xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //===-- lib/CodeGen/GlobalISel/GICombinerHelper.cpp -----------------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 #include "llvm/CodeGen/GlobalISel/CombinerHelper.h"
9 #include "llvm/ADT/APFloat.h"
10 #include "llvm/ADT/STLExtras.h"
11 #include "llvm/ADT/SetVector.h"
12 #include "llvm/ADT/SmallBitVector.h"
13 #include "llvm/Analysis/CmpInstAnalysis.h"
14 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
15 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
16 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
17 #include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
18 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
19 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
20 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
21 #include "llvm/CodeGen/GlobalISel/Utils.h"
22 #include "llvm/CodeGen/LowLevelTypeUtils.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineInstr.h"
26 #include "llvm/CodeGen/MachineMemOperand.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/RegisterBankInfo.h"
29 #include "llvm/CodeGen/TargetInstrInfo.h"
30 #include "llvm/CodeGen/TargetLowering.h"
31 #include "llvm/CodeGen/TargetOpcodes.h"
32 #include "llvm/IR/ConstantRange.h"
33 #include "llvm/IR/DataLayout.h"
34 #include "llvm/IR/InstrTypes.h"
35 #include "llvm/Support/Casting.h"
36 #include "llvm/Support/DivisionByConstantInfo.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/MathExtras.h"
39 #include "llvm/Target/TargetMachine.h"
40 #include <cmath>
41 #include <optional>
42 #include <tuple>
43 
44 #define DEBUG_TYPE "gi-combiner"
45 
46 using namespace llvm;
47 using namespace MIPatternMatch;
48 
49 // Option to allow testing of the combiner while no targets know about indexed
50 // addressing.
51 static cl::opt<bool>
52     ForceLegalIndexing("force-legal-indexing", cl::Hidden, cl::init(false),
53                        cl::desc("Force all indexed operations to be "
54                                 "legal for the GlobalISel combiner"));
55 
CombinerHelper(GISelChangeObserver & Observer,MachineIRBuilder & B,bool IsPreLegalize,GISelKnownBits * KB,MachineDominatorTree * MDT,const LegalizerInfo * LI)56 CombinerHelper::CombinerHelper(GISelChangeObserver &Observer,
57                                MachineIRBuilder &B, bool IsPreLegalize,
58                                GISelKnownBits *KB, MachineDominatorTree *MDT,
59                                const LegalizerInfo *LI)
60     : Builder(B), MRI(Builder.getMF().getRegInfo()), Observer(Observer), KB(KB),
61       MDT(MDT), IsPreLegalize(IsPreLegalize), LI(LI),
62       RBI(Builder.getMF().getSubtarget().getRegBankInfo()),
63       TRI(Builder.getMF().getSubtarget().getRegisterInfo()) {
64   (void)this->KB;
65 }
66 
getTargetLowering() const67 const TargetLowering &CombinerHelper::getTargetLowering() const {
68   return *Builder.getMF().getSubtarget().getTargetLowering();
69 }
70 
71 /// \returns The little endian in-memory byte position of byte \p I in a
72 /// \p ByteWidth bytes wide type.
73 ///
74 /// E.g. Given a 4-byte type x, x[0] -> byte 0
littleEndianByteAt(const unsigned ByteWidth,const unsigned I)75 static unsigned littleEndianByteAt(const unsigned ByteWidth, const unsigned I) {
76   assert(I < ByteWidth && "I must be in [0, ByteWidth)");
77   return I;
78 }
79 
80 /// Determines the LogBase2 value for a non-null input value using the
81 /// transform: LogBase2(V) = (EltBits - 1) - ctlz(V).
buildLogBase2(Register V,MachineIRBuilder & MIB)82 static Register buildLogBase2(Register V, MachineIRBuilder &MIB) {
83   auto &MRI = *MIB.getMRI();
84   LLT Ty = MRI.getType(V);
85   auto Ctlz = MIB.buildCTLZ(Ty, V);
86   auto Base = MIB.buildConstant(Ty, Ty.getScalarSizeInBits() - 1);
87   return MIB.buildSub(Ty, Base, Ctlz).getReg(0);
88 }
89 
90 /// \returns The big endian in-memory byte position of byte \p I in a
91 /// \p ByteWidth bytes wide type.
92 ///
93 /// E.g. Given a 4-byte type x, x[0] -> byte 3
bigEndianByteAt(const unsigned ByteWidth,const unsigned I)94 static unsigned bigEndianByteAt(const unsigned ByteWidth, const unsigned I) {
95   assert(I < ByteWidth && "I must be in [0, ByteWidth)");
96   return ByteWidth - I - 1;
97 }
98 
99 /// Given a map from byte offsets in memory to indices in a load/store,
100 /// determine if that map corresponds to a little or big endian byte pattern.
101 ///
102 /// \param MemOffset2Idx maps memory offsets to address offsets.
103 /// \param LowestIdx is the lowest index in \p MemOffset2Idx.
104 ///
105 /// \returns true if the map corresponds to a big endian byte pattern, false if
106 /// it corresponds to a little endian byte pattern, and std::nullopt otherwise.
107 ///
108 /// E.g. given a 32-bit type x, and x[AddrOffset], the in-memory byte patterns
109 /// are as follows:
110 ///
111 /// AddrOffset   Little endian    Big endian
112 /// 0            0                3
113 /// 1            1                2
114 /// 2            2                1
115 /// 3            3                0
116 static std::optional<bool>
isBigEndian(const SmallDenseMap<int64_t,int64_t,8> & MemOffset2Idx,int64_t LowestIdx)117 isBigEndian(const SmallDenseMap<int64_t, int64_t, 8> &MemOffset2Idx,
118             int64_t LowestIdx) {
119   // Need at least two byte positions to decide on endianness.
120   unsigned Width = MemOffset2Idx.size();
121   if (Width < 2)
122     return std::nullopt;
123   bool BigEndian = true, LittleEndian = true;
124   for (unsigned MemOffset = 0; MemOffset < Width; ++ MemOffset) {
125     auto MemOffsetAndIdx = MemOffset2Idx.find(MemOffset);
126     if (MemOffsetAndIdx == MemOffset2Idx.end())
127       return std::nullopt;
128     const int64_t Idx = MemOffsetAndIdx->second - LowestIdx;
129     assert(Idx >= 0 && "Expected non-negative byte offset?");
130     LittleEndian &= Idx == littleEndianByteAt(Width, MemOffset);
131     BigEndian &= Idx == bigEndianByteAt(Width, MemOffset);
132     if (!BigEndian && !LittleEndian)
133       return std::nullopt;
134   }
135 
136   assert((BigEndian != LittleEndian) &&
137          "Pattern cannot be both big and little endian!");
138   return BigEndian;
139 }
140 
isPreLegalize() const141 bool CombinerHelper::isPreLegalize() const { return IsPreLegalize; }
142 
isLegal(const LegalityQuery & Query) const143 bool CombinerHelper::isLegal(const LegalityQuery &Query) const {
144   assert(LI && "Must have LegalizerInfo to query isLegal!");
145   return LI->getAction(Query).Action == LegalizeActions::Legal;
146 }
147 
isLegalOrBeforeLegalizer(const LegalityQuery & Query) const148 bool CombinerHelper::isLegalOrBeforeLegalizer(
149     const LegalityQuery &Query) const {
150   return isPreLegalize() || isLegal(Query);
151 }
152 
isConstantLegalOrBeforeLegalizer(const LLT Ty) const153 bool CombinerHelper::isConstantLegalOrBeforeLegalizer(const LLT Ty) const {
154   if (!Ty.isVector())
155     return isLegalOrBeforeLegalizer({TargetOpcode::G_CONSTANT, {Ty}});
156   // Vector constants are represented as a G_BUILD_VECTOR of scalar G_CONSTANTs.
157   if (isPreLegalize())
158     return true;
159   LLT EltTy = Ty.getElementType();
160   return isLegal({TargetOpcode::G_BUILD_VECTOR, {Ty, EltTy}}) &&
161          isLegal({TargetOpcode::G_CONSTANT, {EltTy}});
162 }
163 
replaceRegWith(MachineRegisterInfo & MRI,Register FromReg,Register ToReg) const164 void CombinerHelper::replaceRegWith(MachineRegisterInfo &MRI, Register FromReg,
165                                     Register ToReg) const {
166   Observer.changingAllUsesOfReg(MRI, FromReg);
167 
168   if (MRI.constrainRegAttrs(ToReg, FromReg))
169     MRI.replaceRegWith(FromReg, ToReg);
170   else
171     Builder.buildCopy(ToReg, FromReg);
172 
173   Observer.finishedChangingAllUsesOfReg();
174 }
175 
replaceRegOpWith(MachineRegisterInfo & MRI,MachineOperand & FromRegOp,Register ToReg) const176 void CombinerHelper::replaceRegOpWith(MachineRegisterInfo &MRI,
177                                       MachineOperand &FromRegOp,
178                                       Register ToReg) const {
179   assert(FromRegOp.getParent() && "Expected an operand in an MI");
180   Observer.changingInstr(*FromRegOp.getParent());
181 
182   FromRegOp.setReg(ToReg);
183 
184   Observer.changedInstr(*FromRegOp.getParent());
185 }
186 
replaceOpcodeWith(MachineInstr & FromMI,unsigned ToOpcode) const187 void CombinerHelper::replaceOpcodeWith(MachineInstr &FromMI,
188                                        unsigned ToOpcode) const {
189   Observer.changingInstr(FromMI);
190 
191   FromMI.setDesc(Builder.getTII().get(ToOpcode));
192 
193   Observer.changedInstr(FromMI);
194 }
195 
getRegBank(Register Reg) const196 const RegisterBank *CombinerHelper::getRegBank(Register Reg) const {
197   return RBI->getRegBank(Reg, MRI, *TRI);
198 }
199 
setRegBank(Register Reg,const RegisterBank * RegBank)200 void CombinerHelper::setRegBank(Register Reg, const RegisterBank *RegBank) {
201   if (RegBank)
202     MRI.setRegBank(Reg, *RegBank);
203 }
204 
tryCombineCopy(MachineInstr & MI)205 bool CombinerHelper::tryCombineCopy(MachineInstr &MI) {
206   if (matchCombineCopy(MI)) {
207     applyCombineCopy(MI);
208     return true;
209   }
210   return false;
211 }
matchCombineCopy(MachineInstr & MI)212 bool CombinerHelper::matchCombineCopy(MachineInstr &MI) {
213   if (MI.getOpcode() != TargetOpcode::COPY)
214     return false;
215   Register DstReg = MI.getOperand(0).getReg();
216   Register SrcReg = MI.getOperand(1).getReg();
217   return canReplaceReg(DstReg, SrcReg, MRI);
218 }
applyCombineCopy(MachineInstr & MI)219 void CombinerHelper::applyCombineCopy(MachineInstr &MI) {
220   Register DstReg = MI.getOperand(0).getReg();
221   Register SrcReg = MI.getOperand(1).getReg();
222   MI.eraseFromParent();
223   replaceRegWith(MRI, DstReg, SrcReg);
224 }
225 
matchFreezeOfSingleMaybePoisonOperand(MachineInstr & MI,BuildFnTy & MatchInfo)226 bool CombinerHelper::matchFreezeOfSingleMaybePoisonOperand(
227     MachineInstr &MI, BuildFnTy &MatchInfo) {
228   // Ported from InstCombinerImpl::pushFreezeToPreventPoisonFromPropagating.
229   Register DstOp = MI.getOperand(0).getReg();
230   Register OrigOp = MI.getOperand(1).getReg();
231 
232   if (!MRI.hasOneNonDBGUse(OrigOp))
233     return false;
234 
235   MachineInstr *OrigDef = MRI.getUniqueVRegDef(OrigOp);
236   // Even if only a single operand of the PHI is not guaranteed non-poison,
237   // moving freeze() backwards across a PHI can cause optimization issues for
238   // other users of that operand.
239   //
240   // Moving freeze() from one of the output registers of a G_UNMERGE_VALUES to
241   // the source register is unprofitable because it makes the freeze() more
242   // strict than is necessary (it would affect the whole register instead of
243   // just the subreg being frozen).
244   if (OrigDef->isPHI() || isa<GUnmerge>(OrigDef))
245     return false;
246 
247   if (canCreateUndefOrPoison(OrigOp, MRI,
248                              /*ConsiderFlagsAndMetadata=*/false))
249     return false;
250 
251   std::optional<MachineOperand> MaybePoisonOperand;
252   for (MachineOperand &Operand : OrigDef->uses()) {
253     if (!Operand.isReg())
254       return false;
255 
256     if (isGuaranteedNotToBeUndefOrPoison(Operand.getReg(), MRI))
257       continue;
258 
259     if (!MaybePoisonOperand)
260       MaybePoisonOperand = Operand;
261     else {
262       // We have more than one maybe-poison operand. Moving the freeze is
263       // unsafe.
264       return false;
265     }
266   }
267 
268   // Eliminate freeze if all operands are guaranteed non-poison.
269   if (!MaybePoisonOperand) {
270     MatchInfo = [=](MachineIRBuilder &B) {
271       Observer.changingInstr(*OrigDef);
272       cast<GenericMachineInstr>(OrigDef)->dropPoisonGeneratingFlags();
273       Observer.changedInstr(*OrigDef);
274       B.buildCopy(DstOp, OrigOp);
275     };
276     return true;
277   }
278 
279   Register MaybePoisonOperandReg = MaybePoisonOperand->getReg();
280   LLT MaybePoisonOperandRegTy = MRI.getType(MaybePoisonOperandReg);
281 
282   MatchInfo = [=](MachineIRBuilder &B) mutable {
283     Observer.changingInstr(*OrigDef);
284     cast<GenericMachineInstr>(OrigDef)->dropPoisonGeneratingFlags();
285     Observer.changedInstr(*OrigDef);
286     B.setInsertPt(*OrigDef->getParent(), OrigDef->getIterator());
287     auto Freeze = B.buildFreeze(MaybePoisonOperandRegTy, MaybePoisonOperandReg);
288     replaceRegOpWith(
289         MRI, *OrigDef->findRegisterUseOperand(MaybePoisonOperandReg, TRI),
290         Freeze.getReg(0));
291     replaceRegWith(MRI, DstOp, OrigOp);
292   };
293   return true;
294 }
295 
matchCombineConcatVectors(MachineInstr & MI,SmallVector<Register> & Ops)296 bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI,
297                                                SmallVector<Register> &Ops) {
298   assert(MI.getOpcode() == TargetOpcode::G_CONCAT_VECTORS &&
299          "Invalid instruction");
300   bool IsUndef = true;
301   MachineInstr *Undef = nullptr;
302 
303   // Walk over all the operands of concat vectors and check if they are
304   // build_vector themselves or undef.
305   // Then collect their operands in Ops.
306   for (const MachineOperand &MO : MI.uses()) {
307     Register Reg = MO.getReg();
308     MachineInstr *Def = MRI.getVRegDef(Reg);
309     assert(Def && "Operand not defined");
310     if (!MRI.hasOneNonDBGUse(Reg))
311       return false;
312     switch (Def->getOpcode()) {
313     case TargetOpcode::G_BUILD_VECTOR:
314       IsUndef = false;
315       // Remember the operands of the build_vector to fold
316       // them into the yet-to-build flattened concat vectors.
317       for (const MachineOperand &BuildVecMO : Def->uses())
318         Ops.push_back(BuildVecMO.getReg());
319       break;
320     case TargetOpcode::G_IMPLICIT_DEF: {
321       LLT OpType = MRI.getType(Reg);
322       // Keep one undef value for all the undef operands.
323       if (!Undef) {
324         Builder.setInsertPt(*MI.getParent(), MI);
325         Undef = Builder.buildUndef(OpType.getScalarType());
326       }
327       assert(MRI.getType(Undef->getOperand(0).getReg()) ==
328                  OpType.getScalarType() &&
329              "All undefs should have the same type");
330       // Break the undef vector in as many scalar elements as needed
331       // for the flattening.
332       for (unsigned EltIdx = 0, EltEnd = OpType.getNumElements();
333            EltIdx != EltEnd; ++EltIdx)
334         Ops.push_back(Undef->getOperand(0).getReg());
335       break;
336     }
337     default:
338       return false;
339     }
340   }
341 
342   // Check if the combine is illegal
343   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
344   if (!isLegalOrBeforeLegalizer(
345           {TargetOpcode::G_BUILD_VECTOR, {DstTy, MRI.getType(Ops[0])}})) {
346     return false;
347   }
348 
349   if (IsUndef)
350     Ops.clear();
351 
352   return true;
353 }
applyCombineConcatVectors(MachineInstr & MI,SmallVector<Register> & Ops)354 void CombinerHelper::applyCombineConcatVectors(MachineInstr &MI,
355                                                SmallVector<Register> &Ops) {
356   // We determined that the concat_vectors can be flatten.
357   // Generate the flattened build_vector.
358   Register DstReg = MI.getOperand(0).getReg();
359   Builder.setInsertPt(*MI.getParent(), MI);
360   Register NewDstReg = MRI.cloneVirtualRegister(DstReg);
361 
362   // Note: IsUndef is sort of redundant. We could have determine it by
363   // checking that at all Ops are undef.  Alternatively, we could have
364   // generate a build_vector of undefs and rely on another combine to
365   // clean that up.  For now, given we already gather this information
366   // in matchCombineConcatVectors, just save compile time and issue the
367   // right thing.
368   if (Ops.empty())
369     Builder.buildUndef(NewDstReg);
370   else
371     Builder.buildBuildVector(NewDstReg, Ops);
372   MI.eraseFromParent();
373   replaceRegWith(MRI, DstReg, NewDstReg);
374 }
375 
matchCombineShuffleConcat(MachineInstr & MI,SmallVector<Register> & Ops)376 bool CombinerHelper::matchCombineShuffleConcat(MachineInstr &MI,
377                                                SmallVector<Register> &Ops) {
378   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
379   auto ConcatMI1 =
380       dyn_cast<GConcatVectors>(MRI.getVRegDef(MI.getOperand(1).getReg()));
381   auto ConcatMI2 =
382       dyn_cast<GConcatVectors>(MRI.getVRegDef(MI.getOperand(2).getReg()));
383   if (!ConcatMI1 || !ConcatMI2)
384     return false;
385 
386   // Check that the sources of the Concat instructions have the same type
387   if (MRI.getType(ConcatMI1->getSourceReg(0)) !=
388       MRI.getType(ConcatMI2->getSourceReg(0)))
389     return false;
390 
391   LLT ConcatSrcTy = MRI.getType(ConcatMI1->getReg(1));
392   LLT ShuffleSrcTy1 = MRI.getType(MI.getOperand(1).getReg());
393   unsigned ConcatSrcNumElt = ConcatSrcTy.getNumElements();
394   for (unsigned i = 0; i < Mask.size(); i += ConcatSrcNumElt) {
395     // Check if the index takes a whole source register from G_CONCAT_VECTORS
396     // Assumes that all Sources of G_CONCAT_VECTORS are the same type
397     if (Mask[i] == -1) {
398       for (unsigned j = 1; j < ConcatSrcNumElt; j++) {
399         if (i + j >= Mask.size())
400           return false;
401         if (Mask[i + j] != -1)
402           return false;
403       }
404       if (!isLegalOrBeforeLegalizer(
405               {TargetOpcode::G_IMPLICIT_DEF, {ConcatSrcTy}}))
406         return false;
407       Ops.push_back(0);
408     } else if (Mask[i] % ConcatSrcNumElt == 0) {
409       for (unsigned j = 1; j < ConcatSrcNumElt; j++) {
410         if (i + j >= Mask.size())
411           return false;
412         if (Mask[i + j] != Mask[i] + static_cast<int>(j))
413           return false;
414       }
415       // Retrieve the source register from its respective G_CONCAT_VECTORS
416       // instruction
417       if (Mask[i] < ShuffleSrcTy1.getNumElements()) {
418         Ops.push_back(ConcatMI1->getSourceReg(Mask[i] / ConcatSrcNumElt));
419       } else {
420         Ops.push_back(ConcatMI2->getSourceReg(Mask[i] / ConcatSrcNumElt -
421                                               ConcatMI1->getNumSources()));
422       }
423     } else {
424       return false;
425     }
426   }
427 
428   if (!isLegalOrBeforeLegalizer(
429           {TargetOpcode::G_CONCAT_VECTORS,
430            {MRI.getType(MI.getOperand(0).getReg()), ConcatSrcTy}}))
431     return false;
432 
433   return !Ops.empty();
434 }
435 
applyCombineShuffleConcat(MachineInstr & MI,SmallVector<Register> & Ops)436 void CombinerHelper::applyCombineShuffleConcat(MachineInstr &MI,
437                                                SmallVector<Register> &Ops) {
438   LLT SrcTy = MRI.getType(Ops[0]);
439   Register UndefReg = 0;
440 
441   for (Register &Reg : Ops) {
442     if (Reg == 0) {
443       if (UndefReg == 0)
444         UndefReg = Builder.buildUndef(SrcTy).getReg(0);
445       Reg = UndefReg;
446     }
447   }
448 
449   if (Ops.size() > 1)
450     Builder.buildConcatVectors(MI.getOperand(0).getReg(), Ops);
451   else
452     Builder.buildCopy(MI.getOperand(0).getReg(), Ops[0]);
453   MI.eraseFromParent();
454 }
455 
tryCombineShuffleVector(MachineInstr & MI)456 bool CombinerHelper::tryCombineShuffleVector(MachineInstr &MI) {
457   SmallVector<Register, 4> Ops;
458   if (matchCombineShuffleVector(MI, Ops)) {
459     applyCombineShuffleVector(MI, Ops);
460     return true;
461   }
462   return false;
463 }
464 
matchCombineShuffleVector(MachineInstr & MI,SmallVectorImpl<Register> & Ops)465 bool CombinerHelper::matchCombineShuffleVector(MachineInstr &MI,
466                                                SmallVectorImpl<Register> &Ops) {
467   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR &&
468          "Invalid instruction kind");
469   LLT DstType = MRI.getType(MI.getOperand(0).getReg());
470   Register Src1 = MI.getOperand(1).getReg();
471   LLT SrcType = MRI.getType(Src1);
472   // As bizarre as it may look, shuffle vector can actually produce
473   // scalar! This is because at the IR level a <1 x ty> shuffle
474   // vector is perfectly valid.
475   unsigned DstNumElts = DstType.isVector() ? DstType.getNumElements() : 1;
476   unsigned SrcNumElts = SrcType.isVector() ? SrcType.getNumElements() : 1;
477 
478   // If the resulting vector is smaller than the size of the source
479   // vectors being concatenated, we won't be able to replace the
480   // shuffle vector into a concat_vectors.
481   //
482   // Note: We may still be able to produce a concat_vectors fed by
483   //       extract_vector_elt and so on. It is less clear that would
484   //       be better though, so don't bother for now.
485   //
486   // If the destination is a scalar, the size of the sources doesn't
487   // matter. we will lower the shuffle to a plain copy. This will
488   // work only if the source and destination have the same size. But
489   // that's covered by the next condition.
490   //
491   // TODO: If the size between the source and destination don't match
492   //       we could still emit an extract vector element in that case.
493   if (DstNumElts < 2 * SrcNumElts && DstNumElts != 1)
494     return false;
495 
496   // Check that the shuffle mask can be broken evenly between the
497   // different sources.
498   if (DstNumElts % SrcNumElts != 0)
499     return false;
500 
501   // Mask length is a multiple of the source vector length.
502   // Check if the shuffle is some kind of concatenation of the input
503   // vectors.
504   unsigned NumConcat = DstNumElts / SrcNumElts;
505   SmallVector<int, 8> ConcatSrcs(NumConcat, -1);
506   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
507   for (unsigned i = 0; i != DstNumElts; ++i) {
508     int Idx = Mask[i];
509     // Undef value.
510     if (Idx < 0)
511       continue;
512     // Ensure the indices in each SrcType sized piece are sequential and that
513     // the same source is used for the whole piece.
514     if ((Idx % SrcNumElts != (i % SrcNumElts)) ||
515         (ConcatSrcs[i / SrcNumElts] >= 0 &&
516          ConcatSrcs[i / SrcNumElts] != (int)(Idx / SrcNumElts)))
517       return false;
518     // Remember which source this index came from.
519     ConcatSrcs[i / SrcNumElts] = Idx / SrcNumElts;
520   }
521 
522   // The shuffle is concatenating multiple vectors together.
523   // Collect the different operands for that.
524   Register UndefReg;
525   Register Src2 = MI.getOperand(2).getReg();
526   for (auto Src : ConcatSrcs) {
527     if (Src < 0) {
528       if (!UndefReg) {
529         Builder.setInsertPt(*MI.getParent(), MI);
530         UndefReg = Builder.buildUndef(SrcType).getReg(0);
531       }
532       Ops.push_back(UndefReg);
533     } else if (Src == 0)
534       Ops.push_back(Src1);
535     else
536       Ops.push_back(Src2);
537   }
538   return true;
539 }
540 
applyCombineShuffleVector(MachineInstr & MI,const ArrayRef<Register> Ops)541 void CombinerHelper::applyCombineShuffleVector(MachineInstr &MI,
542                                                const ArrayRef<Register> Ops) {
543   Register DstReg = MI.getOperand(0).getReg();
544   Builder.setInsertPt(*MI.getParent(), MI);
545   Register NewDstReg = MRI.cloneVirtualRegister(DstReg);
546 
547   if (Ops.size() == 1)
548     Builder.buildCopy(NewDstReg, Ops[0]);
549   else
550     Builder.buildMergeLikeInstr(NewDstReg, Ops);
551 
552   MI.eraseFromParent();
553   replaceRegWith(MRI, DstReg, NewDstReg);
554 }
555 
matchShuffleToExtract(MachineInstr & MI)556 bool CombinerHelper::matchShuffleToExtract(MachineInstr &MI) {
557   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR &&
558          "Invalid instruction kind");
559 
560   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
561   return Mask.size() == 1;
562 }
563 
applyShuffleToExtract(MachineInstr & MI)564 void CombinerHelper::applyShuffleToExtract(MachineInstr &MI) {
565   Register DstReg = MI.getOperand(0).getReg();
566   Builder.setInsertPt(*MI.getParent(), MI);
567 
568   int I = MI.getOperand(3).getShuffleMask()[0];
569   Register Src1 = MI.getOperand(1).getReg();
570   LLT Src1Ty = MRI.getType(Src1);
571   int Src1NumElts = Src1Ty.isVector() ? Src1Ty.getNumElements() : 1;
572   Register SrcReg;
573   if (I >= Src1NumElts) {
574     SrcReg = MI.getOperand(2).getReg();
575     I -= Src1NumElts;
576   } else if (I >= 0)
577     SrcReg = Src1;
578 
579   if (I < 0)
580     Builder.buildUndef(DstReg);
581   else if (!MRI.getType(SrcReg).isVector())
582     Builder.buildCopy(DstReg, SrcReg);
583   else
584     Builder.buildExtractVectorElementConstant(DstReg, SrcReg, I);
585 
586   MI.eraseFromParent();
587 }
588 
589 namespace {
590 
591 /// Select a preference between two uses. CurrentUse is the current preference
592 /// while *ForCandidate is attributes of the candidate under consideration.
ChoosePreferredUse(MachineInstr & LoadMI,PreferredTuple & CurrentUse,const LLT TyForCandidate,unsigned OpcodeForCandidate,MachineInstr * MIForCandidate)593 PreferredTuple ChoosePreferredUse(MachineInstr &LoadMI,
594                                   PreferredTuple &CurrentUse,
595                                   const LLT TyForCandidate,
596                                   unsigned OpcodeForCandidate,
597                                   MachineInstr *MIForCandidate) {
598   if (!CurrentUse.Ty.isValid()) {
599     if (CurrentUse.ExtendOpcode == OpcodeForCandidate ||
600         CurrentUse.ExtendOpcode == TargetOpcode::G_ANYEXT)
601       return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
602     return CurrentUse;
603   }
604 
605   // We permit the extend to hoist through basic blocks but this is only
606   // sensible if the target has extending loads. If you end up lowering back
607   // into a load and extend during the legalizer then the end result is
608   // hoisting the extend up to the load.
609 
610   // Prefer defined extensions to undefined extensions as these are more
611   // likely to reduce the number of instructions.
612   if (OpcodeForCandidate == TargetOpcode::G_ANYEXT &&
613       CurrentUse.ExtendOpcode != TargetOpcode::G_ANYEXT)
614     return CurrentUse;
615   else if (CurrentUse.ExtendOpcode == TargetOpcode::G_ANYEXT &&
616            OpcodeForCandidate != TargetOpcode::G_ANYEXT)
617     return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
618 
619   // Prefer sign extensions to zero extensions as sign-extensions tend to be
620   // more expensive. Don't do this if the load is already a zero-extend load
621   // though, otherwise we'll rewrite a zero-extend load into a sign-extend
622   // later.
623   if (!isa<GZExtLoad>(LoadMI) && CurrentUse.Ty == TyForCandidate) {
624     if (CurrentUse.ExtendOpcode == TargetOpcode::G_SEXT &&
625         OpcodeForCandidate == TargetOpcode::G_ZEXT)
626       return CurrentUse;
627     else if (CurrentUse.ExtendOpcode == TargetOpcode::G_ZEXT &&
628              OpcodeForCandidate == TargetOpcode::G_SEXT)
629       return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
630   }
631 
632   // This is potentially target specific. We've chosen the largest type
633   // because G_TRUNC is usually free. One potential catch with this is that
634   // some targets have a reduced number of larger registers than smaller
635   // registers and this choice potentially increases the live-range for the
636   // larger value.
637   if (TyForCandidate.getSizeInBits() > CurrentUse.Ty.getSizeInBits()) {
638     return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
639   }
640   return CurrentUse;
641 }
642 
643 /// Find a suitable place to insert some instructions and insert them. This
644 /// function accounts for special cases like inserting before a PHI node.
645 /// The current strategy for inserting before PHI's is to duplicate the
646 /// instructions for each predecessor. However, while that's ok for G_TRUNC
647 /// on most targets since it generally requires no code, other targets/cases may
648 /// want to try harder to find a dominating block.
InsertInsnsWithoutSideEffectsBeforeUse(MachineIRBuilder & Builder,MachineInstr & DefMI,MachineOperand & UseMO,std::function<void (MachineBasicBlock *,MachineBasicBlock::iterator,MachineOperand & UseMO)> Inserter)649 static void InsertInsnsWithoutSideEffectsBeforeUse(
650     MachineIRBuilder &Builder, MachineInstr &DefMI, MachineOperand &UseMO,
651     std::function<void(MachineBasicBlock *, MachineBasicBlock::iterator,
652                        MachineOperand &UseMO)>
653         Inserter) {
654   MachineInstr &UseMI = *UseMO.getParent();
655 
656   MachineBasicBlock *InsertBB = UseMI.getParent();
657 
658   // If the use is a PHI then we want the predecessor block instead.
659   if (UseMI.isPHI()) {
660     MachineOperand *PredBB = std::next(&UseMO);
661     InsertBB = PredBB->getMBB();
662   }
663 
664   // If the block is the same block as the def then we want to insert just after
665   // the def instead of at the start of the block.
666   if (InsertBB == DefMI.getParent()) {
667     MachineBasicBlock::iterator InsertPt = &DefMI;
668     Inserter(InsertBB, std::next(InsertPt), UseMO);
669     return;
670   }
671 
672   // Otherwise we want the start of the BB
673   Inserter(InsertBB, InsertBB->getFirstNonPHI(), UseMO);
674 }
675 } // end anonymous namespace
676 
tryCombineExtendingLoads(MachineInstr & MI)677 bool CombinerHelper::tryCombineExtendingLoads(MachineInstr &MI) {
678   PreferredTuple Preferred;
679   if (matchCombineExtendingLoads(MI, Preferred)) {
680     applyCombineExtendingLoads(MI, Preferred);
681     return true;
682   }
683   return false;
684 }
685 
getExtLoadOpcForExtend(unsigned ExtOpc)686 static unsigned getExtLoadOpcForExtend(unsigned ExtOpc) {
687   unsigned CandidateLoadOpc;
688   switch (ExtOpc) {
689   case TargetOpcode::G_ANYEXT:
690     CandidateLoadOpc = TargetOpcode::G_LOAD;
691     break;
692   case TargetOpcode::G_SEXT:
693     CandidateLoadOpc = TargetOpcode::G_SEXTLOAD;
694     break;
695   case TargetOpcode::G_ZEXT:
696     CandidateLoadOpc = TargetOpcode::G_ZEXTLOAD;
697     break;
698   default:
699     llvm_unreachable("Unexpected extend opc");
700   }
701   return CandidateLoadOpc;
702 }
703 
matchCombineExtendingLoads(MachineInstr & MI,PreferredTuple & Preferred)704 bool CombinerHelper::matchCombineExtendingLoads(MachineInstr &MI,
705                                                 PreferredTuple &Preferred) {
706   // We match the loads and follow the uses to the extend instead of matching
707   // the extends and following the def to the load. This is because the load
708   // must remain in the same position for correctness (unless we also add code
709   // to find a safe place to sink it) whereas the extend is freely movable.
710   // It also prevents us from duplicating the load for the volatile case or just
711   // for performance.
712   GAnyLoad *LoadMI = dyn_cast<GAnyLoad>(&MI);
713   if (!LoadMI)
714     return false;
715 
716   Register LoadReg = LoadMI->getDstReg();
717 
718   LLT LoadValueTy = MRI.getType(LoadReg);
719   if (!LoadValueTy.isScalar())
720     return false;
721 
722   // Most architectures are going to legalize <s8 loads into at least a 1 byte
723   // load, and the MMOs can only describe memory accesses in multiples of bytes.
724   // If we try to perform extload combining on those, we can end up with
725   // %a(s8) = extload %ptr (load 1 byte from %ptr)
726   // ... which is an illegal extload instruction.
727   if (LoadValueTy.getSizeInBits() < 8)
728     return false;
729 
730   // For non power-of-2 types, they will very likely be legalized into multiple
731   // loads. Don't bother trying to match them into extending loads.
732   if (!llvm::has_single_bit<uint32_t>(LoadValueTy.getSizeInBits()))
733     return false;
734 
735   // Find the preferred type aside from the any-extends (unless it's the only
736   // one) and non-extending ops. We'll emit an extending load to that type and
737   // and emit a variant of (extend (trunc X)) for the others according to the
738   // relative type sizes. At the same time, pick an extend to use based on the
739   // extend involved in the chosen type.
740   unsigned PreferredOpcode =
741       isa<GLoad>(&MI)
742           ? TargetOpcode::G_ANYEXT
743           : isa<GSExtLoad>(&MI) ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
744   Preferred = {LLT(), PreferredOpcode, nullptr};
745   for (auto &UseMI : MRI.use_nodbg_instructions(LoadReg)) {
746     if (UseMI.getOpcode() == TargetOpcode::G_SEXT ||
747         UseMI.getOpcode() == TargetOpcode::G_ZEXT ||
748         (UseMI.getOpcode() == TargetOpcode::G_ANYEXT)) {
749       const auto &MMO = LoadMI->getMMO();
750       // Don't do anything for atomics.
751       if (MMO.isAtomic())
752         continue;
753       // Check for legality.
754       if (!isPreLegalize()) {
755         LegalityQuery::MemDesc MMDesc(MMO);
756         unsigned CandidateLoadOpc = getExtLoadOpcForExtend(UseMI.getOpcode());
757         LLT UseTy = MRI.getType(UseMI.getOperand(0).getReg());
758         LLT SrcTy = MRI.getType(LoadMI->getPointerReg());
759         if (LI->getAction({CandidateLoadOpc, {UseTy, SrcTy}, {MMDesc}})
760                 .Action != LegalizeActions::Legal)
761           continue;
762       }
763       Preferred = ChoosePreferredUse(MI, Preferred,
764                                      MRI.getType(UseMI.getOperand(0).getReg()),
765                                      UseMI.getOpcode(), &UseMI);
766     }
767   }
768 
769   // There were no extends
770   if (!Preferred.MI)
771     return false;
772   // It should be impossible to chose an extend without selecting a different
773   // type since by definition the result of an extend is larger.
774   assert(Preferred.Ty != LoadValueTy && "Extending to same type?");
775 
776   LLVM_DEBUG(dbgs() << "Preferred use is: " << *Preferred.MI);
777   return true;
778 }
779 
applyCombineExtendingLoads(MachineInstr & MI,PreferredTuple & Preferred)780 void CombinerHelper::applyCombineExtendingLoads(MachineInstr &MI,
781                                                 PreferredTuple &Preferred) {
782   // Rewrite the load to the chosen extending load.
783   Register ChosenDstReg = Preferred.MI->getOperand(0).getReg();
784 
785   // Inserter to insert a truncate back to the original type at a given point
786   // with some basic CSE to limit truncate duplication to one per BB.
787   DenseMap<MachineBasicBlock *, MachineInstr *> EmittedInsns;
788   auto InsertTruncAt = [&](MachineBasicBlock *InsertIntoBB,
789                            MachineBasicBlock::iterator InsertBefore,
790                            MachineOperand &UseMO) {
791     MachineInstr *PreviouslyEmitted = EmittedInsns.lookup(InsertIntoBB);
792     if (PreviouslyEmitted) {
793       Observer.changingInstr(*UseMO.getParent());
794       UseMO.setReg(PreviouslyEmitted->getOperand(0).getReg());
795       Observer.changedInstr(*UseMO.getParent());
796       return;
797     }
798 
799     Builder.setInsertPt(*InsertIntoBB, InsertBefore);
800     Register NewDstReg = MRI.cloneVirtualRegister(MI.getOperand(0).getReg());
801     MachineInstr *NewMI = Builder.buildTrunc(NewDstReg, ChosenDstReg);
802     EmittedInsns[InsertIntoBB] = NewMI;
803     replaceRegOpWith(MRI, UseMO, NewDstReg);
804   };
805 
806   Observer.changingInstr(MI);
807   unsigned LoadOpc = getExtLoadOpcForExtend(Preferred.ExtendOpcode);
808   MI.setDesc(Builder.getTII().get(LoadOpc));
809 
810   // Rewrite all the uses to fix up the types.
811   auto &LoadValue = MI.getOperand(0);
812   SmallVector<MachineOperand *, 4> Uses;
813   for (auto &UseMO : MRI.use_operands(LoadValue.getReg()))
814     Uses.push_back(&UseMO);
815 
816   for (auto *UseMO : Uses) {
817     MachineInstr *UseMI = UseMO->getParent();
818 
819     // If the extend is compatible with the preferred extend then we should fix
820     // up the type and extend so that it uses the preferred use.
821     if (UseMI->getOpcode() == Preferred.ExtendOpcode ||
822         UseMI->getOpcode() == TargetOpcode::G_ANYEXT) {
823       Register UseDstReg = UseMI->getOperand(0).getReg();
824       MachineOperand &UseSrcMO = UseMI->getOperand(1);
825       const LLT UseDstTy = MRI.getType(UseDstReg);
826       if (UseDstReg != ChosenDstReg) {
827         if (Preferred.Ty == UseDstTy) {
828           // If the use has the same type as the preferred use, then merge
829           // the vregs and erase the extend. For example:
830           //    %1:_(s8) = G_LOAD ...
831           //    %2:_(s32) = G_SEXT %1(s8)
832           //    %3:_(s32) = G_ANYEXT %1(s8)
833           //    ... = ... %3(s32)
834           // rewrites to:
835           //    %2:_(s32) = G_SEXTLOAD ...
836           //    ... = ... %2(s32)
837           replaceRegWith(MRI, UseDstReg, ChosenDstReg);
838           Observer.erasingInstr(*UseMO->getParent());
839           UseMO->getParent()->eraseFromParent();
840         } else if (Preferred.Ty.getSizeInBits() < UseDstTy.getSizeInBits()) {
841           // If the preferred size is smaller, then keep the extend but extend
842           // from the result of the extending load. For example:
843           //    %1:_(s8) = G_LOAD ...
844           //    %2:_(s32) = G_SEXT %1(s8)
845           //    %3:_(s64) = G_ANYEXT %1(s8)
846           //    ... = ... %3(s64)
847           /// rewrites to:
848           //    %2:_(s32) = G_SEXTLOAD ...
849           //    %3:_(s64) = G_ANYEXT %2:_(s32)
850           //    ... = ... %3(s64)
851           replaceRegOpWith(MRI, UseSrcMO, ChosenDstReg);
852         } else {
853           // If the preferred size is large, then insert a truncate. For
854           // example:
855           //    %1:_(s8) = G_LOAD ...
856           //    %2:_(s64) = G_SEXT %1(s8)
857           //    %3:_(s32) = G_ZEXT %1(s8)
858           //    ... = ... %3(s32)
859           /// rewrites to:
860           //    %2:_(s64) = G_SEXTLOAD ...
861           //    %4:_(s8) = G_TRUNC %2:_(s32)
862           //    %3:_(s64) = G_ZEXT %2:_(s8)
863           //    ... = ... %3(s64)
864           InsertInsnsWithoutSideEffectsBeforeUse(Builder, MI, *UseMO,
865                                                  InsertTruncAt);
866         }
867         continue;
868       }
869       // The use is (one of) the uses of the preferred use we chose earlier.
870       // We're going to update the load to def this value later so just erase
871       // the old extend.
872       Observer.erasingInstr(*UseMO->getParent());
873       UseMO->getParent()->eraseFromParent();
874       continue;
875     }
876 
877     // The use isn't an extend. Truncate back to the type we originally loaded.
878     // This is free on many targets.
879     InsertInsnsWithoutSideEffectsBeforeUse(Builder, MI, *UseMO, InsertTruncAt);
880   }
881 
882   MI.getOperand(0).setReg(ChosenDstReg);
883   Observer.changedInstr(MI);
884 }
885 
matchCombineLoadWithAndMask(MachineInstr & MI,BuildFnTy & MatchInfo)886 bool CombinerHelper::matchCombineLoadWithAndMask(MachineInstr &MI,
887                                                  BuildFnTy &MatchInfo) {
888   assert(MI.getOpcode() == TargetOpcode::G_AND);
889 
890   // If we have the following code:
891   //  %mask = G_CONSTANT 255
892   //  %ld   = G_LOAD %ptr, (load s16)
893   //  %and  = G_AND %ld, %mask
894   //
895   // Try to fold it into
896   //   %ld = G_ZEXTLOAD %ptr, (load s8)
897 
898   Register Dst = MI.getOperand(0).getReg();
899   if (MRI.getType(Dst).isVector())
900     return false;
901 
902   auto MaybeMask =
903       getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
904   if (!MaybeMask)
905     return false;
906 
907   APInt MaskVal = MaybeMask->Value;
908 
909   if (!MaskVal.isMask())
910     return false;
911 
912   Register SrcReg = MI.getOperand(1).getReg();
913   // Don't use getOpcodeDef() here since intermediate instructions may have
914   // multiple users.
915   GAnyLoad *LoadMI = dyn_cast<GAnyLoad>(MRI.getVRegDef(SrcReg));
916   if (!LoadMI || !MRI.hasOneNonDBGUse(LoadMI->getDstReg()))
917     return false;
918 
919   Register LoadReg = LoadMI->getDstReg();
920   LLT RegTy = MRI.getType(LoadReg);
921   Register PtrReg = LoadMI->getPointerReg();
922   unsigned RegSize = RegTy.getSizeInBits();
923   LocationSize LoadSizeBits = LoadMI->getMemSizeInBits();
924   unsigned MaskSizeBits = MaskVal.countr_one();
925 
926   // The mask may not be larger than the in-memory type, as it might cover sign
927   // extended bits
928   if (MaskSizeBits > LoadSizeBits.getValue())
929     return false;
930 
931   // If the mask covers the whole destination register, there's nothing to
932   // extend
933   if (MaskSizeBits >= RegSize)
934     return false;
935 
936   // Most targets cannot deal with loads of size < 8 and need to re-legalize to
937   // at least byte loads. Avoid creating such loads here
938   if (MaskSizeBits < 8 || !isPowerOf2_32(MaskSizeBits))
939     return false;
940 
941   const MachineMemOperand &MMO = LoadMI->getMMO();
942   LegalityQuery::MemDesc MemDesc(MMO);
943 
944   // Don't modify the memory access size if this is atomic/volatile, but we can
945   // still adjust the opcode to indicate the high bit behavior.
946   if (LoadMI->isSimple())
947     MemDesc.MemoryTy = LLT::scalar(MaskSizeBits);
948   else if (LoadSizeBits.getValue() > MaskSizeBits ||
949            LoadSizeBits.getValue() == RegSize)
950     return false;
951 
952   // TODO: Could check if it's legal with the reduced or original memory size.
953   if (!isLegalOrBeforeLegalizer(
954           {TargetOpcode::G_ZEXTLOAD, {RegTy, MRI.getType(PtrReg)}, {MemDesc}}))
955     return false;
956 
957   MatchInfo = [=](MachineIRBuilder &B) {
958     B.setInstrAndDebugLoc(*LoadMI);
959     auto &MF = B.getMF();
960     auto PtrInfo = MMO.getPointerInfo();
961     auto *NewMMO = MF.getMachineMemOperand(&MMO, PtrInfo, MemDesc.MemoryTy);
962     B.buildLoadInstr(TargetOpcode::G_ZEXTLOAD, Dst, PtrReg, *NewMMO);
963     LoadMI->eraseFromParent();
964   };
965   return true;
966 }
967 
isPredecessor(const MachineInstr & DefMI,const MachineInstr & UseMI)968 bool CombinerHelper::isPredecessor(const MachineInstr &DefMI,
969                                    const MachineInstr &UseMI) {
970   assert(!DefMI.isDebugInstr() && !UseMI.isDebugInstr() &&
971          "shouldn't consider debug uses");
972   assert(DefMI.getParent() == UseMI.getParent());
973   if (&DefMI == &UseMI)
974     return true;
975   const MachineBasicBlock &MBB = *DefMI.getParent();
976   auto DefOrUse = find_if(MBB, [&DefMI, &UseMI](const MachineInstr &MI) {
977     return &MI == &DefMI || &MI == &UseMI;
978   });
979   if (DefOrUse == MBB.end())
980     llvm_unreachable("Block must contain both DefMI and UseMI!");
981   return &*DefOrUse == &DefMI;
982 }
983 
dominates(const MachineInstr & DefMI,const MachineInstr & UseMI)984 bool CombinerHelper::dominates(const MachineInstr &DefMI,
985                                const MachineInstr &UseMI) {
986   assert(!DefMI.isDebugInstr() && !UseMI.isDebugInstr() &&
987          "shouldn't consider debug uses");
988   if (MDT)
989     return MDT->dominates(&DefMI, &UseMI);
990   else if (DefMI.getParent() != UseMI.getParent())
991     return false;
992 
993   return isPredecessor(DefMI, UseMI);
994 }
995 
matchSextTruncSextLoad(MachineInstr & MI)996 bool CombinerHelper::matchSextTruncSextLoad(MachineInstr &MI) {
997   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
998   Register SrcReg = MI.getOperand(1).getReg();
999   Register LoadUser = SrcReg;
1000 
1001   if (MRI.getType(SrcReg).isVector())
1002     return false;
1003 
1004   Register TruncSrc;
1005   if (mi_match(SrcReg, MRI, m_GTrunc(m_Reg(TruncSrc))))
1006     LoadUser = TruncSrc;
1007 
1008   uint64_t SizeInBits = MI.getOperand(2).getImm();
1009   // If the source is a G_SEXTLOAD from the same bit width, then we don't
1010   // need any extend at all, just a truncate.
