoad(DstReg, PtrReg, MMO);
1699 MI.eraseFromParent();
1703 if (DstTy.isScalar()) {
1705 MRI.createGenericVirtualRegister(LLT::scalar(MMO.getSizeInBits()));
1706 MIRBuilder.buildLoad(TmpReg, PtrReg, MMO);
1707 switch (MI.getOpcode()) {
1709 llvm_unreachable("Unexpected opcode");
1710 case TargetOpcode::G_LOAD:
1711 MIRBuilder.buildAnyExt(DstReg, TmpReg);
1713 case TargetOpcode::G_SEXTLOAD:
1714 MIRBuilder.buildSExt(DstReg, TmpReg);
1716 case TargetOpcode::G_ZEXTLOAD:
1717 MIRBuilder.buildZExt(DstReg, TmpReg);
1720 MI.eraseFromParent();
1724 return UnableToLegalize;
1726 case TargetOpcode::G_CTLZ_ZERO_UNDEF:
1727 case TargetOpcode::G_CTTZ_ZERO_UNDEF:
1728 case TargetOpcode::G_CTLZ:
1729 case TargetOpcode::G_CTTZ:
1730 case TargetOpcode::G_CTPOP:
1731 return lowerBitCount(MI, TypeIdx, Ty);
1733 Register Res = MI.getOperand(0).getReg();
1734 Register CarryOut = MI.getOperand(1).getReg();
1735 Register LHS = MI.getOperand(2).getReg();
1736 Register RHS = MI.getOperand(3).getReg();
1738 MIRBuilder.buildAdd(Res, LHS, RHS);
1739 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS);
1741 MI.eraseFromParent();
1745 Register Res = MI.getOperand(0).getReg();
1746 Register CarryOut = MI.getOperand(1).getReg();
1747 Register LHS = MI.getOperand(2).getReg();
1748 Register RHS = MI.getOperand(3).getReg();
1749 Register CarryIn = MI.getOperand(4).getReg();
1751 Register TmpRes = MRI.createGenericVirtualRegister(Ty);
1752 Register ZExtCarryIn = MRI.createGenericVirtualRegister(Ty);
1754 MIRBuilder.buildAdd(TmpRes, LHS, RHS);
1755 MIRBuilder.buildZExt(ZExtCarryIn, CarryIn);
1756 MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn);
1757 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, LHS);
1759 MI.eraseFromParent();
1763 Register Res = MI.getOperand(0).getReg();
1764 Register BorrowOut = MI.getOperand(1).getReg();
1765 Register LHS = MI.getOperand(2).getReg();
1766 Register RHS = MI.getOperand(3).getReg();
1768 MIRBuilder.buildSub(Res, LHS, RHS);
1769 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS);
1771 MI.eraseFromParent();
1775 Register Res = MI.getOperand(0).getReg();
1776 Register BorrowOut = MI.getOperand(1).getReg();
1777 Register LHS = MI.getOperand(2).getReg();
1778 Register RHS = MI.getOperand(3).getReg();
1779 Register BorrowIn = MI.getOperand(4).getReg();
1781 Register TmpRes = MRI.createGenericVirtualRegister(Ty);
1782 Register ZExtBorrowIn = MRI.createGenericVirtualRegister(Ty);
1783 Register LHS_EQ_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1));
1784 Register LHS_ULT_RHS = MRI.createGenericVirtualRegister(LLT::scalar(1));
1786 MIRBuilder.buildSub(TmpRes, LHS, RHS);
1787 MIRBuilder.buildZExt(ZExtBorrowIn, BorrowIn);
1788 MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn);
1789 MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LHS_EQ_RHS, LHS, RHS);
1790 MIRBuilder.buildICmp(CmpInst::ICMP_ULT, LHS_ULT_RHS, LHS, RHS);
1791 MIRBuilder.buildSelect(BorrowOut, LHS_EQ_RHS, BorrowIn, LHS_ULT_RHS);
1793 MI.eraseFromParent();
1797 return lowerUITOFP(MI, TypeIdx, Ty);
1799 return lowerSITOFP(MI, TypeIdx, Ty);
1804 return lowerMinMax(MI, TypeIdx, Ty);
1806 return lowerFCopySign(MI, TypeIdx, Ty);
1809 return lowerFMinNumMaxNum(MI);
1813 LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorImplicitDef(
1814 MachineInstr &MI, unsigned TypeIdx, LLT NarrowTy) {
1815 SmallVector<Register, 2> DstRegs;
1817 unsigned NarrowSize = NarrowTy.getSizeInBits();
1818 Register DstReg = MI.getOperand(0).getReg();
1819 unsigned Size = MRI.getType(DstReg).getSizeInBits();
1820 int NumParts = Size / NarrowSize;
1821 // FIXME: Don't know how to handle the situation where the small vectors
1822 // aren't all the same size yet.
1823 if (Size % NarrowSize != 0)
1824 return UnableToLegalize;
1826 for (int i = 0; i < NumParts; ++i) {
1827 Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
1828 MIRBuilder.buildUndef(TmpReg);
1829 DstRegs.push_back(TmpReg);
1832 if (NarrowTy.isVector())
1833 MIRBuilder.buildConcatVectors(DstReg, DstRegs);
1835 MIRBuilder.buildBuildVector(DstReg, DstRegs);
1837 MI.eraseFromParent();
1841 LegalizerHelper::LegalizeResult
1842 LegalizerHelper::fewerElementsVectorBasic(MachineInstr &MI, unsigned TypeIdx,
1844 const unsigned Opc = MI.getOpcode();
1845 const unsigned NumOps = MI.getNumOperands() - 1;
1846 const unsigned NarrowSize = NarrowTy.getSizeInBits();
1847 const Register DstReg = MI.getOperand(0).getReg();
1848 const unsigned Flags = MI.getFlags();
1849 const LLT DstTy = MRI.getType(DstReg);
1850 const unsigned Size = DstTy.getSizeInBits();
1851 const int NumParts = Size / NarrowSize;
1852 const LLT EltTy = DstTy.getElementType();
1853 const unsigned EltSize = EltTy.getSizeInBits();
1854 const unsigned BitsForNumParts = NarrowSize * NumParts;
1856 // Check if we have any leftovers. If we do, then only handle the case where
1857 // the leftover is one element.
1858 if (BitsForNumParts != Size && BitsForNumParts + EltSize != Size)
1859 return UnableToLegalize;
1861 if (BitsForNumParts != Size) {
1862 Register AccumDstReg = MRI.createGenericVirtualRegister(DstTy);
1863 MIRBuilder.buildUndef(AccumDstReg);
1865 // Handle the pieces which evenly divide into the requested type with
1866 // extract/op/insert sequence.
1867 for (unsigned Offset = 0; Offset < BitsForNumParts; Offset += NarrowSize) {
1868 SmallVector<SrcOp, 4> SrcOps;
1869 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) {
1870 Register PartOpReg = MRI.createGenericVirtualRegister(NarrowTy);
1871 MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(), Offset);
1872 SrcOps.push_back(PartOpReg);
1875 Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy);
1876 MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags);
1878 Register PartInsertReg = MRI.createGenericVirtualRegister(DstTy);
1879 MIRBuilder.buildInsert(PartInsertReg, AccumDstReg, PartDstReg, Offset);
1880 AccumDstReg = PartInsertReg;
1883 // Handle the remaining element sized leftover piece.
1884 SmallVector<SrcOp, 4> SrcOps;
1885 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) {
1886 Register PartOpReg = MRI.createGenericVirtualRegister(EltTy);
1887 MIRBuilder.buildExtract(PartOpReg, MI.getOperand(I).getReg(),
1889 SrcOps.push_back(PartOpReg);
1892 Register PartDstReg = MRI.createGenericVirtualRegister(EltTy);
1893 MIRBuilder.buildInstr(Opc, {PartDstReg}, SrcOps, Flags);
1894 MIRBuilder.buildInsert(DstReg, AccumDstReg, PartDstReg, BitsForNumParts);
1895 MI.eraseFromParent();
1900 SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs;
1902 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, Src0Regs);
1905 extractParts(MI.getOperand(2).getReg(), NarrowTy, NumParts, Src1Regs);
1908 extractParts(MI.getOperand(3).getReg(), NarrowTy, NumParts, Src2Regs);
1910 for (int i = 0; i < NumParts; ++i) {
1911 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
1914 MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i]}, Flags);
1915 else if (NumOps == 2) {
1916 MIRBuilder.buildInstr(Opc, {DstReg}, {Src0Regs[i], Src1Regs[i]}, Flags);
1917 } else if (NumOps == 3) {
1918 MIRBuilder.buildInstr(Opc, {DstReg},
1919 {Src0Regs[i], Src1Regs[i], Src2Regs[i]}, Flags);
1922 DstRegs.push_back(DstReg);
1925 if (NarrowTy.isVector())
1926 MIRBuilder.buildConcatVectors(DstReg, DstRegs);
1928 MIRBuilder.buildBuildVector(DstReg, DstRegs);
1930 MI.eraseFromParent();
1934 // Handle splitting vector operations which need to have the same number of
1935 // elements in each type index, but each type index may have a different element
1938 // e.g. <4 x s64> = G_SHL <4 x s64>, <4 x s32> ->
1939 // <2 x s64> = G_SHL <2 x s64>, <2 x s32>
1940 // <2 x s64> = G_SHL <2 x s64>, <2 x s32>
1942 // Also handles some irregular breakdown cases, e.g.
1943 // e.g. <3 x s64> = G_SHL <3 x s64>, <3 x s32> ->
1944 // <2 x s64> = G_SHL <2 x s64>, <2 x s32>
1945 // s64 = G_SHL s64, s32
1946 LegalizerHelper::LegalizeResult
1947 LegalizerHelper::fewerElementsVectorMultiEltType(
1948 MachineInstr &MI, unsigned TypeIdx, LLT NarrowTyArg) {
1950 return UnableToLegalize;
1952 const LLT NarrowTy0 = NarrowTyArg;
1953 const unsigned NewNumElts =
1954 NarrowTy0.isVector() ? NarrowTy0.getNumElements() : 1;
1956 const Register DstReg = MI.getOperand(0).getReg();
1957 LLT DstTy = MRI.getType(DstReg);
1960 // All of the operands need to have the same number of elements, so if we can
1961 // determine a type breakdown for the result type, we can for all of the
1963 int NumParts = getNarrowTypeBreakDown(DstTy, NarrowTy0, LeftoverTy0).first;
1965 return UnableToLegalize;
1967 SmallVector<MachineInstrBuilder, 4> NewInsts;
1969 SmallVector<Register, 4> DstRegs, LeftoverDstRegs;
1970 SmallVector<Register, 4> PartRegs, LeftoverRegs;
1972 for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I) {
1974 Register SrcReg = MI.getOperand(I).getReg();
1975 LLT SrcTyI = MRI.getType(SrcReg);
1976 LLT NarrowTyI = LLT::scalarOrVector(NewNumElts, SrcTyI.getScalarType());
1979 // Split this operand into the requested typed registers, and any leftover
1980 // required to reproduce the original type.
1981 if (!extractParts(SrcReg, SrcTyI, NarrowTyI, LeftoverTyI, PartRegs,
1983 return UnableToLegalize;
1986 // For the first operand, create an instruction for each part and setup
1988 for (Register PartReg : PartRegs) {
1989 Register PartDstReg = MRI.createGenericVirtualRegister(NarrowTy0);
1990 NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode())
1993 DstRegs.push_back(PartDstReg);
1996 for (Register LeftoverReg : LeftoverRegs) {
1997 Register PartDstReg = MRI.createGenericVirtualRegister(LeftoverTy0);
1998 NewInsts.push_back(MIRBuilder.buildInstrNoInsert(MI.getOpcode())
2000 .addUse(LeftoverReg));
2001 LeftoverDstRegs.push_back(PartDstReg);
2004 assert(NewInsts.size() == PartRegs.size() + LeftoverRegs.size());
2006 // Add the newly created operand splits to the existing instructions. The
2007 // odd-sized pieces are ordered after the requested NarrowTyArg sized
2009 unsigned InstCount = 0;
2010 for (unsigned J = 0, JE = PartRegs.size(); J != JE; ++J)
2011 NewInsts[InstCount++].addUse(PartRegs[J]);
2012 for (unsigned J = 0, JE = LeftoverRegs.size(); J != JE; ++J)
2013 NewInsts[InstCount++].addUse(LeftoverRegs[J]);
2017 LeftoverRegs.clear();
2020 // Insert the newly built operations and rebuild the result register.
2021 for (auto &MIB : NewInsts)
2022 MIRBuilder.insertInstr(MIB);
2024 insertParts(DstReg, DstTy, NarrowTy0, DstRegs, LeftoverTy0, LeftoverDstRegs);
2026 MI.eraseFromParent();
2030 LegalizerHelper::LegalizeResult
2031 LegalizerHelper::fewerElementsVectorCasts(MachineInstr &MI, unsigned TypeIdx,
2034 return UnableToLegalize;
2036 Register DstReg = MI.getOperand(0).getReg();
2037 Register SrcReg = MI.getOperand(1).getReg();
2038 LLT DstTy = MRI.getType(DstReg);
2039 LLT SrcTy = MRI.getType(SrcReg);
2041 LLT NarrowTy0 = NarrowTy;
2045 if (NarrowTy.isVector()) {
2046 // Uneven breakdown not handled.
2047 NumParts = DstTy.getNumElements() / NarrowTy.getNumElements();
2048 if (NumParts * NarrowTy.getNumElements() != DstTy.getNumElements())
2049 return UnableToLegalize;
2051 NarrowTy1 = LLT::vector(NumParts, SrcTy.getElementType().getSizeInBits());
2053 NumParts = DstTy.getNumElements();
2054 NarrowTy1 = SrcTy.getElementType();
2057 SmallVector<Register, 4> SrcRegs, DstRegs;
2058 extractParts(SrcReg, NarrowTy1, NumParts, SrcRegs);
2060 for (unsigned I = 0; I < NumParts; ++I) {
2061 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0);
2062 MachineInstr *NewInst = MIRBuilder.buildInstr(MI.getOpcode())
2064 .addUse(SrcRegs[I]);
2066 NewInst->setFlags(MI.getFlags());
2067 DstRegs.push_back(DstReg);
2070 if (NarrowTy.isVector())
2071 MIRBuilder.buildConcatVectors(DstReg, DstRegs);
2073 MIRBuilder.buildBuildVector(DstReg, DstRegs);
2075 MI.eraseFromParent();
2079 LegalizerHelper::LegalizeResult
2080 LegalizerHelper::fewerElementsVectorCmp(MachineInstr &MI, unsigned TypeIdx,
2082 Register DstReg = MI.getOperand(0).getReg();
2083 Register Src0Reg = MI.getOperand(2).getReg();
2084 LLT DstTy = MRI.getType(DstReg);
2085 LLT SrcTy = MRI.getType(Src0Reg);
2088 LLT NarrowTy0, NarrowTy1;
2091 unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1;
2092 unsigned OldElts = DstTy.getNumElements();
2094 NarrowTy0 = NarrowTy;
2095 NumParts = NarrowTy.isVector() ? (OldElts / NewElts) : DstTy.getNumElements();
2096 NarrowTy1 = NarrowTy.isVector() ?
2097 LLT::vector(NarrowTy.getNumElements(), SrcTy.getScalarSizeInBits()) :
2098 SrcTy.getElementType();
2101 unsigned NewElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1;
2102 unsigned OldElts = SrcTy.getNumElements();
2104 NumParts = NarrowTy.isVector() ? (OldElts / NewElts) :
2105 NarrowTy.getNumElements();
2106 NarrowTy0 = LLT::vector(NarrowTy.getNumElements(),
2107 DstTy.getScalarSizeInBits());
2108 NarrowTy1 = NarrowTy;
2111 // FIXME: Don't know how to handle the situation where the small vectors
2112 // aren't all the same size yet.
2113 if (NarrowTy1.isVector() &&
2114 NarrowTy1.getNumElements() * NumParts != DstTy.getNumElements())
2115 return UnableToLegalize;
2117 CmpInst::Predicate Pred
2118 = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
2120 SmallVector<Register, 2> Src1Regs, Src2Regs, DstRegs;
2121 extractParts(MI.getOperand(2).getReg(), NarrowTy1, NumParts, Src1Regs);
2122 extractParts(MI.getOperand(3).getReg(), NarrowTy1, NumParts, Src2Regs);
2124 for (unsigned I = 0; I < NumParts; ++I) {
2125 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0);
2126 DstRegs.push_back(DstReg);
2128 if (MI.getOpcode() == TargetOpcode::G_ICMP)
2129 MIRBuilder.buildICmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]);
2131 MachineInstr *NewCmp
2132 = MIRBuilder.buildFCmp(Pred, DstReg, Src1Regs[I], Src2Regs[I]);
2133 NewCmp->setFlags(MI.getFlags());
2137 if (NarrowTy1.isVector())
2138 MIRBuilder.buildConcatVectors(DstReg, DstRegs);
2140 MIRBuilder.buildBuildVector(DstReg, DstRegs);
2142 MI.eraseFromParent();
2146 LegalizerHelper::LegalizeResult
2147 LegalizerHelper::fewerElementsVectorSelect(MachineInstr &MI, unsigned TypeIdx,
2149 Register DstReg = MI.getOperand(0).getReg();
2150 Register CondReg = MI.getOperand(1).getReg();
2152 unsigned NumParts = 0;
2153 LLT NarrowTy0, NarrowTy1;
2155 LLT DstTy = MRI.getType(DstReg);
2156 LLT CondTy = MRI.getType(CondReg);
2157 unsigned Size = DstTy.getSizeInBits();
2159 assert(TypeIdx == 0 || CondTy.isVector());
2162 NarrowTy0 = NarrowTy;
2165 unsigned NarrowSize = NarrowTy0.getSizeInBits();
2166 // FIXME: Don't know how to handle the situation where the small vectors
2167 // aren't all the same size yet.