1011   if (auto *LoadMI = getOpcodeDef<GSExtLoad>(LoadUser, MRI)) {
1012     // If truncating more than the original extended value, abort.
1013     auto LoadSizeBits = LoadMI->getMemSizeInBits();
1014     if (TruncSrc &&
1015         MRI.getType(TruncSrc).getSizeInBits() < LoadSizeBits.getValue())
1016       return false;
1017     if (LoadSizeBits == SizeInBits)
1018       return true;
1019   }
1020   return false;
1021 }
1022 
applySextTruncSextLoad(MachineInstr & MI)1023 void CombinerHelper::applySextTruncSextLoad(MachineInstr &MI) {
1024   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
1025   Builder.buildCopy(MI.getOperand(0).getReg(), MI.getOperand(1).getReg());
1026   MI.eraseFromParent();
1027 }
1028 
matchSextInRegOfLoad(MachineInstr & MI,std::tuple<Register,unsigned> & MatchInfo)1029 bool CombinerHelper::matchSextInRegOfLoad(
1030     MachineInstr &MI, std::tuple<Register, unsigned> &MatchInfo) {
1031   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
1032 
1033   Register DstReg = MI.getOperand(0).getReg();
1034   LLT RegTy = MRI.getType(DstReg);
1035 
1036   // Only supports scalars for now.
1037   if (RegTy.isVector())
1038     return false;
1039 
1040   Register SrcReg = MI.getOperand(1).getReg();
1041   auto *LoadDef = getOpcodeDef<GLoad>(SrcReg, MRI);
1042   if (!LoadDef || !MRI.hasOneNonDBGUse(DstReg))
1043     return false;
1044 
1045   uint64_t MemBits = LoadDef->getMemSizeInBits().getValue();
1046 
1047   // If the sign extend extends from a narrower width than the load's width,
1048   // then we can narrow the load width when we combine to a G_SEXTLOAD.
1049   // Avoid widening the load at all.
1050   unsigned NewSizeBits = std::min((uint64_t)MI.getOperand(2).getImm(), MemBits);
1051 
1052   // Don't generate G_SEXTLOADs with a < 1 byte width.
1053   if (NewSizeBits < 8)
1054     return false;
1055   // Don't bother creating a non-power-2 sextload, it will likely be broken up
1056   // anyway for most targets.
1057   if (!isPowerOf2_32(NewSizeBits))
1058     return false;
1059 
1060   const MachineMemOperand &MMO = LoadDef->getMMO();
1061   LegalityQuery::MemDesc MMDesc(MMO);
1062 
1063   // Don't modify the memory access size if this is atomic/volatile, but we can
1064   // still adjust the opcode to indicate the high bit behavior.
1065   if (LoadDef->isSimple())
1066     MMDesc.MemoryTy = LLT::scalar(NewSizeBits);
1067   else if (MemBits > NewSizeBits || MemBits == RegTy.getSizeInBits())
1068     return false;
1069 
1070   // TODO: Could check if it's legal with the reduced or original memory size.
1071   if (!isLegalOrBeforeLegalizer({TargetOpcode::G_SEXTLOAD,
1072                                  {MRI.getType(LoadDef->getDstReg()),
1073                                   MRI.getType(LoadDef->getPointerReg())},
1074                                  {MMDesc}}))
1075     return false;
1076 
1077   MatchInfo = std::make_tuple(LoadDef->getDstReg(), NewSizeBits);
1078   return true;
1079 }
1080 
applySextInRegOfLoad(MachineInstr & MI,std::tuple<Register,unsigned> & MatchInfo)1081 void CombinerHelper::applySextInRegOfLoad(
1082     MachineInstr &MI, std::tuple<Register, unsigned> &MatchInfo) {
1083   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
1084   Register LoadReg;
1085   unsigned ScalarSizeBits;
1086   std::tie(LoadReg, ScalarSizeBits) = MatchInfo;
1087   GLoad *LoadDef = cast<GLoad>(MRI.getVRegDef(LoadReg));
1088 
1089   // If we have the following:
1090   // %ld = G_LOAD %ptr, (load 2)
1091   // %ext = G_SEXT_INREG %ld, 8
1092   //    ==>
1093   // %ld = G_SEXTLOAD %ptr (load 1)
1094 
1095   auto &MMO = LoadDef->getMMO();
1096   Builder.setInstrAndDebugLoc(*LoadDef);
1097   auto &MF = Builder.getMF();
1098   auto PtrInfo = MMO.getPointerInfo();
1099   auto *NewMMO = MF.getMachineMemOperand(&MMO, PtrInfo, ScalarSizeBits / 8);
1100   Builder.buildLoadInstr(TargetOpcode::G_SEXTLOAD, MI.getOperand(0).getReg(),
1101                          LoadDef->getPointerReg(), *NewMMO);
1102   MI.eraseFromParent();
1103 }
1104 
1105 /// Return true if 'MI' is a load or a store that may be fold it's address
1106 /// operand into the load / store addressing mode.
canFoldInAddressingMode(GLoadStore * MI,const TargetLowering & TLI,MachineRegisterInfo & MRI)1107 static bool canFoldInAddressingMode(GLoadStore *MI, const TargetLowering &TLI,
1108                                     MachineRegisterInfo &MRI) {
1109   TargetLowering::AddrMode AM;
1110   auto *MF = MI->getMF();
1111   auto *Addr = getOpcodeDef<GPtrAdd>(MI->getPointerReg(), MRI);
1112   if (!Addr)
1113     return false;
1114 
1115   AM.HasBaseReg = true;
1116   if (auto CstOff = getIConstantVRegVal(Addr->getOffsetReg(), MRI))
1117     AM.BaseOffs = CstOff->getSExtValue(); // [reg +/- imm]
1118   else
1119     AM.Scale = 1; // [reg +/- reg]
1120 
1121   return TLI.isLegalAddressingMode(
1122       MF->getDataLayout(), AM,
1123       getTypeForLLT(MI->getMMO().getMemoryType(),
1124                     MF->getFunction().getContext()),
1125       MI->getMMO().getAddrSpace());
1126 }
1127 
getIndexedOpc(unsigned LdStOpc)1128 static unsigned getIndexedOpc(unsigned LdStOpc) {
1129   switch (LdStOpc) {
1130   case TargetOpcode::G_LOAD:
1131     return TargetOpcode::G_INDEXED_LOAD;
1132   case TargetOpcode::G_STORE:
1133     return TargetOpcode::G_INDEXED_STORE;
1134   case TargetOpcode::G_ZEXTLOAD:
1135     return TargetOpcode::G_INDEXED_ZEXTLOAD;
1136   case TargetOpcode::G_SEXTLOAD:
1137     return TargetOpcode::G_INDEXED_SEXTLOAD;
1138   default:
1139     llvm_unreachable("Unexpected opcode");
1140   }
1141 }
1142 
isIndexedLoadStoreLegal(GLoadStore & LdSt) const1143 bool CombinerHelper::isIndexedLoadStoreLegal(GLoadStore &LdSt) const {
1144   // Check for legality.
1145   LLT PtrTy = MRI.getType(LdSt.getPointerReg());
1146   LLT Ty = MRI.getType(LdSt.getReg(0));
1147   LLT MemTy = LdSt.getMMO().getMemoryType();
1148   SmallVector<LegalityQuery::MemDesc, 2> MemDescrs(
1149       {{MemTy, MemTy.getSizeInBits().getKnownMinValue(),
1150         AtomicOrdering::NotAtomic}});
1151   unsigned IndexedOpc = getIndexedOpc(LdSt.getOpcode());
1152   SmallVector<LLT> OpTys;
1153   if (IndexedOpc == TargetOpcode::G_INDEXED_STORE)
1154     OpTys = {PtrTy, Ty, Ty};
1155   else
1156     OpTys = {Ty, PtrTy}; // For G_INDEXED_LOAD, G_INDEXED_[SZ]EXTLOAD
1157 
1158   LegalityQuery Q(IndexedOpc, OpTys, MemDescrs);
1159   return isLegal(Q);
1160 }
1161 
1162 static cl::opt<unsigned> PostIndexUseThreshold(
1163     "post-index-use-threshold", cl::Hidden, cl::init(32),
1164     cl::desc("Number of uses of a base pointer to check before it is no longer "
1165              "considered for post-indexing."));
1166 
findPostIndexCandidate(GLoadStore & LdSt,Register & Addr,Register & Base,Register & Offset,bool & RematOffset)1167 bool CombinerHelper::findPostIndexCandidate(GLoadStore &LdSt, Register &Addr,
1168                                             Register &Base, Register &Offset,
1169                                             bool &RematOffset) {
1170   // We're looking for the following pattern, for either load or store:
1171   // %baseptr:_(p0) = ...
1172   // G_STORE %val(s64), %baseptr(p0)
1173   // %offset:_(s64) = G_CONSTANT i64 -256
1174   // %new_addr:_(p0) = G_PTR_ADD %baseptr, %offset(s64)
1175   const auto &TLI = getTargetLowering();
1176 
1177   Register Ptr = LdSt.getPointerReg();
1178   // If the store is the only use, don't bother.
1179   if (MRI.hasOneNonDBGUse(Ptr))
1180     return false;
1181 
1182   if (!isIndexedLoadStoreLegal(LdSt))
1183     return false;
1184 
1185   if (getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Ptr, MRI))
1186     return false;
1187 
1188   MachineInstr *StoredValDef = getDefIgnoringCopies(LdSt.getReg(0), MRI);
1189   auto *PtrDef = MRI.getVRegDef(Ptr);
1190 
1191   unsigned NumUsesChecked = 0;
1192   for (auto &Use : MRI.use_nodbg_instructions(Ptr)) {
1193     if (++NumUsesChecked > PostIndexUseThreshold)
1194       return false; // Try to avoid exploding compile time.
1195 
1196     auto *PtrAdd = dyn_cast<GPtrAdd>(&Use);
1197     // The use itself might be dead. This can happen during combines if DCE
1198     // hasn't had a chance to run yet. Don't allow it to form an indexed op.
1199     if (!PtrAdd || MRI.use_nodbg_empty(PtrAdd->getReg(0)))
1200       continue;
1201 
1202     // Check the user of this isn't the store, otherwise we'd be generate a
1203     // indexed store defining its own use.
1204     if (StoredValDef == &Use)
1205       continue;
1206 
1207     Offset = PtrAdd->getOffsetReg();
1208     if (!ForceLegalIndexing &&
1209         !TLI.isIndexingLegal(LdSt, PtrAdd->getBaseReg(), Offset,
1210                              /*IsPre*/ false, MRI))
1211       continue;
1212 
1213     // Make sure the offset calculation is before the potentially indexed op.
1214     MachineInstr *OffsetDef = MRI.getVRegDef(Offset);
1215     RematOffset = false;
1216     if (!dominates(*OffsetDef, LdSt)) {
1217       // If the offset however is just a G_CONSTANT, we can always just
1218       // rematerialize it where we need it.
1219       if (OffsetDef->getOpcode() != TargetOpcode::G_CONSTANT)
1220         continue;
1221       RematOffset = true;
1222     }
1223 
1224     for (auto &BasePtrUse : MRI.use_nodbg_instructions(PtrAdd->getBaseReg())) {
1225       if (&BasePtrUse == PtrDef)
1226         continue;
1227 
1228       // If the user is a later load/store that can be post-indexed, then don't
1229       // combine this one.
1230       auto *BasePtrLdSt = dyn_cast<GLoadStore>(&BasePtrUse);
1231       if (BasePtrLdSt && BasePtrLdSt != &LdSt &&
1232           dominates(LdSt, *BasePtrLdSt) &&
1233           isIndexedLoadStoreLegal(*BasePtrLdSt))
1234         return false;
1235 
1236       // Now we're looking for the key G_PTR_ADD instruction, which contains
1237       // the offset add that we want to fold.
1238       if (auto *BasePtrUseDef = dyn_cast<GPtrAdd>(&BasePtrUse)) {
1239         Register PtrAddDefReg = BasePtrUseDef->getReg(0);
1240         for (auto &BaseUseUse : MRI.use_nodbg_instructions(PtrAddDefReg)) {
1241           // If the use is in a different block, then we may produce worse code
1242           // due to the extra register pressure.
1243           if (BaseUseUse.getParent() != LdSt.getParent())
1244             return false;
1245 
1246           if (auto *UseUseLdSt = dyn_cast<GLoadStore>(&BaseUseUse))
1247             if (canFoldInAddressingMode(UseUseLdSt, TLI, MRI))
1248               return false;
1249         }
1250         if (!dominates(LdSt, BasePtrUse))
1251           return false; // All use must be dominated by the load/store.
1252       }
1253     }
1254 
1255     Addr = PtrAdd->getReg(0);
1256     Base = PtrAdd->getBaseReg();
1257     return true;
1258   }
1259 
1260   return false;
1261 }
1262 
findPreIndexCandidate(GLoadStore & LdSt,Register & Addr,Register & Base,Register & Offset)1263 bool CombinerHelper::findPreIndexCandidate(GLoadStore &LdSt, Register &Addr,
1264                                            Register &Base, Register &Offset) {
1265   auto &MF = *LdSt.getParent()->getParent();
1266   const auto &TLI = *MF.getSubtarget().getTargetLowering();
1267 
1268   Addr = LdSt.getPointerReg();
1269   if (!mi_match(Addr, MRI, m_GPtrAdd(m_Reg(Base), m_Reg(Offset))) ||
1270       MRI.hasOneNonDBGUse(Addr))
1271     return false;
1272 
1273   if (!ForceLegalIndexing &&
1274       !TLI.isIndexingLegal(LdSt, Base, Offset, /*IsPre*/ true, MRI))
1275     return false;
1276 
1277   if (!isIndexedLoadStoreLegal(LdSt))
1278     return false;
1279 
1280   MachineInstr *BaseDef = getDefIgnoringCopies(Base, MRI);
1281   if (BaseDef->getOpcode() == TargetOpcode::G_FRAME_INDEX)
1282     return false;
1283 
1284   if (auto *St = dyn_cast<GStore>(&LdSt)) {
1285     // Would require a copy.
1286     if (Base == St->getValueReg())
1287       return false;
1288 
1289     // We're expecting one use of Addr in MI, but it could also be the
1290     // value stored, which isn't actually dominated by the instruction.
1291     if (St->getValueReg() == Addr)
1292       return false;
1293   }
1294 
1295   // Avoid increasing cross-block register pressure.
1296   for (auto &AddrUse : MRI.use_nodbg_instructions(Addr))
1297     if (AddrUse.getParent() != LdSt.getParent())
1298       return false;
1299 
1300   // FIXME: check whether all uses of the base pointer are constant PtrAdds.
1301   // That might allow us to end base's liveness here by adjusting the constant.
1302   bool RealUse = false;
1303   for (auto &AddrUse : MRI.use_nodbg_instructions(Addr)) {
1304     if (!dominates(LdSt, AddrUse))
1305       return false; // All use must be dominated by the load/store.
1306 
1307     // If Ptr may be folded in addressing mode of other use, then it's
1308     // not profitable to do this transformation.
1309     if (auto *UseLdSt = dyn_cast<GLoadStore>(&AddrUse)) {
1310       if (!canFoldInAddressingMode(UseLdSt, TLI, MRI))
1311         RealUse = true;
1312     } else {
1313       RealUse = true;
1314     }
1315   }
1316   return RealUse;
1317 }
1318 
matchCombineExtractedVectorLoad(MachineInstr & MI,BuildFnTy & MatchInfo)1319 bool CombinerHelper::matchCombineExtractedVectorLoad(MachineInstr &MI,
1320                                                      BuildFnTy &MatchInfo) {
1321   assert(MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT);
1322 
1323   // Check if there is a load that defines the vector being extracted from.
1324   auto *LoadMI = getOpcodeDef<GLoad>(MI.getOperand(1).getReg(), MRI);
1325   if (!LoadMI)
1326     return false;
1327 
1328   Register Vector = MI.getOperand(1).getReg();
1329   LLT VecEltTy = MRI.getType(Vector).getElementType();
1330 
1331   assert(MRI.getType(MI.getOperand(0).getReg()) == VecEltTy);
1332 
1333   // Checking whether we should reduce the load width.
1334   if (!MRI.hasOneNonDBGUse(Vector))
1335     return false;
1336 
1337   // Check if the defining load is simple.
1338   if (!LoadMI->isSimple())
1339     return false;
1340 
1341   // If the vector element type is not a multiple of a byte then we are unable
1342   // to correctly compute an address to load only the extracted element as a
1343   // scalar.
1344   if (!VecEltTy.isByteSized())
1345     return false;
1346 
1347   // Check for load fold barriers between the extraction and the load.
1348   if (MI.getParent() != LoadMI->getParent())
1349     return false;
1350   const unsigned MaxIter = 20;
1351   unsigned Iter = 0;
1352   for (auto II = LoadMI->getIterator(), IE = MI.getIterator(); II != IE; ++II) {
1353     if (II->isLoadFoldBarrier())
1354       return false;
1355     if (Iter++ == MaxIter)
1356       return false;
1357   }
1358 
1359   // Check if the new load that we are going to create is legal
1360   // if we are in the post-legalization phase.
1361   MachineMemOperand MMO = LoadMI->getMMO();
1362   Align Alignment = MMO.getAlign();
1363   MachinePointerInfo PtrInfo;
1364   uint64_t Offset;
1365 
1366   // Finding the appropriate PtrInfo if offset is a known constant.
1367   // This is required to create the memory operand for the narrowed load.
1368   // This machine memory operand object helps us infer about legality
1369   // before we proceed to combine the instruction.
1370   if (auto CVal = getIConstantVRegVal(Vector, MRI)) {
1371     int Elt = CVal->getZExtValue();
1372     // FIXME: should be (ABI size)*Elt.
1373     Offset = VecEltTy.getSizeInBits() * Elt / 8;
1374     PtrInfo = MMO.getPointerInfo().getWithOffset(Offset);
1375   } else {
1376     // Discard the pointer info except the address space because the memory
1377     // operand can't represent this new access since the offset is variable.
1378     Offset = VecEltTy.getSizeInBits() / 8;
1379     PtrInfo = MachinePointerInfo(MMO.getPointerInfo().getAddrSpace());
1380   }
1381 
1382   Alignment = commonAlignment(Alignment, Offset);
1383 
1384   Register VecPtr = LoadMI->getPointerReg();
1385   LLT PtrTy = MRI.getType(VecPtr);
1386 
1387   MachineFunction &MF = *MI.getMF();
1388   auto *NewMMO = MF.getMachineMemOperand(&MMO, PtrInfo, VecEltTy);
1389 
1390   LegalityQuery::MemDesc MMDesc(*NewMMO);
1391 
1392   LegalityQuery Q = {TargetOpcode::G_LOAD, {VecEltTy, PtrTy}, {MMDesc}};
1393 
1394   if (!isLegalOrBeforeLegalizer(Q))
1395     return false;
1396 
1397   // Load must be allowed and fast on the target.
1398   LLVMContext &C = MF.getFunction().getContext();
1399   auto &DL = MF.getDataLayout();
1400   unsigned Fast = 0;
1401   if (!getTargetLowering().allowsMemoryAccess(C, DL, VecEltTy, *NewMMO,
1402                                               &Fast) ||
1403       !Fast)
1404     return false;
1405 
1406   Register Result = MI.getOperand(0).getReg();
1407   Register Index = MI.getOperand(2).getReg();
1408 
1409   MatchInfo = [=](MachineIRBuilder &B) {
1410     GISelObserverWrapper DummyObserver;
1411     LegalizerHelper Helper(B.getMF(), DummyObserver, B);
1412     //// Get pointer to the vector element.
1413     Register finalPtr = Helper.getVectorElementPointer(
1414         LoadMI->getPointerReg(), MRI.getType(LoadMI->getOperand(0).getReg()),
1415         Index);
1416     // New G_LOAD instruction.
1417     B.buildLoad(Result, finalPtr, PtrInfo, Alignment);
1418     // Remove original GLOAD instruction.
1419     LoadMI->eraseFromParent();
1420   };
1421 
1422   return true;
1423 }
1424 
matchCombineIndexedLoadStore(MachineInstr & MI,IndexedLoadStoreMatchInfo & MatchInfo)1425 bool CombinerHelper::matchCombineIndexedLoadStore(
1426     MachineInstr &MI, IndexedLoadStoreMatchInfo &MatchInfo) {
1427   auto &LdSt = cast<GLoadStore>(MI);
1428 
1429   if (LdSt.isAtomic())
1430     return false;
1431 
1432   MatchInfo.IsPre = findPreIndexCandidate(LdSt, MatchInfo.Addr, MatchInfo.Base,
1433                                           MatchInfo.Offset);
1434   if (!MatchInfo.IsPre &&
1435       !findPostIndexCandidate(LdSt, MatchInfo.Addr, MatchInfo.Base,
1436                               MatchInfo.Offset, MatchInfo.RematOffset))
1437     return false;
1438 
1439   return true;
1440 }
1441 
applyCombineIndexedLoadStore(MachineInstr & MI,IndexedLoadStoreMatchInfo & MatchInfo)1442 void CombinerHelper::applyCombineIndexedLoadStore(
1443     MachineInstr &MI, IndexedLoadStoreMatchInfo &MatchInfo) {
1444   MachineInstr &AddrDef = *MRI.getUniqueVRegDef(MatchInfo.Addr);
1445   unsigned Opcode = MI.getOpcode();
1446   bool IsStore = Opcode == TargetOpcode::G_STORE;
1447   unsigned NewOpcode = getIndexedOpc(Opcode);
1448 
1449   // If the offset constant didn't happen to dominate the load/store, we can
1450   // just clone it as needed.
1451   if (MatchInfo.RematOffset) {
1452     auto *OldCst = MRI.getVRegDef(MatchInfo.Offset);
1453     auto NewCst = Builder.buildConstant(MRI.getType(MatchInfo.Offset),
1454                                         *OldCst->getOperand(1).getCImm());
1455     MatchInfo.Offset = NewCst.getReg(0);
1456   }
1457 
1458   auto MIB = Builder.buildInstr(NewOpcode);
1459   if (IsStore) {
1460     MIB.addDef(MatchInfo.Addr);
1461     MIB.addUse(MI.getOperand(0).getReg());
1462   } else {
1463     MIB.addDef(MI.getOperand(0).getReg());
1464     MIB.addDef(MatchInfo.Addr);
1465   }
1466 
1467   MIB.addUse(MatchInfo.Base);
1468   MIB.addUse(MatchInfo.Offset);
1469   MIB.addImm(MatchInfo.IsPre);
1470   MIB->cloneMemRefs(*MI.getMF(), MI);
1471   MI.eraseFromParent();
1472   AddrDef.eraseFromParent();
1473 
1474   LLVM_DEBUG(dbgs() << "    Combinined to indexed operation");
1475 }
1476 
matchCombineDivRem(MachineInstr & MI,MachineInstr * & OtherMI)1477 bool CombinerHelper::matchCombineDivRem(MachineInstr &MI,
1478                                         MachineInstr *&OtherMI) {
1479   unsigned Opcode = MI.getOpcode();
1480   bool IsDiv, IsSigned;
1481 
1482   switch (Opcode) {
1483   default:
1484     llvm_unreachable("Unexpected opcode!");
1485   case TargetOpcode::G_SDIV:
1486   case TargetOpcode::G_UDIV: {
1487     IsDiv = true;
1488     IsSigned = Opcode == TargetOpcode::G_SDIV;
1489     break;
1490   }
1491   case TargetOpcode::G_SREM:
1492   case TargetOpcode::G_UREM: {
1493     IsDiv = false;
1494     IsSigned = Opcode == TargetOpcode::G_SREM;
1495     break;
1496   }
1497   }
1498 
1499   Register Src1 = MI.getOperand(1).getReg();
1500   unsigned DivOpcode, RemOpcode, DivremOpcode;
1501   if (IsSigned) {
1502     DivOpcode = TargetOpcode::G_SDIV;
1503     RemOpcode = TargetOpcode::G_SREM;
1504     DivremOpcode = TargetOpcode::G_SDIVREM;
1505   } else {
1506     DivOpcode = TargetOpcode::G_UDIV;
1507     RemOpcode = TargetOpcode::G_UREM;
1508     DivremOpcode = TargetOpcode::G_UDIVREM;
1509   }
1510 
1511   if (!isLegalOrBeforeLegalizer({DivremOpcode, {MRI.getType(Src1)}}))
1512     return false;
1513 
1514   // Combine:
1515   //   %div:_ = G_[SU]DIV %src1:_, %src2:_
1516   //   %rem:_ = G_[SU]REM %src1:_, %src2:_
1517   // into:
1518   //  %div:_, %rem:_ = G_[SU]DIVREM %src1:_, %src2:_
1519 
1520   // Combine:
1521   //   %rem:_ = G_[SU]REM %src1:_, %src2:_
1522   //   %div:_ = G_[SU]DIV %src1:_, %src2:_
1523   // into:
1524   //  %div:_, %rem:_ = G_[SU]DIVREM %src1:_, %src2:_
1525 
1526   for (auto &UseMI : MRI.use_nodbg_instructions(Src1)) {
1527     if (MI.getParent() == UseMI.getParent() &&
1528         ((IsDiv && UseMI.getOpcode() == RemOpcode) ||
1529          (!IsDiv && UseMI.getOpcode() == DivOpcode)) &&
1530         matchEqualDefs(MI.getOperand(2), UseMI.getOperand(2)) &&
1531         matchEqualDefs(MI.getOperand(1), UseMI.getOperand(1))) {
1532       OtherMI = &UseMI;
1533       return true;
1534     }
1535   }
1536 
1537   return false;
1538 }
1539 
applyCombineDivRem(MachineInstr & MI,MachineInstr * & OtherMI)1540 void CombinerHelper::applyCombineDivRem(MachineInstr &MI,
1541                                         MachineInstr *&OtherMI) {
1542   unsigned Opcode = MI.getOpcode();
1543   assert(OtherMI && "OtherMI shouldn't be empty.");
1544 
1545   Register DestDivReg, DestRemReg;
1546   if (Opcode == TargetOpcode::G_SDIV || Opcode == TargetOpcode::G_UDIV) {
1547     DestDivReg = MI.getOperand(0).getReg();
1548     DestRemReg = OtherMI->getOperand(0).getReg();
1549   } else {
1550     DestDivReg = OtherMI->getOperand(0).getReg();
1551     DestRemReg = MI.getOperand(0).getReg();
1552   }
1553 
1554   bool IsSigned =
1555       Opcode == TargetOpcode::G_SDIV || Opcode == TargetOpcode::G_SREM;
1556 
1557   // Check which instruction is first in the block so we don't break def-use
1558   // deps by "moving" the instruction incorrectly. Also keep track of which
1559   // instruction is first so we pick it's operands, avoiding use-before-def
1560   // bugs.
1561   MachineInstr *FirstInst = dominates(MI, *OtherMI) ? &MI : OtherMI;
1562   Builder.setInstrAndDebugLoc(*FirstInst);
1563 
1564   Builder.buildInstr(IsSigned ? TargetOpcode::G_SDIVREM
1565                               : TargetOpcode::G_UDIVREM,
1566                      {DestDivReg, DestRemReg},
1567                      { FirstInst->getOperand(1), FirstInst->getOperand(2) });
1568   MI.eraseFromParent();
1569   OtherMI->eraseFromParent();
1570 }
1571 
matchOptBrCondByInvertingCond(MachineInstr & MI,MachineInstr * & BrCond)1572 bool CombinerHelper::matchOptBrCondByInvertingCond(MachineInstr &MI,
1573                                                    MachineInstr *&BrCond) {
1574   assert(MI.getOpcode() == TargetOpcode::G_BR);
1575 
1576   // Try to match the following:
1577   // bb1:
1578   //   G_BRCOND %c1, %bb2
1579   //   G_BR %bb3
1580   // bb2:
1581   // ...
1582   // bb3:
1583 
1584   // The above pattern does not have a fall through to the successor bb2, always
1585   // resulting in a branch no matter which path is taken. Here we try to find
1586   // and replace that pattern with conditional branch to bb3 and otherwise
1587   // fallthrough to bb2. This is generally better for branch predictors.
1588 
1589   MachineBasicBlock *MBB = MI.getParent();
1590   MachineBasicBlock::iterator BrIt(MI);
1591   if (BrIt == MBB->begin())
1592     return false;
1593   assert(std::next(BrIt) == MBB->end() && "expected G_BR to be a terminator");
1594 
1595   BrCond = &*std::prev(BrIt);
1596   if (BrCond->getOpcode() != TargetOpcode::G_BRCOND)
1597     return false;
1598 
1599   // Check that the next block is the conditional branch target. Also make sure
1600   // that it isn't the same as the G_BR's target (otherwise, this will loop.)
1601   MachineBasicBlock *BrCondTarget = BrCond->getOperand(1).getMBB();
1602   return BrCondTarget != MI.getOperand(0).getMBB() &&
1603          MBB->isLayoutSuccessor(BrCondTarget);
1604 }
1605 
applyOptBrCondByInvertingCond(MachineInstr & MI,MachineInstr * & BrCond)1606 void CombinerHelper::applyOptBrCondByInvertingCond(MachineInstr &MI,
1607                                                    MachineInstr *&BrCond) {
1608   MachineBasicBlock *BrTarget = MI.getOperand(0).getMBB();
1609   Builder.setInstrAndDebugLoc(*BrCond);
1610   LLT Ty = MRI.getType(BrCond->getOperand(0).getReg());
1611   // FIXME: Does int/fp matter for this? If so, we might need to restrict
1612   // this to i1 only since we might not know for sure what kind of
1613   // compare generated the condition value.
1614   auto True = Builder.buildConstant(
1615       Ty, getICmpTrueVal(getTargetLowering(), false, false));
1616   auto Xor = Builder.buildXor(Ty, BrCond->getOperand(0), True);
1617 
1618   auto *FallthroughBB = BrCond->getOperand(1).getMBB();
1619   Observer.changingInstr(MI);
1620   MI.getOperand(0).setMBB(FallthroughBB);
1621   Observer.changedInstr(MI);
1622 
1623   // Change the conditional branch to use the inverted condition and
1624   // new target block.
1625   Observer.changingInstr(*BrCond);
1626   BrCond->getOperand(0).setReg(Xor.getReg(0));
1627   BrCond->getOperand(1).setMBB(BrTarget);
1628   Observer.changedInstr(*BrCond);
1629 }
1630 
1631 
tryEmitMemcpyInline(MachineInstr & MI)1632 bool CombinerHelper::tryEmitMemcpyInline(MachineInstr &MI) {
1633   MachineIRBuilder HelperBuilder(MI);
1634   GISelObserverWrapper DummyObserver;
1635   LegalizerHelper Helper(HelperBuilder.getMF(), DummyObserver, HelperBuilder);
1636   return Helper.lowerMemcpyInline(MI) ==
1637          LegalizerHelper::LegalizeResult::Legalized;
1638 }
1639 
tryCombineMemCpyFamily(MachineInstr & MI,unsigned MaxLen)1640 bool CombinerHelper::tryCombineMemCpyFamily(MachineInstr &MI, unsigned MaxLen) {
1641   MachineIRBuilder HelperBuilder(MI);
1642   GISelObserverWrapper DummyObserver;
1643   LegalizerHelper Helper(HelperBuilder.getMF(), DummyObserver, HelperBuilder);
1644   return Helper.lowerMemCpyFamily(MI, MaxLen) ==
1645          LegalizerHelper::LegalizeResult::Legalized;
1646 }
1647 
constantFoldFpUnary(const MachineInstr & MI,const MachineRegisterInfo & MRI,const APFloat & Val)1648 static APFloat constantFoldFpUnary(const MachineInstr &MI,
1649                                    const MachineRegisterInfo &MRI,
1650                                    const APFloat &Val) {
1651   APFloat Result(Val);
1652   switch (MI.getOpcode()) {
1653   default:
1654     llvm_unreachable("Unexpected opcode!");
1655   case TargetOpcode::G_FNEG: {
1656     Result.changeSign();
1657     return Result;
1658   }
1659   case TargetOpcode::G_FABS: {
1660     Result.clearSign();
1661     return Result;
1662   }
1663   case TargetOpcode::G_FPTRUNC: {
1664     bool Unused;
1665     LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
1666     Result.convert(getFltSemanticForLLT(DstTy), APFloat::rmNearestTiesToEven,
1667                    &Unused);
1668     return Result;
1669   }
1670   case TargetOpcode::G_FSQRT: {
1671     bool Unused;
1672     Result.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
1673                    &Unused);
1674     Result = APFloat(sqrt(Result.convertToDouble()));
1675     break;
1676   }
1677   case TargetOpcode::G_FLOG2: {
1678     bool Unused;
1679     Result.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
1680                    &Unused);
1681     Result = APFloat(log2(Result.convertToDouble()));
1682     break;
1683   }
1684   }
1685   // Convert `APFloat` to appropriate IEEE type depending on `DstTy`. Otherwise,
1686   // `buildFConstant` will assert on size mismatch. Only `G_FSQRT`, and
1687   // `G_FLOG2` reach here.
1688   bool Unused;
1689   Result.convert(Val.getSemantics(), APFloat::rmNearestTiesToEven, &Unused);
1690   return Result;
1691 }
1692 
applyCombineConstantFoldFpUnary(MachineInstr & MI,const ConstantFP * Cst)1693 void CombinerHelper::applyCombineConstantFoldFpUnary(MachineInstr &MI,
1694                                                      const ConstantFP *Cst) {
1695   APFloat Folded = constantFoldFpUnary(MI, MRI, Cst->getValue());
1696   const ConstantFP *NewCst = ConstantFP::get(Builder.getContext(), Folded);
1697   Builder.buildFConstant(MI.getOperand(0), *NewCst);
1698   MI.eraseFromParent();
1699 }
1700 
matchPtrAddImmedChain(MachineInstr & MI,PtrAddChain & MatchInfo)1701 bool CombinerHelper::matchPtrAddImmedChain(MachineInstr &MI,
1702                                            PtrAddChain &MatchInfo) {
1703   // We're trying to match the following pattern:
1704   //   %t1 = G_PTR_ADD %base, G_CONSTANT imm1
1705   //   %root = G_PTR_ADD %t1, G_CONSTANT imm2
1706   // -->
1707   //   %root = G_PTR_ADD %base, G_CONSTANT (imm1 + imm2)
1708 
1709   if (MI.getOpcode() != TargetOpcode::G_PTR_ADD)
1710     return false;
1711 
1712   Register Add2 = MI.getOperand(1).getReg();
1713   Register Imm1 = MI.getOperand(2).getReg();
1714   auto MaybeImmVal = getIConstantVRegValWithLookThrough(Imm1, MRI);
1715   if (!MaybeImmVal)
1716     return false;
1717 
1718   MachineInstr *Add2Def = MRI.getVRegDef(Add2);
1719   if (!Add2Def || Add2Def->getOpcode() != TargetOpcode::G_PTR_ADD)
1720     return false;
1721 
1722   Register Base = Add2Def->getOperand(1).getReg();
1723   Register Imm2 = Add2Def->getOperand(2).getReg();
1724   auto MaybeImm2Val = getIConstantVRegValWithLookThrough(Imm2, MRI);
1725   if (!MaybeImm2Val)
1726     return false;
1727 
1728   // Check if the new combined immediate forms an illegal addressing mode.
1729   // Do not combine if it was legal before but would get illegal.
1730   // To do so, we need to find a load/store user of the pointer to get
1731   // the access type.
1732   Type *AccessTy = nullptr;
1733   auto &MF = *MI.getMF();
1734   for (auto &UseMI : MRI.use_nodbg_instructions(MI.getOperand(0).getReg())) {
1735     if (auto *LdSt = dyn_cast<GLoadStore>(&UseMI)) {
1736       AccessTy = getTypeForLLT(MRI.getType(LdSt->getReg(0)),
1737                                MF.getFunction().getContext());
1738       break;
1739     }
1740   }
1741   TargetLoweringBase::AddrMode AMNew;
1742   APInt CombinedImm = MaybeImmVal->Value + MaybeImm2Val->Value;
1743   AMNew.BaseOffs = CombinedImm.getSExtValue();
1744   if (AccessTy) {
1745     AMNew.HasBaseReg = true;
1746     TargetLoweringBase::AddrMode AMOld;
1747     AMOld.BaseOffs = MaybeImmVal->Value.getSExtValue();
1748     AMOld.HasBaseReg = true;
1749     unsigned AS = MRI.getType(Add2).getAddressSpace();
1750     const auto &TLI = *MF.getSubtarget().getTargetLowering();
1751     if (TLI.isLegalAddressingMode(MF.getDataLayout(), AMOld, AccessTy, AS) &&
1752         !TLI.isLegalAddressingMode(MF.getDataLayout(), AMNew, AccessTy, AS))
1753       return false;
1754   }
1755 
1756   // Pass the combined immediate to the apply function.
1757   MatchInfo.Imm = AMNew.BaseOffs;
1758   MatchInfo.Base = Base;
1759   MatchInfo.Bank = getRegBank(Imm2);
1760   return true;
1761 }
1762 
applyPtrAddImmedChain(MachineInstr & MI,PtrAddChain & MatchInfo)1763 void CombinerHelper::applyPtrAddImmedChain(MachineInstr &MI,
1764                                            PtrAddChain &MatchInfo) {
1765   assert(MI.getOpcode() == TargetOpcode::G_PTR_ADD && "Expected G_PTR_ADD");
1766   MachineIRBuilder MIB(MI);
1767   LLT OffsetTy = MRI.getType(MI.getOperand(2).getReg());
1768   auto NewOffset = MIB.buildConstant(OffsetTy, MatchInfo.Imm);
1769   setRegBank(NewOffset.getReg(0), MatchInfo.Bank);
1770   Observer.changingInstr(MI);
1771   MI.getOperand(1).setReg(MatchInfo.Base);
1772   MI.getOperand(2).setReg(NewOffset.getReg(0));
1773   Observer.changedInstr(MI);
1774 }
1775 
matchShiftImmedChain(MachineInstr & MI,RegisterImmPair & MatchInfo)1776 bool CombinerHelper::matchShiftImmedChain(MachineInstr &MI,
1777                                           RegisterImmPair &MatchInfo) {
1778   // We're trying to match the following pattern with any of
1779   // G_SHL/G_ASHR/G_LSHR/G_SSHLSAT/G_USHLSAT shift instructions:
1780   //   %t1 = SHIFT %base, G_CONSTANT imm1
1781   //   %root = SHIFT %t1, G_CONSTANT imm2
1782   // -->
1783   //   %root = SHIFT %base, G_CONSTANT (imm1 + imm2)
1784 
1785   unsigned Opcode = MI.getOpcode();
1786   assert((Opcode == TargetOpcode::G_SHL || Opcode == TargetOpcode::G_ASHR ||
1787           Opcode == TargetOpcode::G_LSHR || Opcode == TargetOpcode::G_SSHLSAT ||
1788           Opcode == TargetOpcode::G_USHLSAT) &&
1789          "Expected G_SHL, G_ASHR, G_LSHR, G_SSHLSAT or G_USHLSAT");
1790 
1791   Register Shl2 = MI.getOperand(1).getReg();
1792   Register Imm1 = MI.getOperand(2).getReg();
1793   auto MaybeImmVal = getIConstantVRegValWithLookThrough(Imm1, MRI);
1794   if (!MaybeImmVal)
1795     return false;
1796 
1797   MachineInstr *Shl2Def = MRI.getUniqueVRegDef(Shl2);
1798   if (Shl2Def->getOpcode() != Opcode)
1799     return false;
1800 
1801   Register Base = Shl2Def->getOperand(1).getReg();
1802   Register Imm2 = Shl2Def->getOperand(2).getReg();
1803   auto MaybeImm2Val = getIConstantVRegValWithLookThrough(Imm2, MRI);
1804   if (!MaybeImm2Val)
1805     return false;
1806 
1807   // Pass the combined immediate to the apply function.
1808   MatchInfo.Imm =
1809       (MaybeImmVal->Value.getZExtValue() + MaybeImm2Val->Value).getZExtValue();
1810   MatchInfo.Reg = Base;
1811 
1812   // There is no simple replacement for a saturating unsigned left shift that
1813   // exceeds the scalar size.
1814   if (Opcode == TargetOpcode::G_USHLSAT &&
1815       MatchInfo.Imm >= MRI.getType(Shl2).getScalarSizeInBits())
1816     return false;
1817 
1818   return true;
1819 }
1820 
applyShiftImmedChain(MachineInstr & MI,RegisterImmPair & MatchInfo)1821 void CombinerHelper::applyShiftImmedChain(MachineInstr &MI,
1822                                           RegisterImmPair &MatchInfo) {
1823   unsigned Opcode = MI.getOpcode();
1824   assert((Opcode == TargetOpcode::G_SHL || Opcode == TargetOpcode::G_ASHR ||
1825           Opcode == TargetOpcode::G_LSHR || Opcode == TargetOpcode::G_SSHLSAT ||
1826           Opcode == TargetOpcode::G_USHLSAT) &&
1827          "Expected G_SHL, G_ASHR, G_LSHR, G_SSHLSAT or G_USHLSAT");
1828 
1829   LLT Ty = MRI.getType(MI.getOperand(1).getReg());
1830   unsigned const ScalarSizeInBits = Ty.getScalarSizeInBits();
1831   auto Imm = MatchInfo.Imm;
1832 
1833   if (Imm >= ScalarSizeInBits) {
1834     // Any logical shift that exceeds scalar size will produce zero.
1835     if (Opcode == TargetOpcode::G_SHL || Opcode == TargetOpcode::G_LSHR) {
1836       Builder.buildConstant(MI.getOperand(0), 0);
1837       MI.eraseFromParent();
1838       return;
1839     }
1840     // Arithmetic shift and saturating signed left shift have no effect beyond
1841     // scalar size.
1842     Imm = ScalarSizeInBits - 1;
1843   }
1844 
1845   LLT ImmTy = MRI.getType(MI.getOperand(2).getReg());
1846   Register NewImm = Builder.buildConstant(ImmTy, Imm).getReg(0);
1847   Observer.changingInstr(MI);
1848   MI.getOperand(1).setReg(MatchInfo.Reg);
1849   MI.getOperand(2).setReg(NewImm);
1850   Observer.changedInstr(MI);
1851 }
1852 
matchShiftOfShiftedLogic(MachineInstr & MI,ShiftOfShiftedLogic & MatchInfo)1853 bool CombinerHelper::matchShiftOfShiftedLogic(MachineInstr &MI,
1854                                               ShiftOfShiftedLogic &MatchInfo) {
1855   // We're trying to match the following pattern with any of
1856   // G_SHL/G_ASHR/G_LSHR/G_USHLSAT/G_SSHLSAT shift instructions in combination
1857   // with any of G_AND/G_OR/G_XOR logic instructions.
1858   //   %t1 = SHIFT %X, G_CONSTANT C0
1859   //   %t2 = LOGIC %t1, %Y
1860   //   %root = SHIFT %t2, G_CONSTANT C1
1861   // -->
1862   //   %t3 = SHIFT %X, G_CONSTANT (C0+C1)
1863   //   %t4 = SHIFT %Y, G_CONSTANT C1
1864   //   %root = LOGIC %t3, %t4
1865   unsigned ShiftOpcode = MI.getOpcode();
1866   assert((ShiftOpcode == TargetOpcode::G_SHL ||
1867           ShiftOpcode == TargetOpcode::G_ASHR ||
1868           ShiftOpcode == TargetOpcode::G_LSHR ||
1869           ShiftOpcode == TargetOpcode::G_USHLSAT ||
1870           ShiftOpcode == TargetOpcode::G_SSHLSAT) &&
1871          "Expected G_SHL, G_ASHR, G_LSHR, G_USHLSAT and G_SSHLSAT");
1872 
1873   // Match a one-use bitwise logic op.
1874   Register LogicDest = MI.getOperand(1).getReg();
1875   if (!MRI.hasOneNonDBGUse(LogicDest))
1876     return false;
1877 
1878   MachineInstr *LogicMI = MRI.getUniqueVRegDef(LogicDest);
1879   unsigned LogicOpcode = LogicMI->getOpcode();
1880   if (LogicOpcode != TargetOpcode::G_AND && LogicOpcode != TargetOpcode::G_OR &&
1881       LogicOpcode != TargetOpcode::G_XOR)
1882     return false;
1883 
1884   // Find a matching one-use shift by constant.
1885   const Register C1 = MI.getOperand(2).getReg();
1886   auto MaybeImmVal = getIConstantVRegValWithLookThrough(C1, MRI);
1887   if (!MaybeImmVal || MaybeImmVal->Value == 0)
1888     return false;
1889 
1890   const uint64_t C1Val = MaybeImmVal->Value.getZExtValue();
1891 
1892   auto matchFirstShift = [&](const MachineInstr *MI, uint64_t &ShiftVal) {
1893     // Shift should match previous one and should be a one-use.
1894     if (MI->getOpcode() != ShiftOpcode ||
1895         !MRI.hasOneNonDBGUse(MI->getOperand(0).getReg()))
1896       return false;
1897 
1898     // Must be a constant.