2168 if (Size % NarrowSize != 0)
2169 return UnableToLegalize;
2171 NumParts = Size / NarrowSize;
2173 // Need to break down the condition type
2174 if (CondTy.isVector()) {
2175 if (CondTy.getNumElements() == NumParts)
2176 NarrowTy1 = CondTy.getElementType();
2178 NarrowTy1 = LLT::vector(CondTy.getNumElements() / NumParts,
2179 CondTy.getScalarSizeInBits());
2182 NumParts = CondTy.getNumElements();
2183 if (NarrowTy.isVector()) {
2184 // TODO: Handle uneven breakdown.
2185 if (NumParts * NarrowTy.getNumElements() != CondTy.getNumElements())
2186 return UnableToLegalize;
2188 return UnableToLegalize;
2190 NarrowTy0 = DstTy.getElementType();
2191 NarrowTy1 = NarrowTy;
2195 SmallVector<Register, 2> DstRegs, Src0Regs, Src1Regs, Src2Regs;
2196 if (CondTy.isVector())
2197 extractParts(MI.getOperand(1).getReg(), NarrowTy1, NumParts, Src0Regs);
2199 extractParts(MI.getOperand(2).getReg(), NarrowTy0, NumParts, Src1Regs);
2200 extractParts(MI.getOperand(3).getReg(), NarrowTy0, NumParts, Src2Regs);
2202 for (unsigned i = 0; i < NumParts; ++i) {
2203 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy0);
2204 MIRBuilder.buildSelect(DstReg, CondTy.isVector() ? Src0Regs[i] : CondReg,
2205 Src1Regs[i], Src2Regs[i]);
2206 DstRegs.push_back(DstReg);
2209 if (NarrowTy0.isVector())
2210 MIRBuilder.buildConcatVectors(DstReg, DstRegs);
2212 MIRBuilder.buildBuildVector(DstReg, DstRegs);
2214 MI.eraseFromParent();
2218 LegalizerHelper::LegalizeResult
2219 LegalizerHelper::fewerElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx,
2221 const Register DstReg = MI.getOperand(0).getReg();
2222 LLT PhiTy = MRI.getType(DstReg);
2225 // All of the operands need to have the same number of elements, so if we can
2226 // determine a type breakdown for the result type, we can for all of the
2228 int NumParts, NumLeftover;
2229 std::tie(NumParts, NumLeftover)
2230 = getNarrowTypeBreakDown(PhiTy, NarrowTy, LeftoverTy);
2232 return UnableToLegalize;
2234 SmallVector<Register, 4> DstRegs, LeftoverDstRegs;
2235 SmallVector<MachineInstrBuilder, 4> NewInsts;
2237 const int TotalNumParts = NumParts + NumLeftover;
2239 // Insert the new phis in the result block first.
2240 for (int I = 0; I != TotalNumParts; ++I) {
2241 LLT Ty = I < NumParts ? NarrowTy : LeftoverTy;
2242 Register PartDstReg = MRI.createGenericVirtualRegister(Ty);
2243 NewInsts.push_back(MIRBuilder.buildInstr(TargetOpcode::G_PHI)
2244 .addDef(PartDstReg));
2246 DstRegs.push_back(PartDstReg);
2248 LeftoverDstRegs.push_back(PartDstReg);
2251 MachineBasicBlock *MBB = MI.getParent();
2252 MIRBuilder.setInsertPt(*MBB, MBB->getFirstNonPHI());
2253 insertParts(DstReg, PhiTy, NarrowTy, DstRegs, LeftoverTy, LeftoverDstRegs);
2255 SmallVector<Register, 4> PartRegs, LeftoverRegs;
2257 // Insert code to extract the incoming values in each predecessor block.
2258 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
2260 LeftoverRegs.clear();
2262 Register SrcReg = MI.getOperand(I).getReg();
2263 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB();
2264 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
2267 if (!extractParts(SrcReg, PhiTy, NarrowTy, Unused, PartRegs,
2269 return UnableToLegalize;
2271 // Add the newly created operand splits to the existing instructions. The
2272 // odd-sized pieces are ordered after the requested NarrowTyArg sized
2274 for (int J = 0; J != TotalNumParts; ++J) {
2275 MachineInstrBuilder MIB = NewInsts[J];
2276 MIB.addUse(J < NumParts ? PartRegs[J] : LeftoverRegs[J - NumParts]);
2281 MI.eraseFromParent();
2285 LegalizerHelper::LegalizeResult
2286 LegalizerHelper::reduceLoadStoreWidth(MachineInstr &MI, unsigned TypeIdx,
2288 // FIXME: Don't know how to handle secondary types yet.
2290 return UnableToLegalize;
2292 MachineMemOperand *MMO = *MI.memoperands_begin();
2294 // This implementation doesn't work for atomics. Give up instead of doing
2295 // something invalid.
2296 if (MMO->getOrdering() != AtomicOrdering::NotAtomic ||
2297 MMO->getFailureOrdering() != AtomicOrdering::NotAtomic)
2298 return UnableToLegalize;
2300 bool IsLoad = MI.getOpcode() == TargetOpcode::G_LOAD;
2301 Register ValReg = MI.getOperand(0).getReg();
2302 Register AddrReg = MI.getOperand(1).getReg();
2303 LLT ValTy = MRI.getType(ValReg);
2306 int NumLeftover = -1;
2308 SmallVector<Register, 8> NarrowRegs, NarrowLeftoverRegs;
2310 std::tie(NumParts, NumLeftover) = getNarrowTypeBreakDown(ValTy, NarrowTy, LeftoverTy);
2312 if (extractParts(ValReg, ValTy, NarrowTy, LeftoverTy, NarrowRegs,
2313 NarrowLeftoverRegs)) {
2314 NumParts = NarrowRegs.size();
2315 NumLeftover = NarrowLeftoverRegs.size();
2320 return UnableToLegalize;
2322 const LLT OffsetTy = LLT::scalar(MRI.getType(AddrReg).getScalarSizeInBits());
2324 unsigned TotalSize = ValTy.getSizeInBits();
2326 // Split the load/store into PartTy sized pieces starting at Offset. If this
2327 // is a load, return the new registers in ValRegs. For a store, each elements
2328 // of ValRegs should be PartTy. Returns the next offset that needs to be
2330 auto splitTypePieces = [=](LLT PartTy, SmallVectorImpl<Register> &ValRegs,
2331 unsigned Offset) -> unsigned {
2332 MachineFunction &MF = MIRBuilder.getMF();
2333 unsigned PartSize = PartTy.getSizeInBits();
2334 for (unsigned Idx = 0, E = NumParts; Idx != E && Offset < TotalSize;
2335 Offset += PartSize, ++Idx) {
2336 unsigned ByteSize = PartSize / 8;
2337 unsigned ByteOffset = Offset / 8;
2338 Register NewAddrReg;
2340 MIRBuilder.materializeGEP(NewAddrReg, AddrReg, OffsetTy, ByteOffset);
2342 MachineMemOperand *NewMMO =
2343 MF.getMachineMemOperand(MMO, ByteOffset, ByteSize);
2346 Register Dst = MRI.createGenericVirtualRegister(PartTy);
2347 ValRegs.push_back(Dst);
2348 MIRBuilder.buildLoad(Dst, NewAddrReg, *NewMMO);
2350 MIRBuilder.buildStore(ValRegs[Idx], NewAddrReg, *NewMMO);
2357 unsigned HandledOffset = splitTypePieces(NarrowTy, NarrowRegs, 0);
2359 // Handle the rest of the register if this isn't an even type breakdown.
2360 if (LeftoverTy.isValid())
2361 splitTypePieces(LeftoverTy, NarrowLeftoverRegs, HandledOffset);
2364 insertParts(ValReg, ValTy, NarrowTy, NarrowRegs,
2365 LeftoverTy, NarrowLeftoverRegs);
2368 MI.eraseFromParent();
2372 LegalizerHelper::LegalizeResult
2373 LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
2375 using namespace TargetOpcode;
2377 MIRBuilder.setInstr(MI);
2378 switch (MI.getOpcode()) {
2379 case G_IMPLICIT_DEF:
2380 return fewerElementsVectorImplicitDef(MI, TypeIdx, NarrowTy);
2394 case G_FCANONICALIZE:
2408 case G_INTRINSIC_ROUND:
2409 case G_INTRINSIC_TRUNC:
2421 case G_FMINNUM_IEEE:
2422 case G_FMAXNUM_IEEE:
2425 return fewerElementsVectorBasic(MI, TypeIdx, NarrowTy);
2430 case G_CTLZ_ZERO_UNDEF:
2432 case G_CTTZ_ZERO_UNDEF:
2435 return fewerElementsVectorMultiEltType(MI, TypeIdx, NarrowTy);
2447 case G_ADDRSPACE_CAST:
2448 return fewerElementsVectorCasts(MI, TypeIdx, NarrowTy);
2451 return fewerElementsVectorCmp(MI, TypeIdx, NarrowTy);
2453 return fewerElementsVectorSelect(MI, TypeIdx, NarrowTy);
2455 return fewerElementsVectorPhi(MI, TypeIdx, NarrowTy);
2458 return reduceLoadStoreWidth(MI, TypeIdx, NarrowTy);
2460 return UnableToLegalize;
2464 LegalizerHelper::LegalizeResult
2465 LegalizerHelper::narrowScalarShiftByConstant(MachineInstr &MI, const APInt &Amt,
2466 const LLT HalfTy, const LLT AmtTy) {
2468 Register InL = MRI.createGenericVirtualRegister(HalfTy);
2469 Register InH = MRI.createGenericVirtualRegister(HalfTy);
2470 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1).getReg());
2472 if (Amt.isNullValue()) {
2473 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {InL, InH});
2474 MI.eraseFromParent();
2479 unsigned NVTBits = HalfTy.getSizeInBits();
2480 unsigned VTBits = 2 * NVTBits;
2482 SrcOp Lo(Register(0)), Hi(Register(0));
2483 if (MI.getOpcode() == TargetOpcode::G_SHL) {
2484 if (Amt.ugt(VTBits)) {
2485 Lo = Hi = MIRBuilder.buildConstant(NVT, 0);
2486 } else if (Amt.ugt(NVTBits)) {
2487 Lo = MIRBuilder.buildConstant(NVT, 0);
2488 Hi = MIRBuilder.buildShl(NVT, InL,
2489 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits));
2490 } else if (Amt == NVTBits) {
2491 Lo = MIRBuilder.buildConstant(NVT, 0);
2494 Lo = MIRBuilder.buildShl(NVT, InL, MIRBuilder.buildConstant(AmtTy, Amt));
2496 MIRBuilder.buildShl(NVT, InH, MIRBuilder.buildConstant(AmtTy, Amt));
2497 auto OrRHS = MIRBuilder.buildLShr(
2498 NVT, InL, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits));
2499 Hi = MIRBuilder.buildOr(NVT, OrLHS, OrRHS);
2501 } else if (MI.getOpcode() == TargetOpcode::G_LSHR) {
2502 if (Amt.ugt(VTBits)) {
2503 Lo = Hi = MIRBuilder.buildConstant(NVT, 0);
2504 } else if (Amt.ugt(NVTBits)) {
2505 Lo = MIRBuilder.buildLShr(NVT, InH,
2506 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits));
2507 Hi = MIRBuilder.buildConstant(NVT, 0);
2508 } else if (Amt == NVTBits) {
2510 Hi = MIRBuilder.buildConstant(NVT, 0);
2512 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt);
2514 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst);
2515 auto OrRHS = MIRBuilder.buildShl(
2516 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits));
2518 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS);
2519 Hi = MIRBuilder.buildLShr(NVT, InH, ShiftAmtConst);
2522 if (Amt.ugt(VTBits)) {
2523 Hi = Lo = MIRBuilder.buildAShr(
2524 NVT, InH, MIRBuilder.buildConstant(AmtTy, NVTBits - 1));
2525 } else if (Amt.ugt(NVTBits)) {
2526 Lo = MIRBuilder.buildAShr(NVT, InH,
2527 MIRBuilder.buildConstant(AmtTy, Amt - NVTBits));
2528 Hi = MIRBuilder.buildAShr(NVT, InH,
2529 MIRBuilder.buildConstant(AmtTy, NVTBits - 1));
2530 } else if (Amt == NVTBits) {
2532 Hi = MIRBuilder.buildAShr(NVT, InH,
2533 MIRBuilder.buildConstant(AmtTy, NVTBits - 1));
2535 auto ShiftAmtConst = MIRBuilder.buildConstant(AmtTy, Amt);
2537 auto OrLHS = MIRBuilder.buildLShr(NVT, InL, ShiftAmtConst);
2538 auto OrRHS = MIRBuilder.buildShl(
2539 NVT, InH, MIRBuilder.buildConstant(AmtTy, -Amt + NVTBits));
2541 Lo = MIRBuilder.buildOr(NVT, OrLHS, OrRHS);
2542 Hi = MIRBuilder.buildAShr(NVT, InH, ShiftAmtConst);
2546 MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {Lo.getReg(), Hi.getReg()});
2547 MI.eraseFromParent();
2552 // TODO: Optimize if constant shift amount.
2553 LegalizerHelper::LegalizeResult
2554 LegalizerHelper::narrowScalarShift(MachineInstr &MI, unsigned TypeIdx,
2557 Observer.changingInstr(MI);
2558 narrowScalarSrc(MI, RequestedTy, 2);
2559 Observer.changedInstr(MI);
2563 Register DstReg = MI.getOperand(0).getReg();
2564 LLT DstTy = MRI.getType(DstReg);
2565 if (DstTy.isVector())
2566 return UnableToLegalize;
2568 Register Amt = MI.getOperand(2).getReg();
2569 LLT ShiftAmtTy = MRI.getType(Amt);
2570 const unsigned DstEltSize = DstTy.getScalarSizeInBits();
2571 if (DstEltSize % 2 != 0)
2572 return UnableToLegalize;
2574 // Ignore the input type. We can only go to exactly half the size of the
2575 // input. If that isn't small enough, the resulting pieces will be further
2577 const unsigned NewBitSize = DstEltSize / 2;
2578 const LLT HalfTy = LLT::scalar(NewBitSize);
2579 const LLT CondTy = LLT::scalar(1);
2581 if (const MachineInstr *KShiftAmt =
2582 getOpcodeDef(TargetOpcode::G_CONSTANT, Amt, MRI)) {
2583 return narrowScalarShiftByConstant(
2584 MI, KShiftAmt->getOperand(1).getCImm()->getValue(), HalfTy, ShiftAmtTy);
2587 // TODO: Expand with known bits.
2589 // Handle the fully general expansion by an unknown amount.
2590 auto NewBits = MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize);
2592 Register InL = MRI.createGenericVirtualRegister(HalfTy);
2593 Register InH = MRI.createGenericVirtualRegister(HalfTy);
2594 MIRBuilder.buildUnmerge({InL, InH}, MI.getOperand(1).getReg());
2596 auto AmtExcess = MIRBuilder.buildSub(ShiftAmtTy, Amt, NewBits);
2597 auto AmtLack = MIRBuilder.buildSub(ShiftAmtTy, NewBits, Amt);
2599 auto Zero = MIRBuilder.buildConstant(ShiftAmtTy, 0);
2600 auto IsShort = MIRBuilder.buildICmp(ICmpInst::ICMP_ULT, CondTy, Amt, NewBits);
2601 auto IsZero = MIRBuilder.buildICmp(ICmpInst::ICMP_EQ, CondTy, Amt, Zero);
2603 Register ResultRegs[2];
2604 switch (MI.getOpcode()) {
2605 case TargetOpcode::G_SHL: {
2606 // Short: ShAmt < NewBitSize
2607 auto LoS = MIRBuilder.buildShl(HalfTy, InH, Amt);
2609 auto OrLHS = MIRBuilder.buildShl(HalfTy, InH, Amt);
2610 auto OrRHS = MIRBuilder.buildLShr(HalfTy, InL, AmtLack);
2611 auto HiS = MIRBuilder.buildOr(HalfTy, OrLHS, OrRHS);
2613 // Long: ShAmt >= NewBitSize
2614 auto LoL = MIRBuilder.buildConstant(HalfTy, 0); // Lo part is zero.
2615 auto HiL = MIRBuilder.buildShl(HalfTy, InL, AmtExcess); // Hi from Lo part.
2617 auto Lo = MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL);
2618 auto Hi = MIRBuilder.buildSelect(
2619 HalfTy, IsZero, InH, MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL));
2621 ResultRegs[0] = Lo.getReg(0);
2622 ResultRegs[1] = Hi.getReg(0);
2625 case TargetOpcode::G_LSHR: {
2626 // Short: ShAmt < NewBitSize
2627 auto HiS = MIRBuilder.buildLShr(HalfTy, InH, Amt);
2629 auto OrLHS = MIRBuilder.buildLShr(HalfTy, InL, Amt);
2630 auto OrRHS = MIRBuilder.buildShl(HalfTy, InH, AmtLack);
2631 auto LoS = MIRBuilder.buildOr(HalfTy, OrLHS, OrRHS);
2633 // Long: ShAmt >= NewBitSize
2634 auto HiL = MIRBuilder.buildConstant(HalfTy, 0); // Hi part is zero.