1899     auto MaybeImmVal =
1900         getIConstantVRegValWithLookThrough(MI->getOperand(2).getReg(), MRI);
1901     if (!MaybeImmVal)
1902       return false;
1903 
1904     ShiftVal = MaybeImmVal->Value.getSExtValue();
1905     return true;
1906   };
1907 
1908   // Logic ops are commutative, so check each operand for a match.
1909   Register LogicMIReg1 = LogicMI->getOperand(1).getReg();
1910   MachineInstr *LogicMIOp1 = MRI.getUniqueVRegDef(LogicMIReg1);
1911   Register LogicMIReg2 = LogicMI->getOperand(2).getReg();
1912   MachineInstr *LogicMIOp2 = MRI.getUniqueVRegDef(LogicMIReg2);
1913   uint64_t C0Val;
1914 
1915   if (matchFirstShift(LogicMIOp1, C0Val)) {
1916     MatchInfo.LogicNonShiftReg = LogicMIReg2;
1917     MatchInfo.Shift2 = LogicMIOp1;
1918   } else if (matchFirstShift(LogicMIOp2, C0Val)) {
1919     MatchInfo.LogicNonShiftReg = LogicMIReg1;
1920     MatchInfo.Shift2 = LogicMIOp2;
1921   } else
1922     return false;
1923 
1924   MatchInfo.ValSum = C0Val + C1Val;
1925 
1926   // The fold is not valid if the sum of the shift values exceeds bitwidth.
1927   if (MatchInfo.ValSum >= MRI.getType(LogicDest).getScalarSizeInBits())
1928     return false;
1929 
1930   MatchInfo.Logic = LogicMI;
1931   return true;
1932 }
1933 
applyShiftOfShiftedLogic(MachineInstr & MI,ShiftOfShiftedLogic & MatchInfo)1934 void CombinerHelper::applyShiftOfShiftedLogic(MachineInstr &MI,
1935                                               ShiftOfShiftedLogic &MatchInfo) {
1936   unsigned Opcode = MI.getOpcode();
1937   assert((Opcode == TargetOpcode::G_SHL || Opcode == TargetOpcode::G_ASHR ||
1938           Opcode == TargetOpcode::G_LSHR || Opcode == TargetOpcode::G_USHLSAT ||
1939           Opcode == TargetOpcode::G_SSHLSAT) &&
1940          "Expected G_SHL, G_ASHR, G_LSHR, G_USHLSAT and G_SSHLSAT");
1941 
1942   LLT ShlType = MRI.getType(MI.getOperand(2).getReg());
1943   LLT DestType = MRI.getType(MI.getOperand(0).getReg());
1944 
1945   Register Const = Builder.buildConstant(ShlType, MatchInfo.ValSum).getReg(0);
1946 
1947   Register Shift1Base = MatchInfo.Shift2->getOperand(1).getReg();
1948   Register Shift1 =
1949       Builder.buildInstr(Opcode, {DestType}, {Shift1Base, Const}).getReg(0);
1950 
1951   // If LogicNonShiftReg is the same to Shift1Base, and shift1 const is the same
1952   // to MatchInfo.Shift2 const, CSEMIRBuilder will reuse the old shift1 when
1953   // build shift2. So, if we erase MatchInfo.Shift2 at the end, actually we
1954   // remove old shift1. And it will cause crash later. So erase it earlier to
1955   // avoid the crash.
1956   MatchInfo.Shift2->eraseFromParent();
1957 
1958   Register Shift2Const = MI.getOperand(2).getReg();
1959   Register Shift2 = Builder
1960                         .buildInstr(Opcode, {DestType},
1961                                     {MatchInfo.LogicNonShiftReg, Shift2Const})
1962                         .getReg(0);
1963 
1964   Register Dest = MI.getOperand(0).getReg();
1965   Builder.buildInstr(MatchInfo.Logic->getOpcode(), {Dest}, {Shift1, Shift2});
1966 
1967   // This was one use so it's safe to remove it.
1968   MatchInfo.Logic->eraseFromParent();
1969 
1970   MI.eraseFromParent();
1971 }
1972 
matchCommuteShift(MachineInstr & MI,BuildFnTy & MatchInfo)1973 bool CombinerHelper::matchCommuteShift(MachineInstr &MI, BuildFnTy &MatchInfo) {
1974   assert(MI.getOpcode() == TargetOpcode::G_SHL && "Expected G_SHL");
1975   // Combine (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
1976   // Combine (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2)
1977   auto &Shl = cast<GenericMachineInstr>(MI);
1978   Register DstReg = Shl.getReg(0);
1979   Register SrcReg = Shl.getReg(1);
1980   Register ShiftReg = Shl.getReg(2);
1981   Register X, C1;
1982 
1983   if (!getTargetLowering().isDesirableToCommuteWithShift(MI, !isPreLegalize()))
1984     return false;
1985 
1986   if (!mi_match(SrcReg, MRI,
1987                 m_OneNonDBGUse(m_any_of(m_GAdd(m_Reg(X), m_Reg(C1)),
1988                                         m_GOr(m_Reg(X), m_Reg(C1))))))
1989     return false;
1990 
1991   APInt C1Val, C2Val;
1992   if (!mi_match(C1, MRI, m_ICstOrSplat(C1Val)) ||
1993       !mi_match(ShiftReg, MRI, m_ICstOrSplat(C2Val)))
1994     return false;
1995 
1996   auto *SrcDef = MRI.getVRegDef(SrcReg);
1997   assert((SrcDef->getOpcode() == TargetOpcode::G_ADD ||
1998           SrcDef->getOpcode() == TargetOpcode::G_OR) && "Unexpected op");
1999   LLT SrcTy = MRI.getType(SrcReg);
2000   MatchInfo = [=](MachineIRBuilder &B) {
2001     auto S1 = B.buildShl(SrcTy, X, ShiftReg);
2002     auto S2 = B.buildShl(SrcTy, C1, ShiftReg);
2003     B.buildInstr(SrcDef->getOpcode(), {DstReg}, {S1, S2});
2004   };
2005   return true;
2006 }
2007 
matchCombineMulToShl(MachineInstr & MI,unsigned & ShiftVal)2008 bool CombinerHelper::matchCombineMulToShl(MachineInstr &MI,
2009                                           unsigned &ShiftVal) {
2010   assert(MI.getOpcode() == TargetOpcode::G_MUL && "Expected a G_MUL");
2011   auto MaybeImmVal =
2012       getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
2013   if (!MaybeImmVal)
2014     return false;
2015 
2016   ShiftVal = MaybeImmVal->Value.exactLogBase2();
2017   return (static_cast<int32_t>(ShiftVal) != -1);
2018 }
2019 
applyCombineMulToShl(MachineInstr & MI,unsigned & ShiftVal)2020 void CombinerHelper::applyCombineMulToShl(MachineInstr &MI,
2021                                           unsigned &ShiftVal) {
2022   assert(MI.getOpcode() == TargetOpcode::G_MUL && "Expected a G_MUL");
2023   MachineIRBuilder MIB(MI);
2024   LLT ShiftTy = MRI.getType(MI.getOperand(0).getReg());
2025   auto ShiftCst = MIB.buildConstant(ShiftTy, ShiftVal);
2026   Observer.changingInstr(MI);
2027   MI.setDesc(MIB.getTII().get(TargetOpcode::G_SHL));
2028   MI.getOperand(2).setReg(ShiftCst.getReg(0));
2029   Observer.changedInstr(MI);
2030 }
2031 
2032 // shl ([sza]ext x), y => zext (shl x, y), if shift does not overflow source
matchCombineShlOfExtend(MachineInstr & MI,RegisterImmPair & MatchData)2033 bool CombinerHelper::matchCombineShlOfExtend(MachineInstr &MI,
2034                                              RegisterImmPair &MatchData) {
2035   assert(MI.getOpcode() == TargetOpcode::G_SHL && KB);
2036   if (!getTargetLowering().isDesirableToPullExtFromShl(MI))
2037     return false;
2038 
2039   Register LHS = MI.getOperand(1).getReg();
2040 
2041   Register ExtSrc;
2042   if (!mi_match(LHS, MRI, m_GAnyExt(m_Reg(ExtSrc))) &&
2043       !mi_match(LHS, MRI, m_GZExt(m_Reg(ExtSrc))) &&
2044       !mi_match(LHS, MRI, m_GSExt(m_Reg(ExtSrc))))
2045     return false;
2046 
2047   Register RHS = MI.getOperand(2).getReg();
2048   MachineInstr *MIShiftAmt = MRI.getVRegDef(RHS);
2049   auto MaybeShiftAmtVal = isConstantOrConstantSplatVector(*MIShiftAmt, MRI);
2050   if (!MaybeShiftAmtVal)
2051     return false;
2052 
2053   if (LI) {
2054     LLT SrcTy = MRI.getType(ExtSrc);
2055 
2056     // We only really care about the legality with the shifted value. We can
2057     // pick any type the constant shift amount, so ask the target what to
2058     // use. Otherwise we would have to guess and hope it is reported as legal.
2059     LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(SrcTy);
2060     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_SHL, {SrcTy, ShiftAmtTy}}))
2061       return false;
2062   }
2063 
2064   int64_t ShiftAmt = MaybeShiftAmtVal->getSExtValue();
2065   MatchData.Reg = ExtSrc;
2066   MatchData.Imm = ShiftAmt;
2067 
2068   unsigned MinLeadingZeros = KB->getKnownZeroes(ExtSrc).countl_one();
2069   unsigned SrcTySize = MRI.getType(ExtSrc).getScalarSizeInBits();
2070   return MinLeadingZeros >= ShiftAmt && ShiftAmt < SrcTySize;
2071 }
2072 
applyCombineShlOfExtend(MachineInstr & MI,const RegisterImmPair & MatchData)2073 void CombinerHelper::applyCombineShlOfExtend(MachineInstr &MI,
2074                                              const RegisterImmPair &MatchData) {
2075   Register ExtSrcReg = MatchData.Reg;
2076   int64_t ShiftAmtVal = MatchData.Imm;
2077 
2078   LLT ExtSrcTy = MRI.getType(ExtSrcReg);
2079   auto ShiftAmt = Builder.buildConstant(ExtSrcTy, ShiftAmtVal);
2080   auto NarrowShift =
2081       Builder.buildShl(ExtSrcTy, ExtSrcReg, ShiftAmt, MI.getFlags());
2082   Builder.buildZExt(MI.getOperand(0), NarrowShift);
2083   MI.eraseFromParent();
2084 }
2085 
matchCombineMergeUnmerge(MachineInstr & MI,Register & MatchInfo)2086 bool CombinerHelper::matchCombineMergeUnmerge(MachineInstr &MI,
2087                                               Register &MatchInfo) {
2088   GMerge &Merge = cast<GMerge>(MI);
2089   SmallVector<Register, 16> MergedValues;
2090   for (unsigned I = 0; I < Merge.getNumSources(); ++I)
2091     MergedValues.emplace_back(Merge.getSourceReg(I));
2092 
2093   auto *Unmerge = getOpcodeDef<GUnmerge>(MergedValues[0], MRI);
2094   if (!Unmerge || Unmerge->getNumDefs() != Merge.getNumSources())
2095     return false;
2096 
2097   for (unsigned I = 0; I < MergedValues.size(); ++I)
2098     if (MergedValues[I] != Unmerge->getReg(I))
2099       return false;
2100 
2101   MatchInfo = Unmerge->getSourceReg();
2102   return true;
2103 }
2104 
peekThroughBitcast(Register Reg,const MachineRegisterInfo & MRI)2105 static Register peekThroughBitcast(Register Reg,
2106                                    const MachineRegisterInfo &MRI) {
2107   while (mi_match(Reg, MRI, m_GBitcast(m_Reg(Reg))))
2108     ;
2109 
2110   return Reg;
2111 }
2112 
matchCombineUnmergeMergeToPlainValues(MachineInstr & MI,SmallVectorImpl<Register> & Operands)2113 bool CombinerHelper::matchCombineUnmergeMergeToPlainValues(
2114     MachineInstr &MI, SmallVectorImpl<Register> &Operands) {
2115   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2116          "Expected an unmerge");
2117   auto &Unmerge = cast<GUnmerge>(MI);
2118   Register SrcReg = peekThroughBitcast(Unmerge.getSourceReg(), MRI);
2119 
2120   auto *SrcInstr = getOpcodeDef<GMergeLikeInstr>(SrcReg, MRI);
2121   if (!SrcInstr)
2122     return false;
2123 
2124   // Check the source type of the merge.
2125   LLT SrcMergeTy = MRI.getType(SrcInstr->getSourceReg(0));
2126   LLT Dst0Ty = MRI.getType(Unmerge.getReg(0));
2127   bool SameSize = Dst0Ty.getSizeInBits() == SrcMergeTy.getSizeInBits();
2128   if (SrcMergeTy != Dst0Ty && !SameSize)
2129     return false;
2130   // They are the same now (modulo a bitcast).
2131   // We can collect all the src registers.
2132   for (unsigned Idx = 0; Idx < SrcInstr->getNumSources(); ++Idx)
2133     Operands.push_back(SrcInstr->getSourceReg(Idx));
2134   return true;
2135 }
2136 
applyCombineUnmergeMergeToPlainValues(MachineInstr & MI,SmallVectorImpl<Register> & Operands)2137 void CombinerHelper::applyCombineUnmergeMergeToPlainValues(
2138     MachineInstr &MI, SmallVectorImpl<Register> &Operands) {
2139   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2140          "Expected an unmerge");
2141   assert((MI.getNumOperands() - 1 == Operands.size()) &&
2142          "Not enough operands to replace all defs");
2143   unsigned NumElems = MI.getNumOperands() - 1;
2144 
2145   LLT SrcTy = MRI.getType(Operands[0]);
2146   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
2147   bool CanReuseInputDirectly = DstTy == SrcTy;
2148   for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
2149     Register DstReg = MI.getOperand(Idx).getReg();
2150     Register SrcReg = Operands[Idx];
2151 
2152     // This combine may run after RegBankSelect, so we need to be aware of
2153     // register banks.
2154     const auto &DstCB = MRI.getRegClassOrRegBank(DstReg);
2155     if (!DstCB.isNull() && DstCB != MRI.getRegClassOrRegBank(SrcReg)) {
2156       SrcReg = Builder.buildCopy(MRI.getType(SrcReg), SrcReg).getReg(0);
2157       MRI.setRegClassOrRegBank(SrcReg, DstCB);
2158     }
2159 
2160     if (CanReuseInputDirectly)
2161       replaceRegWith(MRI, DstReg, SrcReg);
2162     else
2163       Builder.buildCast(DstReg, SrcReg);
2164   }
2165   MI.eraseFromParent();
2166 }
2167 
matchCombineUnmergeConstant(MachineInstr & MI,SmallVectorImpl<APInt> & Csts)2168 bool CombinerHelper::matchCombineUnmergeConstant(MachineInstr &MI,
2169                                                  SmallVectorImpl<APInt> &Csts) {
2170   unsigned SrcIdx = MI.getNumOperands() - 1;
2171   Register SrcReg = MI.getOperand(SrcIdx).getReg();
2172   MachineInstr *SrcInstr = MRI.getVRegDef(SrcReg);
2173   if (SrcInstr->getOpcode() != TargetOpcode::G_CONSTANT &&
2174       SrcInstr->getOpcode() != TargetOpcode::G_FCONSTANT)
2175     return false;
2176   // Break down the big constant in smaller ones.
2177   const MachineOperand &CstVal = SrcInstr->getOperand(1);
2178   APInt Val = SrcInstr->getOpcode() == TargetOpcode::G_CONSTANT
2179                   ? CstVal.getCImm()->getValue()
2180                   : CstVal.getFPImm()->getValueAPF().bitcastToAPInt();
2181 
2182   LLT Dst0Ty = MRI.getType(MI.getOperand(0).getReg());
2183   unsigned ShiftAmt = Dst0Ty.getSizeInBits();
2184   // Unmerge a constant.
2185   for (unsigned Idx = 0; Idx != SrcIdx; ++Idx) {
2186     Csts.emplace_back(Val.trunc(ShiftAmt));
2187     Val = Val.lshr(ShiftAmt);
2188   }
2189 
2190   return true;
2191 }
2192 
applyCombineUnmergeConstant(MachineInstr & MI,SmallVectorImpl<APInt> & Csts)2193 void CombinerHelper::applyCombineUnmergeConstant(MachineInstr &MI,
2194                                                  SmallVectorImpl<APInt> &Csts) {
2195   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2196          "Expected an unmerge");
2197   assert((MI.getNumOperands() - 1 == Csts.size()) &&
2198          "Not enough operands to replace all defs");
2199   unsigned NumElems = MI.getNumOperands() - 1;
2200   for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
2201     Register DstReg = MI.getOperand(Idx).getReg();
2202     Builder.buildConstant(DstReg, Csts[Idx]);
2203   }
2204 
2205   MI.eraseFromParent();
2206 }
2207 
matchCombineUnmergeUndef(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)2208 bool CombinerHelper::matchCombineUnmergeUndef(
2209     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
2210   unsigned SrcIdx = MI.getNumOperands() - 1;
2211   Register SrcReg = MI.getOperand(SrcIdx).getReg();
2212   MatchInfo = [&MI](MachineIRBuilder &B) {
2213     unsigned NumElems = MI.getNumOperands() - 1;
2214     for (unsigned Idx = 0; Idx < NumElems; ++Idx) {
2215       Register DstReg = MI.getOperand(Idx).getReg();
2216       B.buildUndef(DstReg);
2217     }
2218   };
2219   return isa<GImplicitDef>(MRI.getVRegDef(SrcReg));
2220 }
2221 
matchCombineUnmergeWithDeadLanesToTrunc(MachineInstr & MI)2222 bool CombinerHelper::matchCombineUnmergeWithDeadLanesToTrunc(MachineInstr &MI) {
2223   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2224          "Expected an unmerge");
2225   if (MRI.getType(MI.getOperand(0).getReg()).isVector() ||
2226       MRI.getType(MI.getOperand(MI.getNumDefs()).getReg()).isVector())
2227     return false;
2228   // Check that all the lanes are dead except the first one.
2229   for (unsigned Idx = 1, EndIdx = MI.getNumDefs(); Idx != EndIdx; ++Idx) {
2230     if (!MRI.use_nodbg_empty(MI.getOperand(Idx).getReg()))
2231       return false;
2232   }
2233   return true;
2234 }
2235 
applyCombineUnmergeWithDeadLanesToTrunc(MachineInstr & MI)2236 void CombinerHelper::applyCombineUnmergeWithDeadLanesToTrunc(MachineInstr &MI) {
2237   Register SrcReg = MI.getOperand(MI.getNumDefs()).getReg();
2238   Register Dst0Reg = MI.getOperand(0).getReg();
2239   Builder.buildTrunc(Dst0Reg, SrcReg);
2240   MI.eraseFromParent();
2241 }
2242 
matchCombineUnmergeZExtToZExt(MachineInstr & MI)2243 bool CombinerHelper::matchCombineUnmergeZExtToZExt(MachineInstr &MI) {
2244   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2245          "Expected an unmerge");
2246   Register Dst0Reg = MI.getOperand(0).getReg();
2247   LLT Dst0Ty = MRI.getType(Dst0Reg);
2248   // G_ZEXT on vector applies to each lane, so it will
2249   // affect all destinations. Therefore we won't be able
2250   // to simplify the unmerge to just the first definition.
2251   if (Dst0Ty.isVector())
2252     return false;
2253   Register SrcReg = MI.getOperand(MI.getNumDefs()).getReg();
2254   LLT SrcTy = MRI.getType(SrcReg);
2255   if (SrcTy.isVector())
2256     return false;
2257 
2258   Register ZExtSrcReg;
2259   if (!mi_match(SrcReg, MRI, m_GZExt(m_Reg(ZExtSrcReg))))
2260     return false;
2261 
2262   // Finally we can replace the first definition with
2263   // a zext of the source if the definition is big enough to hold
2264   // all of ZExtSrc bits.
2265   LLT ZExtSrcTy = MRI.getType(ZExtSrcReg);
2266   return ZExtSrcTy.getSizeInBits() <= Dst0Ty.getSizeInBits();
2267 }
2268 
applyCombineUnmergeZExtToZExt(MachineInstr & MI)2269 void CombinerHelper::applyCombineUnmergeZExtToZExt(MachineInstr &MI) {
2270   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES &&
2271          "Expected an unmerge");
2272 
2273   Register Dst0Reg = MI.getOperand(0).getReg();
2274 
2275   MachineInstr *ZExtInstr =
2276       MRI.getVRegDef(MI.getOperand(MI.getNumDefs()).getReg());
2277   assert(ZExtInstr && ZExtInstr->getOpcode() == TargetOpcode::G_ZEXT &&
2278          "Expecting a G_ZEXT");
2279 
2280   Register ZExtSrcReg = ZExtInstr->getOperand(1).getReg();
2281   LLT Dst0Ty = MRI.getType(Dst0Reg);
2282   LLT ZExtSrcTy = MRI.getType(ZExtSrcReg);
2283 
2284   if (Dst0Ty.getSizeInBits() > ZExtSrcTy.getSizeInBits()) {
2285     Builder.buildZExt(Dst0Reg, ZExtSrcReg);
2286   } else {
2287     assert(Dst0Ty.getSizeInBits() == ZExtSrcTy.getSizeInBits() &&
2288            "ZExt src doesn't fit in destination");
2289     replaceRegWith(MRI, Dst0Reg, ZExtSrcReg);
2290   }
2291 
2292   Register ZeroReg;
2293   for (unsigned Idx = 1, EndIdx = MI.getNumDefs(); Idx != EndIdx; ++Idx) {
2294     if (!ZeroReg)
2295       ZeroReg = Builder.buildConstant(Dst0Ty, 0).getReg(0);
2296     replaceRegWith(MRI, MI.getOperand(Idx).getReg(), ZeroReg);
2297   }
2298   MI.eraseFromParent();
2299 }
2300 
matchCombineShiftToUnmerge(MachineInstr & MI,unsigned TargetShiftSize,unsigned & ShiftVal)2301 bool CombinerHelper::matchCombineShiftToUnmerge(MachineInstr &MI,
2302                                                 unsigned TargetShiftSize,
2303                                                 unsigned &ShiftVal) {
2304   assert((MI.getOpcode() == TargetOpcode::G_SHL ||
2305           MI.getOpcode() == TargetOpcode::G_LSHR ||
2306           MI.getOpcode() == TargetOpcode::G_ASHR) && "Expected a shift");
2307 
2308   LLT Ty = MRI.getType(MI.getOperand(0).getReg());
2309   if (Ty.isVector()) // TODO:
2310     return false;
2311 
2312   // Don't narrow further than the requested size.
2313   unsigned Size = Ty.getSizeInBits();
2314   if (Size <= TargetShiftSize)
2315     return false;
2316 
2317   auto MaybeImmVal =
2318       getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
2319   if (!MaybeImmVal)
2320     return false;
2321 
2322   ShiftVal = MaybeImmVal->Value.getSExtValue();
2323   return ShiftVal >= Size / 2 && ShiftVal < Size;
2324 }
2325 
applyCombineShiftToUnmerge(MachineInstr & MI,const unsigned & ShiftVal)2326 void CombinerHelper::applyCombineShiftToUnmerge(MachineInstr &MI,
2327                                                 const unsigned &ShiftVal) {
2328   Register DstReg = MI.getOperand(0).getReg();
2329   Register SrcReg = MI.getOperand(1).getReg();
2330   LLT Ty = MRI.getType(SrcReg);
2331   unsigned Size = Ty.getSizeInBits();
2332   unsigned HalfSize = Size / 2;
2333   assert(ShiftVal >= HalfSize);
2334 
2335   LLT HalfTy = LLT::scalar(HalfSize);
2336 
2337   auto Unmerge = Builder.buildUnmerge(HalfTy, SrcReg);
2338   unsigned NarrowShiftAmt = ShiftVal - HalfSize;
2339 
2340   if (MI.getOpcode() == TargetOpcode::G_LSHR) {
2341     Register Narrowed = Unmerge.getReg(1);
2342 
2343     //  dst = G_LSHR s64:x, C for C >= 32
2344     // =>
2345     //   lo, hi = G_UNMERGE_VALUES x
2346     //   dst = G_MERGE_VALUES (G_LSHR hi, C - 32), 0
2347 
2348     if (NarrowShiftAmt != 0) {
2349       Narrowed = Builder.buildLShr(HalfTy, Narrowed,
2350         Builder.buildConstant(HalfTy, NarrowShiftAmt)).getReg(0);
2351     }
2352 
2353     auto Zero = Builder.buildConstant(HalfTy, 0);
2354     Builder.buildMergeLikeInstr(DstReg, {Narrowed, Zero});
2355   } else if (MI.getOpcode() == TargetOpcode::G_SHL) {
2356     Register Narrowed = Unmerge.getReg(0);
2357     //  dst = G_SHL s64:x, C for C >= 32
2358     // =>
2359     //   lo, hi = G_UNMERGE_VALUES x
2360     //   dst = G_MERGE_VALUES 0, (G_SHL hi, C - 32)
2361     if (NarrowShiftAmt != 0) {
2362       Narrowed = Builder.buildShl(HalfTy, Narrowed,
2363         Builder.buildConstant(HalfTy, NarrowShiftAmt)).getReg(0);
2364     }
2365 
2366     auto Zero = Builder.buildConstant(HalfTy, 0);
2367     Builder.buildMergeLikeInstr(DstReg, {Zero, Narrowed});
2368   } else {
2369     assert(MI.getOpcode() == TargetOpcode::G_ASHR);
2370     auto Hi = Builder.buildAShr(
2371       HalfTy, Unmerge.getReg(1),
2372       Builder.buildConstant(HalfTy, HalfSize - 1));
2373 
2374     if (ShiftVal == HalfSize) {
2375       // (G_ASHR i64:x, 32) ->
2376       //   G_MERGE_VALUES hi_32(x), (G_ASHR hi_32(x), 31)
2377       Builder.buildMergeLikeInstr(DstReg, {Unmerge.getReg(1), Hi});
2378     } else if (ShiftVal == Size - 1) {
2379       // Don't need a second shift.
2380       // (G_ASHR i64:x, 63) ->
2381       //   %narrowed = (G_ASHR hi_32(x), 31)
2382       //   G_MERGE_VALUES %narrowed, %narrowed
2383       Builder.buildMergeLikeInstr(DstReg, {Hi, Hi});
2384     } else {
2385       auto Lo = Builder.buildAShr(
2386         HalfTy, Unmerge.getReg(1),
2387         Builder.buildConstant(HalfTy, ShiftVal - HalfSize));
2388 
2389       // (G_ASHR i64:x, C) ->, for C >= 32
2390       //   G_MERGE_VALUES (G_ASHR hi_32(x), C - 32), (G_ASHR hi_32(x), 31)
2391       Builder.buildMergeLikeInstr(DstReg, {Lo, Hi});
2392     }
2393   }
2394 
2395   MI.eraseFromParent();
2396 }
2397 
tryCombineShiftToUnmerge(MachineInstr & MI,unsigned TargetShiftAmount)2398 bool CombinerHelper::tryCombineShiftToUnmerge(MachineInstr &MI,
2399                                               unsigned TargetShiftAmount) {
2400   unsigned ShiftAmt;
2401   if (matchCombineShiftToUnmerge(MI, TargetShiftAmount, ShiftAmt)) {
2402     applyCombineShiftToUnmerge(MI, ShiftAmt);
2403     return true;
2404   }
2405 
2406   return false;
2407 }
2408 
matchCombineI2PToP2I(MachineInstr & MI,Register & Reg)2409 bool CombinerHelper::matchCombineI2PToP2I(MachineInstr &MI, Register &Reg) {
2410   assert(MI.getOpcode() == TargetOpcode::G_INTTOPTR && "Expected a G_INTTOPTR");
2411   Register DstReg = MI.getOperand(0).getReg();
2412   LLT DstTy = MRI.getType(DstReg);
2413   Register SrcReg = MI.getOperand(1).getReg();
2414   return mi_match(SrcReg, MRI,
2415                   m_GPtrToInt(m_all_of(m_SpecificType(DstTy), m_Reg(Reg))));
2416 }
2417 
applyCombineI2PToP2I(MachineInstr & MI,Register & Reg)2418 void CombinerHelper::applyCombineI2PToP2I(MachineInstr &MI, Register &Reg) {
2419   assert(MI.getOpcode() == TargetOpcode::G_INTTOPTR && "Expected a G_INTTOPTR");
2420   Register DstReg = MI.getOperand(0).getReg();
2421   Builder.buildCopy(DstReg, Reg);
2422   MI.eraseFromParent();
2423 }
2424 
applyCombineP2IToI2P(MachineInstr & MI,Register & Reg)2425 void CombinerHelper::applyCombineP2IToI2P(MachineInstr &MI, Register &Reg) {
2426   assert(MI.getOpcode() == TargetOpcode::G_PTRTOINT && "Expected a G_PTRTOINT");
2427   Register DstReg = MI.getOperand(0).getReg();
2428   Builder.buildZExtOrTrunc(DstReg, Reg);
2429   MI.eraseFromParent();
2430 }
2431 
matchCombineAddP2IToPtrAdd(MachineInstr & MI,std::pair<Register,bool> & PtrReg)2432 bool CombinerHelper::matchCombineAddP2IToPtrAdd(
2433     MachineInstr &MI, std::pair<Register, bool> &PtrReg) {
2434   assert(MI.getOpcode() == TargetOpcode::G_ADD);
2435   Register LHS = MI.getOperand(1).getReg();
2436   Register RHS = MI.getOperand(2).getReg();
2437   LLT IntTy = MRI.getType(LHS);
2438 
2439   // G_PTR_ADD always has the pointer in the LHS, so we may need to commute the
2440   // instruction.
2441   PtrReg.second = false;
2442   for (Register SrcReg : {LHS, RHS}) {
2443     if (mi_match(SrcReg, MRI, m_GPtrToInt(m_Reg(PtrReg.first)))) {
2444       // Don't handle cases where the integer is implicitly converted to the
2445       // pointer width.
2446       LLT PtrTy = MRI.getType(PtrReg.first);
2447       if (PtrTy.getScalarSizeInBits() == IntTy.getScalarSizeInBits())
2448         return true;
2449     }
2450 
2451     PtrReg.second = true;
2452   }
2453 
2454   return false;
2455 }
2456 
applyCombineAddP2IToPtrAdd(MachineInstr & MI,std::pair<Register,bool> & PtrReg)2457 void CombinerHelper::applyCombineAddP2IToPtrAdd(
2458     MachineInstr &MI, std::pair<Register, bool> &PtrReg) {
2459   Register Dst = MI.getOperand(0).getReg();
2460   Register LHS = MI.getOperand(1).getReg();
2461   Register RHS = MI.getOperand(2).getReg();
2462 
2463   const bool DoCommute = PtrReg.second;
2464   if (DoCommute)
2465     std::swap(LHS, RHS);
2466   LHS = PtrReg.first;
2467 
2468   LLT PtrTy = MRI.getType(LHS);
2469 
2470   auto PtrAdd = Builder.buildPtrAdd(PtrTy, LHS, RHS);
2471   Builder.buildPtrToInt(Dst, PtrAdd);
2472   MI.eraseFromParent();
2473 }
2474 
matchCombineConstPtrAddToI2P(MachineInstr & MI,APInt & NewCst)2475 bool CombinerHelper::matchCombineConstPtrAddToI2P(MachineInstr &MI,
2476                                                   APInt &NewCst) {
2477   auto &PtrAdd = cast<GPtrAdd>(MI);
2478   Register LHS = PtrAdd.getBaseReg();
2479   Register RHS = PtrAdd.getOffsetReg();
2480   MachineRegisterInfo &MRI = Builder.getMF().getRegInfo();
2481 
2482   if (auto RHSCst = getIConstantVRegVal(RHS, MRI)) {
2483     APInt Cst;
2484     if (mi_match(LHS, MRI, m_GIntToPtr(m_ICst(Cst)))) {
2485       auto DstTy = MRI.getType(PtrAdd.getReg(0));
2486       // G_INTTOPTR uses zero-extension
2487       NewCst = Cst.zextOrTrunc(DstTy.getSizeInBits());
2488       NewCst += RHSCst->sextOrTrunc(DstTy.getSizeInBits());
2489       return true;
2490     }
2491   }
2492 
2493   return false;
2494 }
2495 
applyCombineConstPtrAddToI2P(MachineInstr & MI,APInt & NewCst)2496 void CombinerHelper::applyCombineConstPtrAddToI2P(MachineInstr &MI,
2497                                                   APInt &NewCst) {
2498   auto &PtrAdd = cast<GPtrAdd>(MI);
2499   Register Dst = PtrAdd.getReg(0);
2500 
2501   Builder.buildConstant(Dst, NewCst);
2502   PtrAdd.eraseFromParent();
2503 }
2504 
matchCombineAnyExtTrunc(MachineInstr & MI,Register & Reg)2505 bool CombinerHelper::matchCombineAnyExtTrunc(MachineInstr &MI, Register &Reg) {
2506   assert(MI.getOpcode() == TargetOpcode::G_ANYEXT && "Expected a G_ANYEXT");
2507   Register DstReg = MI.getOperand(0).getReg();
2508   Register SrcReg = MI.getOperand(1).getReg();
2509   Register OriginalSrcReg = getSrcRegIgnoringCopies(SrcReg, MRI);
2510   if (OriginalSrcReg.isValid())
2511     SrcReg = OriginalSrcReg;
2512   LLT DstTy = MRI.getType(DstReg);
2513   return mi_match(SrcReg, MRI,
2514                   m_GTrunc(m_all_of(m_Reg(Reg), m_SpecificType(DstTy))));
2515 }
2516 
matchCombineZextTrunc(MachineInstr & MI,Register & Reg)2517 bool CombinerHelper::matchCombineZextTrunc(MachineInstr &MI, Register &Reg) {
2518   assert(MI.getOpcode() == TargetOpcode::G_ZEXT && "Expected a G_ZEXT");
2519   Register DstReg = MI.getOperand(0).getReg();
2520   Register SrcReg = MI.getOperand(1).getReg();
2521   LLT DstTy = MRI.getType(DstReg);
2522   if (mi_match(SrcReg, MRI,
2523                m_GTrunc(m_all_of(m_Reg(Reg), m_SpecificType(DstTy))))) {
2524     unsigned DstSize = DstTy.getScalarSizeInBits();
2525     unsigned SrcSize = MRI.getType(SrcReg).getScalarSizeInBits();
2526     return KB->getKnownBits(Reg).countMinLeadingZeros() >= DstSize - SrcSize;
2527   }
2528   return false;
2529 }
2530 
matchCombineExtOfExt(MachineInstr & MI,std::tuple<Register,unsigned> & MatchInfo)2531 bool CombinerHelper::matchCombineExtOfExt(
2532     MachineInstr &MI, std::tuple<Register, unsigned> &MatchInfo) {
2533   assert((MI.getOpcode() == TargetOpcode::G_ANYEXT ||
2534           MI.getOpcode() == TargetOpcode::G_SEXT ||
2535           MI.getOpcode() == TargetOpcode::G_ZEXT) &&
2536          "Expected a G_[ASZ]EXT");
2537   Register SrcReg = MI.getOperand(1).getReg();
2538   Register OriginalSrcReg = getSrcRegIgnoringCopies(SrcReg, MRI);
2539   if (OriginalSrcReg.isValid())
2540     SrcReg = OriginalSrcReg;
2541   MachineInstr *SrcMI = MRI.getVRegDef(SrcReg);
2542   // Match exts with the same opcode, anyext([sz]ext) and sext(zext).
2543   unsigned Opc = MI.getOpcode();
2544   unsigned SrcOpc = SrcMI->getOpcode();
2545   if (Opc == SrcOpc ||
2546       (Opc == TargetOpcode::G_ANYEXT &&
2547        (SrcOpc == TargetOpcode::G_SEXT || SrcOpc == TargetOpcode::G_ZEXT)) ||
2548       (Opc == TargetOpcode::G_SEXT && SrcOpc == TargetOpcode::G_ZEXT)) {
2549     MatchInfo = std::make_tuple(SrcMI->getOperand(1).getReg(), SrcOpc);
2550     return true;
2551   }
2552   return false;
2553 }
2554 
applyCombineExtOfExt(MachineInstr & MI,std::tuple<Register,unsigned> & MatchInfo)2555 void CombinerHelper::applyCombineExtOfExt(
2556     MachineInstr &MI, std::tuple<Register, unsigned> &MatchInfo) {
2557   assert((MI.getOpcode() == TargetOpcode::G_ANYEXT ||
2558           MI.getOpcode() == TargetOpcode::G_SEXT ||
2559           MI.getOpcode() == TargetOpcode::G_ZEXT) &&
2560          "Expected a G_[ASZ]EXT");
2561 
2562   Register Reg = std::get<0>(MatchInfo);
2563   unsigned SrcExtOp = std::get<1>(MatchInfo);
2564 
2565   // Combine exts with the same opcode.
2566   if (MI.getOpcode() == SrcExtOp) {
2567     Observer.changingInstr(MI);
2568     MI.getOperand(1).setReg(Reg);
2569     Observer.changedInstr(MI);
2570     return;
2571   }
2572 
2573   // Combine:
2574   // - anyext([sz]ext x) to [sz]ext x
2575   // - sext(zext x) to zext x
2576   if (MI.getOpcode() == TargetOpcode::G_ANYEXT ||
2577       (MI.getOpcode() == TargetOpcode::G_SEXT &&
2578        SrcExtOp == TargetOpcode::G_ZEXT)) {
2579     Register DstReg = MI.getOperand(0).getReg();
2580     Builder.buildInstr(SrcExtOp, {DstReg}, {Reg});
2581     MI.eraseFromParent();
2582   }
2583 }
2584 
matchCombineTruncOfExt(MachineInstr & MI,std::pair<Register,unsigned> & MatchInfo)2585 bool CombinerHelper::matchCombineTruncOfExt(
2586     MachineInstr &MI, std::pair<Register, unsigned> &MatchInfo) {
2587   assert(MI.getOpcode() == TargetOpcode::G_TRUNC && "Expected a G_TRUNC");
2588   Register SrcReg = MI.getOperand(1).getReg();
2589   MachineInstr *SrcMI = MRI.getVRegDef(SrcReg);
2590   unsigned SrcOpc = SrcMI->getOpcode();
2591   if (SrcOpc == TargetOpcode::G_ANYEXT || SrcOpc == TargetOpcode::G_SEXT ||
2592       SrcOpc == TargetOpcode::G_ZEXT) {
2593     MatchInfo = std::make_pair(SrcMI->getOperand(1).getReg(), SrcOpc);
2594     return true;
2595   }
2596   return false;
2597 }
2598 
applyCombineTruncOfExt(MachineInstr & MI,std::pair<Register,unsigned> & MatchInfo)2599 void CombinerHelper::applyCombineTruncOfExt(
2600     MachineInstr &MI, std::pair<Register, unsigned> &MatchInfo) {
2601   assert(MI.getOpcode() == TargetOpcode::G_TRUNC && "Expected a G_TRUNC");
2602   Register SrcReg = MatchInfo.first;
2603   unsigned SrcExtOp = MatchInfo.second;
2604   Register DstReg = MI.getOperand(0).getReg();
2605   LLT SrcTy = MRI.getType(SrcReg);
2606   LLT DstTy = MRI.getType(DstReg);
2607   if (SrcTy == DstTy) {
2608     MI.eraseFromParent();
2609     replaceRegWith(MRI, DstReg, SrcReg);
2610     return;
2611   }
2612   if (SrcTy.getSizeInBits() < DstTy.getSizeInBits())
2613     Builder.buildInstr(SrcExtOp, {DstReg}, {SrcReg});
2614   else
2615     Builder.buildTrunc(DstReg, SrcReg);
2616   MI.eraseFromParent();
2617 }
2618 
getMidVTForTruncRightShiftCombine(LLT ShiftTy,LLT TruncTy)2619 static LLT getMidVTForTruncRightShiftCombine(LLT ShiftTy, LLT TruncTy) {
2620   const unsigned ShiftSize = ShiftTy.getScalarSizeInBits();
2621   const unsigned TruncSize = TruncTy.getScalarSizeInBits();
2622 
2623   // ShiftTy > 32 > TruncTy -> 32
2624   if (ShiftSize > 32 && TruncSize < 32)
2625     return ShiftTy.changeElementSize(32);
2626 
2627   // TODO: We could also reduce to 16 bits, but that's more target-dependent.
2628   //  Some targets like it, some don't, some only like it under certain
2629   //  conditions/processor versions, etc.
2630   //  A TL hook might be needed for this.
2631 
2632   // Don't combine
2633   return ShiftTy;
2634 }
2635 
matchCombineTruncOfShift(MachineInstr & MI,std::pair<MachineInstr *,LLT> & MatchInfo)2636 bool CombinerHelper::matchCombineTruncOfShift(
2637     MachineInstr &MI, std::pair<MachineInstr *, LLT> &MatchInfo) {
2638   assert(MI.getOpcode() == TargetOpcode::G_TRUNC && "Expected a G_TRUNC");
2639   Register DstReg = MI.getOperand(0).getReg();
2640   Register SrcReg = MI.getOperand(1).getReg();
2641 
2642   if (!MRI.hasOneNonDBGUse(SrcReg))
2643     return false;
2644 
2645   LLT SrcTy = MRI.getType(SrcReg);
2646   LLT DstTy = MRI.getType(DstReg);
2647 
2648   MachineInstr *SrcMI = getDefIgnoringCopies(SrcReg, MRI);
2649   const auto &TL = getTargetLowering();
2650 
2651   LLT NewShiftTy;
2652   switch (SrcMI->getOpcode()) {
2653   default:
2654     return false;
2655   case TargetOpcode::G_SHL: {
2656     NewShiftTy = DstTy;
2657 
2658     // Make sure new shift amount is legal.
2659     KnownBits Known = KB->getKnownBits(SrcMI->getOperand(2).getReg());
2660     if (Known.getMaxValue().uge(NewShiftTy.getScalarSizeInBits()))
2661       return false;
2662     break;
2663   }
2664   case TargetOpcode::G_LSHR:
2665   case TargetOpcode::G_ASHR: {
2666     // For right shifts, we conservatively do not do the transform if the TRUNC
2667     // has any STORE users. The reason is that if we change the type of the
2668     // shift, we may break the truncstore combine.
2669     //
2670     // TODO: Fix truncstore combine to handle (trunc(lshr (trunc x), k)).
2671     for (auto &User : MRI.use_instructions(DstReg))
2672       if (User.getOpcode() == TargetOpcode::G_STORE)
2673         return false;
2674 
2675     NewShiftTy = getMidVTForTruncRightShiftCombine(SrcTy, DstTy);
2676     if (NewShiftTy == SrcTy)
2677       return false;
2678 
2679     // Make sure we won't lose information by truncating the high bits.
2680     KnownBits Known = KB->getKnownBits(SrcMI->getOperand(2).getReg());
2681     if (Known.getMaxValue().ugt(NewShiftTy.getScalarSizeInBits() -
2682                                 DstTy.getScalarSizeInBits()))
2683       return false;
2684     break;
2685   }
2686   }
2687 
2688   if (!isLegalOrBeforeLegalizer(
2689           {SrcMI->getOpcode(),
2690            {NewShiftTy, TL.getPreferredShiftAmountTy(NewShiftTy)}}))
2691     return false;
2692 
2693   MatchInfo = std::make_pair(SrcMI, NewShiftTy);
2694   return true;
2695 }
2696 
applyCombineTruncOfShift(MachineInstr & MI,std::pair<MachineInstr *,LLT> & MatchInfo)2697 void CombinerHelper::applyCombineTruncOfShift(
2698     MachineInstr &MI, std::pair<MachineInstr *, LLT> &MatchInfo) {
2699   MachineInstr *ShiftMI = MatchInfo.first;
2700   LLT NewShiftTy = MatchInfo.second;
2701 
2702   Register Dst = MI.getOperand(0).getReg();
2703   LLT DstTy = MRI.getType(Dst);
2704 
2705   Register ShiftAmt = ShiftMI->getOperand(2).getReg();
2706   Register ShiftSrc = ShiftMI->getOperand(1).getReg();
2707   ShiftSrc = Builder.buildTrunc(NewShiftTy, ShiftSrc).getReg(0);
2708 
2709   Register NewShift =
2710       Builder
2711           .buildInstr(ShiftMI->getOpcode(), {NewShiftTy}, {ShiftSrc, ShiftAmt})
2712           .getReg(0);
2713 
2714   if (NewShiftTy == DstTy)
2715     replaceRegWith(MRI, Dst, NewShift);
2716   else
2717     Builder.buildTrunc(Dst, NewShift);
2718 
2719   eraseInst(MI);
2720 }
2721 
matchAnyExplicitUseIsUndef(MachineInstr & MI)2722 bool CombinerHelper::matchAnyExplicitUseIsUndef(MachineInstr &MI) {
2723   return any_of(MI.explicit_uses(), [this](const MachineOperand &MO) {
2724     return MO.isReg() &&
2725            getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MO.getReg(), MRI);
2726   });
2727 }
2728 
matchAllExplicitUsesAreUndef(MachineInstr & MI)2729 bool CombinerHelper::matchAllExplicitUsesAreUndef(MachineInstr &MI) {
2730   return all_of(MI.explicit_uses(), [this](const MachineOperand &MO) {
2731     return !MO.isReg() ||
2732            getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MO.getReg(), MRI);
2733   });
2734 }
2735 
matchUndefShuffleVectorMask(MachineInstr & MI)2736 bool CombinerHelper::matchUndefShuffleVectorMask(MachineInstr &MI) {
2737   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
2738   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
2739   return all_of(Mask, [](int Elt) { return Elt < 0; });
2740 }
2741 
matchUndefStore(MachineInstr & MI)2742 bool CombinerHelper::matchUndefStore(MachineInstr &MI) {
2743   assert(MI.getOpcode() == TargetOpcode::G_STORE);
2744   return getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MI.getOperand(0).getReg(),
2745                       MRI);
2746 }
2747 
matchUndefSelectCmp(MachineInstr & MI)2748 bool CombinerHelper::matchUndefSelectCmp(MachineInstr &MI) {
2749   assert(MI.getOpcode() == TargetOpcode::G_SELECT);
2750   return getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MI.getOperand(1).getReg(),
2751                       MRI);
2752 }
2753 
matchInsertExtractVecEltOutOfBounds(MachineInstr & MI)2754 bool CombinerHelper::matchInsertExtractVecEltOutOfBounds(MachineInstr &MI) {
2755   assert((MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT ||
2756           MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT) &&
2757          "Expected an insert/extract element op");
2758   LLT VecTy = MRI.getType(MI.getOperand(1).getReg());
2759   unsigned IdxIdx =
2760       MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3;
2761   auto Idx = getIConstantVRegVal(MI.getOperand(IdxIdx).getReg(), MRI);
2762   if (!Idx)
2763     return false;
2764   return Idx->getZExtValue() >= VecTy.getNumElements();
2765 }
2766 
matchConstantSelectCmp(MachineInstr & MI,unsigned & OpIdx)2767 bool CombinerHelper::matchConstantSelectCmp(MachineInstr &MI, unsigned &OpIdx) {
2768   GSelect &SelMI = cast<GSelect>(MI);
2769   auto Cst =
2770       isConstantOrConstantSplatVector(*MRI.getVRegDef(SelMI.getCondReg()), MRI);
2771   if (!Cst)
2772     return false;
2773   OpIdx = Cst->isZero() ? 3 : 2;
2774   return true;
2775 }
2776 
eraseInst(MachineInstr & MI)2777 void CombinerHelper::eraseInst(MachineInstr &MI) { MI.eraseFromParent(); }
2778 
matchEqualDefs(const MachineOperand & MOP1,const MachineOperand & MOP2)2779 bool CombinerHelper::matchEqualDefs(const MachineOperand &MOP1,
2780                                     const MachineOperand &MOP2) {
2781   if (!MOP1.isReg() || !MOP2.isReg())
2782     return false;
2783   auto InstAndDef1 = getDefSrcRegIgnoringCopies(MOP1.getReg(), MRI);
2784   if (!InstAndDef1)
2785     return false;
2786   auto InstAndDef2 = getDefSrcRegIgnoringCopies(MOP2.getReg(), MRI);
2787   if (!InstAndDef2)
2788     return false;
2789   MachineInstr *I1 = InstAndDef1->MI;
2790   MachineInstr *I2 = InstAndDef2->MI;
2791 
2792   // Handle a case like this:
2793   //
2794   // %0:_(s64), %1:_(s64) = G_UNMERGE_VALUES %2:_(<2 x s64>)
2795   //
2796   // Even though %0 and %1 are produced by the same instruction they are not
2797   // the same values.