2635 auto LoL = MIRBuilder.buildLShr(HalfTy, InH, AmtExcess); // Lo from Hi part.
2637 auto Lo = MIRBuilder.buildSelect(
2638 HalfTy, IsZero, InL, MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL));
2639 auto Hi = MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL);
2641 ResultRegs[0] = Lo.getReg(0);
2642 ResultRegs[1] = Hi.getReg(0);
2645 case TargetOpcode::G_ASHR: {
2646 // Short: ShAmt < NewBitSize
2647 auto HiS = MIRBuilder.buildAShr(HalfTy, InH, Amt);
2649 auto OrLHS = MIRBuilder.buildLShr(HalfTy, InL, Amt);
2650 auto OrRHS = MIRBuilder.buildLShr(HalfTy, InH, AmtLack);
2651 auto LoS = MIRBuilder.buildOr(HalfTy, OrLHS, OrRHS);
2653 // Long: ShAmt >= NewBitSize
2656 auto HiL = MIRBuilder.buildAShr(
2657 HalfTy, InH, MIRBuilder.buildConstant(ShiftAmtTy, NewBitSize - 1));
2659 auto LoL = MIRBuilder.buildAShr(HalfTy, InH, AmtExcess); // Lo from Hi part.
2661 auto Lo = MIRBuilder.buildSelect(
2662 HalfTy, IsZero, InL, MIRBuilder.buildSelect(HalfTy, IsShort, LoS, LoL));
2664 auto Hi = MIRBuilder.buildSelect(HalfTy, IsShort, HiS, HiL);
2666 ResultRegs[0] = Lo.getReg(0);
2667 ResultRegs[1] = Hi.getReg(0);
2671 llvm_unreachable("not a shift");
2674 MIRBuilder.buildMerge(DstReg, ResultRegs);
2675 MI.eraseFromParent();
2679 LegalizerHelper::LegalizeResult
2680 LegalizerHelper::moreElementsVectorPhi(MachineInstr &MI, unsigned TypeIdx,
2682 assert(TypeIdx == 0 && "Expecting only Idx 0");
2684 Observer.changingInstr(MI);
2685 for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
2686 MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB();
2687 MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminator());
2688 moreElementsVectorSrc(MI, MoreTy, I);
2691 MachineBasicBlock &MBB = *MI.getParent();
2692 MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI());
2693 moreElementsVectorDst(MI, MoreTy, 0);
2694 Observer.changedInstr(MI);
2698 LegalizerHelper::LegalizeResult
2699 LegalizerHelper::moreElementsVector(MachineInstr &MI, unsigned TypeIdx,
2701 MIRBuilder.setInstr(MI);
2702 unsigned Opc = MI.getOpcode();
2704 case TargetOpcode::G_IMPLICIT_DEF: {
2705 Observer.changingInstr(MI);
2706 moreElementsVectorDst(MI, MoreTy, 0);
2707 Observer.changedInstr(MI);
2710 case TargetOpcode::G_AND:
2711 case TargetOpcode::G_OR:
2712 case TargetOpcode::G_XOR:
2713 case TargetOpcode::G_SMIN:
2714 case TargetOpcode::G_SMAX:
2715 case TargetOpcode::G_UMIN:
2716 case TargetOpcode::G_UMAX: {
2717 Observer.changingInstr(MI);
2718 moreElementsVectorSrc(MI, MoreTy, 1);
2719 moreElementsVectorSrc(MI, MoreTy, 2);
2720 moreElementsVectorDst(MI, MoreTy, 0);
2721 Observer.changedInstr(MI);
2724 case TargetOpcode::G_EXTRACT:
2726 return UnableToLegalize;
2727 Observer.changingInstr(MI);
2728 moreElementsVectorSrc(MI, MoreTy, 1);
2729 Observer.changedInstr(MI);
2731 case TargetOpcode::G_INSERT:
2733 return UnableToLegalize;
2734 Observer.changingInstr(MI);
2735 moreElementsVectorSrc(MI, MoreTy, 1);
2736 moreElementsVectorDst(MI, MoreTy, 0);
2737 Observer.changedInstr(MI);
2739 case TargetOpcode::G_SELECT:
2741 return UnableToLegalize;
2742 if (MRI.getType(MI.getOperand(1).getReg()).isVector())
2743 return UnableToLegalize;
2745 Observer.changingInstr(MI);
2746 moreElementsVectorSrc(MI, MoreTy, 2);
2747 moreElementsVectorSrc(MI, MoreTy, 3);
2748 moreElementsVectorDst(MI, MoreTy, 0);
2749 Observer.changedInstr(MI);
2751 case TargetOpcode::G_PHI:
2752 return moreElementsVectorPhi(MI, TypeIdx, MoreTy);
2754 return UnableToLegalize;
2758 void LegalizerHelper::multiplyRegisters(SmallVectorImpl<Register> &DstRegs,
2759 ArrayRef<Register> Src1Regs,
2760 ArrayRef<Register> Src2Regs,
2762 MachineIRBuilder &B = MIRBuilder;
2763 unsigned SrcParts = Src1Regs.size();
2764 unsigned DstParts = DstRegs.size();
2766 unsigned DstIdx = 0; // Low bits of the result.
2767 Register FactorSum =
2768 B.buildMul(NarrowTy, Src1Regs[DstIdx], Src2Regs[DstIdx]).getReg(0);
2769 DstRegs[DstIdx] = FactorSum;
2771 unsigned CarrySumPrevDstIdx;
2772 SmallVector<Register, 4> Factors;
2774 for (DstIdx = 1; DstIdx < DstParts; DstIdx++) {
2775 // Collect low parts of muls for DstIdx.
2776 for (unsigned i = DstIdx + 1 < SrcParts ? 0 : DstIdx - SrcParts + 1;
2777 i <= std::min(DstIdx, SrcParts - 1); ++i) {
2778 MachineInstrBuilder Mul =
2779 B.buildMul(NarrowTy, Src1Regs[DstIdx - i], Src2Regs[i]);
2780 Factors.push_back(Mul.getReg(0));
2782 // Collect high parts of muls from previous DstIdx.
2783 for (unsigned i = DstIdx < SrcParts ? 0 : DstIdx - SrcParts;
2784 i <= std::min(DstIdx - 1, SrcParts - 1); ++i) {
2785 MachineInstrBuilder Umulh =
2786 B.buildUMulH(NarrowTy, Src1Regs[DstIdx - 1 - i], Src2Regs[i]);
2787 Factors.push_back(Umulh.getReg(0));
2789 // Add CarrySum from additons calculated for previous DstIdx.
2791 Factors.push_back(CarrySumPrevDstIdx);
2795 // Add all factors and accumulate all carries into CarrySum.
2796 if (DstIdx != DstParts - 1) {
2797 MachineInstrBuilder Uaddo =
2798 B.buildUAddo(NarrowTy, LLT::scalar(1), Factors[0], Factors[1]);
2799 FactorSum = Uaddo.getReg(0);
2800 CarrySum = B.buildZExt(NarrowTy, Uaddo.getReg(1)).getReg(0);
2801 for (unsigned i = 2; i < Factors.size(); ++i) {
2802 MachineInstrBuilder Uaddo =
2803 B.buildUAddo(NarrowTy, LLT::scalar(1), FactorSum, Factors[i]);
2804 FactorSum = Uaddo.getReg(0);
2805 MachineInstrBuilder Carry = B.buildZExt(NarrowTy, Uaddo.getReg(1));
2806 CarrySum = B.buildAdd(NarrowTy, CarrySum, Carry).getReg(0);
2809 // Since value for the next index is not calculated, neither is CarrySum.