2798   if (I1 == I2)
2799     return MOP1.getReg() == MOP2.getReg();
2800 
2801   // If we have an instruction which loads or stores, we can't guarantee that
2802   // it is identical.
2803   //
2804   // For example, we may have
2805   //
2806   // %x1 = G_LOAD %addr (load N from @somewhere)
2807   // ...
2808   // call @foo
2809   // ...
2810   // %x2 = G_LOAD %addr (load N from @somewhere)
2811   // ...
2812   // %or = G_OR %x1, %x2
2813   //
2814   // It's possible that @foo will modify whatever lives at the address we're
2815   // loading from. To be safe, let's just assume that all loads and stores
2816   // are different (unless we have something which is guaranteed to not
2817   // change.)
2818   if (I1->mayLoadOrStore() && !I1->isDereferenceableInvariantLoad())
2819     return false;
2820 
2821   // If both instructions are loads or stores, they are equal only if both
2822   // are dereferenceable invariant loads with the same number of bits.
2823   if (I1->mayLoadOrStore() && I2->mayLoadOrStore()) {
2824     GLoadStore *LS1 = dyn_cast<GLoadStore>(I1);
2825     GLoadStore *LS2 = dyn_cast<GLoadStore>(I2);
2826     if (!LS1 || !LS2)
2827       return false;
2828 
2829     if (!I2->isDereferenceableInvariantLoad() ||
2830         (LS1->getMemSizeInBits() != LS2->getMemSizeInBits()))
2831       return false;
2832   }
2833 
2834   // Check for physical registers on the instructions first to avoid cases
2835   // like this:
2836   //
2837   // %a = COPY $physreg
2838   // ...
2839   // SOMETHING implicit-def $physreg
2840   // ...
2841   // %b = COPY $physreg
2842   //
2843   // These copies are not equivalent.
2844   if (any_of(I1->uses(), [](const MachineOperand &MO) {
2845         return MO.isReg() && MO.getReg().isPhysical();
2846       })) {
2847     // Check if we have a case like this:
2848     //
2849     // %a = COPY $physreg
2850     // %b = COPY %a
2851     //
2852     // In this case, I1 and I2 will both be equal to %a = COPY $physreg.
2853     // From that, we know that they must have the same value, since they must
2854     // have come from the same COPY.
2855     return I1->isIdenticalTo(*I2);
2856   }
2857 
2858   // We don't have any physical registers, so we don't necessarily need the
2859   // same vreg defs.
2860   //
2861   // On the off-chance that there's some target instruction feeding into the
2862   // instruction, let's use produceSameValue instead of isIdenticalTo.
2863   if (Builder.getTII().produceSameValue(*I1, *I2, &MRI)) {
2864     // Handle instructions with multiple defs that produce same values. Values
2865     // are same for operands with same index.
2866     // %0:_(s8), %1:_(s8), %2:_(s8), %3:_(s8) = G_UNMERGE_VALUES %4:_(<4 x s8>)
2867     // %5:_(s8), %6:_(s8), %7:_(s8), %8:_(s8) = G_UNMERGE_VALUES %4:_(<4 x s8>)
2868     // I1 and I2 are different instructions but produce same values,
2869     // %1 and %6 are same, %1 and %7 are not the same value.
2870     return I1->findRegisterDefOperandIdx(InstAndDef1->Reg, /*TRI=*/nullptr) ==
2871            I2->findRegisterDefOperandIdx(InstAndDef2->Reg, /*TRI=*/nullptr);
2872   }
2873   return false;
2874 }
2875 
matchConstantOp(const MachineOperand & MOP,int64_t C)2876 bool CombinerHelper::matchConstantOp(const MachineOperand &MOP, int64_t C) {
2877   if (!MOP.isReg())
2878     return false;
2879   auto *MI = MRI.getVRegDef(MOP.getReg());
2880   auto MaybeCst = isConstantOrConstantSplatVector(*MI, MRI);
2881   return MaybeCst && MaybeCst->getBitWidth() <= 64 &&
2882          MaybeCst->getSExtValue() == C;
2883 }
2884 
matchConstantFPOp(const MachineOperand & MOP,double C)2885 bool CombinerHelper::matchConstantFPOp(const MachineOperand &MOP, double C) {
2886   if (!MOP.isReg())
2887     return false;
2888   std::optional<FPValueAndVReg> MaybeCst;
2889   if (!mi_match(MOP.getReg(), MRI, m_GFCstOrSplat(MaybeCst)))
2890     return false;
2891 
2892   return MaybeCst->Value.isExactlyValue(C);
2893 }
2894 
replaceSingleDefInstWithOperand(MachineInstr & MI,unsigned OpIdx)2895 void CombinerHelper::replaceSingleDefInstWithOperand(MachineInstr &MI,
2896                                                      unsigned OpIdx) {
2897   assert(MI.getNumExplicitDefs() == 1 && "Expected one explicit def?");
2898   Register OldReg = MI.getOperand(0).getReg();
2899   Register Replacement = MI.getOperand(OpIdx).getReg();
2900   assert(canReplaceReg(OldReg, Replacement, MRI) && "Cannot replace register?");
2901   MI.eraseFromParent();
2902   replaceRegWith(MRI, OldReg, Replacement);
2903 }
2904 
replaceSingleDefInstWithReg(MachineInstr & MI,Register Replacement)2905 void CombinerHelper::replaceSingleDefInstWithReg(MachineInstr &MI,
2906                                                  Register Replacement) {
2907   assert(MI.getNumExplicitDefs() == 1 && "Expected one explicit def?");
2908   Register OldReg = MI.getOperand(0).getReg();
2909   assert(canReplaceReg(OldReg, Replacement, MRI) && "Cannot replace register?");
2910   MI.eraseFromParent();
2911   replaceRegWith(MRI, OldReg, Replacement);
2912 }
2913 
matchConstantLargerBitWidth(MachineInstr & MI,unsigned ConstIdx)2914 bool CombinerHelper::matchConstantLargerBitWidth(MachineInstr &MI,
2915                                                  unsigned ConstIdx) {
2916   Register ConstReg = MI.getOperand(ConstIdx).getReg();
2917   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
2918 
2919   // Get the shift amount
2920   auto VRegAndVal = getIConstantVRegValWithLookThrough(ConstReg, MRI);
2921   if (!VRegAndVal)
2922     return false;
2923 
2924   // Return true of shift amount >= Bitwidth
2925   return (VRegAndVal->Value.uge(DstTy.getSizeInBits()));
2926 }
2927 
applyFunnelShiftConstantModulo(MachineInstr & MI)2928 void CombinerHelper::applyFunnelShiftConstantModulo(MachineInstr &MI) {
2929   assert((MI.getOpcode() == TargetOpcode::G_FSHL ||
2930           MI.getOpcode() == TargetOpcode::G_FSHR) &&
2931          "This is not a funnel shift operation");
2932 
2933   Register ConstReg = MI.getOperand(3).getReg();
2934   LLT ConstTy = MRI.getType(ConstReg);
2935   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
2936 
2937   auto VRegAndVal = getIConstantVRegValWithLookThrough(ConstReg, MRI);
2938   assert((VRegAndVal) && "Value is not a constant");
2939 
2940   // Calculate the new Shift Amount = Old Shift Amount % BitWidth
2941   APInt NewConst = VRegAndVal->Value.urem(
2942       APInt(ConstTy.getSizeInBits(), DstTy.getScalarSizeInBits()));
2943 
2944   auto NewConstInstr = Builder.buildConstant(ConstTy, NewConst.getZExtValue());
2945   Builder.buildInstr(
2946       MI.getOpcode(), {MI.getOperand(0)},
2947       {MI.getOperand(1), MI.getOperand(2), NewConstInstr.getReg(0)});
2948 
2949   MI.eraseFromParent();
2950 }
2951 
matchSelectSameVal(MachineInstr & MI)2952 bool CombinerHelper::matchSelectSameVal(MachineInstr &MI) {
2953   assert(MI.getOpcode() == TargetOpcode::G_SELECT);
2954   // Match (cond ? x : x)
2955   return matchEqualDefs(MI.getOperand(2), MI.getOperand(3)) &&
2956          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(2).getReg(),
2957                        MRI);
2958 }
2959 
matchBinOpSameVal(MachineInstr & MI)2960 bool CombinerHelper::matchBinOpSameVal(MachineInstr &MI) {
2961   return matchEqualDefs(MI.getOperand(1), MI.getOperand(2)) &&
2962          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(1).getReg(),
2963                        MRI);
2964 }
2965 
matchOperandIsZero(MachineInstr & MI,unsigned OpIdx)2966 bool CombinerHelper::matchOperandIsZero(MachineInstr &MI, unsigned OpIdx) {
2967   return matchConstantOp(MI.getOperand(OpIdx), 0) &&
2968          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(OpIdx).getReg(),
2969                        MRI);
2970 }
2971 
matchOperandIsUndef(MachineInstr & MI,unsigned OpIdx)2972 bool CombinerHelper::matchOperandIsUndef(MachineInstr &MI, unsigned OpIdx) {
2973   MachineOperand &MO = MI.getOperand(OpIdx);
2974   return MO.isReg() &&
2975          getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MO.getReg(), MRI);
2976 }
2977 
matchOperandIsKnownToBeAPowerOfTwo(MachineInstr & MI,unsigned OpIdx)2978 bool CombinerHelper::matchOperandIsKnownToBeAPowerOfTwo(MachineInstr &MI,
2979                                                         unsigned OpIdx) {
2980   MachineOperand &MO = MI.getOperand(OpIdx);
2981   return isKnownToBeAPowerOfTwo(MO.getReg(), MRI, KB);
2982 }
2983 
replaceInstWithFConstant(MachineInstr & MI,double C)2984 void CombinerHelper::replaceInstWithFConstant(MachineInstr &MI, double C) {
2985   assert(MI.getNumDefs() == 1 && "Expected only one def?");
2986   Builder.buildFConstant(MI.getOperand(0), C);
2987   MI.eraseFromParent();
2988 }
2989 
replaceInstWithConstant(MachineInstr & MI,int64_t C)2990 void CombinerHelper::replaceInstWithConstant(MachineInstr &MI, int64_t C) {
2991   assert(MI.getNumDefs() == 1 && "Expected only one def?");
2992   Builder.buildConstant(MI.getOperand(0), C);
2993   MI.eraseFromParent();
2994 }
2995 
replaceInstWithConstant(MachineInstr & MI,APInt C)2996 void CombinerHelper::replaceInstWithConstant(MachineInstr &MI, APInt C) {
2997   assert(MI.getNumDefs() == 1 && "Expected only one def?");
2998   Builder.buildConstant(MI.getOperand(0), C);
2999   MI.eraseFromParent();
3000 }
3001 
replaceInstWithFConstant(MachineInstr & MI,ConstantFP * CFP)3002 void CombinerHelper::replaceInstWithFConstant(MachineInstr &MI,
3003                                               ConstantFP *CFP) {
3004   assert(MI.getNumDefs() == 1 && "Expected only one def?");
3005   Builder.buildFConstant(MI.getOperand(0), CFP->getValueAPF());
3006   MI.eraseFromParent();
3007 }
3008 
replaceInstWithUndef(MachineInstr & MI)3009 void CombinerHelper::replaceInstWithUndef(MachineInstr &MI) {
3010   assert(MI.getNumDefs() == 1 && "Expected only one def?");
3011   Builder.buildUndef(MI.getOperand(0));
3012   MI.eraseFromParent();
3013 }
3014 
matchSimplifyAddToSub(MachineInstr & MI,std::tuple<Register,Register> & MatchInfo)3015 bool CombinerHelper::matchSimplifyAddToSub(
3016     MachineInstr &MI, std::tuple<Register, Register> &MatchInfo) {
3017   Register LHS = MI.getOperand(1).getReg();
3018   Register RHS = MI.getOperand(2).getReg();
3019   Register &NewLHS = std::get<0>(MatchInfo);
3020   Register &NewRHS = std::get<1>(MatchInfo);
3021 
3022   // Helper lambda to check for opportunities for
3023   // ((0-A) + B) -> B - A
3024   // (A + (0-B)) -> A - B
3025   auto CheckFold = [&](Register &MaybeSub, Register &MaybeNewLHS) {
3026     if (!mi_match(MaybeSub, MRI, m_Neg(m_Reg(NewRHS))))
3027       return false;
3028     NewLHS = MaybeNewLHS;
3029     return true;
3030   };
3031 
3032   return CheckFold(LHS, RHS) || CheckFold(RHS, LHS);
3033 }
3034 
matchCombineInsertVecElts(MachineInstr & MI,SmallVectorImpl<Register> & MatchInfo)3035 bool CombinerHelper::matchCombineInsertVecElts(
3036     MachineInstr &MI, SmallVectorImpl<Register> &MatchInfo) {
3037   assert(MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT &&
3038          "Invalid opcode");
3039   Register DstReg = MI.getOperand(0).getReg();
3040   LLT DstTy = MRI.getType(DstReg);
3041   assert(DstTy.isVector() && "Invalid G_INSERT_VECTOR_ELT?");
3042   unsigned NumElts = DstTy.getNumElements();
3043   // If this MI is part of a sequence of insert_vec_elts, then
3044   // don't do the combine in the middle of the sequence.
3045   if (MRI.hasOneUse(DstReg) && MRI.use_instr_begin(DstReg)->getOpcode() ==
3046                                    TargetOpcode::G_INSERT_VECTOR_ELT)
3047     return false;
3048   MachineInstr *CurrInst = &MI;
3049   MachineInstr *TmpInst;
3050   int64_t IntImm;
3051   Register TmpReg;
3052   MatchInfo.resize(NumElts);
3053   while (mi_match(
3054       CurrInst->getOperand(0).getReg(), MRI,
3055       m_GInsertVecElt(m_MInstr(TmpInst), m_Reg(TmpReg), m_ICst(IntImm)))) {
3056     if (IntImm >= NumElts || IntImm < 0)
3057       return false;
3058     if (!MatchInfo[IntImm])
3059       MatchInfo[IntImm] = TmpReg;
3060     CurrInst = TmpInst;
3061   }
3062   // Variable index.
3063   if (CurrInst->getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT)
3064     return false;
3065   if (TmpInst->getOpcode() == TargetOpcode::G_BUILD_VECTOR) {
3066     for (unsigned I = 1; I < TmpInst->getNumOperands(); ++I) {
3067       if (!MatchInfo[I - 1].isValid())
3068         MatchInfo[I - 1] = TmpInst->getOperand(I).getReg();
3069     }
3070     return true;
3071   }
3072   // If we didn't end in a G_IMPLICIT_DEF and the source is not fully
3073   // overwritten, bail out.
3074   return TmpInst->getOpcode() == TargetOpcode::G_IMPLICIT_DEF ||
3075          all_of(MatchInfo, [](Register Reg) { return !!Reg; });
3076 }
3077 
applyCombineInsertVecElts(MachineInstr & MI,SmallVectorImpl<Register> & MatchInfo)3078 void CombinerHelper::applyCombineInsertVecElts(
3079     MachineInstr &MI, SmallVectorImpl<Register> &MatchInfo) {
3080   Register UndefReg;
3081   auto GetUndef = [&]() {
3082     if (UndefReg)
3083       return UndefReg;
3084     LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
3085     UndefReg = Builder.buildUndef(DstTy.getScalarType()).getReg(0);
3086     return UndefReg;
3087   };
3088   for (Register &Reg : MatchInfo) {
3089     if (!Reg)
3090       Reg = GetUndef();
3091   }
3092   Builder.buildBuildVector(MI.getOperand(0).getReg(), MatchInfo);
3093   MI.eraseFromParent();
3094 }
3095 
applySimplifyAddToSub(MachineInstr & MI,std::tuple<Register,Register> & MatchInfo)3096 void CombinerHelper::applySimplifyAddToSub(
3097     MachineInstr &MI, std::tuple<Register, Register> &MatchInfo) {
3098   Register SubLHS, SubRHS;
3099   std::tie(SubLHS, SubRHS) = MatchInfo;
3100   Builder.buildSub(MI.getOperand(0).getReg(), SubLHS, SubRHS);
3101   MI.eraseFromParent();
3102 }
3103 
matchHoistLogicOpWithSameOpcodeHands(MachineInstr & MI,InstructionStepsMatchInfo & MatchInfo)3104 bool CombinerHelper::matchHoistLogicOpWithSameOpcodeHands(
3105     MachineInstr &MI, InstructionStepsMatchInfo &MatchInfo) {
3106   // Matches: logic (hand x, ...), (hand y, ...) -> hand (logic x, y), ...
3107   //
3108   // Creates the new hand + logic instruction (but does not insert them.)
3109   //
3110   // On success, MatchInfo is populated with the new instructions. These are
3111   // inserted in applyHoistLogicOpWithSameOpcodeHands.
3112   unsigned LogicOpcode = MI.getOpcode();
3113   assert(LogicOpcode == TargetOpcode::G_AND ||
3114          LogicOpcode == TargetOpcode::G_OR ||
3115          LogicOpcode == TargetOpcode::G_XOR);
3116   MachineIRBuilder MIB(MI);
3117   Register Dst = MI.getOperand(0).getReg();
3118   Register LHSReg = MI.getOperand(1).getReg();
3119   Register RHSReg = MI.getOperand(2).getReg();
3120 
3121   // Don't recompute anything.
3122   if (!MRI.hasOneNonDBGUse(LHSReg) || !MRI.hasOneNonDBGUse(RHSReg))
3123     return false;
3124 
3125   // Make sure we have (hand x, ...), (hand y, ...)
3126   MachineInstr *LeftHandInst = getDefIgnoringCopies(LHSReg, MRI);
3127   MachineInstr *RightHandInst = getDefIgnoringCopies(RHSReg, MRI);
3128   if (!LeftHandInst || !RightHandInst)
3129     return false;
3130   unsigned HandOpcode = LeftHandInst->getOpcode();
3131   if (HandOpcode != RightHandInst->getOpcode())
3132     return false;
3133   if (!LeftHandInst->getOperand(1).isReg() ||
3134       !RightHandInst->getOperand(1).isReg())
3135     return false;
3136 
3137   // Make sure the types match up, and if we're doing this post-legalization,
3138   // we end up with legal types.
3139   Register X = LeftHandInst->getOperand(1).getReg();
3140   Register Y = RightHandInst->getOperand(1).getReg();
3141   LLT XTy = MRI.getType(X);
3142   LLT YTy = MRI.getType(Y);
3143   if (!XTy.isValid() || XTy != YTy)
3144     return false;
3145 
3146   // Optional extra source register.
3147   Register ExtraHandOpSrcReg;
3148   switch (HandOpcode) {
3149   default:
3150     return false;
3151   case TargetOpcode::G_ANYEXT:
3152   case TargetOpcode::G_SEXT:
3153   case TargetOpcode::G_ZEXT: {
3154     // Match: logic (ext X), (ext Y) --> ext (logic X, Y)
3155     break;
3156   }
3157   case TargetOpcode::G_TRUNC: {
3158     // Match: logic (trunc X), (trunc Y) -> trunc (logic X, Y)
3159     const MachineFunction *MF = MI.getMF();
3160     const DataLayout &DL = MF->getDataLayout();
3161     LLVMContext &Ctx = MF->getFunction().getContext();
3162 
3163     LLT DstTy = MRI.getType(Dst);
3164     const TargetLowering &TLI = getTargetLowering();
3165 
3166     // Be extra careful sinking truncate. If it's free, there's no benefit in
3167     // widening a binop.
3168     if (TLI.isZExtFree(DstTy, XTy, DL, Ctx) &&
3169         TLI.isTruncateFree(XTy, DstTy, DL, Ctx))
3170       return false;
3171     break;
3172   }
3173   case TargetOpcode::G_AND:
3174   case TargetOpcode::G_ASHR:
3175   case TargetOpcode::G_LSHR:
3176   case TargetOpcode::G_SHL: {
3177     // Match: logic (binop x, z), (binop y, z) -> binop (logic x, y), z
3178     MachineOperand &ZOp = LeftHandInst->getOperand(2);
3179     if (!matchEqualDefs(ZOp, RightHandInst->getOperand(2)))
3180       return false;
3181     ExtraHandOpSrcReg = ZOp.getReg();
3182     break;
3183   }
3184   }
3185 
3186   if (!isLegalOrBeforeLegalizer({LogicOpcode, {XTy, YTy}}))
3187     return false;
3188 
3189   // Record the steps to build the new instructions.
3190   //
3191   // Steps to build (logic x, y)
3192   auto NewLogicDst = MRI.createGenericVirtualRegister(XTy);
3193   OperandBuildSteps LogicBuildSteps = {
3194       [=](MachineInstrBuilder &MIB) { MIB.addDef(NewLogicDst); },
3195       [=](MachineInstrBuilder &MIB) { MIB.addReg(X); },
3196       [=](MachineInstrBuilder &MIB) { MIB.addReg(Y); }};
3197   InstructionBuildSteps LogicSteps(LogicOpcode, LogicBuildSteps);
3198 
3199   // Steps to build hand (logic x, y), ...z
3200   OperandBuildSteps HandBuildSteps = {
3201       [=](MachineInstrBuilder &MIB) { MIB.addDef(Dst); },
3202       [=](MachineInstrBuilder &MIB) { MIB.addReg(NewLogicDst); }};
3203   if (ExtraHandOpSrcReg.isValid())
3204     HandBuildSteps.push_back(
3205         [=](MachineInstrBuilder &MIB) { MIB.addReg(ExtraHandOpSrcReg); });
3206   InstructionBuildSteps HandSteps(HandOpcode, HandBuildSteps);
3207 
3208   MatchInfo = InstructionStepsMatchInfo({LogicSteps, HandSteps});
3209   return true;
3210 }
3211 
applyBuildInstructionSteps(MachineInstr & MI,InstructionStepsMatchInfo & MatchInfo)3212 void CombinerHelper::applyBuildInstructionSteps(
3213     MachineInstr &MI, InstructionStepsMatchInfo &MatchInfo) {
3214   assert(MatchInfo.InstrsToBuild.size() &&
3215          "Expected at least one instr to build?");
3216   for (auto &InstrToBuild : MatchInfo.InstrsToBuild) {
3217     assert(InstrToBuild.Opcode && "Expected a valid opcode?");
3218     assert(InstrToBuild.OperandFns.size() && "Expected at least one operand?");
3219     MachineInstrBuilder Instr = Builder.buildInstr(InstrToBuild.Opcode);
3220     for (auto &OperandFn : InstrToBuild.OperandFns)
3221       OperandFn(Instr);
3222   }
3223   MI.eraseFromParent();
3224 }
3225 
matchAshrShlToSextInreg(MachineInstr & MI,std::tuple<Register,int64_t> & MatchInfo)3226 bool CombinerHelper::matchAshrShlToSextInreg(
3227     MachineInstr &MI, std::tuple<Register, int64_t> &MatchInfo) {
3228   assert(MI.getOpcode() == TargetOpcode::G_ASHR);
3229   int64_t ShlCst, AshrCst;
3230   Register Src;
3231   if (!mi_match(MI.getOperand(0).getReg(), MRI,
3232                 m_GAShr(m_GShl(m_Reg(Src), m_ICstOrSplat(ShlCst)),
3233                         m_ICstOrSplat(AshrCst))))
3234     return false;
3235   if (ShlCst != AshrCst)
3236     return false;
3237   if (!isLegalOrBeforeLegalizer(
3238           {TargetOpcode::G_SEXT_INREG, {MRI.getType(Src)}}))
3239     return false;
3240   MatchInfo = std::make_tuple(Src, ShlCst);
3241   return true;
3242 }
3243 
applyAshShlToSextInreg(MachineInstr & MI,std::tuple<Register,int64_t> & MatchInfo)3244 void CombinerHelper::applyAshShlToSextInreg(
3245     MachineInstr &MI, std::tuple<Register, int64_t> &MatchInfo) {
3246   assert(MI.getOpcode() == TargetOpcode::G_ASHR);
3247   Register Src;
3248   int64_t ShiftAmt;
3249   std::tie(Src, ShiftAmt) = MatchInfo;
3250   unsigned Size = MRI.getType(Src).getScalarSizeInBits();
3251   Builder.buildSExtInReg(MI.getOperand(0).getReg(), Src, Size - ShiftAmt);
3252   MI.eraseFromParent();
3253 }
3254 
3255 /// and(and(x, C1), C2) -> C1&C2 ? and(x, C1&C2) : 0
matchOverlappingAnd(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)3256 bool CombinerHelper::matchOverlappingAnd(
3257     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
3258   assert(MI.getOpcode() == TargetOpcode::G_AND);
3259 
3260   Register Dst = MI.getOperand(0).getReg();
3261   LLT Ty = MRI.getType(Dst);
3262 
3263   Register R;
3264   int64_t C1;
3265   int64_t C2;
3266   if (!mi_match(
3267           Dst, MRI,
3268           m_GAnd(m_GAnd(m_Reg(R), m_ICst(C1)), m_ICst(C2))))
3269     return false;
3270 
3271   MatchInfo = [=](MachineIRBuilder &B) {
3272     if (C1 & C2) {
3273       B.buildAnd(Dst, R, B.buildConstant(Ty, C1 & C2));
3274       return;
3275     }
3276     auto Zero = B.buildConstant(Ty, 0);
3277     replaceRegWith(MRI, Dst, Zero->getOperand(0).getReg());
3278   };
3279   return true;
3280 }
3281 
matchRedundantAnd(MachineInstr & MI,Register & Replacement)3282 bool CombinerHelper::matchRedundantAnd(MachineInstr &MI,
3283                                        Register &Replacement) {
3284   // Given
3285   //
3286   // %y:_(sN) = G_SOMETHING
3287   // %x:_(sN) = G_SOMETHING
3288   // %res:_(sN) = G_AND %x, %y
3289   //
3290   // Eliminate the G_AND when it is known that x & y == x or x & y == y.
3291   //
3292   // Patterns like this can appear as a result of legalization. E.g.
3293   //
3294   // %cmp:_(s32) = G_ICMP intpred(pred), %x(s32), %y
3295   // %one:_(s32) = G_CONSTANT i32 1
3296   // %and:_(s32) = G_AND %cmp, %one
3297   //
3298   // In this case, G_ICMP only produces a single bit, so x & 1 == x.
3299   assert(MI.getOpcode() == TargetOpcode::G_AND);
3300   if (!KB)
3301     return false;
3302 
3303   Register AndDst = MI.getOperand(0).getReg();
3304   Register LHS = MI.getOperand(1).getReg();
3305   Register RHS = MI.getOperand(2).getReg();
3306 
3307   // Check the RHS (maybe a constant) first, and if we have no KnownBits there,
3308   // we can't do anything. If we do, then it depends on whether we have
3309   // KnownBits on the LHS.
3310   KnownBits RHSBits = KB->getKnownBits(RHS);
3311   if (RHSBits.isUnknown())
3312     return false;
3313 
3314   KnownBits LHSBits = KB->getKnownBits(LHS);
3315 
3316   // Check that x & Mask == x.
3317   // x & 1 == x, always
3318   // x & 0 == x, only if x is also 0
3319   // Meaning Mask has no effect if every bit is either one in Mask or zero in x.
3320   //
3321   // Check if we can replace AndDst with the LHS of the G_AND
3322   if (canReplaceReg(AndDst, LHS, MRI) &&
3323       (LHSBits.Zero | RHSBits.One).isAllOnes()) {
3324     Replacement = LHS;
3325     return true;
3326   }
3327 
3328   // Check if we can replace AndDst with the RHS of the G_AND
3329   if (canReplaceReg(AndDst, RHS, MRI) &&
3330       (LHSBits.One | RHSBits.Zero).isAllOnes()) {
3331     Replacement = RHS;
3332     return true;
3333   }
3334 
3335   return false;
3336 }
3337 
matchRedundantOr(MachineInstr & MI,Register & Replacement)3338 bool CombinerHelper::matchRedundantOr(MachineInstr &MI, Register &Replacement) {
3339   // Given
3340   //
3341   // %y:_(sN) = G_SOMETHING
3342   // %x:_(sN) = G_SOMETHING
3343   // %res:_(sN) = G_OR %x, %y
3344   //
3345   // Eliminate the G_OR when it is known that x | y == x or x | y == y.
3346   assert(MI.getOpcode() == TargetOpcode::G_OR);
3347   if (!KB)
3348     return false;
3349 
3350   Register OrDst = MI.getOperand(0).getReg();
3351   Register LHS = MI.getOperand(1).getReg();
3352   Register RHS = MI.getOperand(2).getReg();
3353 
3354   KnownBits LHSBits = KB->getKnownBits(LHS);
3355   KnownBits RHSBits = KB->getKnownBits(RHS);
3356 
3357   // Check that x | Mask == x.
3358   // x | 0 == x, always
3359   // x | 1 == x, only if x is also 1
3360   // Meaning Mask has no effect if every bit is either zero in Mask or one in x.
3361   //
3362   // Check if we can replace OrDst with the LHS of the G_OR
3363   if (canReplaceReg(OrDst, LHS, MRI) &&
3364       (LHSBits.One | RHSBits.Zero).isAllOnes()) {
3365     Replacement = LHS;
3366     return true;
3367   }
3368 
3369   // Check if we can replace OrDst with the RHS of the G_OR
3370   if (canReplaceReg(OrDst, RHS, MRI) &&
3371       (LHSBits.Zero | RHSBits.One).isAllOnes()) {
3372     Replacement = RHS;
3373     return true;
3374   }
3375 
3376   return false;
3377 }
3378 
matchRedundantSExtInReg(MachineInstr & MI)3379 bool CombinerHelper::matchRedundantSExtInReg(MachineInstr &MI) {
3380   // If the input is already sign extended, just drop the extension.
3381   Register Src = MI.getOperand(1).getReg();
3382   unsigned ExtBits = MI.getOperand(2).getImm();
3383   unsigned TypeSize = MRI.getType(Src).getScalarSizeInBits();
3384   return KB->computeNumSignBits(Src) >= (TypeSize - ExtBits + 1);
3385 }
3386 
isConstValidTrue(const TargetLowering & TLI,unsigned ScalarSizeBits,int64_t Cst,bool IsVector,bool IsFP)3387 static bool isConstValidTrue(const TargetLowering &TLI, unsigned ScalarSizeBits,
3388                              int64_t Cst, bool IsVector, bool IsFP) {
3389   // For i1, Cst will always be -1 regardless of boolean contents.
3390   return (ScalarSizeBits == 1 && Cst == -1) ||
3391          isConstTrueVal(TLI, Cst, IsVector, IsFP);
3392 }
3393 
matchNotCmp(MachineInstr & MI,SmallVectorImpl<Register> & RegsToNegate)3394 bool CombinerHelper::matchNotCmp(MachineInstr &MI,
3395                                  SmallVectorImpl<Register> &RegsToNegate) {
3396   assert(MI.getOpcode() == TargetOpcode::G_XOR);
3397   LLT Ty = MRI.getType(MI.getOperand(0).getReg());
3398   const auto &TLI = *Builder.getMF().getSubtarget().getTargetLowering();
3399   Register XorSrc;
3400   Register CstReg;
3401   // We match xor(src, true) here.
3402   if (!mi_match(MI.getOperand(0).getReg(), MRI,
3403                 m_GXor(m_Reg(XorSrc), m_Reg(CstReg))))
3404     return false;
3405 
3406   if (!MRI.hasOneNonDBGUse(XorSrc))
3407     return false;
3408 
3409   // Check that XorSrc is the root of a tree of comparisons combined with ANDs
3410   // and ORs. The suffix of RegsToNegate starting from index I is used a work
3411   // list of tree nodes to visit.
3412   RegsToNegate.push_back(XorSrc);
3413   // Remember whether the comparisons are all integer or all floating point.
3414   bool IsInt = false;
3415   bool IsFP = false;
3416   for (unsigned I = 0; I < RegsToNegate.size(); ++I) {
3417     Register Reg = RegsToNegate[I];
3418     if (!MRI.hasOneNonDBGUse(Reg))
3419       return false;
3420     MachineInstr *Def = MRI.getVRegDef(Reg);
3421     switch (Def->getOpcode()) {
3422     default:
3423       // Don't match if the tree contains anything other than ANDs, ORs and
3424       // comparisons.
3425       return false;
3426     case TargetOpcode::G_ICMP:
3427       if (IsFP)
3428         return false;
3429       IsInt = true;
3430       // When we apply the combine we will invert the predicate.
3431       break;
3432     case TargetOpcode::G_FCMP:
3433       if (IsInt)
3434         return false;
3435       IsFP = true;
3436       // When we apply the combine we will invert the predicate.
3437       break;
3438     case TargetOpcode::G_AND:
3439     case TargetOpcode::G_OR:
3440       // Implement De Morgan's laws:
3441       // ~(x & y) -> ~x | ~y
3442       // ~(x | y) -> ~x & ~y
3443       // When we apply the combine we will change the opcode and recursively
3444       // negate the operands.
3445       RegsToNegate.push_back(Def->getOperand(1).getReg());
3446       RegsToNegate.push_back(Def->getOperand(2).getReg());
3447       break;
3448     }
3449   }
3450 
3451   // Now we know whether the comparisons are integer or floating point, check
3452   // the constant in the xor.
3453   int64_t Cst;
3454   if (Ty.isVector()) {
3455     MachineInstr *CstDef = MRI.getVRegDef(CstReg);
3456     auto MaybeCst = getIConstantSplatSExtVal(*CstDef, MRI);
3457     if (!MaybeCst)
3458       return false;
3459     if (!isConstValidTrue(TLI, Ty.getScalarSizeInBits(), *MaybeCst, true, IsFP))
3460       return false;
3461   } else {
3462     if (!mi_match(CstReg, MRI, m_ICst(Cst)))
3463       return false;
3464     if (!isConstValidTrue(TLI, Ty.getSizeInBits(), Cst, false, IsFP))
3465       return false;
3466   }
3467 
3468   return true;
3469 }
3470 
applyNotCmp(MachineInstr & MI,SmallVectorImpl<Register> & RegsToNegate)3471 void CombinerHelper::applyNotCmp(MachineInstr &MI,
3472                                  SmallVectorImpl<Register> &RegsToNegate) {
3473   for (Register Reg : RegsToNegate) {
3474     MachineInstr *Def = MRI.getVRegDef(Reg);
3475     Observer.changingInstr(*Def);
3476     // For each comparison, invert the opcode. For each AND and OR, change the
3477     // opcode.
3478     switch (Def->getOpcode()) {
3479     default:
3480       llvm_unreachable("Unexpected opcode");
3481     case TargetOpcode::G_ICMP:
3482     case TargetOpcode::G_FCMP: {
3483       MachineOperand &PredOp = Def->getOperand(1);
3484       CmpInst::Predicate NewP = CmpInst::getInversePredicate(
3485           (CmpInst::Predicate)PredOp.getPredicate());
3486       PredOp.setPredicate(NewP);
3487       break;
3488     }
3489     case TargetOpcode::G_AND:
3490       Def->setDesc(Builder.getTII().get(TargetOpcode::G_OR));
3491       break;
3492     case TargetOpcode::G_OR:
3493       Def->setDesc(Builder.getTII().get(TargetOpcode::G_AND));
3494       break;
3495     }
3496     Observer.changedInstr(*Def);
3497   }
3498 
3499   replaceRegWith(MRI, MI.getOperand(0).getReg(), MI.getOperand(1).getReg());
3500   MI.eraseFromParent();
3501 }
3502 
matchXorOfAndWithSameReg(MachineInstr & MI,std::pair<Register,Register> & MatchInfo)3503 bool CombinerHelper::matchXorOfAndWithSameReg(
3504     MachineInstr &MI, std::pair<Register, Register> &MatchInfo) {
3505   // Match (xor (and x, y), y) (or any of its commuted cases)
3506   assert(MI.getOpcode() == TargetOpcode::G_XOR);
3507   Register &X = MatchInfo.first;
3508   Register &Y = MatchInfo.second;
3509   Register AndReg = MI.getOperand(1).getReg();
3510   Register SharedReg = MI.getOperand(2).getReg();
3511 
3512   // Find a G_AND on either side of the G_XOR.
3513   // Look for one of
3514   //
3515   // (xor (and x, y), SharedReg)
3516   // (xor SharedReg, (and x, y))
3517   if (!mi_match(AndReg, MRI, m_GAnd(m_Reg(X), m_Reg(Y)))) {
3518     std::swap(AndReg, SharedReg);
3519     if (!mi_match(AndReg, MRI, m_GAnd(m_Reg(X), m_Reg(Y))))
3520       return false;
3521   }
3522 
3523   // Only do this if we'll eliminate the G_AND.
3524   if (!MRI.hasOneNonDBGUse(AndReg))
3525     return false;
3526 
3527   // We can combine if SharedReg is the same as either the LHS or RHS of the
3528   // G_AND.
3529   if (Y != SharedReg)
3530     std::swap(X, Y);
3531   return Y == SharedReg;
3532 }
3533 
applyXorOfAndWithSameReg(MachineInstr & MI,std::pair<Register,Register> & MatchInfo)3534 void CombinerHelper::applyXorOfAndWithSameReg(
3535     MachineInstr &MI, std::pair<Register, Register> &MatchInfo) {
3536   // Fold (xor (and x, y), y) -> (and (not x), y)
3537   Register X, Y;
3538   std::tie(X, Y) = MatchInfo;
3539   auto Not = Builder.buildNot(MRI.getType(X), X);
3540   Observer.changingInstr(MI);
3541   MI.setDesc(Builder.getTII().get(TargetOpcode::G_AND));
3542   MI.getOperand(1).setReg(Not->getOperand(0).getReg());
3543   MI.getOperand(2).setReg(Y);
3544   Observer.changedInstr(MI);
3545 }
3546 
matchPtrAddZero(MachineInstr & MI)3547 bool CombinerHelper::matchPtrAddZero(MachineInstr &MI) {
3548   auto &PtrAdd = cast<GPtrAdd>(MI);
3549   Register DstReg = PtrAdd.getReg(0);
3550   LLT Ty = MRI.getType(DstReg);
3551   const DataLayout &DL = Builder.getMF().getDataLayout();
3552 
3553   if (DL.isNonIntegralAddressSpace(Ty.getScalarType().getAddressSpace()))
3554     return false;
3555 
3556   if (Ty.isPointer()) {
3557     auto ConstVal = getIConstantVRegVal(PtrAdd.getBaseReg(), MRI);
3558     return ConstVal && *ConstVal == 0;
3559   }
3560 
3561   assert(Ty.isVector() && "Expecting a vector type");
3562   const MachineInstr *VecMI = MRI.getVRegDef(PtrAdd.getBaseReg());
3563   return isBuildVectorAllZeros(*VecMI, MRI);
3564 }
3565 
applyPtrAddZero(MachineInstr & MI)3566 void CombinerHelper::applyPtrAddZero(MachineInstr &MI) {
3567   auto &PtrAdd = cast<GPtrAdd>(MI);
3568   Builder.buildIntToPtr(PtrAdd.getReg(0), PtrAdd.getOffsetReg());
3569   PtrAdd.eraseFromParent();
3570 }
3571 
3572 /// The second source operand is known to be a power of 2.
applySimplifyURemByPow2(MachineInstr & MI)3573 void CombinerHelper::applySimplifyURemByPow2(MachineInstr &MI) {
3574   Register DstReg = MI.getOperand(0).getReg();
3575   Register Src0 = MI.getOperand(1).getReg();
3576   Register Pow2Src1 = MI.getOperand(2).getReg();
3577   LLT Ty = MRI.getType(DstReg);
3578 
3579   // Fold (urem x, pow2) -> (and x, pow2-1)
3580   auto NegOne = Builder.buildConstant(Ty, -1);
3581   auto Add = Builder.buildAdd(Ty, Pow2Src1, NegOne);
3582   Builder.buildAnd(DstReg, Src0, Add);
3583   MI.eraseFromParent();
3584 }
3585 
matchFoldBinOpIntoSelect(MachineInstr & MI,unsigned & SelectOpNo)3586 bool CombinerHelper::matchFoldBinOpIntoSelect(MachineInstr &MI,
3587                                               unsigned &SelectOpNo) {
3588   Register LHS = MI.getOperand(1).getReg();
3589   Register RHS = MI.getOperand(2).getReg();
3590 
3591   Register OtherOperandReg = RHS;
3592   SelectOpNo = 1;
3593   MachineInstr *Select = MRI.getVRegDef(LHS);
3594 
3595   // Don't do this unless the old select is going away. We want to eliminate the
3596   // binary operator, not replace a binop with a select.
3597   if (Select->getOpcode() != TargetOpcode::G_SELECT ||
3598       !MRI.hasOneNonDBGUse(LHS)) {
3599     OtherOperandReg = LHS;
3600     SelectOpNo = 2;
3601     Select = MRI.getVRegDef(RHS);
3602     if (Select->getOpcode() != TargetOpcode::G_SELECT ||
3603         !MRI.hasOneNonDBGUse(RHS))
3604       return false;
3605   }
3606 
3607   MachineInstr *SelectLHS = MRI.getVRegDef(Select->getOperand(2).getReg());
3608   MachineInstr *SelectRHS = MRI.getVRegDef(Select->getOperand(3).getReg());
3609 
3610   if (!isConstantOrConstantVector(*SelectLHS, MRI,
3611                                   /*AllowFP*/ true,
3612                                   /*AllowOpaqueConstants*/ false))
3613     return false;
3614   if (!isConstantOrConstantVector(*SelectRHS, MRI,
3615                                   /*AllowFP*/ true,
3616                                   /*AllowOpaqueConstants*/ false))
3617     return false;
3618 
3619   unsigned BinOpcode = MI.getOpcode();
3620 
3621   // We know that one of the operands is a select of constants. Now verify that
3622   // the other binary operator operand is either a constant, or we can handle a
3623   // variable.
3624   bool CanFoldNonConst =
3625       (BinOpcode == TargetOpcode::G_AND || BinOpcode == TargetOpcode::G_OR) &&
3626       (isNullOrNullSplat(*SelectLHS, MRI) ||
3627        isAllOnesOrAllOnesSplat(*SelectLHS, MRI)) &&
3628       (isNullOrNullSplat(*SelectRHS, MRI) ||
3629        isAllOnesOrAllOnesSplat(*SelectRHS, MRI));
3630   if (CanFoldNonConst)
3631     return true;
3632 
3633   return isConstantOrConstantVector(*MRI.getVRegDef(OtherOperandReg), MRI,
3634                                     /*AllowFP*/ true,
3635                                     /*AllowOpaqueConstants*/ false);
3636 }
3637 
3638 /// \p SelectOperand is the operand in binary operator \p MI that is the select
3639 /// to fold.
applyFoldBinOpIntoSelect(MachineInstr & MI,const unsigned & SelectOperand)3640 void CombinerHelper::applyFoldBinOpIntoSelect(MachineInstr &MI,
3641                                               const unsigned &SelectOperand) {
3642   Register Dst = MI.getOperand(0).getReg();
3643   Register LHS = MI.getOperand(1).getReg();
3644   Register RHS = MI.getOperand(2).getReg();
3645   MachineInstr *Select = MRI.getVRegDef(MI.getOperand(SelectOperand).getReg());
3646 
3647   Register SelectCond = Select->getOperand(1).getReg();
3648   Register SelectTrue = Select->getOperand(2).getReg();
3649   Register SelectFalse = Select->getOperand(3).getReg();
3650 
3651   LLT Ty = MRI.getType(Dst);
3652   unsigned BinOpcode = MI.getOpcode();
3653 
3654   Register FoldTrue, FoldFalse;
3655 
3656   // We have a select-of-constants followed by a binary operator with a
3657   // constant. Eliminate the binop by pulling the constant math into the select.
3658   // Example: add (select Cond, CT, CF), CBO --> select Cond, CT + CBO, CF + CBO
3659   if (SelectOperand == 1) {
3660     // TODO: SelectionDAG verifies this actually constant folds before
3661     // committing to the combine.