2810 FactorSum = B.buildAdd(NarrowTy, Factors[0], Factors[1]).getReg(0);
2811 for (unsigned i = 2; i < Factors.size(); ++i)
2812 FactorSum = B.buildAdd(NarrowTy, FactorSum, Factors[i]).getReg(0);
2815 CarrySumPrevDstIdx = CarrySum;
2816 DstRegs[DstIdx] = FactorSum;
2821 LegalizerHelper::LegalizeResult
2822 LegalizerHelper::narrowScalarMul(MachineInstr &MI, LLT NarrowTy) {
2823 Register DstReg = MI.getOperand(0).getReg();
2824 Register Src1 = MI.getOperand(1).getReg();
2825 Register Src2 = MI.getOperand(2).getReg();
2827 LLT Ty = MRI.getType(DstReg);
2829 return UnableToLegalize;
2831 unsigned SrcSize = MRI.getType(Src1).getSizeInBits();
2832 unsigned DstSize = Ty.getSizeInBits();
2833 unsigned NarrowSize = NarrowTy.getSizeInBits();
2834 if (DstSize % NarrowSize != 0 || SrcSize % NarrowSize != 0)
2835 return UnableToLegalize;
2837 unsigned NumDstParts = DstSize / NarrowSize;
2838 unsigned NumSrcParts = SrcSize / NarrowSize;
2839 bool IsMulHigh = MI.getOpcode() == TargetOpcode::G_UMULH;
2840 unsigned DstTmpParts = NumDstParts * (IsMulHigh ? 2 : 1);
2842 SmallVector<Register, 2> Src1Parts, Src2Parts, DstTmpRegs;
2843 extractParts(Src1, NarrowTy, NumSrcParts, Src1Parts);
2844 extractParts(Src2, NarrowTy, NumSrcParts, Src2Parts);
2845 DstTmpRegs.resize(DstTmpParts);
2846 multiplyRegisters(DstTmpRegs, Src1Parts, Src2Parts, NarrowTy);
2848 // Take only high half of registers if this is high mul.
2849 ArrayRef<Register> DstRegs(
2850 IsMulHigh ? &DstTmpRegs[DstTmpParts / 2] : &DstTmpRegs[0], NumDstParts);
2851 MIRBuilder.buildMerge(DstReg, DstRegs);
2852 MI.eraseFromParent();
2856 LegalizerHelper::LegalizeResult
2857 LegalizerHelper::narrowScalarExtract(MachineInstr &MI, unsigned TypeIdx,
2860 return UnableToLegalize;
2862 uint64_t NarrowSize = NarrowTy.getSizeInBits();
2864 int64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
2865 // FIXME: add support for when SizeOp1 isn't an exact multiple of
2867 if (SizeOp1 % NarrowSize != 0)
2868 return UnableToLegalize;
2869 int NumParts = SizeOp1 / NarrowSize;
2871 SmallVector<Register, 2> SrcRegs, DstRegs;
2872 SmallVector<uint64_t, 2> Indexes;
2873 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs);
2875 Register OpReg = MI.getOperand(0).getReg();
2876 uint64_t OpStart = MI.getOperand(2).getImm();
2877 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits();
2878 for (int i = 0; i < NumParts; ++i) {
2879 unsigned SrcStart = i * NarrowSize;
2881 if (SrcStart + NarrowSize <= OpStart || SrcStart >= OpStart + OpSize) {
2882 // No part of the extract uses this subregister, ignore it.
2884 } else if (SrcStart == OpStart && NarrowTy == MRI.getType(OpReg)) {
2885 // The entire subregister is extracted, forward the value.
2886 DstRegs.push_back(SrcRegs[i]);
2890 // OpSegStart is where this destination segment would start in OpReg if it
2891 // extended infinitely in both directions.
2892 int64_t ExtractOffset;
2894 if (OpStart < SrcStart) {
2896 SegSize = std::min(NarrowSize, OpStart + OpSize - SrcStart);
2898 ExtractOffset = OpStart - SrcStart;
2899 SegSize = std::min(SrcStart + NarrowSize - OpStart, OpSize);
2902 Register SegReg = SrcRegs[i];
2903 if (ExtractOffset != 0 || SegSize != NarrowSize) {
2904 // A genuine extract is needed.
2905 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize));
2906 MIRBuilder.buildExtract(SegReg, SrcRegs[i], ExtractOffset);
2909 DstRegs.push_back(SegReg);
2912 Register DstReg = MI.getOperand(0).getReg();
2913 if(MRI.getType(DstReg).isVector())
2914 MIRBuilder.buildBuildVector(DstReg, DstRegs);
2916 MIRBuilder.buildMerge(DstReg, DstRegs);
2917 MI.eraseFromParent();
2921 LegalizerHelper::LegalizeResult
2922 LegalizerHelper::narrowScalarInsert(MachineInstr &MI, unsigned TypeIdx,
2924 // FIXME: Don't know how to handle secondary types yet.
2926 return UnableToLegalize;
2928 uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
2929 uint64_t NarrowSize = NarrowTy.getSizeInBits();
2931 // FIXME: add support for when SizeOp0 isn't an exact multiple of
2933 if (SizeOp0 % NarrowSize != 0)
2934 return UnableToLegalize;
2936 int NumParts = SizeOp0 / NarrowSize;
2938 SmallVector<Register, 2> SrcRegs, DstRegs;
2939 SmallVector<uint64_t, 2> Indexes;
2940 extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs);
2942 Register OpReg = MI.getOperand(2).getReg();
2943 uint64_t OpStart = MI.getOperand(3).getImm();
2944 uint64_t OpSize = MRI.getType(OpReg).getSizeInBits();
2945 for (int i = 0; i < NumParts; ++i) {
2946 unsigned DstStart = i * NarrowSize;
2948 if (DstStart + NarrowSize <= OpStart || DstStart >= OpStart + OpSize) {
2949 // No part of the insert affects this subregister, forward the original.
2950 DstRegs.push_back(SrcRegs[i]);
2952 } else if (DstStart == OpStart && NarrowTy == MRI.getType(OpReg)) {
2953 // The entire subregister is defined by this insert, forward the new
2955 DstRegs.push_back(OpReg);
2959 // OpSegStart is where this destination segment would start in OpReg if it
2960 // extended infinitely in both directions.
2961 int64_t ExtractOffset, InsertOffset;
2963 if (OpStart < DstStart) {
2965 ExtractOffset = DstStart - OpStart;
2966 SegSize = std::min(NarrowSize, OpStart + OpSize - DstStart);
2968 InsertOffset = OpStart - DstStart;
2971 std::min(NarrowSize - InsertOffset, OpStart + OpSize - DstStart);
2974 Register SegReg = OpReg;
2975 if (ExtractOffset != 0 || SegSize != OpSize) {
2976 // A genuine extract is needed.
2977 SegReg = MRI.createGenericVirtualRegister(LLT::scalar(SegSize));
2978 MIRBuilder.buildExtract(SegReg, OpReg, ExtractOffset);
2981 Register DstReg = MRI.createGenericVirtualRegister(NarrowTy);
2982 MIRBuilder.buildInsert(DstReg, SrcRegs[i], SegReg, InsertOffset);
2983 DstRegs.push_back(DstReg);
2986 assert(DstRegs.size() == (unsigned)NumParts && "not all parts covered");
2987 Register DstReg = MI.getOperand(0).getReg();
2988 if(MRI.getType(DstReg).isVector())
2989 MIRBuilder.buildBuildVector(DstReg, DstRegs);
2991 MIRBuilder.buildMerge(DstReg, DstRegs);
2992 MI.eraseFromParent();
2996 LegalizerHelper::LegalizeResult
2997 LegalizerHelper::narrowScalarBasic(MachineInstr &MI, unsigned TypeIdx,
2999 Register DstReg = MI.getOperand(0).getReg();
3000 LLT DstTy = MRI.getType(DstReg);
3002 assert(MI.getNumOperands() == 3 && TypeIdx == 0);
3004 SmallVector<Register, 4> DstRegs, DstLeftoverRegs;
3005 SmallVector<Register, 4> Src0Regs, Src0LeftoverRegs;
3006 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs;
3008 if (!extractParts(MI.getOperand(1).getReg(), DstTy, NarrowTy, LeftoverTy,
3009 Src0Regs, Src0LeftoverRegs))
3010 return UnableToLegalize;
3013 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, Unused,
3014 Src1Regs, Src1LeftoverRegs))
3015 llvm_unreachable("inconsistent extractParts result");
3017 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) {
3018 auto Inst = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy},
3019 {Src0Regs[I], Src1Regs[I]});
3020 DstRegs.push_back(Inst->getOperand(0).getReg());
3023 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) {
3024 auto Inst = MIRBuilder.buildInstr(
3026 {LeftoverTy}, {Src0LeftoverRegs[I], Src1LeftoverRegs[I]});
3027 DstLeftoverRegs.push_back(Inst->getOperand(0).getReg());
3030 insertParts(DstReg, DstTy, NarrowTy, DstRegs,
3031 LeftoverTy, DstLeftoverRegs);
3033 MI.eraseFromParent();
3037 LegalizerHelper::LegalizeResult
3038 LegalizerHelper::narrowScalarSelect(MachineInstr &MI, unsigned TypeIdx,
3041 return UnableToLegalize;
3043 Register CondReg = MI.getOperand(1).getReg();
3044 LLT CondTy = MRI.getType(CondReg);
3045 if (CondTy.isVector()) // TODO: Handle vselect
3046 return UnableToLegalize;
3048 Register DstReg = MI.getOperand(0).getReg();
3049 LLT DstTy = MRI.getType(DstReg);
3051 SmallVector<Register, 4> DstRegs, DstLeftoverRegs;
3052 SmallVector<Register, 4> Src1Regs, Src1LeftoverRegs;
3053 SmallVector<Register, 4> Src2Regs, Src2LeftoverRegs;
3055 if (!extractParts(MI.getOperand(2).getReg(), DstTy, NarrowTy, LeftoverTy,
3056 Src1Regs, Src1LeftoverRegs))
3057 return UnableToLegalize;
3060 if (!extractParts(MI.getOperand(3).getReg(), DstTy, NarrowTy, Unused,
3061 Src2Regs, Src2LeftoverRegs))
3062 llvm_unreachable("inconsistent extractParts result");
3064 for (unsigned I = 0, E = Src1Regs.size(); I != E; ++I) {
3065 auto Select = MIRBuilder.buildSelect(NarrowTy,
3066 CondReg, Src1Regs[I], Src2Regs[I]);
3067 DstRegs.push_back(Select->getOperand(0).getReg());
3070 for (unsigned I = 0, E = Src1LeftoverRegs.size(); I != E; ++I) {
3071 auto Select = MIRBuilder.buildSelect(
3072 LeftoverTy, CondReg, Src1LeftoverRegs[I], Src2LeftoverRegs[I]);
3073 DstLeftoverRegs.push_back(Select->getOperand(0).getReg());
3076 insertParts(DstReg, DstTy, NarrowTy, DstRegs,
3077 LeftoverTy, DstLeftoverRegs);
3079 MI.eraseFromParent();
3083 LegalizerHelper::LegalizeResult
3084 LegalizerHelper::lowerBitCount(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
3085 unsigned Opc = MI.getOpcode();
3086 auto &TII = *MI.getMF()->getSubtarget().getInstrInfo();
3087 auto isSupported = [this](const LegalityQuery &Q) {
3088 auto QAction = LI.getAction(Q).Action;
3089 return QAction == Legal || QAction == Libcall || QAction == Custom;
3093 return UnableToLegalize;
3094 case TargetOpcode::G_CTLZ_ZERO_UNDEF: {
3095 // This trivially expands to CTLZ.
3096 Observer.changingInstr(MI);
3097 MI.setDesc(TII.get(TargetOpcode::G_CTLZ));
3098 Observer.changedInstr(MI);
3101 case TargetOpcode::G_CTLZ: {
3102 Register SrcReg = MI.getOperand(1).getReg();
3103 unsigned Len = Ty.getSizeInBits();
3104 if (isSupported({TargetOpcode::G_CTLZ_ZERO_UNDEF, {Ty, Ty}})) {
3105 // If CTLZ_ZERO_UNDEF is supported, emit that and a select for zero.
3106 auto MIBCtlzZU = MIRBuilder.buildInstr(TargetOpcode::G_CTLZ_ZERO_UNDEF,
3108 auto MIBZero = MIRBuilder.buildConstant(Ty, 0);
3109 auto MIBLen = MIRBuilder.buildConstant(Ty, Len);
3110 auto MIBICmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1),
3112 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), MIBICmp, MIBLen,
3114 MI.eraseFromParent();
3117 // for now, we do this:
3118 // NewLen = NextPowerOf2(Len);
3119 // x = x | (x >> 1);
3120 // x = x | (x >> 2);
3122 // x = x | (x >>16);
3123 // x = x | (x >>32); // for 64-bit input
3125 // return Len - popcount(x);
3127 // Ref: "Hacker's Delight" by Henry Warren
3128 Register Op = SrcReg;
3129 unsigned NewLen = PowerOf2Ceil(Len);
3130 for (unsigned i = 0; (1U << i) <= (NewLen / 2); ++i) {
3131 auto MIBShiftAmt = MIRBuilder.buildConstant(Ty, 1ULL << i);
3132 auto MIBOp = MIRBuilder.buildInstr(
3133 TargetOpcode::G_OR, {Ty},
3134 {Op, MIRBuilder.buildInstr(TargetOpcode::G_LSHR, {Ty},
3135 {Op, MIBShiftAmt})});
3136 Op = MIBOp->getOperand(0).getReg();
3138 auto MIBPop = MIRBuilder.buildInstr(TargetOpcode::G_CTPOP, {Ty}, {Op});
3139 MIRBuilder.buildInstr(TargetOpcode::G_SUB, {MI.getOperand(0).getReg()},
3140 {MIRBuilder.buildConstant(Ty, Len), MIBPop});
3141 MI.eraseFromParent();
3144 case TargetOpcode::G_CTTZ_ZERO_UNDEF: {
3145 // This trivially expands to CTTZ.
3146 Observer.changingInstr(MI);
3147 MI.setDesc(TII.get(TargetOpcode::G_CTTZ));
3148 Observer.changedInstr(MI);
3151 case TargetOpcode::G_CTTZ: {
3152 Register SrcReg = MI.getOperand(1).getReg();
3153 unsigned Len = Ty.getSizeInBits();
3154 if (isSupported({TargetOpcode::G_CTTZ_ZERO_UNDEF, {Ty, Ty}})) {
3155 // If CTTZ_ZERO_UNDEF is legal or custom, emit that and a select with
3157 auto MIBCttzZU = MIRBuilder.buildInstr(TargetOpcode::G_CTTZ_ZERO_UNDEF,
3159 auto MIBZero = MIRBuilder.buildConstant(Ty, 0);
3160 auto MIBLen = MIRBuilder.buildConstant(Ty, Len);
3161 auto MIBICmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, LLT::scalar(1),
3163 MIRBuilder.buildSelect(MI.getOperand(0).getReg(), MIBICmp, MIBLen,
3165 MI.eraseFromParent();
3168 // for now, we use: { return popcount(~x & (x - 1)); }
3169 // unless the target has ctlz but not ctpop, in which case we use:
3170 // { return 32 - nlz(~x & (x-1)); }
3171 // Ref: "Hacker's Delight" by Henry Warren
3172 auto MIBCstNeg1 = MIRBuilder.buildConstant(Ty, -1);
3174 MIRBuilder.buildInstr(TargetOpcode::G_XOR, {Ty}, {SrcReg, MIBCstNeg1});
3175 auto MIBTmp = MIRBuilder.buildInstr(
3176 TargetOpcode::G_AND, {Ty},
3177 {MIBNot, MIRBuilder.buildInstr(TargetOpcode::G_ADD, {Ty},
3178 {SrcReg, MIBCstNeg1})});
3179 if (!isSupported({TargetOpcode::G_CTPOP, {Ty, Ty}}) &&
3180 isSupported({TargetOpcode::G_CTLZ, {Ty, Ty}})) {
3181 auto MIBCstLen = MIRBuilder.buildConstant(Ty, Len);
3182 MIRBuilder.buildInstr(
3183 TargetOpcode::G_SUB, {MI.getOperand(0).getReg()},
3185 MIRBuilder.buildInstr(TargetOpcode::G_CTLZ, {Ty}, {MIBTmp})});
3186 MI.eraseFromParent();
3189 MI.setDesc(TII.get(TargetOpcode::G_CTPOP));
3190 MI.getOperand(1).setReg(MIBTmp->getOperand(0).getReg());
3196 // Expand s32 = G_UITOFP s64 using bit operations to an IEEE float
3198 LegalizerHelper::LegalizeResult
3199 LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) {
3200 Register Dst = MI.getOperand(0).getReg();
3201 Register Src = MI.getOperand(1).getReg();
3202 const LLT S64 = LLT::scalar(64);
3203 const LLT S32 = LLT::scalar(32);
3204 const LLT S1 = LLT::scalar(1);
3206 assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32);
3208 // unsigned cul2f(ulong u) {
3209 // uint lz = clz(u);
3210 // uint e = (u != 0) ? 127U + 63U - lz : 0;
3211 // u = (u << lz) & 0x7fffffffffffffffUL;
3212 // ulong t = u & 0xffffffffffUL;
3213 // uint v = (e << 23) | (uint)(u >> 40);
3214 // uint r = t > 0x8000000000UL ? 1U : (t == 0x8000000000UL ? v & 1U : 0U);
3215 // return as_float(v + r);
3218 auto Zero32 = MIRBuilder.buildConstant(S32, 0);
3219 auto Zero64 = MIRBuilder.buildConstant(S64, 0);
3221 auto LZ = MIRBuilder.buildCTLZ_ZERO_UNDEF(S32, Src);
3223 auto K = MIRBuilder.buildConstant(S32, 127U + 63U);
3224 auto Sub = MIRBuilder.buildSub(S32, K, LZ);
3226 auto NotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, Src, Zero64);
3227 auto E = MIRBuilder.buildSelect(S32, NotZero, Sub, Zero32);
3229 auto Mask0 = MIRBuilder.buildConstant(S64, (-1ULL) >> 1);
3230 auto ShlLZ = MIRBuilder.buildShl(S64, Src, LZ);
3232 auto U = MIRBuilder.buildAnd(S64, ShlLZ, Mask0);
3234 auto Mask1 = MIRBuilder.buildConstant(S64, 0xffffffffffULL);
3235 auto T = MIRBuilder.buildAnd(S64, U, Mask1);
3237 auto UShl = MIRBuilder.buildLShr(S64, U, MIRBuilder.buildConstant(S64, 40));
3238 auto ShlE = MIRBuilder.buildShl(S32, E, MIRBuilder.buildConstant(S32, 23));
3239 auto V = MIRBuilder.buildOr(S32, ShlE, MIRBuilder.buildTrunc(S32, UShl));
3241 auto C = MIRBuilder.buildConstant(S64, 0x8000000000ULL);
3242 auto RCmp = MIRBuilder.buildICmp(CmpInst::ICMP_UGT, S1, T, C);
3243 auto TCmp = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, S1, T, C);
3244 auto One = MIRBuilder.buildConstant(S32, 1);
3246 auto VTrunc1 = MIRBuilder.buildAnd(S32, V, One);
3247 auto Select0 = MIRBuilder.buildSelect(S32, TCmp, VTrunc1, Zero32);
3248 auto R = MIRBuilder.buildSelect(S32, RCmp, One, Select0);
3249 MIRBuilder.buildAdd(Dst, V, R);
3254 LegalizerHelper::LegalizeResult
3255 LegalizerHelper::lowerUITOFP(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
3256 Register Dst = MI.getOperand(0).getReg();
3257 Register Src = MI.getOperand(1).getReg();
3258 LLT DstTy = MRI.getType(Dst);
3259 LLT SrcTy = MRI.getType(Src);
3261 if (SrcTy != LLT::scalar(64))
3262 return UnableToLegalize;
3264 if (DstTy == LLT::scalar(32)) {
3265 // TODO: SelectionDAG has several alternative expansions to port which may
3266 // be more reasonble depending on the available instructions. If a target
3267 // has sitofp, does not have CTLZ, or can efficiently use f64 as an
3268 // intermediate type, this is probably worse.