3662 
3663     FoldTrue = Builder.buildInstr(BinOpcode, {Ty}, {SelectTrue, RHS}).getReg(0);
3664     FoldFalse =
3665         Builder.buildInstr(BinOpcode, {Ty}, {SelectFalse, RHS}).getReg(0);
3666   } else {
3667     FoldTrue = Builder.buildInstr(BinOpcode, {Ty}, {LHS, SelectTrue}).getReg(0);
3668     FoldFalse =
3669         Builder.buildInstr(BinOpcode, {Ty}, {LHS, SelectFalse}).getReg(0);
3670   }
3671 
3672   Builder.buildSelect(Dst, SelectCond, FoldTrue, FoldFalse, MI.getFlags());
3673   MI.eraseFromParent();
3674 }
3675 
3676 std::optional<SmallVector<Register, 8>>
findCandidatesForLoadOrCombine(const MachineInstr * Root) const3677 CombinerHelper::findCandidatesForLoadOrCombine(const MachineInstr *Root) const {
3678   assert(Root->getOpcode() == TargetOpcode::G_OR && "Expected G_OR only!");
3679   // We want to detect if Root is part of a tree which represents a bunch
3680   // of loads being merged into a larger load. We'll try to recognize patterns
3681   // like, for example:
3682   //
3683   //  Reg   Reg
3684   //   \    /
3685   //    OR_1   Reg
3686   //     \    /
3687   //      OR_2
3688   //        \     Reg
3689   //         .. /
3690   //        Root
3691   //
3692   //  Reg   Reg   Reg   Reg
3693   //     \ /       \   /
3694   //     OR_1      OR_2
3695   //       \       /
3696   //        \    /
3697   //         ...
3698   //         Root
3699   //
3700   // Each "Reg" may have been produced by a load + some arithmetic. This
3701   // function will save each of them.
3702   SmallVector<Register, 8> RegsToVisit;
3703   SmallVector<const MachineInstr *, 7> Ors = {Root};
3704 
3705   // In the "worst" case, we're dealing with a load for each byte. So, there
3706   // are at most #bytes - 1 ORs.
3707   const unsigned MaxIter =
3708       MRI.getType(Root->getOperand(0).getReg()).getSizeInBytes() - 1;
3709   for (unsigned Iter = 0; Iter < MaxIter; ++Iter) {
3710     if (Ors.empty())
3711       break;
3712     const MachineInstr *Curr = Ors.pop_back_val();
3713     Register OrLHS = Curr->getOperand(1).getReg();
3714     Register OrRHS = Curr->getOperand(2).getReg();
3715 
3716     // In the combine, we want to elimate the entire tree.
3717     if (!MRI.hasOneNonDBGUse(OrLHS) || !MRI.hasOneNonDBGUse(OrRHS))
3718       return std::nullopt;
3719 
3720     // If it's a G_OR, save it and continue to walk. If it's not, then it's
3721     // something that may be a load + arithmetic.
3722     if (const MachineInstr *Or = getOpcodeDef(TargetOpcode::G_OR, OrLHS, MRI))
3723       Ors.push_back(Or);
3724     else
3725       RegsToVisit.push_back(OrLHS);
3726     if (const MachineInstr *Or = getOpcodeDef(TargetOpcode::G_OR, OrRHS, MRI))
3727       Ors.push_back(Or);
3728     else
3729       RegsToVisit.push_back(OrRHS);
3730   }
3731 
3732   // We're going to try and merge each register into a wider power-of-2 type,
3733   // so we ought to have an even number of registers.
3734   if (RegsToVisit.empty() || RegsToVisit.size() % 2 != 0)
3735     return std::nullopt;
3736   return RegsToVisit;
3737 }
3738 
3739 /// Helper function for findLoadOffsetsForLoadOrCombine.
3740 ///
3741 /// Check if \p Reg is the result of loading a \p MemSizeInBits wide value,
3742 /// and then moving that value into a specific byte offset.
3743 ///
3744 /// e.g. x[i] << 24
3745 ///
3746 /// \returns The load instruction and the byte offset it is moved into.
3747 static std::optional<std::pair<GZExtLoad *, int64_t>>
matchLoadAndBytePosition(Register Reg,unsigned MemSizeInBits,const MachineRegisterInfo & MRI)3748 matchLoadAndBytePosition(Register Reg, unsigned MemSizeInBits,
3749                          const MachineRegisterInfo &MRI) {
3750   assert(MRI.hasOneNonDBGUse(Reg) &&
3751          "Expected Reg to only have one non-debug use?");
3752   Register MaybeLoad;
3753   int64_t Shift;
3754   if (!mi_match(Reg, MRI,
3755                 m_OneNonDBGUse(m_GShl(m_Reg(MaybeLoad), m_ICst(Shift))))) {
3756     Shift = 0;
3757     MaybeLoad = Reg;
3758   }
3759 
3760   if (Shift % MemSizeInBits != 0)
3761     return std::nullopt;
3762 
3763   // TODO: Handle other types of loads.
3764   auto *Load = getOpcodeDef<GZExtLoad>(MaybeLoad, MRI);
3765   if (!Load)
3766     return std::nullopt;
3767 
3768   if (!Load->isUnordered() || Load->getMemSizeInBits() != MemSizeInBits)
3769     return std::nullopt;
3770 
3771   return std::make_pair(Load, Shift / MemSizeInBits);
3772 }
3773 
3774 std::optional<std::tuple<GZExtLoad *, int64_t, GZExtLoad *>>
findLoadOffsetsForLoadOrCombine(SmallDenseMap<int64_t,int64_t,8> & MemOffset2Idx,const SmallVector<Register,8> & RegsToVisit,const unsigned MemSizeInBits)3775 CombinerHelper::findLoadOffsetsForLoadOrCombine(
3776     SmallDenseMap<int64_t, int64_t, 8> &MemOffset2Idx,
3777     const SmallVector<Register, 8> &RegsToVisit, const unsigned MemSizeInBits) {
3778 
3779   // Each load found for the pattern. There should be one for each RegsToVisit.
3780   SmallSetVector<const MachineInstr *, 8> Loads;
3781 
3782   // The lowest index used in any load. (The lowest "i" for each x[i].)
3783   int64_t LowestIdx = INT64_MAX;
3784 
3785   // The load which uses the lowest index.
3786   GZExtLoad *LowestIdxLoad = nullptr;
3787 
3788   // Keeps track of the load indices we see. We shouldn't see any indices twice.
3789   SmallSet<int64_t, 8> SeenIdx;
3790 
3791   // Ensure each load is in the same MBB.
3792   // TODO: Support multiple MachineBasicBlocks.
3793   MachineBasicBlock *MBB = nullptr;
3794   const MachineMemOperand *MMO = nullptr;
3795 
3796   // Earliest instruction-order load in the pattern.
3797   GZExtLoad *EarliestLoad = nullptr;
3798 
3799   // Latest instruction-order load in the pattern.
3800   GZExtLoad *LatestLoad = nullptr;
3801 
3802   // Base pointer which every load should share.
3803   Register BasePtr;
3804 
3805   // We want to find a load for each register. Each load should have some
3806   // appropriate bit twiddling arithmetic. During this loop, we will also keep
3807   // track of the load which uses the lowest index. Later, we will check if we
3808   // can use its pointer in the final, combined load.
3809   for (auto Reg : RegsToVisit) {
3810     // Find the load, and find the position that it will end up in (e.g. a
3811     // shifted) value.
3812     auto LoadAndPos = matchLoadAndBytePosition(Reg, MemSizeInBits, MRI);
3813     if (!LoadAndPos)
3814       return std::nullopt;
3815     GZExtLoad *Load;
3816     int64_t DstPos;
3817     std::tie(Load, DstPos) = *LoadAndPos;
3818 
3819     // TODO: Handle multiple MachineBasicBlocks. Currently not handled because
3820     // it is difficult to check for stores/calls/etc between loads.
3821     MachineBasicBlock *LoadMBB = Load->getParent();
3822     if (!MBB)
3823       MBB = LoadMBB;
3824     if (LoadMBB != MBB)
3825       return std::nullopt;
3826 
3827     // Make sure that the MachineMemOperands of every seen load are compatible.
3828     auto &LoadMMO = Load->getMMO();
3829     if (!MMO)
3830       MMO = &LoadMMO;
3831     if (MMO->getAddrSpace() != LoadMMO.getAddrSpace())
3832       return std::nullopt;
3833 
3834     // Find out what the base pointer and index for the load is.
3835     Register LoadPtr;
3836     int64_t Idx;
3837     if (!mi_match(Load->getOperand(1).getReg(), MRI,
3838                   m_GPtrAdd(m_Reg(LoadPtr), m_ICst(Idx)))) {
3839       LoadPtr = Load->getOperand(1).getReg();
3840       Idx = 0;
3841     }
3842 
3843     // Don't combine things like a[i], a[i] -> a bigger load.
3844     if (!SeenIdx.insert(Idx).second)
3845       return std::nullopt;
3846 
3847     // Every load must share the same base pointer; don't combine things like:
3848     //
3849     // a[i], b[i + 1] -> a bigger load.
3850     if (!BasePtr.isValid())
3851       BasePtr = LoadPtr;
3852     if (BasePtr != LoadPtr)
3853       return std::nullopt;
3854 
3855     if (Idx < LowestIdx) {
3856       LowestIdx = Idx;
3857       LowestIdxLoad = Load;
3858     }
3859 
3860     // Keep track of the byte offset that this load ends up at. If we have seen
3861     // the byte offset, then stop here. We do not want to combine:
3862     //
3863     // a[i] << 16, a[i + k] << 16 -> a bigger load.
3864     if (!MemOffset2Idx.try_emplace(DstPos, Idx).second)
3865       return std::nullopt;
3866     Loads.insert(Load);
3867 
3868     // Keep track of the position of the earliest/latest loads in the pattern.
3869     // We will check that there are no load fold barriers between them later
3870     // on.
3871     //
3872     // FIXME: Is there a better way to check for load fold barriers?
3873     if (!EarliestLoad || dominates(*Load, *EarliestLoad))
3874       EarliestLoad = Load;
3875     if (!LatestLoad || dominates(*LatestLoad, *Load))
3876       LatestLoad = Load;
3877   }
3878 
3879   // We found a load for each register. Let's check if each load satisfies the
3880   // pattern.
3881   assert(Loads.size() == RegsToVisit.size() &&
3882          "Expected to find a load for each register?");
3883   assert(EarliestLoad != LatestLoad && EarliestLoad &&
3884          LatestLoad && "Expected at least two loads?");
3885 
3886   // Check if there are any stores, calls, etc. between any of the loads. If
3887   // there are, then we can't safely perform the combine.
3888   //
3889   // MaxIter is chosen based off the (worst case) number of iterations it
3890   // typically takes to succeed in the LLVM test suite plus some padding.
3891   //
3892   // FIXME: Is there a better way to check for load fold barriers?
3893   const unsigned MaxIter = 20;
3894   unsigned Iter = 0;
3895   for (const auto &MI : instructionsWithoutDebug(EarliestLoad->getIterator(),
3896                                                  LatestLoad->getIterator())) {
3897     if (Loads.count(&MI))
3898       continue;
3899     if (MI.isLoadFoldBarrier())
3900       return std::nullopt;
3901     if (Iter++ == MaxIter)
3902       return std::nullopt;
3903   }
3904 
3905   return std::make_tuple(LowestIdxLoad, LowestIdx, LatestLoad);
3906 }
3907 
matchLoadOrCombine(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)3908 bool CombinerHelper::matchLoadOrCombine(
3909     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
3910   assert(MI.getOpcode() == TargetOpcode::G_OR);
3911   MachineFunction &MF = *MI.getMF();
3912   // Assuming a little-endian target, transform:
3913   //  s8 *a = ...
3914   //  s32 val = a[0] | (a[1] << 8) | (a[2] << 16) | (a[3] << 24)
3915   // =>
3916   //  s32 val = *((i32)a)
3917   //
3918   //  s8 *a = ...
3919   //  s32 val = (a[0] << 24) | (a[1] << 16) | (a[2] << 8) | a[3]
3920   // =>
3921   //  s32 val = BSWAP(*((s32)a))
3922   Register Dst = MI.getOperand(0).getReg();
3923   LLT Ty = MRI.getType(Dst);
3924   if (Ty.isVector())
3925     return false;
3926 
3927   // We need to combine at least two loads into this type. Since the smallest
3928   // possible load is into a byte, we need at least a 16-bit wide type.
3929   const unsigned WideMemSizeInBits = Ty.getSizeInBits();
3930   if (WideMemSizeInBits < 16 || WideMemSizeInBits % 8 != 0)
3931     return false;
3932 
3933   // Match a collection of non-OR instructions in the pattern.
3934   auto RegsToVisit = findCandidatesForLoadOrCombine(&MI);
3935   if (!RegsToVisit)
3936     return false;
3937 
3938   // We have a collection of non-OR instructions. Figure out how wide each of
3939   // the small loads should be based off of the number of potential loads we
3940   // found.
3941   const unsigned NarrowMemSizeInBits = WideMemSizeInBits / RegsToVisit->size();
3942   if (NarrowMemSizeInBits % 8 != 0)
3943     return false;
3944 
3945   // Check if each register feeding into each OR is a load from the same
3946   // base pointer + some arithmetic.
3947   //
3948   // e.g. a[0], a[1] << 8, a[2] << 16, etc.
3949   //
3950   // Also verify that each of these ends up putting a[i] into the same memory
3951   // offset as a load into a wide type would.
3952   SmallDenseMap<int64_t, int64_t, 8> MemOffset2Idx;
3953   GZExtLoad *LowestIdxLoad, *LatestLoad;
3954   int64_t LowestIdx;
3955   auto MaybeLoadInfo = findLoadOffsetsForLoadOrCombine(
3956       MemOffset2Idx, *RegsToVisit, NarrowMemSizeInBits);
3957   if (!MaybeLoadInfo)
3958     return false;
3959   std::tie(LowestIdxLoad, LowestIdx, LatestLoad) = *MaybeLoadInfo;
3960 
3961   // We have a bunch of loads being OR'd together. Using the addresses + offsets
3962   // we found before, check if this corresponds to a big or little endian byte
3963   // pattern. If it does, then we can represent it using a load + possibly a
3964   // BSWAP.
3965   bool IsBigEndianTarget = MF.getDataLayout().isBigEndian();
3966   std::optional<bool> IsBigEndian = isBigEndian(MemOffset2Idx, LowestIdx);
3967   if (!IsBigEndian)
3968     return false;
3969   bool NeedsBSwap = IsBigEndianTarget != *IsBigEndian;
3970   if (NeedsBSwap && !isLegalOrBeforeLegalizer({TargetOpcode::G_BSWAP, {Ty}}))
3971     return false;
3972 
3973   // Make sure that the load from the lowest index produces offset 0 in the
3974   // final value.
3975   //
3976   // This ensures that we won't combine something like this:
3977   //
3978   // load x[i] -> byte 2
3979   // load x[i+1] -> byte 0 ---> wide_load x[i]
3980   // load x[i+2] -> byte 1
3981   const unsigned NumLoadsInTy = WideMemSizeInBits / NarrowMemSizeInBits;
3982   const unsigned ZeroByteOffset =
3983       *IsBigEndian
3984           ? bigEndianByteAt(NumLoadsInTy, 0)
3985           : littleEndianByteAt(NumLoadsInTy, 0);
3986   auto ZeroOffsetIdx = MemOffset2Idx.find(ZeroByteOffset);
3987   if (ZeroOffsetIdx == MemOffset2Idx.end() ||
3988       ZeroOffsetIdx->second != LowestIdx)
3989     return false;
3990 
3991   // We wil reuse the pointer from the load which ends up at byte offset 0. It
3992   // may not use index 0.
3993   Register Ptr = LowestIdxLoad->getPointerReg();
3994   const MachineMemOperand &MMO = LowestIdxLoad->getMMO();
3995   LegalityQuery::MemDesc MMDesc(MMO);
3996   MMDesc.MemoryTy = Ty;
3997   if (!isLegalOrBeforeLegalizer(
3998           {TargetOpcode::G_LOAD, {Ty, MRI.getType(Ptr)}, {MMDesc}}))
3999     return false;
4000   auto PtrInfo = MMO.getPointerInfo();
4001   auto *NewMMO = MF.getMachineMemOperand(&MMO, PtrInfo, WideMemSizeInBits / 8);
4002 
4003   // Load must be allowed and fast on the target.
4004   LLVMContext &C = MF.getFunction().getContext();
4005   auto &DL = MF.getDataLayout();
4006   unsigned Fast = 0;
4007   if (!getTargetLowering().allowsMemoryAccess(C, DL, Ty, *NewMMO, &Fast) ||
4008       !Fast)
4009     return false;
4010 
4011   MatchInfo = [=](MachineIRBuilder &MIB) {
4012     MIB.setInstrAndDebugLoc(*LatestLoad);
4013     Register LoadDst = NeedsBSwap ? MRI.cloneVirtualRegister(Dst) : Dst;
4014     MIB.buildLoad(LoadDst, Ptr, *NewMMO);
4015     if (NeedsBSwap)
4016       MIB.buildBSwap(Dst, LoadDst);
4017   };
4018   return true;
4019 }
4020 
matchExtendThroughPhis(MachineInstr & MI,MachineInstr * & ExtMI)4021 bool CombinerHelper::matchExtendThroughPhis(MachineInstr &MI,
4022                                             MachineInstr *&ExtMI) {
4023   auto &PHI = cast<GPhi>(MI);
4024   Register DstReg = PHI.getReg(0);
4025 
4026   // TODO: Extending a vector may be expensive, don't do this until heuristics
4027   // are better.
4028   if (MRI.getType(DstReg).isVector())
4029     return false;
4030 
4031   // Try to match a phi, whose only use is an extend.
4032   if (!MRI.hasOneNonDBGUse(DstReg))
4033     return false;
4034   ExtMI = &*MRI.use_instr_nodbg_begin(DstReg);
4035   switch (ExtMI->getOpcode()) {
4036   case TargetOpcode::G_ANYEXT:
4037     return true; // G_ANYEXT is usually free.
4038   case TargetOpcode::G_ZEXT:
4039   case TargetOpcode::G_SEXT:
4040     break;
4041   default:
4042     return false;
4043   }
4044 
4045   // If the target is likely to fold this extend away, don't propagate.
4046   if (Builder.getTII().isExtendLikelyToBeFolded(*ExtMI, MRI))
4047     return false;
4048 
4049   // We don't want to propagate the extends unless there's a good chance that
4050   // they'll be optimized in some way.
4051   // Collect the unique incoming values.
4052   SmallPtrSet<MachineInstr *, 4> InSrcs;
4053   for (unsigned I = 0; I < PHI.getNumIncomingValues(); ++I) {
4054     auto *DefMI = getDefIgnoringCopies(PHI.getIncomingValue(I), MRI);
4055     switch (DefMI->getOpcode()) {
4056     case TargetOpcode::G_LOAD:
4057     case TargetOpcode::G_TRUNC:
4058     case TargetOpcode::G_SEXT:
4059     case TargetOpcode::G_ZEXT:
4060     case TargetOpcode::G_ANYEXT:
4061     case TargetOpcode::G_CONSTANT:
4062       InSrcs.insert(DefMI);
4063       // Don't try to propagate if there are too many places to create new
4064       // extends, chances are it'll increase code size.
4065       if (InSrcs.size() > 2)
4066         return false;
4067       break;
4068     default:
4069       return false;
4070     }
4071   }
4072   return true;
4073 }
4074 
applyExtendThroughPhis(MachineInstr & MI,MachineInstr * & ExtMI)4075 void CombinerHelper::applyExtendThroughPhis(MachineInstr &MI,
4076                                             MachineInstr *&ExtMI) {
4077   auto &PHI = cast<GPhi>(MI);
4078   Register DstReg = ExtMI->getOperand(0).getReg();
4079   LLT ExtTy = MRI.getType(DstReg);
4080 
4081   // Propagate the extension into the block of each incoming reg's block.
4082   // Use a SetVector here because PHIs can have duplicate edges, and we want
4083   // deterministic iteration order.
4084   SmallSetVector<MachineInstr *, 8> SrcMIs;
4085   SmallDenseMap<MachineInstr *, MachineInstr *, 8> OldToNewSrcMap;
4086   for (unsigned I = 0; I < PHI.getNumIncomingValues(); ++I) {
4087     auto SrcReg = PHI.getIncomingValue(I);
4088     auto *SrcMI = MRI.getVRegDef(SrcReg);
4089     if (!SrcMIs.insert(SrcMI))
4090       continue;
4091 
4092     // Build an extend after each src inst.
4093     auto *MBB = SrcMI->getParent();
4094     MachineBasicBlock::iterator InsertPt = ++SrcMI->getIterator();
4095     if (InsertPt != MBB->end() && InsertPt->isPHI())
4096       InsertPt = MBB->getFirstNonPHI();
4097 
4098     Builder.setInsertPt(*SrcMI->getParent(), InsertPt);
4099     Builder.setDebugLoc(MI.getDebugLoc());
4100     auto NewExt = Builder.buildExtOrTrunc(ExtMI->getOpcode(), ExtTy, SrcReg);
4101     OldToNewSrcMap[SrcMI] = NewExt;
4102   }
4103 
4104   // Create a new phi with the extended inputs.
4105   Builder.setInstrAndDebugLoc(MI);
4106   auto NewPhi = Builder.buildInstrNoInsert(TargetOpcode::G_PHI);
4107   NewPhi.addDef(DstReg);
4108   for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) {
4109     if (!MO.isReg()) {
4110       NewPhi.addMBB(MO.getMBB());
4111       continue;
4112     }
4113     auto *NewSrc = OldToNewSrcMap[MRI.getVRegDef(MO.getReg())];
4114     NewPhi.addUse(NewSrc->getOperand(0).getReg());
4115   }
4116   Builder.insertInstr(NewPhi);
4117   ExtMI->eraseFromParent();
4118 }
4119 
matchExtractVecEltBuildVec(MachineInstr & MI,Register & Reg)4120 bool CombinerHelper::matchExtractVecEltBuildVec(MachineInstr &MI,
4121                                                 Register &Reg) {
4122   assert(MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT);
4123   // If we have a constant index, look for a G_BUILD_VECTOR source
4124   // and find the source register that the index maps to.
4125   Register SrcVec = MI.getOperand(1).getReg();
4126   LLT SrcTy = MRI.getType(SrcVec);
4127 
4128   auto Cst = getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
4129   if (!Cst || Cst->Value.getZExtValue() >= SrcTy.getNumElements())
4130     return false;
4131 
4132   unsigned VecIdx = Cst->Value.getZExtValue();
4133 
4134   // Check if we have a build_vector or build_vector_trunc with an optional
4135   // trunc in front.
4136   MachineInstr *SrcVecMI = MRI.getVRegDef(SrcVec);
4137   if (SrcVecMI->getOpcode() == TargetOpcode::G_TRUNC) {
4138     SrcVecMI = MRI.getVRegDef(SrcVecMI->getOperand(1).getReg());
4139   }
4140 
4141   if (SrcVecMI->getOpcode() != TargetOpcode::G_BUILD_VECTOR &&
4142       SrcVecMI->getOpcode() != TargetOpcode::G_BUILD_VECTOR_TRUNC)
4143     return false;
4144 
4145   EVT Ty(getMVTForLLT(SrcTy));
4146   if (!MRI.hasOneNonDBGUse(SrcVec) &&
4147       !getTargetLowering().aggressivelyPreferBuildVectorSources(Ty))
4148     return false;
4149 
4150   Reg = SrcVecMI->getOperand(VecIdx + 1).getReg();
4151   return true;
4152 }
4153 
applyExtractVecEltBuildVec(MachineInstr & MI,Register & Reg)4154 void CombinerHelper::applyExtractVecEltBuildVec(MachineInstr &MI,
4155                                                 Register &Reg) {
4156   // Check the type of the register, since it may have come from a
4157   // G_BUILD_VECTOR_TRUNC.
4158   LLT ScalarTy = MRI.getType(Reg);
4159   Register DstReg = MI.getOperand(0).getReg();
4160   LLT DstTy = MRI.getType(DstReg);
4161 
4162   if (ScalarTy != DstTy) {
4163     assert(ScalarTy.getSizeInBits() > DstTy.getSizeInBits());
4164     Builder.buildTrunc(DstReg, Reg);
4165     MI.eraseFromParent();
4166     return;
4167   }
4168   replaceSingleDefInstWithReg(MI, Reg);
4169 }
4170 
matchExtractAllEltsFromBuildVector(MachineInstr & MI,SmallVectorImpl<std::pair<Register,MachineInstr * >> & SrcDstPairs)4171 bool CombinerHelper::matchExtractAllEltsFromBuildVector(
4172     MachineInstr &MI,
4173     SmallVectorImpl<std::pair<Register, MachineInstr *>> &SrcDstPairs) {
4174   assert(MI.getOpcode() == TargetOpcode::G_BUILD_VECTOR);
4175   // This combine tries to find build_vector's which have every source element
4176   // extracted using G_EXTRACT_VECTOR_ELT. This can happen when transforms like
4177   // the masked load scalarization is run late in the pipeline. There's already
4178   // a combine for a similar pattern starting from the extract, but that
4179   // doesn't attempt to do it if there are multiple uses of the build_vector,
4180   // which in this case is true. Starting the combine from the build_vector
4181   // feels more natural than trying to find sibling nodes of extracts.
4182   // E.g.
4183   //  %vec(<4 x s32>) = G_BUILD_VECTOR %s1(s32), %s2, %s3, %s4
4184   //  %ext1 = G_EXTRACT_VECTOR_ELT %vec, 0
4185   //  %ext2 = G_EXTRACT_VECTOR_ELT %vec, 1
4186   //  %ext3 = G_EXTRACT_VECTOR_ELT %vec, 2
4187   //  %ext4 = G_EXTRACT_VECTOR_ELT %vec, 3
4188   // ==>
4189   // replace ext{1,2,3,4} with %s{1,2,3,4}
4190 
4191   Register DstReg = MI.getOperand(0).getReg();
4192   LLT DstTy = MRI.getType(DstReg);
4193   unsigned NumElts = DstTy.getNumElements();
4194 
4195   SmallBitVector ExtractedElts(NumElts);
4196   for (MachineInstr &II : MRI.use_nodbg_instructions(DstReg)) {
4197     if (II.getOpcode() != TargetOpcode::G_EXTRACT_VECTOR_ELT)
4198       return false;
4199     auto Cst = getIConstantVRegVal(II.getOperand(2).getReg(), MRI);
4200     if (!Cst)
4201       return false;
4202     unsigned Idx = Cst->getZExtValue();
4203     if (Idx >= NumElts)
4204       return false; // Out of range.
4205     ExtractedElts.set(Idx);
4206     SrcDstPairs.emplace_back(
4207         std::make_pair(MI.getOperand(Idx + 1).getReg(), &II));
4208   }
4209   // Match if every element was extracted.
4210   return ExtractedElts.all();
4211 }
4212 
applyExtractAllEltsFromBuildVector(MachineInstr & MI,SmallVectorImpl<std::pair<Register,MachineInstr * >> & SrcDstPairs)4213 void CombinerHelper::applyExtractAllEltsFromBuildVector(
4214     MachineInstr &MI,
4215     SmallVectorImpl<std::pair<Register, MachineInstr *>> &SrcDstPairs) {
4216   assert(MI.getOpcode() == TargetOpcode::G_BUILD_VECTOR);
4217   for (auto &Pair : SrcDstPairs) {
4218     auto *ExtMI = Pair.second;
4219     replaceRegWith(MRI, ExtMI->getOperand(0).getReg(), Pair.first);
4220     ExtMI->eraseFromParent();
4221   }
4222   MI.eraseFromParent();
4223 }
4224 
applyBuildFn(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4225 void CombinerHelper::applyBuildFn(
4226     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4227   applyBuildFnNoErase(MI, MatchInfo);
4228   MI.eraseFromParent();
4229 }
4230 
applyBuildFnNoErase(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4231 void CombinerHelper::applyBuildFnNoErase(
4232     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4233   MatchInfo(Builder);
4234 }
4235 
matchOrShiftToFunnelShift(MachineInstr & MI,BuildFnTy & MatchInfo)4236 bool CombinerHelper::matchOrShiftToFunnelShift(MachineInstr &MI,
4237                                                BuildFnTy &MatchInfo) {
4238   assert(MI.getOpcode() == TargetOpcode::G_OR);
4239 
4240   Register Dst = MI.getOperand(0).getReg();
4241   LLT Ty = MRI.getType(Dst);
4242   unsigned BitWidth = Ty.getScalarSizeInBits();
4243 
4244   Register ShlSrc, ShlAmt, LShrSrc, LShrAmt, Amt;
4245   unsigned FshOpc = 0;
4246 
4247   // Match (or (shl ...), (lshr ...)).
4248   if (!mi_match(Dst, MRI,
4249                 // m_GOr() handles the commuted version as well.
4250                 m_GOr(m_GShl(m_Reg(ShlSrc), m_Reg(ShlAmt)),
4251                       m_GLShr(m_Reg(LShrSrc), m_Reg(LShrAmt)))))
4252     return false;
4253 
4254   // Given constants C0 and C1 such that C0 + C1 is bit-width:
4255   // (or (shl x, C0), (lshr y, C1)) -> (fshl x, y, C0) or (fshr x, y, C1)
4256   int64_t CstShlAmt, CstLShrAmt;
4257   if (mi_match(ShlAmt, MRI, m_ICstOrSplat(CstShlAmt)) &&
4258       mi_match(LShrAmt, MRI, m_ICstOrSplat(CstLShrAmt)) &&
4259       CstShlAmt + CstLShrAmt == BitWidth) {
4260     FshOpc = TargetOpcode::G_FSHR;
4261     Amt = LShrAmt;
4262 
4263   } else if (mi_match(LShrAmt, MRI,
4264                       m_GSub(m_SpecificICstOrSplat(BitWidth), m_Reg(Amt))) &&
4265              ShlAmt == Amt) {
4266     // (or (shl x, amt), (lshr y, (sub bw, amt))) -> (fshl x, y, amt)
4267     FshOpc = TargetOpcode::G_FSHL;
4268 
4269   } else if (mi_match(ShlAmt, MRI,
4270                       m_GSub(m_SpecificICstOrSplat(BitWidth), m_Reg(Amt))) &&
4271              LShrAmt == Amt) {
4272     // (or (shl x, (sub bw, amt)), (lshr y, amt)) -> (fshr x, y, amt)
4273     FshOpc = TargetOpcode::G_FSHR;
4274 
4275   } else {
4276     return false;
4277   }
4278 
4279   LLT AmtTy = MRI.getType(Amt);
4280   if (!isLegalOrBeforeLegalizer({FshOpc, {Ty, AmtTy}}))
4281     return false;
4282 
4283   MatchInfo = [=](MachineIRBuilder &B) {
4284     B.buildInstr(FshOpc, {Dst}, {ShlSrc, LShrSrc, Amt});
4285   };
4286   return true;
4287 }
4288 
4289 /// Match an FSHL or FSHR that can be combined to a ROTR or ROTL rotate.
matchFunnelShiftToRotate(MachineInstr & MI)4290 bool CombinerHelper::matchFunnelShiftToRotate(MachineInstr &MI) {
4291   unsigned Opc = MI.getOpcode();
4292   assert(Opc == TargetOpcode::G_FSHL || Opc == TargetOpcode::G_FSHR);
4293   Register X = MI.getOperand(1).getReg();
4294   Register Y = MI.getOperand(2).getReg();
4295   if (X != Y)
4296     return false;
4297   unsigned RotateOpc =
4298       Opc == TargetOpcode::G_FSHL ? TargetOpcode::G_ROTL : TargetOpcode::G_ROTR;
4299   return isLegalOrBeforeLegalizer({RotateOpc, {MRI.getType(X), MRI.getType(Y)}});
4300 }
4301 
applyFunnelShiftToRotate(MachineInstr & MI)4302 void CombinerHelper::applyFunnelShiftToRotate(MachineInstr &MI) {
4303   unsigned Opc = MI.getOpcode();
4304   assert(Opc == TargetOpcode::G_FSHL || Opc == TargetOpcode::G_FSHR);
4305   bool IsFSHL = Opc == TargetOpcode::G_FSHL;
4306   Observer.changingInstr(MI);
4307   MI.setDesc(Builder.getTII().get(IsFSHL ? TargetOpcode::G_ROTL
4308                                          : TargetOpcode::G_ROTR));
4309   MI.removeOperand(2);
4310   Observer.changedInstr(MI);
4311 }
4312 
4313 // Fold (rot x, c) -> (rot x, c % BitSize)
matchRotateOutOfRange(MachineInstr & MI)4314 bool CombinerHelper::matchRotateOutOfRange(MachineInstr &MI) {
4315   assert(MI.getOpcode() == TargetOpcode::G_ROTL ||
4316          MI.getOpcode() == TargetOpcode::G_ROTR);
4317   unsigned Bitsize =
4318       MRI.getType(MI.getOperand(0).getReg()).getScalarSizeInBits();
4319   Register AmtReg = MI.getOperand(2).getReg();
4320   bool OutOfRange = false;
4321   auto MatchOutOfRange = [Bitsize, &OutOfRange](const Constant *C) {
4322     if (auto *CI = dyn_cast<ConstantInt>(C))
4323       OutOfRange |= CI->getValue().uge(Bitsize);
4324     return true;
4325   };
4326   return matchUnaryPredicate(MRI, AmtReg, MatchOutOfRange) && OutOfRange;
4327 }
4328 
applyRotateOutOfRange(MachineInstr & MI)4329 void CombinerHelper::applyRotateOutOfRange(MachineInstr &MI) {
4330   assert(MI.getOpcode() == TargetOpcode::G_ROTL ||
4331          MI.getOpcode() == TargetOpcode::G_ROTR);
4332   unsigned Bitsize =
4333       MRI.getType(MI.getOperand(0).getReg()).getScalarSizeInBits();
4334   Register Amt = MI.getOperand(2).getReg();
4335   LLT AmtTy = MRI.getType(Amt);
4336   auto Bits = Builder.buildConstant(AmtTy, Bitsize);
4337   Amt = Builder.buildURem(AmtTy, MI.getOperand(2).getReg(), Bits).getReg(0);
4338   Observer.changingInstr(MI);
4339   MI.getOperand(2).setReg(Amt);
4340   Observer.changedInstr(MI);
4341 }
4342 
matchICmpToTrueFalseKnownBits(MachineInstr & MI,int64_t & MatchInfo)4343 bool CombinerHelper::matchICmpToTrueFalseKnownBits(MachineInstr &MI,
4344                                                    int64_t &MatchInfo) {
4345   assert(MI.getOpcode() == TargetOpcode::G_ICMP);
4346   auto Pred = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
4347 
4348   // We want to avoid calling KnownBits on the LHS if possible, as this combine
4349   // has no filter and runs on every G_ICMP instruction. We can avoid calling
4350   // KnownBits on the LHS in two cases:
4351   //
4352   //  - The RHS is unknown: Constants are always on RHS. If the RHS is unknown
4353   //  we cannot do any transforms so we can safely bail out early.
4354   //  - The RHS is zero: we don't need to know the LHS to do unsigned <0 and
4355   //  >=0.
4356   auto KnownRHS = KB->getKnownBits(MI.getOperand(3).getReg());
4357   if (KnownRHS.isUnknown())
4358     return false;
4359 
4360   std::optional<bool> KnownVal;
4361   if (KnownRHS.isZero()) {
4362     // ? uge 0 -> always true
4363     // ? ult 0 -> always false
4364     if (Pred == CmpInst::ICMP_UGE)
4365       KnownVal = true;
4366     else if (Pred == CmpInst::ICMP_ULT)
4367       KnownVal = false;
4368   }
4369 
4370   if (!KnownVal) {
4371     auto KnownLHS = KB->getKnownBits(MI.getOperand(2).getReg());
4372     switch (Pred) {
4373     default:
4374       llvm_unreachable("Unexpected G_ICMP predicate?");
4375     case CmpInst::ICMP_EQ:
4376       KnownVal = KnownBits::eq(KnownLHS, KnownRHS);
4377       break;
4378     case CmpInst::ICMP_NE:
4379       KnownVal = KnownBits::ne(KnownLHS, KnownRHS);
4380       break;
4381     case CmpInst::ICMP_SGE:
4382       KnownVal = KnownBits::sge(KnownLHS, KnownRHS);
4383       break;
4384     case CmpInst::ICMP_SGT:
4385       KnownVal = KnownBits::sgt(KnownLHS, KnownRHS);
4386       break;
4387     case CmpInst::ICMP_SLE:
4388       KnownVal = KnownBits::sle(KnownLHS, KnownRHS);
4389       break;
4390     case CmpInst::ICMP_SLT:
4391       KnownVal = KnownBits::slt(KnownLHS, KnownRHS);
4392       break;
4393     case CmpInst::ICMP_UGE:
4394       KnownVal = KnownBits::uge(KnownLHS, KnownRHS);
4395       break;
4396     case CmpInst::ICMP_UGT:
4397       KnownVal = KnownBits::ugt(KnownLHS, KnownRHS);
4398       break;
4399     case CmpInst::ICMP_ULE:
4400       KnownVal = KnownBits::ule(KnownLHS, KnownRHS);
4401       break;
4402     case CmpInst::ICMP_ULT:
4403       KnownVal = KnownBits::ult(KnownLHS, KnownRHS);
4404       break;
4405     }
4406   }
4407 
4408   if (!KnownVal)
4409     return false;
4410   MatchInfo =
4411       *KnownVal
4412           ? getICmpTrueVal(getTargetLowering(),
4413                            /*IsVector = */
4414                            MRI.getType(MI.getOperand(0).getReg()).isVector(),
4415                            /* IsFP = */ false)
4416           : 0;
4417   return true;
4418 }
4419 
matchICmpToLHSKnownBits(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4420 bool CombinerHelper::matchICmpToLHSKnownBits(
4421     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4422   assert(MI.getOpcode() == TargetOpcode::G_ICMP);
4423   // Given:
4424   //
4425   // %x = G_WHATEVER (... x is known to be 0 or 1 ...)
4426   // %cmp = G_ICMP ne %x, 0
4427   //
4428   // Or:
4429   //
4430   // %x = G_WHATEVER (... x is known to be 0 or 1 ...)
4431   // %cmp = G_ICMP eq %x, 1
4432   //
4433   // We can replace %cmp with %x assuming true is 1 on the target.
4434   auto Pred = static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
4435   if (!CmpInst::isEquality(Pred))
4436     return false;
4437   Register Dst = MI.getOperand(0).getReg();
4438   LLT DstTy = MRI.getType(Dst);
4439   if (getICmpTrueVal(getTargetLowering(), DstTy.isVector(),
4440                      /* IsFP = */ false) != 1)
4441     return false;
4442   int64_t OneOrZero = Pred == CmpInst::ICMP_EQ;
4443   if (!mi_match(MI.getOperand(3).getReg(), MRI, m_SpecificICst(OneOrZero)))
4444     return false;
4445   Register LHS = MI.getOperand(2).getReg();
4446   auto KnownLHS = KB->getKnownBits(LHS);
4447   if (KnownLHS.getMinValue() != 0 || KnownLHS.getMaxValue() != 1)
4448     return false;
4449   // Make sure replacing Dst with the LHS is a legal operation.
4450   LLT LHSTy = MRI.getType(LHS);
4451   unsigned LHSSize = LHSTy.getSizeInBits();
4452   unsigned DstSize = DstTy.getSizeInBits();
4453   unsigned Op = TargetOpcode::COPY;
4454   if (DstSize != LHSSize)
4455     Op = DstSize < LHSSize ? TargetOpcode::G_TRUNC : TargetOpcode::G_ZEXT;
4456   if (!isLegalOrBeforeLegalizer({Op, {DstTy, LHSTy}}))
4457     return false;
4458   MatchInfo = [=](MachineIRBuilder &B) { B.buildInstr(Op, {Dst}, {LHS}); };
4459   return true;
4460 }
4461 
4462 // Replace (and (or x, c1), c2) with (and x, c2) iff c1 & c2 == 0
matchAndOrDisjointMask(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4463 bool CombinerHelper::matchAndOrDisjointMask(
4464     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4465   assert(MI.getOpcode() == TargetOpcode::G_AND);
4466 
4467   // Ignore vector types to simplify matching the two constants.
4468   // TODO: do this for vectors and scalars via a demanded bits analysis.
4469   LLT Ty = MRI.getType(MI.getOperand(0).getReg());
4470   if (Ty.isVector())
4471     return false;
4472 
4473   Register Src;
4474   Register AndMaskReg;
4475   int64_t AndMaskBits;
4476   int64_t OrMaskBits;
4477   if (!mi_match(MI, MRI,
4478                 m_GAnd(m_GOr(m_Reg(Src), m_ICst(OrMaskBits)),
4479                        m_all_of(m_ICst(AndMaskBits), m_Reg(AndMaskReg)))))
4480     return false;
4481 
4482   // Check if OrMask could turn on any bits in Src.
4483   if (AndMaskBits & OrMaskBits)
4484     return false;
4485 
4486   MatchInfo = [=, &MI](MachineIRBuilder &B) {
4487     Observer.changingInstr(MI);
4488     // Canonicalize the result to have the constant on the RHS.
4489     if (MI.getOperand(1).getReg() == AndMaskReg)
4490       MI.getOperand(2).setReg(AndMaskReg);
4491     MI.getOperand(1).setReg(Src);
4492     Observer.changedInstr(MI);
4493   };
4494   return true;
4495 }
4496 
4497 /// Form a G_SBFX from a G_SEXT_INREG fed by a right shift.
matchBitfieldExtractFromSExtInReg(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4498 bool CombinerHelper::matchBitfieldExtractFromSExtInReg(
4499     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4500   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
4501   Register Dst = MI.getOperand(0).getReg();
4502   Register Src = MI.getOperand(1).getReg();
4503   LLT Ty = MRI.getType(Src);
4504   LLT ExtractTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
4505   if (!LI || !LI->isLegalOrCustom({TargetOpcode::G_SBFX, {Ty, ExtractTy}}))
4506     return false;
4507   int64_t Width = MI.getOperand(2).getImm();
4508   Register ShiftSrc;
4509   int64_t ShiftImm;
4510   if (!mi_match(
4511           Src, MRI,
4512           m_OneNonDBGUse(m_any_of(m_GAShr(m_Reg(ShiftSrc), m_ICst(ShiftImm)),
4513                                   m_GLShr(m_Reg(ShiftSrc), m_ICst(ShiftImm))))))
4514     return false;
4515   if (ShiftImm < 0 || ShiftImm + Width > Ty.getScalarSizeInBits())
4516     return false;
4517 
4518   MatchInfo = [=](MachineIRBuilder &B) {
4519     auto Cst1 = B.buildConstant(ExtractTy, ShiftImm);
4520     auto Cst2 = B.buildConstant(ExtractTy, Width);
4521     B.buildSbfx(Dst, ShiftSrc, Cst1, Cst2);
4522   };
4523   return true;
4524 }
4525 
4526 /// Form a G_UBFX from "(a srl b) & mask", where b and mask are constants.
matchBitfieldExtractFromAnd(MachineInstr & MI,BuildFnTy & MatchInfo)4527 bool CombinerHelper::matchBitfieldExtractFromAnd(MachineInstr &MI,
4528                                                  BuildFnTy &MatchInfo) {
4529   GAnd *And = cast<GAnd>(&MI);
4530   Register Dst = And->getReg(0);
4531   LLT Ty = MRI.getType(Dst);
4532   LLT ExtractTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
4533   // Note that isLegalOrBeforeLegalizer is stricter and does not take custom
4534   // into account.
4535   if (LI && !LI->isLegalOrCustom({TargetOpcode::G_UBFX, {Ty, ExtractTy}}))
4536     return false;
4537 
4538   int64_t AndImm, LSBImm;
4539   Register ShiftSrc;
4540   const unsigned Size = Ty.getScalarSizeInBits();
4541   if (!mi_match(And->getReg(0), MRI,
4542                 m_GAnd(m_OneNonDBGUse(m_GLShr(m_Reg(ShiftSrc), m_ICst(LSBImm))),
4543                        m_ICst(AndImm))))
4544     return false;
4545 
4546   // The mask is a mask of the low bits iff imm & (imm+1) == 0.
4547   auto MaybeMask = static_cast<uint64_t>(AndImm);
4548   if (MaybeMask & (MaybeMask + 1))
4549     return false;
4550 
4551   // LSB must fit within the register.