3269 return lowerU64ToF32BitOps(MI);
3272 return UnableToLegalize;
3275 LegalizerHelper::LegalizeResult
3276 LegalizerHelper::lowerSITOFP(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
3277 Register Dst = MI.getOperand(0).getReg();
3278 Register Src = MI.getOperand(1).getReg();
3279 LLT DstTy = MRI.getType(Dst);
3280 LLT SrcTy = MRI.getType(Src);
3282 const LLT S64 = LLT::scalar(64);
3283 const LLT S32 = LLT::scalar(32);
3284 const LLT S1 = LLT::scalar(1);
3287 return UnableToLegalize;
3290 // signed cl2f(long l) {
3291 // long s = l >> 63;
3292 // float r = cul2f((l + s) ^ s);
3293 // return s ? -r : r;
3296 auto SignBit = MIRBuilder.buildConstant(S64, 63);
3297 auto S = MIRBuilder.buildAShr(S64, L, SignBit);
3299 auto LPlusS = MIRBuilder.buildAdd(S64, L, S);
3300 auto Xor = MIRBuilder.buildXor(S64, LPlusS, S);
3301 auto R = MIRBuilder.buildUITOFP(S32, Xor);
3303 auto RNeg = MIRBuilder.buildFNeg(S32, R);
3304 auto SignNotZero = MIRBuilder.buildICmp(CmpInst::ICMP_NE, S1, S,
3305 MIRBuilder.buildConstant(S64, 0));
3306 MIRBuilder.buildSelect(Dst, SignNotZero, RNeg, R);
3310 return UnableToLegalize;
3313 static CmpInst::Predicate minMaxToCompare(unsigned Opc) {
3315 case TargetOpcode::G_SMIN:
3316 return CmpInst::ICMP_SLT;
3317 case TargetOpcode::G_SMAX:
3318 return CmpInst::ICMP_SGT;
3319 case TargetOpcode::G_UMIN:
3320 return CmpInst::ICMP_ULT;
3321 case TargetOpcode::G_UMAX:
3322 return CmpInst::ICMP_UGT;
3324 llvm_unreachable("not in integer min/max");
3328 LegalizerHelper::LegalizeResult
3329 LegalizerHelper::lowerMinMax(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
3330 Register Dst = MI.getOperand(0).getReg();
3331 Register Src0 = MI.getOperand(1).getReg();
3332 Register Src1 = MI.getOperand(2).getReg();
3334 const CmpInst::Predicate Pred = minMaxToCompare(MI.getOpcode());
3335 LLT CmpType = MRI.getType(Dst).changeElementSize(1);
3337 auto Cmp = MIRBuilder.buildICmp(Pred, CmpType, Src0, Src1);
3338 MIRBuilder.buildSelect(Dst, Cmp, Src0, Src1);
3340 MI.eraseFromParent();
3344 LegalizerHelper::LegalizeResult
3345 LegalizerHelper::lowerFCopySign(MachineInstr &MI, unsigned TypeIdx, LLT Ty) {
3346 Register Dst = MI.getOperand(0).getReg();
3347 Register Src0 = MI.getOperand(1).getReg();
3348 Register Src1 = MI.getOperand(2).getReg();
3350 const LLT Src0Ty = MRI.getType(Src0);
3351 const LLT Src1Ty = MRI.getType(Src1);
3353 const int Src0Size = Src0Ty.getScalarSizeInBits();
3354 const int Src1Size = Src1Ty.getScalarSizeInBits();
3356 auto SignBitMask = MIRBuilder.buildConstant(
3357 Src0Ty, APInt::getSignMask(Src0Size));
3359 auto NotSignBitMask = MIRBuilder.buildConstant(
3360 Src0Ty, APInt::getLowBitsSet(Src0Size, Src0Size - 1));
3362 auto And0 = MIRBuilder.buildAnd(Src0Ty, Src0, NotSignBitMask);
3365 if (Src0Ty == Src1Ty) {
3366 auto And1 = MIRBuilder.buildAnd(Src1Ty, Src0, SignBitMask);
3367 Or = MIRBuilder.buildOr(Dst, And0, And1);
3368 } else if (Src0Size > Src1Size) {
3369 auto ShiftAmt = MIRBuilder.buildConstant(Src0Ty, Src0Size - Src1Size);
3370 auto Zext = MIRBuilder.buildZExt(Src0Ty, Src1);
3371 auto Shift = MIRBuilder.buildShl(Src0Ty, Zext, ShiftAmt);
3372 auto And1 = MIRBuilder.buildAnd(Src0Ty, Shift, SignBitMask);
3373 Or = MIRBuilder.buildOr(Dst, And0, And1);
3375 auto ShiftAmt = MIRBuilder.buildConstant(Src1Ty, Src1Size - Src0Size);
3376 auto Shift = MIRBuilder.buildLShr(Src1Ty, Src1, ShiftAmt);
3377 auto Trunc = MIRBuilder.buildTrunc(Src0Ty, Shift);
3378 auto And1 = MIRBuilder.buildAnd(Src0Ty, Trunc, SignBitMask);
3379 Or = MIRBuilder.buildOr(Dst, And0, And1);
3382 // Be careful about setting nsz/nnan/ninf on every instruction, since the
3383 // constants are a nan and -0.0, but the final result should preserve
3385 if (unsigned Flags = MI.getFlags())
3386 Or->setFlags(Flags);
3388 MI.eraseFromParent();
3392 LegalizerHelper::LegalizeResult
3393 LegalizerHelper::lowerFMinNumMaxNum(MachineInstr &MI) {
3394 unsigned NewOp = MI.getOpcode() == TargetOpcode::G_FMINNUM ?
3395 TargetOpcode::G_FMINNUM_IEEE : TargetOpcode::G_FMAXNUM_IEEE;
3397 Register Dst = MI.getOperand(0).getReg();
3398 Register Src0 = MI.getOperand(1).getReg();
3399 Register Src1 = MI.getOperand(2).getReg();
3400 LLT Ty = MRI.getType(Dst);
3402 if (!MI.getFlag(MachineInstr::FmNoNans)) {
3403 // Insert canonicalizes if it's possible we need to quiet to get correct
3406 // Note this must be done here, and not as an optimization combine in the
3407 // absence of a dedicate quiet-snan instruction as we're using an
3408 // omni-purpose G_FCANONICALIZE.
3409 if (!isKnownNeverSNaN(Src0, MRI))
3410 Src0 = MIRBuilder.buildFCanonicalize(Ty, Src0, MI.getFlags()).getReg(0);
3412 if (!isKnownNeverSNaN(Src1, MRI))
3413 Src1 = MIRBuilder.buildFCanonicalize(Ty, Src1, MI.getFlags()).getReg(0);
3416 // If there are no nans, it's safe to simply replace this with the non-IEEE
3418 MIRBuilder.buildInstr(NewOp, {Dst}, {Src0, Src1}, MI.getFlags());
3419 MI.eraseFromParent();