4552   if (static_cast<uint64_t>(LSBImm) >= Size)
4553     return false;
4554 
4555   uint64_t Width = APInt(Size, AndImm).countr_one();
4556   MatchInfo = [=](MachineIRBuilder &B) {
4557     auto WidthCst = B.buildConstant(ExtractTy, Width);
4558     auto LSBCst = B.buildConstant(ExtractTy, LSBImm);
4559     B.buildInstr(TargetOpcode::G_UBFX, {Dst}, {ShiftSrc, LSBCst, WidthCst});
4560   };
4561   return true;
4562 }
4563 
matchBitfieldExtractFromShr(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4564 bool CombinerHelper::matchBitfieldExtractFromShr(
4565     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4566   const unsigned Opcode = MI.getOpcode();
4567   assert(Opcode == TargetOpcode::G_ASHR || Opcode == TargetOpcode::G_LSHR);
4568 
4569   const Register Dst = MI.getOperand(0).getReg();
4570 
4571   const unsigned ExtrOpcode = Opcode == TargetOpcode::G_ASHR
4572                                   ? TargetOpcode::G_SBFX
4573                                   : TargetOpcode::G_UBFX;
4574 
4575   // Check if the type we would use for the extract is legal
4576   LLT Ty = MRI.getType(Dst);
4577   LLT ExtractTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
4578   if (!LI || !LI->isLegalOrCustom({ExtrOpcode, {Ty, ExtractTy}}))
4579     return false;
4580 
4581   Register ShlSrc;
4582   int64_t ShrAmt;
4583   int64_t ShlAmt;
4584   const unsigned Size = Ty.getScalarSizeInBits();
4585 
4586   // Try to match shr (shl x, c1), c2
4587   if (!mi_match(Dst, MRI,
4588                 m_BinOp(Opcode,
4589                         m_OneNonDBGUse(m_GShl(m_Reg(ShlSrc), m_ICst(ShlAmt))),
4590                         m_ICst(ShrAmt))))
4591     return false;
4592 
4593   // Make sure that the shift sizes can fit a bitfield extract
4594   if (ShlAmt < 0 || ShlAmt > ShrAmt || ShrAmt >= Size)
4595     return false;
4596 
4597   // Skip this combine if the G_SEXT_INREG combine could handle it
4598   if (Opcode == TargetOpcode::G_ASHR && ShlAmt == ShrAmt)
4599     return false;
4600 
4601   // Calculate start position and width of the extract
4602   const int64_t Pos = ShrAmt - ShlAmt;
4603   const int64_t Width = Size - ShrAmt;
4604 
4605   MatchInfo = [=](MachineIRBuilder &B) {
4606     auto WidthCst = B.buildConstant(ExtractTy, Width);
4607     auto PosCst = B.buildConstant(ExtractTy, Pos);
4608     B.buildInstr(ExtrOpcode, {Dst}, {ShlSrc, PosCst, WidthCst});
4609   };
4610   return true;
4611 }
4612 
matchBitfieldExtractFromShrAnd(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4613 bool CombinerHelper::matchBitfieldExtractFromShrAnd(
4614     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4615   const unsigned Opcode = MI.getOpcode();
4616   assert(Opcode == TargetOpcode::G_LSHR || Opcode == TargetOpcode::G_ASHR);
4617 
4618   const Register Dst = MI.getOperand(0).getReg();
4619   LLT Ty = MRI.getType(Dst);
4620   LLT ExtractTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
4621   if (LI && !LI->isLegalOrCustom({TargetOpcode::G_UBFX, {Ty, ExtractTy}}))
4622     return false;
4623 
4624   // Try to match shr (and x, c1), c2
4625   Register AndSrc;
4626   int64_t ShrAmt;
4627   int64_t SMask;
4628   if (!mi_match(Dst, MRI,
4629                 m_BinOp(Opcode,
4630                         m_OneNonDBGUse(m_GAnd(m_Reg(AndSrc), m_ICst(SMask))),
4631                         m_ICst(ShrAmt))))
4632     return false;
4633 
4634   const unsigned Size = Ty.getScalarSizeInBits();
4635   if (ShrAmt < 0 || ShrAmt >= Size)
4636     return false;
4637 
4638   // If the shift subsumes the mask, emit the 0 directly.
4639   if (0 == (SMask >> ShrAmt)) {
4640     MatchInfo = [=](MachineIRBuilder &B) {
4641       B.buildConstant(Dst, 0);
4642     };
4643     return true;
4644   }
4645 
4646   // Check that ubfx can do the extraction, with no holes in the mask.
4647   uint64_t UMask = SMask;
4648   UMask |= maskTrailingOnes<uint64_t>(ShrAmt);
4649   UMask &= maskTrailingOnes<uint64_t>(Size);
4650   if (!isMask_64(UMask))
4651     return false;
4652 
4653   // Calculate start position and width of the extract.
4654   const int64_t Pos = ShrAmt;
4655   const int64_t Width = llvm::countr_one(UMask) - ShrAmt;
4656 
4657   // It's preferable to keep the shift, rather than form G_SBFX.
4658   // TODO: remove the G_AND via demanded bits analysis.
4659   if (Opcode == TargetOpcode::G_ASHR && Width + ShrAmt == Size)
4660     return false;
4661 
4662   MatchInfo = [=](MachineIRBuilder &B) {
4663     auto WidthCst = B.buildConstant(ExtractTy, Width);
4664     auto PosCst = B.buildConstant(ExtractTy, Pos);
4665     B.buildInstr(TargetOpcode::G_UBFX, {Dst}, {AndSrc, PosCst, WidthCst});
4666   };
4667   return true;
4668 }
4669 
reassociationCanBreakAddressingModePattern(MachineInstr & MI)4670 bool CombinerHelper::reassociationCanBreakAddressingModePattern(
4671     MachineInstr &MI) {
4672   auto &PtrAdd = cast<GPtrAdd>(MI);
4673 
4674   Register Src1Reg = PtrAdd.getBaseReg();
4675   auto *Src1Def = getOpcodeDef<GPtrAdd>(Src1Reg, MRI);
4676   if (!Src1Def)
4677     return false;
4678 
4679   Register Src2Reg = PtrAdd.getOffsetReg();
4680 
4681   if (MRI.hasOneNonDBGUse(Src1Reg))
4682     return false;
4683 
4684   auto C1 = getIConstantVRegVal(Src1Def->getOffsetReg(), MRI);
4685   if (!C1)
4686     return false;
4687   auto C2 = getIConstantVRegVal(Src2Reg, MRI);
4688   if (!C2)
4689     return false;
4690 
4691   const APInt &C1APIntVal = *C1;
4692   const APInt &C2APIntVal = *C2;
4693   const int64_t CombinedValue = (C1APIntVal + C2APIntVal).getSExtValue();
4694 
4695   for (auto &UseMI : MRI.use_nodbg_instructions(PtrAdd.getReg(0))) {
4696     // This combine may end up running before ptrtoint/inttoptr combines
4697     // manage to eliminate redundant conversions, so try to look through them.
4698     MachineInstr *ConvUseMI = &UseMI;
4699     unsigned ConvUseOpc = ConvUseMI->getOpcode();
4700     while (ConvUseOpc == TargetOpcode::G_INTTOPTR ||
4701            ConvUseOpc == TargetOpcode::G_PTRTOINT) {
4702       Register DefReg = ConvUseMI->getOperand(0).getReg();
4703       if (!MRI.hasOneNonDBGUse(DefReg))
4704         break;
4705       ConvUseMI = &*MRI.use_instr_nodbg_begin(DefReg);
4706       ConvUseOpc = ConvUseMI->getOpcode();
4707     }
4708     auto *LdStMI = dyn_cast<GLoadStore>(ConvUseMI);
4709     if (!LdStMI)
4710       continue;
4711     // Is x[offset2] already not a legal addressing mode? If so then
4712     // reassociating the constants breaks nothing (we test offset2 because
4713     // that's the one we hope to fold into the load or store).
4714     TargetLoweringBase::AddrMode AM;
4715     AM.HasBaseReg = true;
4716     AM.BaseOffs = C2APIntVal.getSExtValue();
4717     unsigned AS = MRI.getType(LdStMI->getPointerReg()).getAddressSpace();
4718     Type *AccessTy = getTypeForLLT(LdStMI->getMMO().getMemoryType(),
4719                                    PtrAdd.getMF()->getFunction().getContext());
4720     const auto &TLI = *PtrAdd.getMF()->getSubtarget().getTargetLowering();
4721     if (!TLI.isLegalAddressingMode(PtrAdd.getMF()->getDataLayout(), AM,
4722                                    AccessTy, AS))
4723       continue;
4724 
4725     // Would x[offset1+offset2] still be a legal addressing mode?
4726     AM.BaseOffs = CombinedValue;
4727     if (!TLI.isLegalAddressingMode(PtrAdd.getMF()->getDataLayout(), AM,
4728                                    AccessTy, AS))
4729       return true;
4730   }
4731 
4732   return false;
4733 }
4734 
matchReassocConstantInnerRHS(GPtrAdd & MI,MachineInstr * RHS,BuildFnTy & MatchInfo)4735 bool CombinerHelper::matchReassocConstantInnerRHS(GPtrAdd &MI,
4736                                                   MachineInstr *RHS,
4737                                                   BuildFnTy &MatchInfo) {
4738   // G_PTR_ADD(BASE, G_ADD(X, C)) -> G_PTR_ADD(G_PTR_ADD(BASE, X), C)
4739   Register Src1Reg = MI.getOperand(1).getReg();
4740   if (RHS->getOpcode() != TargetOpcode::G_ADD)
4741     return false;
4742   auto C2 = getIConstantVRegVal(RHS->getOperand(2).getReg(), MRI);
4743   if (!C2)
4744     return false;
4745 
4746   MatchInfo = [=, &MI](MachineIRBuilder &B) {
4747     LLT PtrTy = MRI.getType(MI.getOperand(0).getReg());
4748 
4749     auto NewBase =
4750         Builder.buildPtrAdd(PtrTy, Src1Reg, RHS->getOperand(1).getReg());
4751     Observer.changingInstr(MI);
4752     MI.getOperand(1).setReg(NewBase.getReg(0));
4753     MI.getOperand(2).setReg(RHS->getOperand(2).getReg());
4754     Observer.changedInstr(MI);
4755   };
4756   return !reassociationCanBreakAddressingModePattern(MI);
4757 }
4758 
matchReassocConstantInnerLHS(GPtrAdd & MI,MachineInstr * LHS,MachineInstr * RHS,BuildFnTy & MatchInfo)4759 bool CombinerHelper::matchReassocConstantInnerLHS(GPtrAdd &MI,
4760                                                   MachineInstr *LHS,
4761                                                   MachineInstr *RHS,
4762                                                   BuildFnTy &MatchInfo) {
4763   // G_PTR_ADD (G_PTR_ADD X, C), Y) -> (G_PTR_ADD (G_PTR_ADD(X, Y), C)
4764   // if and only if (G_PTR_ADD X, C) has one use.
4765   Register LHSBase;
4766   std::optional<ValueAndVReg> LHSCstOff;
4767   if (!mi_match(MI.getBaseReg(), MRI,
4768                 m_OneNonDBGUse(m_GPtrAdd(m_Reg(LHSBase), m_GCst(LHSCstOff)))))
4769     return false;
4770 
4771   auto *LHSPtrAdd = cast<GPtrAdd>(LHS);
4772   MatchInfo = [=, &MI](MachineIRBuilder &B) {
4773     // When we change LHSPtrAdd's offset register we might cause it to use a reg
4774     // before its def. Sink the instruction so the outer PTR_ADD to ensure this
4775     // doesn't happen.
4776     LHSPtrAdd->moveBefore(&MI);
4777     Register RHSReg = MI.getOffsetReg();
4778     // set VReg will cause type mismatch if it comes from extend/trunc
4779     auto NewCst = B.buildConstant(MRI.getType(RHSReg), LHSCstOff->Value);
4780     Observer.changingInstr(MI);
4781     MI.getOperand(2).setReg(NewCst.getReg(0));
4782     Observer.changedInstr(MI);
4783     Observer.changingInstr(*LHSPtrAdd);
4784     LHSPtrAdd->getOperand(2).setReg(RHSReg);
4785     Observer.changedInstr(*LHSPtrAdd);
4786   };
4787   return !reassociationCanBreakAddressingModePattern(MI);
4788 }
4789 
matchReassocFoldConstantsInSubTree(GPtrAdd & MI,MachineInstr * LHS,MachineInstr * RHS,BuildFnTy & MatchInfo)4790 bool CombinerHelper::matchReassocFoldConstantsInSubTree(GPtrAdd &MI,
4791                                                         MachineInstr *LHS,
4792                                                         MachineInstr *RHS,
4793                                                         BuildFnTy &MatchInfo) {
4794   // G_PTR_ADD(G_PTR_ADD(BASE, C1), C2) -> G_PTR_ADD(BASE, C1+C2)
4795   auto *LHSPtrAdd = dyn_cast<GPtrAdd>(LHS);
4796   if (!LHSPtrAdd)
4797     return false;
4798 
4799   Register Src2Reg = MI.getOperand(2).getReg();
4800   Register LHSSrc1 = LHSPtrAdd->getBaseReg();
4801   Register LHSSrc2 = LHSPtrAdd->getOffsetReg();
4802   auto C1 = getIConstantVRegVal(LHSSrc2, MRI);
4803   if (!C1)
4804     return false;
4805   auto C2 = getIConstantVRegVal(Src2Reg, MRI);
4806   if (!C2)
4807     return false;
4808 
4809   MatchInfo = [=, &MI](MachineIRBuilder &B) {
4810     auto NewCst = B.buildConstant(MRI.getType(Src2Reg), *C1 + *C2);
4811     Observer.changingInstr(MI);
4812     MI.getOperand(1).setReg(LHSSrc1);
4813     MI.getOperand(2).setReg(NewCst.getReg(0));
4814     Observer.changedInstr(MI);
4815   };
4816   return !reassociationCanBreakAddressingModePattern(MI);
4817 }
4818 
matchReassocPtrAdd(MachineInstr & MI,BuildFnTy & MatchInfo)4819 bool CombinerHelper::matchReassocPtrAdd(MachineInstr &MI,
4820                                         BuildFnTy &MatchInfo) {
4821   auto &PtrAdd = cast<GPtrAdd>(MI);
4822   // We're trying to match a few pointer computation patterns here for
4823   // re-association opportunities.
4824   // 1) Isolating a constant operand to be on the RHS, e.g.:
4825   // G_PTR_ADD(BASE, G_ADD(X, C)) -> G_PTR_ADD(G_PTR_ADD(BASE, X), C)
4826   //
4827   // 2) Folding two constants in each sub-tree as long as such folding
4828   // doesn't break a legal addressing mode.
4829   // G_PTR_ADD(G_PTR_ADD(BASE, C1), C2) -> G_PTR_ADD(BASE, C1+C2)
4830   //
4831   // 3) Move a constant from the LHS of an inner op to the RHS of the outer.
4832   // G_PTR_ADD (G_PTR_ADD X, C), Y) -> G_PTR_ADD (G_PTR_ADD(X, Y), C)
4833   // iif (G_PTR_ADD X, C) has one use.
4834   MachineInstr *LHS = MRI.getVRegDef(PtrAdd.getBaseReg());
4835   MachineInstr *RHS = MRI.getVRegDef(PtrAdd.getOffsetReg());
4836 
4837   // Try to match example 2.
4838   if (matchReassocFoldConstantsInSubTree(PtrAdd, LHS, RHS, MatchInfo))
4839     return true;
4840 
4841   // Try to match example 3.
4842   if (matchReassocConstantInnerLHS(PtrAdd, LHS, RHS, MatchInfo))
4843     return true;
4844 
4845   // Try to match example 1.
4846   if (matchReassocConstantInnerRHS(PtrAdd, RHS, MatchInfo))
4847     return true;
4848 
4849   return false;
4850 }
tryReassocBinOp(unsigned Opc,Register DstReg,Register OpLHS,Register OpRHS,BuildFnTy & MatchInfo)4851 bool CombinerHelper::tryReassocBinOp(unsigned Opc, Register DstReg,
4852                                      Register OpLHS, Register OpRHS,
4853                                      BuildFnTy &MatchInfo) {
4854   LLT OpRHSTy = MRI.getType(OpRHS);
4855   MachineInstr *OpLHSDef = MRI.getVRegDef(OpLHS);
4856 
4857   if (OpLHSDef->getOpcode() != Opc)
4858     return false;
4859 
4860   MachineInstr *OpRHSDef = MRI.getVRegDef(OpRHS);
4861   Register OpLHSLHS = OpLHSDef->getOperand(1).getReg();
4862   Register OpLHSRHS = OpLHSDef->getOperand(2).getReg();
4863 
4864   // If the inner op is (X op C), pull the constant out so it can be folded with
4865   // other constants in the expression tree. Folding is not guaranteed so we
4866   // might have (C1 op C2). In that case do not pull a constant out because it
4867   // won't help and can lead to infinite loops.
4868   if (isConstantOrConstantSplatVector(*MRI.getVRegDef(OpLHSRHS), MRI) &&
4869       !isConstantOrConstantSplatVector(*MRI.getVRegDef(OpLHSLHS), MRI)) {
4870     if (isConstantOrConstantSplatVector(*OpRHSDef, MRI)) {
4871       // (Opc (Opc X, C1), C2) -> (Opc X, (Opc C1, C2))
4872       MatchInfo = [=](MachineIRBuilder &B) {
4873         auto NewCst = B.buildInstr(Opc, {OpRHSTy}, {OpLHSRHS, OpRHS});
4874         B.buildInstr(Opc, {DstReg}, {OpLHSLHS, NewCst});
4875       };
4876       return true;
4877     }
4878     if (getTargetLowering().isReassocProfitable(MRI, OpLHS, OpRHS)) {
4879       // Reassociate: (op (op x, c1), y) -> (op (op x, y), c1)
4880       //              iff (op x, c1) has one use
4881       MatchInfo = [=](MachineIRBuilder &B) {
4882         auto NewLHSLHS = B.buildInstr(Opc, {OpRHSTy}, {OpLHSLHS, OpRHS});
4883         B.buildInstr(Opc, {DstReg}, {NewLHSLHS, OpLHSRHS});
4884       };
4885       return true;
4886     }
4887   }
4888 
4889   return false;
4890 }
4891 
matchReassocCommBinOp(MachineInstr & MI,BuildFnTy & MatchInfo)4892 bool CombinerHelper::matchReassocCommBinOp(MachineInstr &MI,
4893                                            BuildFnTy &MatchInfo) {
4894   // We don't check if the reassociation will break a legal addressing mode
4895   // here since pointer arithmetic is handled by G_PTR_ADD.
4896   unsigned Opc = MI.getOpcode();
4897   Register DstReg = MI.getOperand(0).getReg();
4898   Register LHSReg = MI.getOperand(1).getReg();
4899   Register RHSReg = MI.getOperand(2).getReg();
4900 
4901   if (tryReassocBinOp(Opc, DstReg, LHSReg, RHSReg, MatchInfo))
4902     return true;
4903   if (tryReassocBinOp(Opc, DstReg, RHSReg, LHSReg, MatchInfo))
4904     return true;
4905   return false;
4906 }
4907 
matchConstantFoldCastOp(MachineInstr & MI,APInt & MatchInfo)4908 bool CombinerHelper::matchConstantFoldCastOp(MachineInstr &MI, APInt &MatchInfo) {
4909   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
4910   Register SrcOp = MI.getOperand(1).getReg();
4911 
4912   if (auto MaybeCst = ConstantFoldCastOp(MI.getOpcode(), DstTy, SrcOp, MRI)) {
4913     MatchInfo = *MaybeCst;
4914     return true;
4915   }
4916 
4917   return false;
4918 }
4919 
matchConstantFoldBinOp(MachineInstr & MI,APInt & MatchInfo)4920 bool CombinerHelper::matchConstantFoldBinOp(MachineInstr &MI, APInt &MatchInfo) {
4921   Register Op1 = MI.getOperand(1).getReg();
4922   Register Op2 = MI.getOperand(2).getReg();
4923   auto MaybeCst = ConstantFoldBinOp(MI.getOpcode(), Op1, Op2, MRI);
4924   if (!MaybeCst)
4925     return false;
4926   MatchInfo = *MaybeCst;
4927   return true;
4928 }
4929 
matchConstantFoldFPBinOp(MachineInstr & MI,ConstantFP * & MatchInfo)4930 bool CombinerHelper::matchConstantFoldFPBinOp(MachineInstr &MI, ConstantFP* &MatchInfo) {
4931   Register Op1 = MI.getOperand(1).getReg();
4932   Register Op2 = MI.getOperand(2).getReg();
4933   auto MaybeCst = ConstantFoldFPBinOp(MI.getOpcode(), Op1, Op2, MRI);
4934   if (!MaybeCst)
4935     return false;
4936   MatchInfo =
4937       ConstantFP::get(MI.getMF()->getFunction().getContext(), *MaybeCst);
4938   return true;
4939 }
4940 
matchConstantFoldFMA(MachineInstr & MI,ConstantFP * & MatchInfo)4941 bool CombinerHelper::matchConstantFoldFMA(MachineInstr &MI,
4942                                           ConstantFP *&MatchInfo) {
4943   assert(MI.getOpcode() == TargetOpcode::G_FMA ||
4944          MI.getOpcode() == TargetOpcode::G_FMAD);
4945   auto [_, Op1, Op2, Op3] = MI.getFirst4Regs();
4946 
4947   const ConstantFP *Op3Cst = getConstantFPVRegVal(Op3, MRI);
4948   if (!Op3Cst)
4949     return false;
4950 
4951   const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI);
4952   if (!Op2Cst)
4953     return false;
4954 
4955   const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI);
4956   if (!Op1Cst)
4957     return false;
4958 
4959   APFloat Op1F = Op1Cst->getValueAPF();
4960   Op1F.fusedMultiplyAdd(Op2Cst->getValueAPF(), Op3Cst->getValueAPF(),
4961                         APFloat::rmNearestTiesToEven);
4962   MatchInfo = ConstantFP::get(MI.getMF()->getFunction().getContext(), Op1F);
4963   return true;
4964 }
4965 
matchNarrowBinopFeedingAnd(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)4966 bool CombinerHelper::matchNarrowBinopFeedingAnd(
4967     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
4968   // Look for a binop feeding into an AND with a mask:
4969   //
4970   // %add = G_ADD %lhs, %rhs
4971   // %and = G_AND %add, 000...11111111
4972   //
4973   // Check if it's possible to perform the binop at a narrower width and zext
4974   // back to the original width like so:
4975   //
4976   // %narrow_lhs = G_TRUNC %lhs
4977   // %narrow_rhs = G_TRUNC %rhs
4978   // %narrow_add = G_ADD %narrow_lhs, %narrow_rhs
4979   // %new_add = G_ZEXT %narrow_add
4980   // %and = G_AND %new_add, 000...11111111
4981   //
4982   // This can allow later combines to eliminate the G_AND if it turns out
4983   // that the mask is irrelevant.
4984   assert(MI.getOpcode() == TargetOpcode::G_AND);
4985   Register Dst = MI.getOperand(0).getReg();
4986   Register AndLHS = MI.getOperand(1).getReg();
4987   Register AndRHS = MI.getOperand(2).getReg();
4988   LLT WideTy = MRI.getType(Dst);
4989 
4990   // If the potential binop has more than one use, then it's possible that one
4991   // of those uses will need its full width.
4992   if (!WideTy.isScalar() || !MRI.hasOneNonDBGUse(AndLHS))
4993     return false;
4994 
4995   // Check if the LHS feeding the AND is impacted by the high bits that we're
4996   // masking out.
4997   //
4998   // e.g. for 64-bit x, y:
4999   //
5000   // add_64(x, y) & 65535 == zext(add_16(trunc(x), trunc(y))) & 65535
5001   MachineInstr *LHSInst = getDefIgnoringCopies(AndLHS, MRI);
5002   if (!LHSInst)
5003     return false;
5004   unsigned LHSOpc = LHSInst->getOpcode();
5005   switch (LHSOpc) {
5006   default:
5007     return false;
5008   case TargetOpcode::G_ADD:
5009   case TargetOpcode::G_SUB:
5010   case TargetOpcode::G_MUL:
5011   case TargetOpcode::G_AND:
5012   case TargetOpcode::G_OR:
5013   case TargetOpcode::G_XOR:
5014     break;
5015   }
5016 
5017   // Find the mask on the RHS.
5018   auto Cst = getIConstantVRegValWithLookThrough(AndRHS, MRI);
5019   if (!Cst)
5020     return false;
5021   auto Mask = Cst->Value;
5022   if (!Mask.isMask())
5023     return false;
5024 
5025   // No point in combining if there's nothing to truncate.
5026   unsigned NarrowWidth = Mask.countr_one();
5027   if (NarrowWidth == WideTy.getSizeInBits())
5028     return false;
5029   LLT NarrowTy = LLT::scalar(NarrowWidth);
5030 
5031   // Check if adding the zext + truncates could be harmful.
5032   auto &MF = *MI.getMF();
5033   const auto &TLI = getTargetLowering();
5034   LLVMContext &Ctx = MF.getFunction().getContext();
5035   auto &DL = MF.getDataLayout();
5036   if (!TLI.isTruncateFree(WideTy, NarrowTy, DL, Ctx) ||
5037       !TLI.isZExtFree(NarrowTy, WideTy, DL, Ctx))
5038     return false;
5039   if (!isLegalOrBeforeLegalizer({TargetOpcode::G_TRUNC, {NarrowTy, WideTy}}) ||
5040       !isLegalOrBeforeLegalizer({TargetOpcode::G_ZEXT, {WideTy, NarrowTy}}))
5041     return false;
5042   Register BinOpLHS = LHSInst->getOperand(1).getReg();
5043   Register BinOpRHS = LHSInst->getOperand(2).getReg();
5044   MatchInfo = [=, &MI](MachineIRBuilder &B) {
5045     auto NarrowLHS = Builder.buildTrunc(NarrowTy, BinOpLHS);
5046     auto NarrowRHS = Builder.buildTrunc(NarrowTy, BinOpRHS);
5047     auto NarrowBinOp =
5048         Builder.buildInstr(LHSOpc, {NarrowTy}, {NarrowLHS, NarrowRHS});
5049     auto Ext = Builder.buildZExt(WideTy, NarrowBinOp);
5050     Observer.changingInstr(MI);
5051     MI.getOperand(1).setReg(Ext.getReg(0));
5052     Observer.changedInstr(MI);
5053   };
5054   return true;
5055 }
5056 
matchMulOBy2(MachineInstr & MI,BuildFnTy & MatchInfo)5057 bool CombinerHelper::matchMulOBy2(MachineInstr &MI, BuildFnTy &MatchInfo) {
5058   unsigned Opc = MI.getOpcode();
5059   assert(Opc == TargetOpcode::G_UMULO || Opc == TargetOpcode::G_SMULO);
5060 
5061   if (!mi_match(MI.getOperand(3).getReg(), MRI, m_SpecificICstOrSplat(2)))
5062     return false;
5063 
5064   MatchInfo = [=, &MI](MachineIRBuilder &B) {
5065     Observer.changingInstr(MI);
5066     unsigned NewOpc = Opc == TargetOpcode::G_UMULO ? TargetOpcode::G_UADDO
5067                                                    : TargetOpcode::G_SADDO;
5068     MI.setDesc(Builder.getTII().get(NewOpc));
5069     MI.getOperand(3).setReg(MI.getOperand(2).getReg());
5070     Observer.changedInstr(MI);
5071   };
5072   return true;
5073 }
5074 
matchMulOBy0(MachineInstr & MI,BuildFnTy & MatchInfo)5075 bool CombinerHelper::matchMulOBy0(MachineInstr &MI, BuildFnTy &MatchInfo) {
5076   // (G_*MULO x, 0) -> 0 + no carry out
5077   assert(MI.getOpcode() == TargetOpcode::G_UMULO ||
5078          MI.getOpcode() == TargetOpcode::G_SMULO);
5079   if (!mi_match(MI.getOperand(3).getReg(), MRI, m_SpecificICstOrSplat(0)))
5080     return false;
5081   Register Dst = MI.getOperand(0).getReg();
5082   Register Carry = MI.getOperand(1).getReg();
5083   if (!isConstantLegalOrBeforeLegalizer(MRI.getType(Dst)) ||
5084       !isConstantLegalOrBeforeLegalizer(MRI.getType(Carry)))
5085     return false;
5086   MatchInfo = [=](MachineIRBuilder &B) {
5087     B.buildConstant(Dst, 0);
5088     B.buildConstant(Carry, 0);
5089   };
5090   return true;
5091 }
5092 
matchAddEToAddO(MachineInstr & MI,BuildFnTy & MatchInfo)5093 bool CombinerHelper::matchAddEToAddO(MachineInstr &MI, BuildFnTy &MatchInfo) {
5094   // (G_*ADDE x, y, 0) -> (G_*ADDO x, y)
5095   // (G_*SUBE x, y, 0) -> (G_*SUBO x, y)
5096   assert(MI.getOpcode() == TargetOpcode::G_UADDE ||
5097          MI.getOpcode() == TargetOpcode::G_SADDE ||
5098          MI.getOpcode() == TargetOpcode::G_USUBE ||
5099          MI.getOpcode() == TargetOpcode::G_SSUBE);
5100   if (!mi_match(MI.getOperand(4).getReg(), MRI, m_SpecificICstOrSplat(0)))
5101     return false;
5102   MatchInfo = [&](MachineIRBuilder &B) {
5103     unsigned NewOpcode;
5104     switch (MI.getOpcode()) {
5105     case TargetOpcode::G_UADDE:
5106       NewOpcode = TargetOpcode::G_UADDO;
5107       break;
5108     case TargetOpcode::G_SADDE:
5109       NewOpcode = TargetOpcode::G_SADDO;
5110       break;
5111     case TargetOpcode::G_USUBE:
5112       NewOpcode = TargetOpcode::G_USUBO;
5113       break;
5114     case TargetOpcode::G_SSUBE:
5115       NewOpcode = TargetOpcode::G_SSUBO;
5116       break;
5117     }
5118     Observer.changingInstr(MI);
5119     MI.setDesc(B.getTII().get(NewOpcode));
5120     MI.removeOperand(4);
5121     Observer.changedInstr(MI);
5122   };
5123   return true;
5124 }
5125 
matchSubAddSameReg(MachineInstr & MI,BuildFnTy & MatchInfo)5126 bool CombinerHelper::matchSubAddSameReg(MachineInstr &MI,
5127                                         BuildFnTy &MatchInfo) {
5128   assert(MI.getOpcode() == TargetOpcode::G_SUB);
5129   Register Dst = MI.getOperand(0).getReg();
5130   // (x + y) - z -> x (if y == z)
5131   // (x + y) - z -> y (if x == z)
5132   Register X, Y, Z;
5133   if (mi_match(Dst, MRI, m_GSub(m_GAdd(m_Reg(X), m_Reg(Y)), m_Reg(Z)))) {
5134     Register ReplaceReg;
5135     int64_t CstX, CstY;
5136     if (Y == Z || (mi_match(Y, MRI, m_ICstOrSplat(CstY)) &&
5137                    mi_match(Z, MRI, m_SpecificICstOrSplat(CstY))))
5138       ReplaceReg = X;
5139     else if (X == Z || (mi_match(X, MRI, m_ICstOrSplat(CstX)) &&
5140                         mi_match(Z, MRI, m_SpecificICstOrSplat(CstX))))
5141       ReplaceReg = Y;
5142     if (ReplaceReg) {
5143       MatchInfo = [=](MachineIRBuilder &B) { B.buildCopy(Dst, ReplaceReg); };
5144       return true;
5145     }
5146   }
5147 
5148   // x - (y + z) -> 0 - y (if x == z)
5149   // x - (y + z) -> 0 - z (if x == y)
5150   if (mi_match(Dst, MRI, m_GSub(m_Reg(X), m_GAdd(m_Reg(Y), m_Reg(Z))))) {
5151     Register ReplaceReg;
5152     int64_t CstX;
5153     if (X == Z || (mi_match(X, MRI, m_ICstOrSplat(CstX)) &&
5154                    mi_match(Z, MRI, m_SpecificICstOrSplat(CstX))))
5155       ReplaceReg = Y;
5156     else if (X == Y || (mi_match(X, MRI, m_ICstOrSplat(CstX)) &&
5157                         mi_match(Y, MRI, m_SpecificICstOrSplat(CstX))))
5158       ReplaceReg = Z;
5159     if (ReplaceReg) {
5160       MatchInfo = [=](MachineIRBuilder &B) {
5161         auto Zero = B.buildConstant(MRI.getType(Dst), 0);
5162         B.buildSub(Dst, Zero, ReplaceReg);
5163       };
5164       return true;
5165     }
5166   }
5167   return false;
5168 }
5169 
buildUDivUsingMul(MachineInstr & MI)5170 MachineInstr *CombinerHelper::buildUDivUsingMul(MachineInstr &MI) {
5171   assert(MI.getOpcode() == TargetOpcode::G_UDIV);
5172   auto &UDiv = cast<GenericMachineInstr>(MI);
5173   Register Dst = UDiv.getReg(0);
5174   Register LHS = UDiv.getReg(1);
5175   Register RHS = UDiv.getReg(2);
5176   LLT Ty = MRI.getType(Dst);
5177   LLT ScalarTy = Ty.getScalarType();
5178   const unsigned EltBits = ScalarTy.getScalarSizeInBits();
5179   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5180   LLT ScalarShiftAmtTy = ShiftAmtTy.getScalarType();
5181 
5182   auto &MIB = Builder;
5183 
5184   bool UseSRL = false;
5185   SmallVector<Register, 16> Shifts, Factors;
5186   auto *RHSDefInstr = cast<GenericMachineInstr>(getDefIgnoringCopies(RHS, MRI));
5187   bool IsSplat = getIConstantSplatVal(*RHSDefInstr, MRI).has_value();
5188 
5189   auto BuildExactUDIVPattern = [&](const Constant *C) {
5190     // Don't recompute inverses for each splat element.
5191     if (IsSplat && !Factors.empty()) {
5192       Shifts.push_back(Shifts[0]);
5193       Factors.push_back(Factors[0]);
5194       return true;
5195     }
5196 
5197     auto *CI = cast<ConstantInt>(C);
5198     APInt Divisor = CI->getValue();
5199     unsigned Shift = Divisor.countr_zero();
5200     if (Shift) {
5201       Divisor.lshrInPlace(Shift);
5202       UseSRL = true;
5203     }
5204 
5205     // Calculate the multiplicative inverse modulo BW.
5206     APInt Factor = Divisor.multiplicativeInverse();
5207     Shifts.push_back(MIB.buildConstant(ScalarShiftAmtTy, Shift).getReg(0));
5208     Factors.push_back(MIB.buildConstant(ScalarTy, Factor).getReg(0));
5209     return true;
5210   };
5211 
5212   if (MI.getFlag(MachineInstr::MIFlag::IsExact)) {
5213     // Collect all magic values from the build vector.
5214     if (!matchUnaryPredicate(MRI, RHS, BuildExactUDIVPattern))
5215       llvm_unreachable("Expected unary predicate match to succeed");
5216 
5217     Register Shift, Factor;
5218     if (Ty.isVector()) {
5219       Shift = MIB.buildBuildVector(ShiftAmtTy, Shifts).getReg(0);
5220       Factor = MIB.buildBuildVector(Ty, Factors).getReg(0);
5221     } else {
5222       Shift = Shifts[0];
5223       Factor = Factors[0];
5224     }
5225 
5226     Register Res = LHS;
5227 
5228     if (UseSRL)
5229       Res = MIB.buildLShr(Ty, Res, Shift, MachineInstr::IsExact).getReg(0);
5230 
5231     return MIB.buildMul(Ty, Res, Factor);
5232   }
5233 
5234   unsigned KnownLeadingZeros =
5235       KB ? KB->getKnownBits(LHS).countMinLeadingZeros() : 0;
5236 
5237   bool UseNPQ = false;
5238   SmallVector<Register, 16> PreShifts, PostShifts, MagicFactors, NPQFactors;
5239   auto BuildUDIVPattern = [&](const Constant *C) {
5240     auto *CI = cast<ConstantInt>(C);
5241     const APInt &Divisor = CI->getValue();
5242 
5243     bool SelNPQ = false;
5244     APInt Magic(Divisor.getBitWidth(), 0);
5245     unsigned PreShift = 0, PostShift = 0;
5246 
5247     // Magic algorithm doesn't work for division by 1. We need to emit a select
5248     // at the end.
5249     // TODO: Use undef values for divisor of 1.
5250     if (!Divisor.isOne()) {
5251 
5252       // UnsignedDivisionByConstantInfo doesn't work correctly if leading zeros
5253       // in the dividend exceeds the leading zeros for the divisor.
5254       UnsignedDivisionByConstantInfo magics =
5255           UnsignedDivisionByConstantInfo::get(
5256               Divisor, std::min(KnownLeadingZeros, Divisor.countl_zero()));
5257 
5258       Magic = std::move(magics.Magic);
5259 
5260       assert(magics.PreShift < Divisor.getBitWidth() &&
5261              "We shouldn't generate an undefined shift!");
5262       assert(magics.PostShift < Divisor.getBitWidth() &&
5263              "We shouldn't generate an undefined shift!");
5264       assert((!magics.IsAdd || magics.PreShift == 0) && "Unexpected pre-shift");
5265       PreShift = magics.PreShift;
5266       PostShift = magics.PostShift;
5267       SelNPQ = magics.IsAdd;
5268     }
5269 
5270     PreShifts.push_back(
5271         MIB.buildConstant(ScalarShiftAmtTy, PreShift).getReg(0));
5272     MagicFactors.push_back(MIB.buildConstant(ScalarTy, Magic).getReg(0));
5273     NPQFactors.push_back(
5274         MIB.buildConstant(ScalarTy,
5275                           SelNPQ ? APInt::getOneBitSet(EltBits, EltBits - 1)
5276                                  : APInt::getZero(EltBits))
5277             .getReg(0));
5278     PostShifts.push_back(
5279         MIB.buildConstant(ScalarShiftAmtTy, PostShift).getReg(0));
5280     UseNPQ |= SelNPQ;
5281     return true;
5282   };
5283 
5284   // Collect the shifts/magic values from each element.
5285   bool Matched = matchUnaryPredicate(MRI, RHS, BuildUDIVPattern);
5286   (void)Matched;
5287   assert(Matched && "Expected unary predicate match to succeed");
5288 
5289   Register PreShift, PostShift, MagicFactor, NPQFactor;
5290   auto *RHSDef = getOpcodeDef<GBuildVector>(RHS, MRI);
5291   if (RHSDef) {
5292     PreShift = MIB.buildBuildVector(ShiftAmtTy, PreShifts).getReg(0);
5293     MagicFactor = MIB.buildBuildVector(Ty, MagicFactors).getReg(0);
5294     NPQFactor = MIB.buildBuildVector(Ty, NPQFactors).getReg(0);
5295     PostShift = MIB.buildBuildVector(ShiftAmtTy, PostShifts).getReg(0);
5296   } else {
5297     assert(MRI.getType(RHS).isScalar() &&
5298            "Non-build_vector operation should have been a scalar");
5299     PreShift = PreShifts[0];
5300     MagicFactor = MagicFactors[0];
5301     PostShift = PostShifts[0];
5302   }
5303 
5304   Register Q = LHS;
5305   Q = MIB.buildLShr(Ty, Q, PreShift).getReg(0);
5306 
5307   // Multiply the numerator (operand 0) by the magic value.
5308   Q = MIB.buildUMulH(Ty, Q, MagicFactor).getReg(0);
5309 
5310   if (UseNPQ) {
5311     Register NPQ = MIB.buildSub(Ty, LHS, Q).getReg(0);
5312 
5313     // For vectors we might have a mix of non-NPQ/NPQ paths, so use
5314     // G_UMULH to act as a SRL-by-1 for NPQ, else multiply by zero.
5315     if (Ty.isVector())
5316       NPQ = MIB.buildUMulH(Ty, NPQ, NPQFactor).getReg(0);
5317     else
5318       NPQ = MIB.buildLShr(Ty, NPQ, MIB.buildConstant(ShiftAmtTy, 1)).getReg(0);
5319 
5320     Q = MIB.buildAdd(Ty, NPQ, Q).getReg(0);
5321   }
5322 
5323   Q = MIB.buildLShr(Ty, Q, PostShift).getReg(0);
5324   auto One = MIB.buildConstant(Ty, 1);
5325   auto IsOne = MIB.buildICmp(
5326       CmpInst::Predicate::ICMP_EQ,
5327       Ty.isScalar() ? LLT::scalar(1) : Ty.changeElementSize(1), RHS, One);
5328   return MIB.buildSelect(Ty, IsOne, LHS, Q);
5329 }
5330 
matchUDivByConst(MachineInstr & MI)5331 bool CombinerHelper::matchUDivByConst(MachineInstr &MI) {
5332   assert(MI.getOpcode() == TargetOpcode::G_UDIV);
5333   Register Dst = MI.getOperand(0).getReg();
5334   Register RHS = MI.getOperand(2).getReg();
5335   LLT DstTy = MRI.getType(Dst);
5336 
5337   auto &MF = *MI.getMF();
5338   AttributeList Attr = MF.getFunction().getAttributes();
5339   const auto &TLI = getTargetLowering();
5340   LLVMContext &Ctx = MF.getFunction().getContext();
5341   auto &DL = MF.getDataLayout();
5342   if (TLI.isIntDivCheap(getApproximateEVTForLLT(DstTy, DL, Ctx), Attr))
5343     return false;
5344 
5345   // Don't do this for minsize because the instruction sequence is usually
5346   // larger.
5347   if (MF.getFunction().hasMinSize())
5348     return false;
5349 
5350   if (MI.getFlag(MachineInstr::MIFlag::IsExact)) {
5351     return matchUnaryPredicate(
5352         MRI, RHS, [](const Constant *C) { return C && !C->isNullValue(); });
5353   }
5354 
5355   auto *RHSDef = MRI.getVRegDef(RHS);
5356   if (!isConstantOrConstantVector(*RHSDef, MRI))
5357     return false;
5358 
5359   // Don't do this if the types are not going to be legal.
5360   if (LI) {
5361     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_MUL, {DstTy, DstTy}}))
5362       return false;
5363     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_UMULH, {DstTy}}))
5364       return false;
5365     if (!isLegalOrBeforeLegalizer(
5366             {TargetOpcode::G_ICMP,
5367              {DstTy.isVector() ? DstTy.changeElementSize(1) : LLT::scalar(1),
5368               DstTy}}))
5369       return false;
5370   }
5371 
5372   return matchUnaryPredicate(
5373       MRI, RHS, [](const Constant *C) { return C && !C->isNullValue(); });
5374 }
5375 
applyUDivByConst(MachineInstr & MI)5376 void CombinerHelper::applyUDivByConst(MachineInstr &MI) {
5377   auto *NewMI = buildUDivUsingMul(MI);
5378   replaceSingleDefInstWithReg(MI, NewMI->getOperand(0).getReg());
5379 }
5380 
matchSDivByConst(MachineInstr & MI)5381 bool CombinerHelper::matchSDivByConst(MachineInstr &MI) {
5382   assert(MI.getOpcode() == TargetOpcode::G_SDIV && "Expected SDIV");
5383   Register Dst = MI.getOperand(0).getReg();
5384   Register RHS = MI.getOperand(2).getReg();
5385   LLT DstTy = MRI.getType(Dst);
5386 
5387   auto &MF = *MI.getMF();
5388   AttributeList Attr = MF.getFunction().getAttributes();
5389   const auto &TLI = getTargetLowering();
5390   LLVMContext &Ctx = MF.getFunction().getContext();
5391   auto &DL = MF.getDataLayout();
5392   if (TLI.isIntDivCheap(getApproximateEVTForLLT(DstTy, DL, Ctx), Attr))
5393     return false;
5394 
5395   // Don't do this for minsize because the instruction sequence is usually
5396   // larger.
5397   if (MF.getFunction().hasMinSize())
5398     return false;
5399 
5400   // If the sdiv has an 'exact' flag we can use a simpler lowering.
5401   if (MI.getFlag(MachineInstr::MIFlag::IsExact)) {
5402     return matchUnaryPredicate(
5403         MRI, RHS, [](const Constant *C) { return C && !C->isNullValue(); });
5404   }
5405 
5406   // Don't support the general case for now.
5407   return false;
5408 }
5409 
applySDivByConst(MachineInstr & MI)5410 void CombinerHelper::applySDivByConst(MachineInstr &MI) {
5411   auto *NewMI = buildSDivUsingMul(MI);
5412   replaceSingleDefInstWithReg(MI, NewMI->getOperand(0).getReg());
5413 }
5414 
buildSDivUsingMul(MachineInstr & MI)5415 MachineInstr *CombinerHelper::buildSDivUsingMul(MachineInstr &MI) {
5416   assert(MI.getOpcode() == TargetOpcode::G_SDIV && "Expected SDIV");
5417   auto &SDiv = cast<GenericMachineInstr>(MI);
5418   Register Dst = SDiv.getReg(0);
5419   Register LHS = SDiv.getReg(1);
5420   Register RHS = SDiv.getReg(2);
5421   LLT Ty = MRI.getType(Dst);
5422   LLT ScalarTy = Ty.getScalarType();
5423   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5424   LLT ScalarShiftAmtTy = ShiftAmtTy.getScalarType();
5425   auto &MIB = Builder;
5426 
5427   bool UseSRA = false;
5428   SmallVector<Register, 16> Shifts, Factors;
5429 
5430   auto *RHSDef = cast<GenericMachineInstr>(getDefIgnoringCopies(RHS, MRI));
5431   bool IsSplat = getIConstantSplatVal(*RHSDef, MRI).has_value();
5432 
5433   auto BuildSDIVPattern = [&](const Constant *C) {
5434     // Don't recompute inverses for each splat element.
5435     if (IsSplat && !Factors.empty()) {
5436       Shifts.push_back(Shifts[0]);
5437       Factors.push_back(Factors[0]);
5438       return true;
5439     }
5440 
5441     auto *CI = cast<ConstantInt>(C);
5442     APInt Divisor = CI->getValue();
5443     unsigned Shift = Divisor.countr_zero();
5444     if (Shift) {
5445       Divisor.ashrInPlace(Shift);
5446       UseSRA = true;
5447     }
5448 
5449     // Calculate the multiplicative inverse modulo BW.
5450     // 2^W requires W + 1 bits, so we have to extend and then truncate.
5451     APInt Factor = Divisor.multiplicativeInverse();
5452     Shifts.push_back(MIB.buildConstant(ScalarShiftAmtTy, Shift).getReg(0));
5453     Factors.push_back(MIB.buildConstant(ScalarTy, Factor).getReg(0));
5454     return true;
5455   };
5456 
5457   // Collect all magic values from the build vector.
5458   bool Matched = matchUnaryPredicate(MRI, RHS, BuildSDIVPattern);
5459   (void)Matched;
5460   assert(Matched && "Expected unary predicate match to succeed");
5461 
5462   Register Shift, Factor;
5463   if (Ty.isVector()) {
5464     Shift = MIB.buildBuildVector(ShiftAmtTy, Shifts).getReg(0);
5465     Factor = MIB.buildBuildVector(Ty, Factors).getReg(0);
5466   } else {
5467     Shift = Shifts[0];
5468     Factor = Factors[0];
5469   }
5470 
5471   Register Res = LHS;
5472 
5473   if (UseSRA)
5474     Res = MIB.buildAShr(Ty, Res, Shift, MachineInstr::IsExact).getReg(0);
5475 
5476   return MIB.buildMul(Ty, Res, Factor);
5477 }
5478 
matchDivByPow2(MachineInstr & MI,bool IsSigned)5479 bool CombinerHelper::matchDivByPow2(MachineInstr &MI, bool IsSigned) {
5480   assert((MI.getOpcode() == TargetOpcode::G_SDIV ||
5481           MI.getOpcode() == TargetOpcode::G_UDIV) &&
5482          "Expected SDIV or UDIV");
5483   auto &Div = cast<GenericMachineInstr>(MI);
5484   Register RHS = Div.getReg(2);
5485   auto MatchPow2 = [&](const Constant *C) {
5486     auto *CI = dyn_cast<ConstantInt>(C);
5487     return CI && (CI->getValue().isPowerOf2() ||
5488                   (IsSigned && CI->getValue().isNegatedPowerOf2()));
5489   };
5490   return matchUnaryPredicate(MRI, RHS, MatchPow2, /*AllowUndefs=*/false);
5491 }
5492 
applySDivByPow2(MachineInstr & MI)5493 void CombinerHelper::applySDivByPow2(MachineInstr &MI) {
5494   assert(MI.getOpcode() == TargetOpcode::G_SDIV && "Expected SDIV");
5495   auto &SDiv = cast<GenericMachineInstr>(MI);
5496   Register Dst = SDiv.getReg(0);
5497   Register LHS = SDiv.getReg(1);
5498   Register RHS = SDiv.getReg(2);
5499   LLT Ty = MRI.getType(Dst);
5500   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5501   LLT CCVT =
5502       Ty.isVector() ? LLT::vector(Ty.getElementCount(), 1) : LLT::scalar(1);
5503 
5504   // Effectively we want to lower G_SDIV %lhs, %rhs, where %rhs is a power of 2,
5505   // to the following version:
5506   //
5507   // %c1 = G_CTTZ %rhs
5508   // %inexact = G_SUB $bitwidth, %c1
5509   // %sign = %G_ASHR %lhs, $(bitwidth - 1)
5510   // %lshr = G_LSHR %sign, %inexact
5511   // %add = G_ADD %lhs, %lshr
5512   // %ashr = G_ASHR %add, %c1
5513   // %ashr = G_SELECT, %isoneorallones, %lhs, %ashr
5514   // %zero = G_CONSTANT $0
5515   // %neg = G_NEG %ashr
5516   // %isneg = G_ICMP SLT %rhs, %zero
5517   // %res = G_SELECT %isneg, %neg, %ashr
5518 
5519   unsigned BitWidth = Ty.getScalarSizeInBits();
5520   auto Zero = Builder.buildConstant(Ty, 0);
5521 
5522   auto Bits = Builder.buildConstant(ShiftAmtTy, BitWidth);
5523   auto C1 = Builder.buildCTTZ(ShiftAmtTy, RHS);
5524   auto Inexact = Builder.buildSub(ShiftAmtTy, Bits, C1);
5525   // Splat the sign bit into the register
5526   auto Sign = Builder.buildAShr(
5527       Ty, LHS, Builder.buildConstant(ShiftAmtTy, BitWidth - 1));
5528 
5529   // Add (LHS < 0) ? abs2 - 1 : 0;
5530   auto LSrl = Builder.buildLShr(Ty, Sign, Inexact);
5531   auto Add = Builder.buildAdd(Ty, LHS, LSrl);
5532   auto AShr = Builder.buildAShr(Ty, Add, C1);
5533 
5534   // Special case: (sdiv X, 1) -> X
5535   // Special Case: (sdiv X, -1) -> 0-X
5536   auto One = Builder.buildConstant(Ty, 1);
5537   auto MinusOne = Builder.buildConstant(Ty, -1);
5538   auto IsOne = Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, One);
5539   auto IsMinusOne =
5540       Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, MinusOne);
5541   auto IsOneOrMinusOne = Builder.buildOr(CCVT, IsOne, IsMinusOne);
5542   AShr = Builder.buildSelect(Ty, IsOneOrMinusOne, LHS, AShr);
5543 
5544   // If divided by a positive value, we're done. Otherwise, the result must be
5545   // negated.
5546   auto Neg = Builder.buildNeg(Ty, AShr);
5547   auto IsNeg = Builder.buildICmp(CmpInst::Predicate::ICMP_SLT, CCVT, RHS, Zero);
5548   Builder.buildSelect(MI.getOperand(0).getReg(), IsNeg, Neg, AShr);
5549   MI.eraseFromParent();
5550 }
5551 
applyUDivByPow2(MachineInstr & MI)5552 void CombinerHelper::applyUDivByPow2(MachineInstr &MI) {
5553   assert(MI.getOpcode() == TargetOpcode::G_UDIV && "Expected UDIV");
5554   auto &UDiv = cast<GenericMachineInstr>(MI);
5555   Register Dst = UDiv.getReg(0);
5556   Register LHS = UDiv.getReg(1);
5557   Register RHS = UDiv.getReg(2);
5558   LLT Ty = MRI.getType(Dst);
5559   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5560 
5561   auto C1 = Builder.buildCTTZ(ShiftAmtTy, RHS);
5562   Builder.buildLShr(MI.getOperand(0).getReg(), LHS, C1);
5563   MI.eraseFromParent();
5564 }
5565 
matchUMulHToLShr(MachineInstr & MI)5566 bool CombinerHelper::matchUMulHToLShr(MachineInstr &MI) {
5567   assert(MI.getOpcode() == TargetOpcode::G_UMULH);
5568   Register RHS = MI.getOperand(2).getReg();
5569   Register Dst = MI.getOperand(0).getReg();
5570   LLT Ty = MRI.getType(Dst);
5571   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5572   auto MatchPow2ExceptOne = [&](const Constant *C) {
5573     if (auto *CI = dyn_cast<ConstantInt>(C))
5574       return CI->getValue().isPowerOf2() && !CI->getValue().isOne();
5575     return false;
5576   };
5577   if (!matchUnaryPredicate(MRI, RHS, MatchPow2ExceptOne, false))
5578     return false;
5579   return isLegalOrBeforeLegalizer({TargetOpcode::G_LSHR, {Ty, ShiftAmtTy}});
5580 }
5581 
applyUMulHToLShr(MachineInstr & MI)5582 void CombinerHelper::applyUMulHToLShr(MachineInstr &MI) {
5583   Register LHS = MI.getOperand(1).getReg();
5584   Register RHS = MI.getOperand(2).getReg();
5585   Register Dst = MI.getOperand(0).getReg();
5586   LLT Ty = MRI.getType(Dst);
5587   LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
5588   unsigned NumEltBits = Ty.getScalarSizeInBits();
5589 
5590   auto LogBase2 = buildLogBase2(RHS, Builder);
5591   auto ShiftAmt =
5592       Builder.buildSub(Ty, Builder.buildConstant(Ty, NumEltBits), LogBase2);
5593   auto Trunc = Builder.buildZExtOrTrunc(ShiftAmtTy, ShiftAmt);
5594   Builder.buildLShr(Dst, LHS, Trunc);
5595   MI.eraseFromParent();
5596 }
5597 
matchRedundantNegOperands(MachineInstr & MI,BuildFnTy & MatchInfo)5598 bool CombinerHelper::matchRedundantNegOperands(MachineInstr &MI,
5599                                                BuildFnTy &MatchInfo) {
5600   unsigned Opc = MI.getOpcode();
5601   assert(Opc == TargetOpcode::G_FADD || Opc == TargetOpcode::G_FSUB ||
5602          Opc == TargetOpcode::G_FMUL || Opc == TargetOpcode::G_FDIV ||
5603          Opc == TargetOpcode::G_FMAD || Opc == TargetOpcode::G_FMA);
5604 
5605   Register Dst = MI.getOperand(0).getReg();
5606   Register X = MI.getOperand(1).getReg();
5607   Register Y = MI.getOperand(2).getReg();
5608   LLT Type = MRI.getType(Dst);
5609 
5610   // fold (fadd x, fneg(y)) -> (fsub x, y)
5611   // fold (fadd fneg(y), x) -> (fsub x, y)
5612   // G_ADD is commutative so both cases are checked by m_GFAdd
5613   if (mi_match(Dst, MRI, m_GFAdd(m_Reg(X), m_GFNeg(m_Reg(Y)))) &&
5614       isLegalOrBeforeLegalizer({TargetOpcode::G_FSUB, {Type}})) {
5615     Opc = TargetOpcode::G_FSUB;
5616   }
5617   /// fold (fsub x, fneg(y)) -> (fadd x, y)
5618   else if (mi_match(Dst, MRI, m_GFSub(m_Reg(X), m_GFNeg(m_Reg(Y)))) &&
5619            isLegalOrBeforeLegalizer({TargetOpcode::G_FADD, {Type}})) {
5620     Opc = TargetOpcode::G_FADD;
5621   }
5622   // fold (fmul fneg(x), fneg(y)) -> (fmul x, y)
5623   // fold (fdiv fneg(x), fneg(y)) -> (fdiv x, y)
5624   // fold (fmad fneg(x), fneg(y), z) -> (fmad x, y, z)
5625   // fold (fma fneg(x), fneg(y), z) -> (fma x, y, z)
5626   else if ((Opc == TargetOpcode::G_FMUL || Opc == TargetOpcode::G_FDIV ||
5627             Opc == TargetOpcode::G_FMAD || Opc == TargetOpcode::G_FMA) &&
5628            mi_match(X, MRI, m_GFNeg(m_Reg(X))) &&
5629            mi_match(Y, MRI, m_GFNeg(m_Reg(Y)))) {
5630     // no opcode change
5631   } else
5632     return false;
5633 
5634   MatchInfo = [=, &MI](MachineIRBuilder &B) {
5635     Observer.changingInstr(MI);
5636     MI.setDesc(B.getTII().get(Opc));
5637     MI.getOperand(1).setReg(X);
5638     MI.getOperand(2).setReg(Y);
5639     Observer.changedInstr(MI);
5640   };
5641   return true;
5642 }
5643 
matchFsubToFneg(MachineInstr & MI,Register & MatchInfo)5644 bool CombinerHelper::matchFsubToFneg(MachineInstr &MI, Register &MatchInfo) {
5645   assert(MI.getOpcode() == TargetOpcode::G_FSUB);
5646 
5647   Register LHS = MI.getOperand(1).getReg();
5648   MatchInfo = MI.getOperand(2).getReg();
5649   LLT Ty = MRI.getType(MI.getOperand(0).getReg());
5650 
5651   const auto LHSCst = Ty.isVector()
5652                           ? getFConstantSplat(LHS, MRI, /* allowUndef */ true)
5653                           : getFConstantVRegValWithLookThrough(LHS, MRI);
5654   if (!LHSCst)
5655     return false;
5656 
5657   // -0.0 is always allowed
5658   if (LHSCst->Value.isNegZero())
5659     return true;
5660 
5661   // +0.0 is only allowed if nsz is set.
5662   if (LHSCst->Value.isPosZero())
5663     return MI.getFlag(MachineInstr::FmNsz);
5664 
5665   return false;
5666 }
5667 
applyFsubToFneg(MachineInstr & MI,Register & MatchInfo)5668 void CombinerHelper::applyFsubToFneg(MachineInstr &MI, Register &MatchInfo) {
5669   Register Dst = MI.getOperand(0).getReg();
5670   Builder.buildFNeg(
5671       Dst, Builder.buildFCanonicalize(MRI.getType(Dst), MatchInfo).getReg(0));
5672   eraseInst(MI);
5673 }
5674 
5675 /// Checks if \p MI is TargetOpcode::G_FMUL and contractable either
5676 /// due to global flags or MachineInstr flags.
isContractableFMul(MachineInstr & MI,bool AllowFusionGlobally)5677 static bool isContractableFMul(MachineInstr &MI, bool AllowFusionGlobally) {
5678   if (MI.getOpcode() != TargetOpcode::G_FMUL)
5679     return false;
5680   return AllowFusionGlobally || MI.getFlag(MachineInstr::MIFlag::FmContract);
5681 }
5682 
hasMoreUses(const MachineInstr & MI0,const MachineInstr & MI1,const MachineRegisterInfo & MRI)5683 static bool hasMoreUses(const MachineInstr &MI0, const MachineInstr &MI1,
5684                         const MachineRegisterInfo &MRI) {
5685   return std::distance(MRI.use_instr_nodbg_begin(MI0.getOperand(0).getReg()),
5686                        MRI.use_instr_nodbg_end()) >
5687          std::distance(MRI.use_instr_nodbg_begin(MI1.getOperand(0).getReg()),
5688                        MRI.use_instr_nodbg_end());
5689 }
5690 
canCombineFMadOrFMA(MachineInstr & MI,bool & AllowFusionGlobally,bool & HasFMAD,bool & Aggressive,bool CanReassociate)5691 bool CombinerHelper::canCombineFMadOrFMA(MachineInstr &MI,
5692                                          bool &AllowFusionGlobally,
5693                                          bool &HasFMAD, bool &Aggressive,
5694                                          bool CanReassociate) {
5695 
5696   auto *MF = MI.getMF();
5697   const auto &TLI = *MF->getSubtarget().getTargetLowering();
5698   const TargetOptions &Options = MF->getTarget().Options;
5699   LLT DstType = MRI.getType(MI.getOperand(0).getReg());
5700 
5701   if (CanReassociate &&
5702       !(Options.UnsafeFPMath || MI.getFlag(MachineInstr::MIFlag::FmReassoc)))
5703     return false;
5704 
5705   // Floating-point multiply-add with intermediate rounding.
5706   HasFMAD = (!isPreLegalize() && TLI.isFMADLegal(MI, DstType));
5707   // Floating-point multiply-add without intermediate rounding.
5708   bool HasFMA = TLI.isFMAFasterThanFMulAndFAdd(*MF, DstType) &&
5709                 isLegalOrBeforeLegalizer({TargetOpcode::G_FMA, {DstType}});
5710   // No valid opcode, do not combine.
5711   if (!HasFMAD && !HasFMA)
5712     return false;
5713 
5714   AllowFusionGlobally = Options.AllowFPOpFusion == FPOpFusion::Fast ||
5715                         Options.UnsafeFPMath || HasFMAD;
5716   // If the addition is not contractable, do not combine.
5717   if (!AllowFusionGlobally && !MI.getFlag(MachineInstr::MIFlag::FmContract))
5718     return false;
5719 
5720   Aggressive = TLI.enableAggressiveFMAFusion(DstType);
5721   return true;
5722 }
5723 
matchCombineFAddFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)5724 bool CombinerHelper::matchCombineFAddFMulToFMadOrFMA(
5725     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
5726   assert(MI.getOpcode() == TargetOpcode::G_FADD);
5727 
5728   bool AllowFusionGlobally, HasFMAD, Aggressive;
5729   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
5730     return false;
5731 
5732   Register Op1 = MI.getOperand(1).getReg();
5733   Register Op2 = MI.getOperand(2).getReg();
5734   DefinitionAndSourceRegister LHS = {MRI.getVRegDef(Op1), Op1};
5735   DefinitionAndSourceRegister RHS = {MRI.getVRegDef(Op2), Op2};
5736   unsigned PreferredFusedOpcode =
5737       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
5738 
5739   // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
5740   // prefer to fold the multiply with fewer uses.
5741   if (Aggressive && isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
5742       isContractableFMul(*RHS.MI, AllowFusionGlobally)) {
5743     if (hasMoreUses(*LHS.MI, *RHS.MI, MRI))
5744       std::swap(LHS, RHS);
5745   }
5746 
5747   // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5748   if (isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
5749       (Aggressive || MRI.hasOneNonDBGUse(LHS.Reg))) {
5750     MatchInfo = [=, &MI](MachineIRBuilder &B) {
5751       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5752                    {LHS.MI->getOperand(1).getReg(),
5753                     LHS.MI->getOperand(2).getReg(), RHS.Reg});
5754     };
5755     return true;
5756   }
5757 
5758   // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
5759   if (isContractableFMul(*RHS.MI, AllowFusionGlobally) &&
5760       (Aggressive || MRI.hasOneNonDBGUse(RHS.Reg))) {
5761     MatchInfo = [=, &MI](MachineIRBuilder &B) {
5762       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5763                    {RHS.MI->getOperand(1).getReg(),
5764                     RHS.MI->getOperand(2).getReg(), LHS.Reg});
5765     };
5766     return true;
5767   }
5768 
5769   return false;
5770 }
5771 
matchCombineFAddFpExtFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)5772 bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMA(
5773     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
5774   assert(MI.getOpcode() == TargetOpcode::G_FADD);
5775 
5776   bool AllowFusionGlobally, HasFMAD, Aggressive;
5777   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
5778     return false;
5779 
5780   const auto &TLI = *MI.getMF()->getSubtarget().getTargetLowering();
5781   Register Op1 = MI.getOperand(1).getReg();
5782   Register Op2 = MI.getOperand(2).getReg();
5783   DefinitionAndSourceRegister LHS = {MRI.getVRegDef(Op1), Op1};
5784   DefinitionAndSourceRegister RHS = {MRI.getVRegDef(Op2), Op2};
5785   LLT DstType = MRI.getType(MI.getOperand(0).getReg());
5786 
5787   unsigned PreferredFusedOpcode =
5788       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
5789 
5790   // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
5791   // prefer to fold the multiply with fewer uses.
5792   if (Aggressive && isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
5793       isContractableFMul(*RHS.MI, AllowFusionGlobally)) {
5794     if (hasMoreUses(*LHS.MI, *RHS.MI, MRI))
5795       std::swap(LHS, RHS);
5796   }
5797 
5798   // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
5799   MachineInstr *FpExtSrc;
5800   if (mi_match(LHS.Reg, MRI, m_GFPExt(m_MInstr(FpExtSrc))) &&
5801       isContractableFMul(*FpExtSrc, AllowFusionGlobally) &&
5802       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
5803                           MRI.getType(FpExtSrc->getOperand(1).getReg()))) {
5804     MatchInfo = [=, &MI](MachineIRBuilder &B) {
5805       auto FpExtX = B.buildFPExt(DstType, FpExtSrc->getOperand(1).getReg());
5806       auto FpExtY = B.buildFPExt(DstType, FpExtSrc->getOperand(2).getReg());
5807       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5808                    {FpExtX.getReg(0), FpExtY.getReg(0), RHS.Reg});
5809     };
5810     return true;
5811   }
5812 
5813   // fold (fadd z, (fpext (fmul x, y))) -> (fma (fpext x), (fpext y), z)
5814   // Note: Commutes FADD operands.
5815   if (mi_match(RHS.Reg, MRI, m_GFPExt(m_MInstr(FpExtSrc))) &&
5816       isContractableFMul(*FpExtSrc, AllowFusionGlobally) &&
5817       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
5818                           MRI.getType(FpExtSrc->getOperand(1).getReg()))) {
5819     MatchInfo = [=, &MI](MachineIRBuilder &B) {
5820       auto FpExtX = B.buildFPExt(DstType, FpExtSrc->getOperand(1).getReg());
5821       auto FpExtY = B.buildFPExt(DstType, FpExtSrc->getOperand(2).getReg());
5822       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5823                    {FpExtX.getReg(0), FpExtY.getReg(0), LHS.Reg});
5824     };
5825     return true;
5826   }
5827 
5828   return false;
5829 }
5830 
matchCombineFAddFMAFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)5831 bool CombinerHelper::matchCombineFAddFMAFMulToFMadOrFMA(
5832     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
5833   assert(MI.getOpcode() == TargetOpcode::G_FADD);
5834 
5835   bool AllowFusionGlobally, HasFMAD, Aggressive;
5836   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive, true))
5837     return false;
5838 
5839   Register Op1 = MI.getOperand(1).getReg();
5840   Register Op2 = MI.getOperand(2).getReg();
5841   DefinitionAndSourceRegister LHS = {MRI.getVRegDef(Op1), Op1};
5842   DefinitionAndSourceRegister RHS = {MRI.getVRegDef(Op2), Op2};
5843   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
5844 
5845   unsigned PreferredFusedOpcode =
5846       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
5847 
5848   // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
5849   // prefer to fold the multiply with fewer uses.
5850   if (Aggressive && isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
5851       isContractableFMul(*RHS.MI, AllowFusionGlobally)) {
5852     if (hasMoreUses(*LHS.MI, *RHS.MI, MRI))
5853       std::swap(LHS, RHS);
5854   }
5855 
5856   MachineInstr *FMA = nullptr;
5857   Register Z;
5858   // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y, (fma u, v, z))
5859   if (LHS.MI->getOpcode() == PreferredFusedOpcode &&
5860       (MRI.getVRegDef(LHS.MI->getOperand(3).getReg())->getOpcode() ==
5861        TargetOpcode::G_FMUL) &&
5862       MRI.hasOneNonDBGUse(LHS.MI->getOperand(0).getReg()) &&
5863       MRI.hasOneNonDBGUse(LHS.MI->getOperand(3).getReg())) {
5864     FMA = LHS.MI;
5865     Z = RHS.Reg;
5866   }
5867   // fold (fadd z, (fma x, y, (fmul u, v))) -> (fma x, y, (fma u, v, z))
5868   else if (RHS.MI->getOpcode() == PreferredFusedOpcode &&
5869            (MRI.getVRegDef(RHS.MI->getOperand(3).getReg())->getOpcode() ==
5870             TargetOpcode::G_FMUL) &&
5871            MRI.hasOneNonDBGUse(RHS.MI->getOperand(0).getReg()) &&
5872            MRI.hasOneNonDBGUse(RHS.MI->getOperand(3).getReg())) {
5873     Z = LHS.Reg;
5874     FMA = RHS.MI;
5875   }
5876 
5877   if (FMA) {
5878     MachineInstr *FMulMI = MRI.getVRegDef(FMA->getOperand(3).getReg());
5879     Register X = FMA->getOperand(1).getReg();
5880     Register Y = FMA->getOperand(2).getReg();
5881     Register U = FMulMI->getOperand(1).getReg();
5882     Register V = FMulMI->getOperand(2).getReg();
5883 
5884     MatchInfo = [=, &MI](MachineIRBuilder &B) {
5885       Register InnerFMA = MRI.createGenericVirtualRegister(DstTy);
5886       B.buildInstr(PreferredFusedOpcode, {InnerFMA}, {U, V, Z});
5887       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5888                    {X, Y, InnerFMA});
5889     };
5890     return true;
5891   }
5892 
5893   return false;
5894 }
5895 
matchCombineFAddFpExtFMulToFMadOrFMAAggressive(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)5896 bool CombinerHelper::matchCombineFAddFpExtFMulToFMadOrFMAAggressive(
5897     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
5898   assert(MI.getOpcode() == TargetOpcode::G_FADD);
5899 
5900   bool AllowFusionGlobally, HasFMAD, Aggressive;
5901   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
5902     return false;
5903 
5904   if (!Aggressive)
5905     return false;
5906 
5907   const auto &TLI = *MI.getMF()->getSubtarget().getTargetLowering();
5908   LLT DstType = MRI.getType(MI.getOperand(0).getReg());
5909   Register Op1 = MI.getOperand(1).getReg();
5910   Register Op2 = MI.getOperand(2).getReg();
5911   DefinitionAndSourceRegister LHS = {MRI.getVRegDef(Op1), Op1};
5912   DefinitionAndSourceRegister RHS = {MRI.getVRegDef(Op2), Op2};
5913 
5914   unsigned PreferredFusedOpcode =
5915       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
5916 
5917   // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
5918   // prefer to fold the multiply with fewer uses.
5919   if (Aggressive && isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
5920       isContractableFMul(*RHS.MI, AllowFusionGlobally)) {
5921     if (hasMoreUses(*LHS.MI, *RHS.MI, MRI))
5922       std::swap(LHS, RHS);
5923   }
5924 
5925   // Builds: (fma x, y, (fma (fpext u), (fpext v), z))
5926   auto buildMatchInfo = [=, &MI](Register U, Register V, Register Z, Register X,
5927                                  Register Y, MachineIRBuilder &B) {
5928     Register FpExtU = B.buildFPExt(DstType, U).getReg(0);
5929     Register FpExtV = B.buildFPExt(DstType, V).getReg(0);
5930     Register InnerFMA =
5931         B.buildInstr(PreferredFusedOpcode, {DstType}, {FpExtU, FpExtV, Z})
5932             .getReg(0);
5933     B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
5934                  {X, Y, InnerFMA});
5935   };
5936 
5937   MachineInstr *FMulMI, *FMAMI;
5938   // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
5939   //   -> (fma x, y, (fma (fpext u), (fpext v), z))
5940   if (LHS.MI->getOpcode() == PreferredFusedOpcode &&
5941       mi_match(LHS.MI->getOperand(3).getReg(), MRI,
5942                m_GFPExt(m_MInstr(FMulMI))) &&
5943       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
5944       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
5945                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
5946     MatchInfo = [=](MachineIRBuilder &B) {
5947       buildMatchInfo(FMulMI->getOperand(1).getReg(),
5948                      FMulMI->getOperand(2).getReg(), RHS.Reg,
5949                      LHS.MI->getOperand(1).getReg(),
5950                      LHS.MI->getOperand(2).getReg(), B);
5951     };
5952     return true;
5953   }
5954 
5955   // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
5956   //   -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
5957   // FIXME: This turns two single-precision and one double-precision
5958   // operation into two double-precision operations, which might not be
5959   // interesting for all targets, especially GPUs.
5960   if (mi_match(LHS.Reg, MRI, m_GFPExt(m_MInstr(FMAMI))) &&
5961       FMAMI->getOpcode() == PreferredFusedOpcode) {
5962     MachineInstr *FMulMI = MRI.getVRegDef(FMAMI->getOperand(3).getReg());
5963     if (isContractableFMul(*FMulMI, AllowFusionGlobally) &&
5964         TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
5965                             MRI.getType(FMAMI->getOperand(0).getReg()))) {
5966       MatchInfo = [=](MachineIRBuilder &B) {
5967         Register X = FMAMI->getOperand(1).getReg();
5968         Register Y = FMAMI->getOperand(2).getReg();
5969         X = B.buildFPExt(DstType, X).getReg(0);
5970         Y = B.buildFPExt(DstType, Y).getReg(0);
5971         buildMatchInfo(FMulMI->getOperand(1).getReg(),
5972                        FMulMI->getOperand(2).getReg(), RHS.Reg, X, Y, B);
5973       };
5974 
5975       return true;
5976     }
5977   }
5978 
5979   // fold (fadd z, (fma x, y, (fpext (fmul u, v)))
5980   //   -> (fma x, y, (fma (fpext u), (fpext v), z))
5981   if (RHS.MI->getOpcode() == PreferredFusedOpcode &&
5982       mi_match(RHS.MI->getOperand(3).getReg(), MRI,
5983                m_GFPExt(m_MInstr(FMulMI))) &&
5984       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
5985       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
5986                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
5987     MatchInfo = [=](MachineIRBuilder &B) {
5988       buildMatchInfo(FMulMI->getOperand(1).getReg(),
5989                      FMulMI->getOperand(2).getReg(), LHS.Reg,
5990                      RHS.MI->getOperand(1).getReg(),
5991                      RHS.MI->getOperand(2).getReg(), B);
5992     };
5993     return true;
5994   }
5995 
5996   // fold (fadd z, (fpext (fma x, y, (fmul u, v)))
5997   //   -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
5998   // FIXME: This turns two single-precision and one double-precision
5999   // operation into two double-precision operations, which might not be
6000   // interesting for all targets, especially GPUs.
6001   if (mi_match(RHS.Reg, MRI, m_GFPExt(m_MInstr(FMAMI))) &&
6002       FMAMI->getOpcode() == PreferredFusedOpcode) {
6003     MachineInstr *FMulMI = MRI.getVRegDef(FMAMI->getOperand(3).getReg());
6004     if (isContractableFMul(*FMulMI, AllowFusionGlobally) &&
6005         TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstType,
6006                             MRI.getType(FMAMI->getOperand(0).getReg()))) {
6007       MatchInfo = [=](MachineIRBuilder &B) {
6008         Register X = FMAMI->getOperand(1).getReg();
6009         Register Y = FMAMI->getOperand(2).getReg();
6010         X = B.buildFPExt(DstType, X).getReg(0);
6011         Y = B.buildFPExt(DstType, Y).getReg(0);
6012         buildMatchInfo(FMulMI->getOperand(1).getReg(),
6013                        FMulMI->getOperand(2).getReg(), LHS.Reg, X, Y, B);
6014       };
6015       return true;
6016     }
6017   }
6018 
6019   return false;
6020 }
6021 
matchCombineFSubFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)6022 bool CombinerHelper::matchCombineFSubFMulToFMadOrFMA(
6023     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
6024   assert(MI.getOpcode() == TargetOpcode::G_FSUB);
6025 
6026   bool AllowFusionGlobally, HasFMAD, Aggressive;
6027   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
6028     return false;
6029 
6030   Register Op1 = MI.getOperand(1).getReg();
6031   Register Op2 = MI.getOperand(2).getReg();
6032   DefinitionAndSourceRegister LHS = {MRI.getVRegDef(Op1), Op1};
6033   DefinitionAndSourceRegister RHS = {MRI.getVRegDef(Op2), Op2};
6034   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
6035 
6036   // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
6037   // prefer to fold the multiply with fewer uses.
6038   int FirstMulHasFewerUses = true;
6039   if (isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
6040       isContractableFMul(*RHS.MI, AllowFusionGlobally) &&
6041       hasMoreUses(*LHS.MI, *RHS.MI, MRI))
6042     FirstMulHasFewerUses = false;
6043 
6044   unsigned PreferredFusedOpcode =
6045       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
6046 
6047   // fold (fsub (fmul x, y), z) -> (fma x, y, -z)
6048   if (FirstMulHasFewerUses &&
6049       (isContractableFMul(*LHS.MI, AllowFusionGlobally) &&
6050        (Aggressive || MRI.hasOneNonDBGUse(LHS.Reg)))) {
6051     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6052       Register NegZ = B.buildFNeg(DstTy, RHS.Reg).getReg(0);
6053       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6054                    {LHS.MI->getOperand(1).getReg(),
6055                     LHS.MI->getOperand(2).getReg(), NegZ});
6056     };
6057     return true;
6058   }
6059   // fold (fsub x, (fmul y, z)) -> (fma -y, z, x)
6060   else if ((isContractableFMul(*RHS.MI, AllowFusionGlobally) &&
6061             (Aggressive || MRI.hasOneNonDBGUse(RHS.Reg)))) {
6062     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6063       Register NegY =
6064           B.buildFNeg(DstTy, RHS.MI->getOperand(1).getReg()).getReg(0);
6065       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6066                    {NegY, RHS.MI->getOperand(2).getReg(), LHS.Reg});
6067     };
6068     return true;
6069   }
6070 
6071   return false;
6072 }
6073 
matchCombineFSubFNegFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)6074 bool CombinerHelper::matchCombineFSubFNegFMulToFMadOrFMA(
6075     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
6076   assert(MI.getOpcode() == TargetOpcode::G_FSUB);
6077 
6078   bool AllowFusionGlobally, HasFMAD, Aggressive;
6079   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
6080     return false;
6081 
6082   Register LHSReg = MI.getOperand(1).getReg();
6083   Register RHSReg = MI.getOperand(2).getReg();
6084   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
6085 
6086   unsigned PreferredFusedOpcode =
6087       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
6088 
6089   MachineInstr *FMulMI;
6090   // fold (fsub (fneg (fmul x, y)), z) -> (fma (fneg x), y, (fneg z))
6091   if (mi_match(LHSReg, MRI, m_GFNeg(m_MInstr(FMulMI))) &&
6092       (Aggressive || (MRI.hasOneNonDBGUse(LHSReg) &&
6093                       MRI.hasOneNonDBGUse(FMulMI->getOperand(0).getReg()))) &&
6094       isContractableFMul(*FMulMI, AllowFusionGlobally)) {
6095     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6096       Register NegX =
6097           B.buildFNeg(DstTy, FMulMI->getOperand(1).getReg()).getReg(0);
6098       Register NegZ = B.buildFNeg(DstTy, RHSReg).getReg(0);
6099       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6100                    {NegX, FMulMI->getOperand(2).getReg(), NegZ});
6101     };
6102     return true;
6103   }
6104 
6105   // fold (fsub x, (fneg (fmul, y, z))) -> (fma y, z, x)
6106   if (mi_match(RHSReg, MRI, m_GFNeg(m_MInstr(FMulMI))) &&
6107       (Aggressive || (MRI.hasOneNonDBGUse(RHSReg) &&
6108                       MRI.hasOneNonDBGUse(FMulMI->getOperand(0).getReg()))) &&
6109       isContractableFMul(*FMulMI, AllowFusionGlobally)) {
6110     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6111       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6112                    {FMulMI->getOperand(1).getReg(),
6113                     FMulMI->getOperand(2).getReg(), LHSReg});
6114     };
6115     return true;
6116   }
6117 
6118   return false;
6119 }
6120 
matchCombineFSubFpExtFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)6121 bool CombinerHelper::matchCombineFSubFpExtFMulToFMadOrFMA(
6122     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
6123   assert(MI.getOpcode() == TargetOpcode::G_FSUB);
6124 
6125   bool AllowFusionGlobally, HasFMAD, Aggressive;
6126   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
6127     return false;
6128 
6129   Register LHSReg = MI.getOperand(1).getReg();
6130   Register RHSReg = MI.getOperand(2).getReg();
6131   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
6132 
6133   unsigned PreferredFusedOpcode =
6134       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
6135 
6136   MachineInstr *FMulMI;
6137   // fold (fsub (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), (fneg z))
6138   if (mi_match(LHSReg, MRI, m_GFPExt(m_MInstr(FMulMI))) &&
6139       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
6140       (Aggressive || MRI.hasOneNonDBGUse(LHSReg))) {
6141     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6142       Register FpExtX =
6143           B.buildFPExt(DstTy, FMulMI->getOperand(1).getReg()).getReg(0);
6144       Register FpExtY =
6145           B.buildFPExt(DstTy, FMulMI->getOperand(2).getReg()).getReg(0);
6146       Register NegZ = B.buildFNeg(DstTy, RHSReg).getReg(0);
6147       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6148                    {FpExtX, FpExtY, NegZ});
6149     };
6150     return true;
6151   }
6152 
6153   // fold (fsub x, (fpext (fmul y, z))) -> (fma (fneg (fpext y)), (fpext z), x)
6154   if (mi_match(RHSReg, MRI, m_GFPExt(m_MInstr(FMulMI))) &&
6155       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
6156       (Aggressive || MRI.hasOneNonDBGUse(RHSReg))) {
6157     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6158       Register FpExtY =
6159           B.buildFPExt(DstTy, FMulMI->getOperand(1).getReg()).getReg(0);
6160       Register NegY = B.buildFNeg(DstTy, FpExtY).getReg(0);
6161       Register FpExtZ =
6162           B.buildFPExt(DstTy, FMulMI->getOperand(2).getReg()).getReg(0);
6163       B.buildInstr(PreferredFusedOpcode, {MI.getOperand(0).getReg()},
6164                    {NegY, FpExtZ, LHSReg});
6165     };
6166     return true;
6167   }
6168 
6169   return false;
6170 }
6171 
matchCombineFSubFpExtFNegFMulToFMadOrFMA(MachineInstr & MI,std::function<void (MachineIRBuilder &)> & MatchInfo)6172 bool CombinerHelper::matchCombineFSubFpExtFNegFMulToFMadOrFMA(
6173     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
6174   assert(MI.getOpcode() == TargetOpcode::G_FSUB);
6175 
6176   bool AllowFusionGlobally, HasFMAD, Aggressive;
6177   if (!canCombineFMadOrFMA(MI, AllowFusionGlobally, HasFMAD, Aggressive))
6178     return false;
6179 
6180   const auto &TLI = *MI.getMF()->getSubtarget().getTargetLowering();
6181   LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
6182   Register LHSReg = MI.getOperand(1).getReg();
6183   Register RHSReg = MI.getOperand(2).getReg();
6184 
6185   unsigned PreferredFusedOpcode =
6186       HasFMAD ? TargetOpcode::G_FMAD : TargetOpcode::G_FMA;
6187 
6188   auto buildMatchInfo = [=](Register Dst, Register X, Register Y, Register Z,
6189                             MachineIRBuilder &B) {
6190     Register FpExtX = B.buildFPExt(DstTy, X).getReg(0);
6191     Register FpExtY = B.buildFPExt(DstTy, Y).getReg(0);
6192     B.buildInstr(PreferredFusedOpcode, {Dst}, {FpExtX, FpExtY, Z});
6193   };
6194 
6195   MachineInstr *FMulMI;
6196   // fold (fsub (fpext (fneg (fmul x, y))), z) ->
6197   //      (fneg (fma (fpext x), (fpext y), z))
6198   // fold (fsub (fneg (fpext (fmul x, y))), z) ->
6199   //      (fneg (fma (fpext x), (fpext y), z))
6200   if ((mi_match(LHSReg, MRI, m_GFPExt(m_GFNeg(m_MInstr(FMulMI)))) ||
6201        mi_match(LHSReg, MRI, m_GFNeg(m_GFPExt(m_MInstr(FMulMI))))) &&
6202       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
6203       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstTy,
6204                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
6205     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6206       Register FMAReg = MRI.createGenericVirtualRegister(DstTy);
6207       buildMatchInfo(FMAReg, FMulMI->getOperand(1).getReg(),
6208                      FMulMI->getOperand(2).getReg(), RHSReg, B);
6209       B.buildFNeg(MI.getOperand(0).getReg(), FMAReg);
6210     };
6211     return true;
6212   }
6213 
6214   // fold (fsub x, (fpext (fneg (fmul y, z)))) -> (fma (fpext y), (fpext z), x)
6215   // fold (fsub x, (fneg (fpext (fmul y, z)))) -> (fma (fpext y), (fpext z), x)
6216   if ((mi_match(RHSReg, MRI, m_GFPExt(m_GFNeg(m_MInstr(FMulMI)))) ||
6217        mi_match(RHSReg, MRI, m_GFNeg(m_GFPExt(m_MInstr(FMulMI))))) &&
6218       isContractableFMul(*FMulMI, AllowFusionGlobally) &&
6219       TLI.isFPExtFoldable(MI, PreferredFusedOpcode, DstTy,
6220                           MRI.getType(FMulMI->getOperand(0).getReg()))) {
6221     MatchInfo = [=, &MI](MachineIRBuilder &B) {
6222       buildMatchInfo(MI.getOperand(0).getReg(), FMulMI->getOperand(1).getReg(),
6223                      FMulMI->getOperand(2).getReg(), LHSReg, B);
6224     };
6225     return true;
6226   }
6227 
6228   return false;
6229 }
6230 
matchCombineFMinMaxNaN(MachineInstr & MI,unsigned & IdxToPropagate)6231 bool CombinerHelper::matchCombineFMinMaxNaN(MachineInstr &MI,
6232                                             unsigned &IdxToPropagate) {
6233   bool PropagateNaN;
6234   switch (MI.getOpcode()) {
6235   default:
6236     return false;
6237   case TargetOpcode::G_FMINNUM:
6238   case TargetOpcode::G_FMAXNUM:
6239     PropagateNaN = false;
6240     break;
6241   case TargetOpcode::G_FMINIMUM:
6242   case TargetOpcode::G_FMAXIMUM:
6243     PropagateNaN = true;
6244     break;
6245   }
6246 
6247   auto MatchNaN = [&](unsigned Idx) {
6248     Register MaybeNaNReg = MI.getOperand(Idx).getReg();
6249     const ConstantFP *MaybeCst = getConstantFPVRegVal(MaybeNaNReg, MRI);
6250     if (!MaybeCst || !MaybeCst->getValueAPF().isNaN())
6251       return false;
6252     IdxToPropagate = PropagateNaN ? Idx : (Idx == 1 ? 2 : 1);
6253     return true;
6254   };
6255 
6256   return MatchNaN(1) || MatchNaN(2);
6257 }
6258 
matchAddSubSameReg(MachineInstr & MI,Register & Src)6259 bool CombinerHelper::matchAddSubSameReg(MachineInstr &MI, Register &Src) {
6260   assert(MI.getOpcode() == TargetOpcode::G_ADD && "Expected a G_ADD");
6261   Register LHS = MI.getOperand(1).getReg();
6262   Register RHS = MI.getOperand(2).getReg();
6263 
6264   // Helper lambda to check for opportunities for
6265   // A + (B - A) -> B
6266   // (B - A) + A -> B
6267   auto CheckFold = [&](Register MaybeSub, Register MaybeSameReg) {
6268     Register Reg;
6269     return mi_match(MaybeSub, MRI, m_GSub(m_Reg(Src), m_Reg(Reg))) &&
6270            Reg == MaybeSameReg;
6271   };
6272   return CheckFold(LHS, RHS) || CheckFold(RHS, LHS);
6273 }
6274 
matchBuildVectorIdentityFold(MachineInstr & MI,Register & MatchInfo)6275 bool CombinerHelper::matchBuildVectorIdentityFold(MachineInstr &MI,
6276                                                   Register &MatchInfo) {
6277   // This combine folds the following patterns:
6278   //
6279   //  G_BUILD_VECTOR_TRUNC (G_BITCAST(x), G_LSHR(G_BITCAST(x), k))
6280   //  G_BUILD_VECTOR(G_TRUNC(G_BITCAST(x)), G_TRUNC(G_LSHR(G_BITCAST(x), k)))
6281   //    into
6282   //      x
6283   //    if
6284   //      k == sizeof(VecEltTy)/2
6285   //      type(x) == type(dst)
6286   //
6287   //  G_BUILD_VECTOR(G_TRUNC(G_BITCAST(x)), undef)
6288   //    into
6289   //      x
6290   //    if
6291   //      type(x) == type(dst)
6292 
6293   LLT DstVecTy = MRI.getType(MI.getOperand(0).getReg());
6294   LLT DstEltTy = DstVecTy.getElementType();
6295 
6296   Register Lo, Hi;
6297 
6298   if (mi_match(
6299           MI, MRI,
6300           m_GBuildVector(m_GTrunc(m_GBitcast(m_Reg(Lo))), m_GImplicitDef()))) {
6301     MatchInfo = Lo;
6302     return MRI.getType(MatchInfo) == DstVecTy;
6303   }
6304 
6305   std::optional<ValueAndVReg> ShiftAmount;
6306   const auto LoPattern = m_GBitcast(m_Reg(Lo));
6307   const auto HiPattern = m_GLShr(m_GBitcast(m_Reg(Hi)), m_GCst(ShiftAmount));
6308   if (mi_match(
6309           MI, MRI,
6310           m_any_of(m_GBuildVectorTrunc(LoPattern, HiPattern),
6311                    m_GBuildVector(m_GTrunc(LoPattern), m_GTrunc(HiPattern))))) {
6312     if (Lo == Hi && ShiftAmount->Value == DstEltTy.getSizeInBits()) {
6313       MatchInfo = Lo;
6314       return MRI.getType(MatchInfo) == DstVecTy;
6315     }
6316   }
6317 
6318   return false;
6319 }
6320 
matchTruncBuildVectorFold(MachineInstr & MI,Register & MatchInfo)6321 bool CombinerHelper::matchTruncBuildVectorFold(MachineInstr &MI,
6322                                                Register &MatchInfo) {
6323   // Replace (G_TRUNC (G_BITCAST (G_BUILD_VECTOR x, y)) with just x
6324   // if type(x) == type(G_TRUNC)
6325   if (!mi_match(MI.getOperand(1).getReg(), MRI,
6326                 m_GBitcast(m_GBuildVector(m_Reg(MatchInfo), m_Reg()))))
6327     return false;
6328 
6329   return MRI.getType(MatchInfo) == MRI.getType(MI.getOperand(0).getReg());
6330 }
6331 
matchTruncLshrBuildVectorFold(MachineInstr & MI,Register & MatchInfo)6332 bool CombinerHelper::matchTruncLshrBuildVectorFold(MachineInstr &MI,
6333                                                    Register &MatchInfo) {
6334   // Replace (G_TRUNC (G_LSHR (G_BITCAST (G_BUILD_VECTOR x, y)), K)) with
6335   //    y if K == size of vector element type
6336   std::optional<ValueAndVReg> ShiftAmt;
6337   if (!mi_match(MI.getOperand(1).getReg(), MRI,
6338                 m_GLShr(m_GBitcast(m_GBuildVector(m_Reg(), m_Reg(MatchInfo))),
6339                         m_GCst(ShiftAmt))))
6340     return false;
6341 
6342   LLT MatchTy = MRI.getType(MatchInfo);
6343   return ShiftAmt->Value.getZExtValue() == MatchTy.getSizeInBits() &&
6344          MatchTy == MRI.getType(MI.getOperand(0).getReg());
6345 }
6346 
getFPMinMaxOpcForSelect(CmpInst::Predicate Pred,LLT DstTy,SelectPatternNaNBehaviour VsNaNRetVal) const6347 unsigned CombinerHelper::getFPMinMaxOpcForSelect(
6348     CmpInst::Predicate Pred, LLT DstTy,
6349     SelectPatternNaNBehaviour VsNaNRetVal) const {
6350   assert(VsNaNRetVal != SelectPatternNaNBehaviour::NOT_APPLICABLE &&
6351          "Expected a NaN behaviour?");
6352   // Choose an opcode based off of legality or the behaviour when one of the
6353   // LHS/RHS may be NaN.
6354   switch (Pred) {
6355   default:
6356     return 0;
6357   case CmpInst::FCMP_UGT:
6358   case CmpInst::FCMP_UGE:
6359   case CmpInst::FCMP_OGT:
6360   case CmpInst::FCMP_OGE:
6361     if (VsNaNRetVal == SelectPatternNaNBehaviour::RETURNS_OTHER)
6362       return TargetOpcode::G_FMAXNUM;
6363     if (VsNaNRetVal == SelectPatternNaNBehaviour::RETURNS_NAN)
6364       return TargetOpcode::G_FMAXIMUM;
6365     if (isLegal({TargetOpcode::G_FMAXNUM, {DstTy}}))
6366       return TargetOpcode::G_FMAXNUM;
6367     if (isLegal({TargetOpcode::G_FMAXIMUM, {DstTy}}))
6368       return TargetOpcode::G_FMAXIMUM;
6369     return 0;
6370   case CmpInst::FCMP_ULT:
6371   case CmpInst::FCMP_ULE:
6372   case CmpInst::FCMP_OLT:
6373   case CmpInst::FCMP_OLE:
6374     if (VsNaNRetVal == SelectPatternNaNBehaviour::RETURNS_OTHER)
6375       return TargetOpcode::G_FMINNUM;
6376     if (VsNaNRetVal == SelectPatternNaNBehaviour::RETURNS_NAN)
6377       return TargetOpcode::G_FMINIMUM;
6378     if (isLegal({TargetOpcode::G_FMINNUM, {DstTy}}))
6379       return TargetOpcode::G_FMINNUM;
6380     if (!isLegal({TargetOpcode::G_FMINIMUM, {DstTy}}))
6381       return 0;
6382     return TargetOpcode::G_FMINIMUM;
6383   }
6384 }
6385 
6386 CombinerHelper::SelectPatternNaNBehaviour
computeRetValAgainstNaN(Register LHS,Register RHS,bool IsOrderedComparison) const6387 CombinerHelper::computeRetValAgainstNaN(Register LHS, Register RHS,
6388                                         bool IsOrderedComparison) const {
6389   bool LHSSafe = isKnownNeverNaN(LHS, MRI);
6390   bool RHSSafe = isKnownNeverNaN(RHS, MRI);
6391   // Completely unsafe.
6392   if (!LHSSafe && !RHSSafe)
6393     return SelectPatternNaNBehaviour::NOT_APPLICABLE;
6394   if (LHSSafe && RHSSafe)
6395     return SelectPatternNaNBehaviour::RETURNS_ANY;
6396   // An ordered comparison will return false when given a NaN, so it
6397   // returns the RHS.
6398   if (IsOrderedComparison)
6399     return LHSSafe ? SelectPatternNaNBehaviour::RETURNS_NAN
6400                    : SelectPatternNaNBehaviour::RETURNS_OTHER;
6401   // An unordered comparison will return true when given a NaN, so it
6402   // returns the LHS.
6403   return LHSSafe ? SelectPatternNaNBehaviour::RETURNS_OTHER
6404                  : SelectPatternNaNBehaviour::RETURNS_NAN;
6405 }
6406 
matchFPSelectToMinMax(Register Dst,Register Cond,Register TrueVal,Register FalseVal,BuildFnTy & MatchInfo)6407 bool CombinerHelper::matchFPSelectToMinMax(Register Dst, Register Cond,
6408                                            Register TrueVal, Register FalseVal,
6409                                            BuildFnTy &MatchInfo) {
6410   // Match: select (fcmp cond x, y) x, y
6411   //        select (fcmp cond x, y) y, x
6412   // And turn it into fminnum/fmaxnum or fmin/fmax based off of the condition.
6413   LLT DstTy = MRI.getType(Dst);
6414   // Bail out early on pointers, since we'll never want to fold to a min/max.
6415   if (DstTy.isPointer())
6416     return false;
6417   // Match a floating point compare with a less-than/greater-than predicate.
6418   // TODO: Allow multiple users of the compare if they are all selects.
6419   CmpInst::Predicate Pred;
6420   Register CmpLHS, CmpRHS;
6421   if (!mi_match(Cond, MRI,
6422                 m_OneNonDBGUse(
6423                     m_GFCmp(m_Pred(Pred), m_Reg(CmpLHS), m_Reg(CmpRHS)))) ||
6424       CmpInst::isEquality(Pred))
6425     return false;
6426   SelectPatternNaNBehaviour ResWithKnownNaNInfo =
6427       computeRetValAgainstNaN(CmpLHS, CmpRHS, CmpInst::isOrdered(Pred));
6428   if (ResWithKnownNaNInfo == SelectPatternNaNBehaviour::NOT_APPLICABLE)
6429     return false;
6430   if (TrueVal == CmpRHS && FalseVal == CmpLHS) {
6431     std::swap(CmpLHS, CmpRHS);
6432     Pred = CmpInst::getSwappedPredicate(Pred);
6433     if (ResWithKnownNaNInfo == SelectPatternNaNBehaviour::RETURNS_NAN)
6434       ResWithKnownNaNInfo = SelectPatternNaNBehaviour::RETURNS_OTHER;
6435     else if (ResWithKnownNaNInfo == SelectPatternNaNBehaviour::RETURNS_OTHER)
6436       ResWithKnownNaNInfo = SelectPatternNaNBehaviour::RETURNS_NAN;
6437   }
6438   if (TrueVal != CmpLHS || FalseVal != CmpRHS)
6439     return false;
6440   // Decide what type of max/min this should be based off of the predicate.
6441   unsigned Opc = getFPMinMaxOpcForSelect(Pred, DstTy, ResWithKnownNaNInfo);
6442   if (!Opc || !isLegal({Opc, {DstTy}}))
6443     return false;
6444   // Comparisons between signed zero and zero may have different results...
6445   // unless we have fmaximum/fminimum. In that case, we know -0 < 0.
6446   if (Opc != TargetOpcode::G_FMAXIMUM && Opc != TargetOpcode::G_FMINIMUM) {
6447     // We don't know if a comparison between two 0s will give us a consistent
6448     // result. Be conservative and only proceed if at least one side is
6449     // non-zero.
6450     auto KnownNonZeroSide = getFConstantVRegValWithLookThrough(CmpLHS, MRI);
6451     if (!KnownNonZeroSide || !KnownNonZeroSide->Value.isNonZero()) {
6452       KnownNonZeroSide = getFConstantVRegValWithLookThrough(CmpRHS, MRI);
6453       if (!KnownNonZeroSide || !KnownNonZeroSide->Value.isNonZero())
6454         return false;
6455     }
6456   }
6457   MatchInfo = [=](MachineIRBuilder &B) {
6458     B.buildInstr(Opc, {Dst}, {CmpLHS, CmpRHS});
6459   };
6460   return true;
6461 }
6462 
matchSimplifySelectToMinMax(MachineInstr & MI,BuildFnTy & MatchInfo)6463 bool CombinerHelper::matchSimplifySelectToMinMax(MachineInstr &MI,
6464                                                  BuildFnTy &MatchInfo) {
6465   // TODO: Handle integer cases.
6466   assert(MI.getOpcode() == TargetOpcode::G_SELECT);
6467   // Condition may be fed by a truncated compare.
6468   Register Cond = MI.getOperand(1).getReg();
6469   Register MaybeTrunc;
6470   if (mi_match(Cond, MRI, m_OneNonDBGUse(m_GTrunc(m_Reg(MaybeTrunc)))))
6471     Cond = MaybeTrunc;
6472   Register Dst = MI.getOperand(0).getReg();
6473   Register TrueVal = MI.getOperand(2).getReg();
6474   Register FalseVal = MI.getOperand(3).getReg();
6475   return matchFPSelectToMinMax(Dst, Cond, TrueVal, FalseVal, MatchInfo);
6476 }
6477 
matchRedundantBinOpInEquality(MachineInstr & MI,BuildFnTy & MatchInfo)6478 bool CombinerHelper::matchRedundantBinOpInEquality(MachineInstr &MI,
6479                                                    BuildFnTy &MatchInfo) {
6480   assert(MI.getOpcode() == TargetOpcode::G_ICMP);
6481   // (X + Y) == X --> Y == 0
6482   // (X + Y) != X --> Y != 0
6483   // (X - Y) == X --> Y == 0
6484   // (X - Y) != X --> Y != 0
6485   // (X ^ Y) == X --> Y == 0
6486   // (X ^ Y) != X --> Y != 0
6487   Register Dst = MI.getOperand(0).getReg();
6488   CmpInst::Predicate Pred;
6489   Register X, Y, OpLHS, OpRHS;
6490   bool MatchedSub = mi_match(
6491       Dst, MRI,
6492       m_c_GICmp(m_Pred(Pred), m_Reg(X), m_GSub(m_Reg(OpLHS), m_Reg(Y))));
6493   if (MatchedSub && X != OpLHS)
6494     return false;
6495   if (!MatchedSub) {
6496     if (!mi_match(Dst, MRI,
6497                   m_c_GICmp(m_Pred(Pred), m_Reg(X),
6498                             m_any_of(m_GAdd(m_Reg(OpLHS), m_Reg(OpRHS)),
6499                                      m_GXor(m_Reg(OpLHS), m_Reg(OpRHS))))))
6500       return false;
6501     Y = X == OpLHS ? OpRHS : X == OpRHS ? OpLHS : Register();
6502   }
6503   MatchInfo = [=](MachineIRBuilder &B) {
6504     auto Zero = B.buildConstant(MRI.getType(Y), 0);
6505     B.buildICmp(Pred, Dst, Y, Zero);
6506   };
6507   return CmpInst::isEquality(Pred) && Y.isValid();
6508 }
6509 
matchShiftsTooBig(MachineInstr & MI)6510 bool CombinerHelper::matchShiftsTooBig(MachineInstr &MI) {
6511   Register ShiftReg = MI.getOperand(2).getReg();
6512   LLT ResTy = MRI.getType(MI.getOperand(0).getReg());
6513   auto IsShiftTooBig = [&](const Constant *C) {
6514     auto *CI = dyn_cast<ConstantInt>(C);
6515     return CI && CI->uge(ResTy.getScalarSizeInBits());
6516   };
6517   return matchUnaryPredicate(MRI, ShiftReg, IsShiftTooBig);
6518 }
6519 
matchCommuteConstantToRHS(MachineInstr & MI)6520 bool CombinerHelper::matchCommuteConstantToRHS(MachineInstr &MI) {
6521   unsigned LHSOpndIdx = 1;
6522   unsigned RHSOpndIdx = 2;
6523   switch (MI.getOpcode()) {
6524   case TargetOpcode::G_UADDO:
6525   case TargetOpcode::G_SADDO:
6526   case TargetOpcode::G_UMULO:
6527   case TargetOpcode::G_SMULO:
6528     LHSOpndIdx = 2;
6529     RHSOpndIdx = 3;
6530     break;
6531   default:
6532     break;
6533   }
6534   Register LHS = MI.getOperand(LHSOpndIdx).getReg();
6535   Register RHS = MI.getOperand(RHSOpndIdx).getReg();
6536   if (!getIConstantVRegVal(LHS, MRI)) {
6537     // Skip commuting if LHS is not a constant. But, LHS may be a
6538     // G_CONSTANT_FOLD_BARRIER. If so we commute as long as we don't already
6539     // have a constant on the RHS.
6540     if (MRI.getVRegDef(LHS)->getOpcode() !=
6541         TargetOpcode::G_CONSTANT_FOLD_BARRIER)
6542       return false;
6543   }
6544   // Commute as long as RHS is not a constant or G_CONSTANT_FOLD_BARRIER.
6545   return MRI.getVRegDef(RHS)->getOpcode() !=
6546              TargetOpcode::G_CONSTANT_FOLD_BARRIER &&
6547          !getIConstantVRegVal(RHS, MRI);
6548 }
6549 
matchCommuteFPConstantToRHS(MachineInstr & MI)6550 bool CombinerHelper::matchCommuteFPConstantToRHS(MachineInstr &MI) {
6551   Register LHS = MI.getOperand(1).getReg();
6552   Register RHS = MI.getOperand(2).getReg();
6553   std::optional<FPValueAndVReg> ValAndVReg;
6554   if (!mi_match(LHS, MRI, m_GFCstOrSplat(ValAndVReg)))
6555     return false;
6556   return !mi_match(RHS, MRI, m_GFCstOrSplat(ValAndVReg));
6557 }
6558 
applyCommuteBinOpOperands(MachineInstr & MI)6559 void CombinerHelper::applyCommuteBinOpOperands(MachineInstr &MI) {
6560   Observer.changingInstr(MI);
6561   unsigned LHSOpndIdx = 1;
6562   unsigned RHSOpndIdx = 2;
6563   switch (MI.getOpcode()) {
6564   case TargetOpcode::G_UADDO:
6565   case TargetOpcode::G_SADDO:
6566   case TargetOpcode::G_UMULO:
6567   case TargetOpcode::G_SMULO:
6568     LHSOpndIdx = 2;
6569     RHSOpndIdx = 3;
6570     break;
6571   default:
6572     break;
6573   }
6574   Register LHSReg = MI.getOperand(LHSOpndIdx).getReg();
6575   Register RHSReg = MI.getOperand(RHSOpndIdx).getReg();
6576   MI.getOperand(LHSOpndIdx).setReg(RHSReg);
6577   MI.getOperand(RHSOpndIdx).setReg(LHSReg);
6578   Observer.changedInstr(MI);
6579 }
6580 
isOneOrOneSplat(Register Src,bool AllowUndefs)6581 bool CombinerHelper::isOneOrOneSplat(Register Src, bool AllowUndefs) {
6582   LLT SrcTy = MRI.getType(Src);
6583   if (SrcTy.isFixedVector())
6584     return isConstantSplatVector(Src, 1, AllowUndefs);
6585   if (SrcTy.isScalar()) {
6586     if (AllowUndefs && getOpcodeDef<GImplicitDef>(Src, MRI) != nullptr)
6587       return true;
6588     auto IConstant = getIConstantVRegValWithLookThrough(Src, MRI);
6589     return IConstant && IConstant->Value == 1;
6590   }
6591   return false; // scalable vector
6592 }
6593 
isZeroOrZeroSplat(Register Src,bool AllowUndefs)6594 bool CombinerHelper::isZeroOrZeroSplat(Register Src, bool AllowUndefs) {
6595   LLT SrcTy = MRI.getType(Src);
6596   if (SrcTy.isFixedVector())
6597     return isConstantSplatVector(Src, 0, AllowUndefs);
6598   if (SrcTy.isScalar()) {
6599     if (AllowUndefs && getOpcodeDef<GImplicitDef>(Src, MRI) != nullptr)
6600       return true;
6601     auto IConstant = getIConstantVRegValWithLookThrough(Src, MRI);
6602     return IConstant && IConstant->Value == 0;
6603   }
6604   return false; // scalable vector
6605 }
6606 
6607 // Ignores COPYs during conformance checks.
6608 // FIXME scalable vectors.
isConstantSplatVector(Register Src,int64_t SplatValue,bool AllowUndefs)6609 bool CombinerHelper::isConstantSplatVector(Register Src, int64_t SplatValue,
6610                                            bool AllowUndefs) {
6611   GBuildVector *BuildVector = getOpcodeDef<GBuildVector>(Src, MRI);
6612   if (!BuildVector)
6613     return false;
6614   unsigned NumSources = BuildVector->getNumSources();
6615 
6616   for (unsigned I = 0; I < NumSources; ++I) {
6617     GImplicitDef *ImplicitDef =
6618         getOpcodeDef<GImplicitDef>(BuildVector->getSourceReg(I), MRI);
6619     if (ImplicitDef && AllowUndefs)
6620       continue;
6621     if (ImplicitDef && !AllowUndefs)
6622       return false;
6623     std::optional<ValueAndVReg> IConstant =
6624         getIConstantVRegValWithLookThrough(BuildVector->getSourceReg(I), MRI);
6625     if (IConstant && IConstant->Value == SplatValue)
6626       continue;
6627     return false;
6628   }
6629   return true;
6630 }
6631 
6632 // Ignores COPYs during lookups.
6633 // FIXME scalable vectors
6634 std::optional<APInt>
getConstantOrConstantSplatVector(Register Src)6635 CombinerHelper::getConstantOrConstantSplatVector(Register Src) {
6636   auto IConstant = getIConstantVRegValWithLookThrough(Src, MRI);
6637   if (IConstant)
6638     return IConstant->Value;
6639 
6640   GBuildVector *BuildVector = getOpcodeDef<GBuildVector>(Src, MRI);
6641   if (!BuildVector)
6642     return std::nullopt;
6643   unsigned NumSources = BuildVector->getNumSources();
6644 
6645   std::optional<APInt> Value = std::nullopt;
6646   for (unsigned I = 0; I < NumSources; ++I) {
6647     std::optional<ValueAndVReg> IConstant =
6648         getIConstantVRegValWithLookThrough(BuildVector->getSourceReg(I), MRI);
6649     if (!IConstant)
6650       return std::nullopt;
6651     if (!Value)
6652       Value = IConstant->Value;
6653     else if (*Value != IConstant->Value)
6654       return std::nullopt;
6655   }
6656   return Value;
6657 }
6658 
6659 // FIXME G_SPLAT_VECTOR
isConstantOrConstantVectorI(Register Src) const6660 bool CombinerHelper::isConstantOrConstantVectorI(Register Src) const {
6661   auto IConstant = getIConstantVRegValWithLookThrough(Src, MRI);
6662   if (IConstant)
6663     return true;
6664 
6665   GBuildVector *BuildVector = getOpcodeDef<GBuildVector>(Src, MRI);
6666   if (!BuildVector)
6667     return false;
6668 
6669   unsigned NumSources = BuildVector->getNumSources();
6670   for (unsigned I = 0; I < NumSources; ++I) {
6671     std::optional<ValueAndVReg> IConstant =
6672         getIConstantVRegValWithLookThrough(BuildVector->getSourceReg(I), MRI);
6673     if (!IConstant)
6674       return false;
6675   }
6676   return true;
6677 }
6678 
6679 // TODO: use knownbits to determine zeros
tryFoldSelectOfConstants(GSelect * Select,BuildFnTy & MatchInfo)6680 bool CombinerHelper::tryFoldSelectOfConstants(GSelect *Select,
6681                                               BuildFnTy &MatchInfo) {
6682   uint32_t Flags = Select->getFlags();
6683   Register Dest = Select->getReg(0);
6684   Register Cond = Select->getCondReg();
6685   Register True = Select->getTrueReg();
6686   Register False = Select->getFalseReg();
6687   LLT CondTy = MRI.getType(Select->getCondReg());
6688   LLT TrueTy = MRI.getType(Select->getTrueReg());
6689 
6690   // We only do this combine for scalar boolean conditions.
6691   if (CondTy != LLT::scalar(1))
6692     return false;
6693 
6694   if (TrueTy.isPointer())
6695     return false;
6696 
6697   // Both are scalars.
6698   std::optional<ValueAndVReg> TrueOpt =
6699       getIConstantVRegValWithLookThrough(True, MRI);
6700   std::optional<ValueAndVReg> FalseOpt =
6701       getIConstantVRegValWithLookThrough(False, MRI);
6702 
6703   if (!TrueOpt || !FalseOpt)
6704     return false;
6705 
6706   APInt TrueValue = TrueOpt->Value;
6707   APInt FalseValue = FalseOpt->Value;
6708 
6709   // select Cond, 1, 0 --> zext (Cond)
6710   if (TrueValue.isOne() && FalseValue.isZero()) {
6711     MatchInfo = [=](MachineIRBuilder &B) {
6712       B.setInstrAndDebugLoc(*Select);
6713       B.buildZExtOrTrunc(Dest, Cond);
6714     };
6715     return true;
6716   }
6717 
6718   // select Cond, -1, 0 --> sext (Cond)
6719   if (TrueValue.isAllOnes() && FalseValue.isZero()) {
6720     MatchInfo = [=](MachineIRBuilder &B) {
6721       B.setInstrAndDebugLoc(*Select);
6722       B.buildSExtOrTrunc(Dest, Cond);
6723     };
6724     return true;
6725   }
6726 
6727   // select Cond, 0, 1 --> zext (!Cond)
6728   if (TrueValue.isZero() && FalseValue.isOne()) {
6729     MatchInfo = [=](MachineIRBuilder &B) {
6730       B.setInstrAndDebugLoc(*Select);
6731       Register Inner = MRI.createGenericVirtualRegister(CondTy);
6732       B.buildNot(Inner, Cond);
6733       B.buildZExtOrTrunc(Dest, Inner);
6734     };
6735     return true;
6736   }
6737 
6738   // select Cond, 0, -1 --> sext (!Cond)
6739   if (TrueValue.isZero() && FalseValue.isAllOnes()) {
6740     MatchInfo = [=](MachineIRBuilder &B) {
6741       B.setInstrAndDebugLoc(*Select);
6742       Register Inner = MRI.createGenericVirtualRegister(CondTy);
6743       B.buildNot(Inner, Cond);
6744       B.buildSExtOrTrunc(Dest, Inner);
6745     };
6746     return true;
6747   }
6748 
6749   // select Cond, C1, C1-1 --> add (zext Cond), C1-1
6750   if (TrueValue - 1 == FalseValue) {
6751     MatchInfo = [=](MachineIRBuilder &B) {
6752       B.setInstrAndDebugLoc(*Select);
6753       Register Inner = MRI.createGenericVirtualRegister(TrueTy);
6754       B.buildZExtOrTrunc(Inner, Cond);
6755       B.buildAdd(Dest, Inner, False);
6756     };
6757     return true;
6758   }
6759 
6760   // select Cond, C1, C1+1 --> add (sext Cond), C1+1
6761   if (TrueValue + 1 == FalseValue) {
6762     MatchInfo = [=](MachineIRBuilder &B) {
6763       B.setInstrAndDebugLoc(*Select);
6764       Register Inner = MRI.createGenericVirtualRegister(TrueTy);
6765       B.buildSExtOrTrunc(Inner, Cond);
6766       B.buildAdd(Dest, Inner, False);
6767     };
6768     return true;
6769   }
6770 
6771   // select Cond, Pow2, 0 --> (zext Cond) << log2(Pow2)
6772   if (TrueValue.isPowerOf2() && FalseValue.isZero()) {
6773     MatchInfo = [=](MachineIRBuilder &B) {
6774       B.setInstrAndDebugLoc(*Select);
6775       Register Inner = MRI.createGenericVirtualRegister(TrueTy);
6776       B.buildZExtOrTrunc(Inner, Cond);
6777       // The shift amount must be scalar.
6778       LLT ShiftTy = TrueTy.isVector() ? TrueTy.getElementType() : TrueTy;
6779       auto ShAmtC = B.buildConstant(ShiftTy, TrueValue.exactLogBase2());
6780       B.buildShl(Dest, Inner, ShAmtC, Flags);
6781     };
6782     return true;
6783   }
6784   // select Cond, -1, C --> or (sext Cond), C
6785   if (TrueValue.isAllOnes()) {
6786     MatchInfo = [=](MachineIRBuilder &B) {
6787       B.setInstrAndDebugLoc(*Select);
6788       Register Inner = MRI.createGenericVirtualRegister(TrueTy);
6789       B.buildSExtOrTrunc(Inner, Cond);
6790       B.buildOr(Dest, Inner, False, Flags);
6791     };
6792     return true;
6793   }
6794 
6795   // select Cond, C, -1 --> or (sext (not Cond)), C
6796   if (FalseValue.isAllOnes()) {
6797     MatchInfo = [=](MachineIRBuilder &B) {
6798       B.setInstrAndDebugLoc(*Select);
6799       Register Not = MRI.createGenericVirtualRegister(CondTy);
6800       B.buildNot(Not, Cond);
6801       Register Inner = MRI.createGenericVirtualRegister(TrueTy);
6802       B.buildSExtOrTrunc(Inner, Not);
6803       B.buildOr(Dest, Inner, True, Flags);
6804     };
6805     return true;
6806   }
6807 
6808   return false;
6809 }
6810 
6811 // TODO: use knownbits to determine zeros
tryFoldBoolSelectToLogic(GSelect * Select,BuildFnTy & MatchInfo)6812 bool CombinerHelper::tryFoldBoolSelectToLogic(GSelect *Select,
6813                                               BuildFnTy &MatchInfo) {
6814   uint32_t Flags = Select->getFlags();
6815   Register DstReg = Select->getReg(0);
6816   Register Cond = Select->getCondReg();
6817   Register True = Select->getTrueReg();
6818   Register False = Select->getFalseReg();
6819   LLT CondTy = MRI.getType(Select->getCondReg());
6820   LLT TrueTy = MRI.getType(Select->getTrueReg());
6821 
6822   // Boolean or fixed vector of booleans.
6823   if (CondTy.isScalableVector() ||
6824       (CondTy.isFixedVector() &&
6825        CondTy.getElementType().getScalarSizeInBits() != 1) ||
6826       CondTy.getScalarSizeInBits() != 1)
6827     return false;
6828 
6829   if (CondTy != TrueTy)
6830     return false;
6831 
6832   // select Cond, Cond, F --> or Cond, F
6833   // select Cond, 1, F    --> or Cond, F
6834   if ((Cond == True) || isOneOrOneSplat(True, /* AllowUndefs */ true)) {
6835     MatchInfo = [=](MachineIRBuilder &B) {
6836       B.setInstrAndDebugLoc(*Select);
6837       Register Ext = MRI.createGenericVirtualRegister(TrueTy);
6838       B.buildZExtOrTrunc(Ext, Cond);
6839       auto FreezeFalse = B.buildFreeze(TrueTy, False);
6840       B.buildOr(DstReg, Ext, FreezeFalse, Flags);
6841     };
6842     return true;
6843   }
6844 
6845   // select Cond, T, Cond --> and Cond, T
6846   // select Cond, T, 0    --> and Cond, T
6847   if ((Cond == False) || isZeroOrZeroSplat(False, /* AllowUndefs */ true)) {
6848     MatchInfo = [=](MachineIRBuilder &B) {
6849       B.setInstrAndDebugLoc(*Select);
6850       Register Ext = MRI.createGenericVirtualRegister(TrueTy);
6851       B.buildZExtOrTrunc(Ext, Cond);
6852       auto FreezeTrue = B.buildFreeze(TrueTy, True);
6853       B.buildAnd(DstReg, Ext, FreezeTrue);
6854     };
6855     return true;
6856   }
6857 
6858   // select Cond, T, 1 --> or (not Cond), T
6859   if (isOneOrOneSplat(False, /* AllowUndefs */ true)) {
6860     MatchInfo = [=](MachineIRBuilder &B) {
6861       B.setInstrAndDebugLoc(*Select);
6862       // First the not.
6863       Register Inner = MRI.createGenericVirtualRegister(CondTy);
6864       B.buildNot(Inner, Cond);
6865       // Then an ext to match the destination register.
6866       Register Ext = MRI.createGenericVirtualRegister(TrueTy);
6867       B.buildZExtOrTrunc(Ext, Inner);
6868       auto FreezeTrue = B.buildFreeze(TrueTy, True);
6869       B.buildOr(DstReg, Ext, FreezeTrue, Flags);
6870     };
6871     return true;
6872   }
6873 
6874   // select Cond, 0, F --> and (not Cond), F
6875   if (isZeroOrZeroSplat(True, /* AllowUndefs */ true)) {
6876     MatchInfo = [=](MachineIRBuilder &B) {
6877       B.setInstrAndDebugLoc(*Select);
6878       // First the not.
6879       Register Inner = MRI.createGenericVirtualRegister(CondTy);
6880       B.buildNot(Inner, Cond);
6881       // Then an ext to match the destination register.
6882       Register Ext = MRI.createGenericVirtualRegister(TrueTy);
6883       B.buildZExtOrTrunc(Ext, Inner);
6884       auto FreezeFalse = B.buildFreeze(TrueTy, False);
6885       B.buildAnd(DstReg, Ext, FreezeFalse);
6886     };
6887     return true;
6888   }
6889 
6890   return false;
6891 }
6892 
matchSelectIMinMax(const MachineOperand & MO,BuildFnTy & MatchInfo)6893 bool CombinerHelper::matchSelectIMinMax(const MachineOperand &MO,
6894                                         BuildFnTy &MatchInfo) {
6895   GSelect *Select = cast<GSelect>(MRI.getVRegDef(MO.getReg()));
6896   GICmp *Cmp = cast<GICmp>(MRI.getVRegDef(Select->getCondReg()));
6897 
6898   Register DstReg = Select->getReg(0);
6899   Register True = Select->getTrueReg();
6900   Register False = Select->getFalseReg();
6901   LLT DstTy = MRI.getType(DstReg);
6902 
6903   if (DstTy.isPointer())
6904     return false;
6905 
6906   // We want to fold the icmp and replace the select.
6907   if (!MRI.hasOneNonDBGUse(Cmp->getReg(0)))
6908     return false;
6909 
6910   CmpInst::Predicate Pred = Cmp->getCond();
6911   // We need a larger or smaller predicate for
6912   // canonicalization.
6913   if (CmpInst::isEquality(Pred))
6914     return false;
6915 
6916   Register CmpLHS = Cmp->getLHSReg();
6917   Register CmpRHS = Cmp->getRHSReg();
6918 
6919   // We can swap CmpLHS and CmpRHS for higher hitrate.
6920   if (True == CmpRHS && False == CmpLHS) {
6921     std::swap(CmpLHS, CmpRHS);
6922     Pred = CmpInst::getSwappedPredicate(Pred);
6923   }
6924 
6925   // (icmp X, Y) ? X : Y -> integer minmax.
6926   // see matchSelectPattern in ValueTracking.
6927   // Legality between G_SELECT and integer minmax can differ.
6928   if (True != CmpLHS || False != CmpRHS)
6929     return false;
6930 
6931   switch (Pred) {
6932   case ICmpInst::ICMP_UGT:
6933   case ICmpInst::ICMP_UGE: {
6934     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_UMAX, DstTy}))
6935       return false;
6936     MatchInfo = [=](MachineIRBuilder &B) { B.buildUMax(DstReg, True, False); };
6937     return true;
6938   }
6939   case ICmpInst::ICMP_SGT:
6940   case ICmpInst::ICMP_SGE: {
6941     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_SMAX, DstTy}))
6942       return false;
6943     MatchInfo = [=](MachineIRBuilder &B) { B.buildSMax(DstReg, True, False); };
6944     return true;
6945   }
6946   case ICmpInst::ICMP_ULT:
6947   case ICmpInst::ICMP_ULE: {
6948     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_UMIN, DstTy}))
6949       return false;
6950     MatchInfo = [=](MachineIRBuilder &B) { B.buildUMin(DstReg, True, False); };
6951     return true;
6952   }
6953   case ICmpInst::ICMP_SLT:
6954   case ICmpInst::ICMP_SLE: {
6955     if (!isLegalOrBeforeLegalizer({TargetOpcode::G_SMIN, DstTy}))
6956       return false;
6957     MatchInfo = [=](MachineIRBuilder &B) { B.buildSMin(DstReg, True, False); };
6958     return true;
6959   }
6960   default:
6961     return false;
6962   }
6963 }
6964 
matchSelect(MachineInstr & MI,BuildFnTy & MatchInfo)6965 bool CombinerHelper::matchSelect(MachineInstr &MI, BuildFnTy &MatchInfo) {
6966   GSelect *Select = cast<GSelect>(&MI);
6967 
6968   if (tryFoldSelectOfConstants(Select, MatchInfo))
6969     return true;
6970 
6971   if (tryFoldBoolSelectToLogic(Select, MatchInfo))
6972     return true;
6973 
6974   return false;
6975 }
6976 
6977 /// Fold (icmp Pred1 V1, C1) && (icmp Pred2 V2, C2)
6978 /// or   (icmp Pred1 V1, C1) || (icmp Pred2 V2, C2)
6979 /// into a single comparison using range-based reasoning.
6980 /// see InstCombinerImpl::foldAndOrOfICmpsUsingRanges.
tryFoldAndOrOrICmpsUsingRanges(GLogicalBinOp * Logic,BuildFnTy & MatchInfo)6981 bool CombinerHelper::tryFoldAndOrOrICmpsUsingRanges(GLogicalBinOp *Logic,
6982                                                     BuildFnTy &MatchInfo) {
6983   assert(Logic->getOpcode() != TargetOpcode::G_XOR && "unexpected xor");
6984   bool IsAnd = Logic->getOpcode() == TargetOpcode::G_AND;
6985   Register DstReg = Logic->getReg(0);
6986   Register LHS = Logic->getLHSReg();
6987   Register RHS = Logic->getRHSReg();
6988   unsigned Flags = Logic->getFlags();
6989 
6990   // We need an G_ICMP on the LHS register.
6991   GICmp *Cmp1 = getOpcodeDef<GICmp>(LHS, MRI);
6992   if (!Cmp1)
6993     return false;
6994 
6995   // We need an G_ICMP on the RHS register.
6996   GICmp *Cmp2 = getOpcodeDef<GICmp>(RHS, MRI);
6997   if (!Cmp2)
6998     return false;
6999 
7000   // We want to fold the icmps.
7001   if (!MRI.hasOneNonDBGUse(Cmp1->getReg(0)) ||
7002       !MRI.hasOneNonDBGUse(Cmp2->getReg(0)))
7003     return false;
7004 
7005   APInt C1;
7006   APInt C2;
7007   std::optional<ValueAndVReg> MaybeC1 =
7008       getIConstantVRegValWithLookThrough(Cmp1->getRHSReg(), MRI);
7009   if (!MaybeC1)
7010     return false;
7011   C1 = MaybeC1->Value;
7012 
7013   std::optional<ValueAndVReg> MaybeC2 =
7014       getIConstantVRegValWithLookThrough(Cmp2->getRHSReg(), MRI);
7015   if (!MaybeC2)
7016     return false;
7017   C2 = MaybeC2->Value;
7018 
7019   Register R1 = Cmp1->getLHSReg();
7020   Register R2 = Cmp2->getLHSReg();
7021   CmpInst::Predicate Pred1 = Cmp1->getCond();
7022   CmpInst::Predicate Pred2 = Cmp2->getCond();
7023   LLT CmpTy = MRI.getType(Cmp1->getReg(0));
7024   LLT CmpOperandTy = MRI.getType(R1);
7025 
7026   if (CmpOperandTy.isPointer())
7027     return false;
7028 
7029   // We build ands, adds, and constants of type CmpOperandTy.
7030   // They must be legal to build.
7031   if (!isLegalOrBeforeLegalizer({TargetOpcode::G_AND, CmpOperandTy}) ||
7032       !isLegalOrBeforeLegalizer({TargetOpcode::G_ADD, CmpOperandTy}) ||
7033       !isConstantLegalOrBeforeLegalizer(CmpOperandTy))
7034     return false;
7035 
7036   // Look through add of a constant offset on R1, R2, or both operands. This
7037   // allows us to interpret the R + C' < C'' range idiom into a proper range.
7038   std::optional<APInt> Offset1;
7039   std::optional<APInt> Offset2;
7040   if (R1 != R2) {
7041     if (GAdd *Add = getOpcodeDef<GAdd>(R1, MRI)) {
7042       std::optional<ValueAndVReg> MaybeOffset1 =
7043           getIConstantVRegValWithLookThrough(Add->getRHSReg(), MRI);
7044       if (MaybeOffset1) {
7045         R1 = Add->getLHSReg();
7046         Offset1 = MaybeOffset1->Value;
7047       }
7048     }
7049     if (GAdd *Add = getOpcodeDef<GAdd>(R2, MRI)) {
7050       std::optional<ValueAndVReg> MaybeOffset2 =
7051           getIConstantVRegValWithLookThrough(Add->getRHSReg(), MRI);
7052       if (MaybeOffset2) {
7053         R2 = Add->getLHSReg();
7054         Offset2 = MaybeOffset2->Value;
7055       }
7056     }
7057   }
7058 
7059   if (R1 != R2)
7060     return false;
7061 
7062   // We calculate the icmp ranges including maybe offsets.
7063   ConstantRange CR1 = ConstantRange::makeExactICmpRegion(
7064       IsAnd ? ICmpInst::getInversePredicate(Pred1) : Pred1, C1);
7065   if (Offset1)
7066     CR1 = CR1.subtract(*Offset1);
7067 
7068   ConstantRange CR2 = ConstantRange::makeExactICmpRegion(
7069       IsAnd ? ICmpInst::getInversePredicate(Pred2) : Pred2, C2);
7070   if (Offset2)
7071     CR2 = CR2.subtract(*Offset2);
7072 
7073   bool CreateMask = false;
7074   APInt LowerDiff;
7075   std::optional<ConstantRange> CR = CR1.exactUnionWith(CR2);
7076   if (!CR) {
7077     // We need non-wrapping ranges.
7078     if (CR1.isWrappedSet() || CR2.isWrappedSet())
7079       return false;
7080 
7081     // Check whether we have equal-size ranges that only differ by one bit.
7082     // In that case we can apply a mask to map one range onto the other.
7083     LowerDiff = CR1.getLower() ^ CR2.getLower();
7084     APInt UpperDiff = (CR1.getUpper() - 1) ^ (CR2.getUpper() - 1);
7085     APInt CR1Size = CR1.getUpper() - CR1.getLower();
7086     if (!LowerDiff.isPowerOf2() || LowerDiff != UpperDiff ||
7087         CR1Size != CR2.getUpper() - CR2.getLower())
7088       return false;
7089 
7090     CR = CR1.getLower().ult(CR2.getLower()) ? CR1 : CR2;
7091     CreateMask = true;
7092   }
7093 
7094   if (IsAnd)
7095     CR = CR->inverse();
7096 
7097   CmpInst::Predicate NewPred;
7098   APInt NewC, Offset;
7099   CR->getEquivalentICmp(NewPred, NewC, Offset);
7100 
7101   // We take the result type of one of the original icmps, CmpTy, for
7102   // the to be build icmp. The operand type, CmpOperandTy, is used for
7103   // the other instructions and constants to be build. The types of
7104   // the parameters and output are the same for add and and.  CmpTy
7105   // and the type of DstReg might differ. That is why we zext or trunc
7106   // the icmp into the destination register.
7107 
7108   MatchInfo = [=](MachineIRBuilder &B) {
7109     if (CreateMask && Offset != 0) {
7110       auto TildeLowerDiff = B.buildConstant(CmpOperandTy, ~LowerDiff);
7111       auto And = B.buildAnd(CmpOperandTy, R1, TildeLowerDiff); // the mask.
7112       auto OffsetC = B.buildConstant(CmpOperandTy, Offset);
7113       auto Add = B.buildAdd(CmpOperandTy, And, OffsetC, Flags);
7114       auto NewCon = B.buildConstant(CmpOperandTy, NewC);
7115       auto ICmp = B.buildICmp(NewPred, CmpTy, Add, NewCon);
7116       B.buildZExtOrTrunc(DstReg, ICmp);
7117     } else if (CreateMask && Offset == 0) {
7118       auto TildeLowerDiff = B.buildConstant(CmpOperandTy, ~LowerDiff);
7119       auto And = B.buildAnd(CmpOperandTy, R1, TildeLowerDiff); // the mask.
7120       auto NewCon = B.buildConstant(CmpOperandTy, NewC);
7121       auto ICmp = B.buildICmp(NewPred, CmpTy, And, NewCon);
7122       B.buildZExtOrTrunc(DstReg, ICmp);
7123     } else if (!CreateMask && Offset != 0) {
7124       auto OffsetC = B.buildConstant(CmpOperandTy, Offset);
7125       auto Add = B.buildAdd(CmpOperandTy, R1, OffsetC, Flags);
7126       auto NewCon = B.buildConstant(CmpOperandTy, NewC);
7127       auto ICmp = B.buildICmp(NewPred, CmpTy, Add, NewCon);
7128       B.buildZExtOrTrunc(DstReg, ICmp);
7129     } else if (!CreateMask && Offset == 0) {
7130       auto NewCon = B.buildConstant(CmpOperandTy, NewC);
7131       auto ICmp = B.buildICmp(NewPred, CmpTy, R1, NewCon);
7132       B.buildZExtOrTrunc(DstReg, ICmp);
7133     } else {
7134       llvm_unreachable("unexpected configuration of CreateMask and Offset");
7135     }
7136   };
7137   return true;
7138 }
7139 
tryFoldLogicOfFCmps(GLogicalBinOp * Logic,BuildFnTy & MatchInfo)7140 bool CombinerHelper::tryFoldLogicOfFCmps(GLogicalBinOp *Logic,
7141                                          BuildFnTy &MatchInfo) {
7142   assert(Logic->getOpcode() != TargetOpcode::G_XOR && "unexpecte xor");
7143   Register DestReg = Logic->getReg(0);
7144   Register LHS = Logic->getLHSReg();
7145   Register RHS = Logic->getRHSReg();
7146   bool IsAnd = Logic->getOpcode() == TargetOpcode::G_AND;
7147 
7148   // We need a compare on the LHS register.
7149   GFCmp *Cmp1 = getOpcodeDef<GFCmp>(LHS, MRI);
7150   if (!Cmp1)
7151     return false;
7152 
7153   // We need a compare on the RHS register.
7154   GFCmp *Cmp2 = getOpcodeDef<GFCmp>(RHS, MRI);
7155   if (!Cmp2)
7156     return false;
7157 
7158   LLT CmpTy = MRI.getType(Cmp1->getReg(0));
7159   LLT CmpOperandTy = MRI.getType(Cmp1->getLHSReg());
7160 
7161   // We build one fcmp, want to fold the fcmps, replace the logic op,
7162   // and the fcmps must have the same shape.
7163   if (!isLegalOrBeforeLegalizer(
7164           {TargetOpcode::G_FCMP, {CmpTy, CmpOperandTy}}) ||
7165       !MRI.hasOneNonDBGUse(Logic->getReg(0)) ||
7166       !MRI.hasOneNonDBGUse(Cmp1->getReg(0)) ||
7167       !MRI.hasOneNonDBGUse(Cmp2->getReg(0)) ||
7168       MRI.getType(Cmp1->getLHSReg()) != MRI.getType(Cmp2->getLHSReg()))
7169     return false;
7170 
7171   CmpInst::Predicate PredL = Cmp1->getCond();
7172   CmpInst::Predicate PredR = Cmp2->getCond();
7173   Register LHS0 = Cmp1->getLHSReg();
7174   Register LHS1 = Cmp1->getRHSReg();
7175   Register RHS0 = Cmp2->getLHSReg();
7176   Register RHS1 = Cmp2->getRHSReg();
7177 
7178   if (LHS0 == RHS1 && LHS1 == RHS0) {
7179     // Swap RHS operands to match LHS.
7180     PredR = CmpInst::getSwappedPredicate(PredR);
7181     std::swap(RHS0, RHS1);
7182   }
7183 
7184   if (LHS0 == RHS0 && LHS1 == RHS1) {
7185     // We determine the new predicate.
7186     unsigned CmpCodeL = getFCmpCode(PredL);
7187     unsigned CmpCodeR = getFCmpCode(PredR);
7188     unsigned NewPred = IsAnd ? CmpCodeL & CmpCodeR : CmpCodeL | CmpCodeR;
7189     unsigned Flags = Cmp1->getFlags() | Cmp2->getFlags();
7190     MatchInfo = [=](MachineIRBuilder &B) {
7191       // The fcmp predicates fill the lower part of the enum.
7192       FCmpInst::Predicate Pred = static_cast<FCmpInst::Predicate>(NewPred);
7193       if (Pred == FCmpInst::FCMP_FALSE &&
7194           isConstantLegalOrBeforeLegalizer(CmpTy)) {
7195         auto False = B.buildConstant(CmpTy, 0);
7196         B.buildZExtOrTrunc(DestReg, False);
7197       } else if (Pred == FCmpInst::FCMP_TRUE &&
7198                  isConstantLegalOrBeforeLegalizer(CmpTy)) {
7199         auto True =
7200             B.buildConstant(CmpTy, getICmpTrueVal(getTargetLowering(),
7201                                                   CmpTy.isVector() /*isVector*/,
7202                                                   true /*isFP*/));
7203         B.buildZExtOrTrunc(DestReg, True);
7204       } else { // We take the predicate without predicate optimizations.
7205         auto Cmp = B.buildFCmp(Pred, CmpTy, LHS0, LHS1, Flags);
7206         B.buildZExtOrTrunc(DestReg, Cmp);
7207       }
7208     };
7209     return true;
7210   }
7211 
7212   return false;
7213 }
7214 
matchAnd(MachineInstr & MI,BuildFnTy & MatchInfo)7215 bool CombinerHelper::matchAnd(MachineInstr &MI, BuildFnTy &MatchInfo) {
7216   GAnd *And = cast<GAnd>(&MI);
7217 
7218   if (tryFoldAndOrOrICmpsUsingRanges(And, MatchInfo))
7219     return true;
7220 
7221   if (tryFoldLogicOfFCmps(And, MatchInfo))
7222     return true;
7223 
7224   return false;
7225 }
7226 
matchOr(MachineInstr & MI,BuildFnTy & MatchInfo)7227 bool CombinerHelper::matchOr(MachineInstr &MI, BuildFnTy &MatchInfo) {
7228   GOr *Or = cast<GOr>(&MI);
7229 
7230   if (tryFoldAndOrOrICmpsUsingRanges(Or, MatchInfo))
7231     return true;
7232 
7233   if (tryFoldLogicOfFCmps(Or, MatchInfo))
7234     return true;
7235 
7236   return false;
7237 }
7238 
matchAddOverflow(MachineInstr & MI,BuildFnTy & MatchInfo)7239 bool CombinerHelper::matchAddOverflow(MachineInstr &MI, BuildFnTy &MatchInfo) {
7240   GAddCarryOut *Add = cast<GAddCarryOut>(&MI);
7241 
7242   // Addo has no flags
7243   Register Dst = Add->getReg(0);
7244   Register Carry = Add->getReg(1);
7245   Register LHS = Add->getLHSReg();
7246   Register RHS = Add->getRHSReg();
7247   bool IsSigned = Add->isSigned();
7248   LLT DstTy = MRI.getType(Dst);
7249   LLT CarryTy = MRI.getType(Carry);
7250 
7251   // Fold addo, if the carry is dead -> add, undef.
7252   if (MRI.use_nodbg_empty(Carry) &&
7253       isLegalOrBeforeLegalizer({TargetOpcode::G_ADD, {DstTy}})) {
7254     MatchInfo = [=](MachineIRBuilder &B) {
7255       B.buildAdd(Dst, LHS, RHS);
7256       B.buildUndef(Carry);
7257     };
7258     return true;
7259   }
7260 
7261   // Canonicalize constant to RHS.
7262   if (isConstantOrConstantVectorI(LHS) && !isConstantOrConstantVectorI(RHS)) {
7263     if (IsSigned) {
7264       MatchInfo = [=](MachineIRBuilder &B) {
7265         B.buildSAddo(Dst, Carry, RHS, LHS);
7266       };
7267       return true;
7268     }
7269     // !IsSigned
7270     MatchInfo = [=](MachineIRBuilder &B) {
7271       B.buildUAddo(Dst, Carry, RHS, LHS);
7272     };
7273     return true;
7274   }
7275 
7276   std::optional<APInt> MaybeLHS = getConstantOrConstantSplatVector(LHS);
7277   std::optional<APInt> MaybeRHS = getConstantOrConstantSplatVector(RHS);
7278 
7279   // Fold addo(c1, c2) -> c3, carry.
7280   if (MaybeLHS && MaybeRHS && isConstantLegalOrBeforeLegalizer(DstTy) &&
7281       isConstantLegalOrBeforeLegalizer(CarryTy)) {
7282     bool Overflow;
7283     APInt Result = IsSigned ? MaybeLHS->sadd_ov(*MaybeRHS, Overflow)
7284                             : MaybeLHS->uadd_ov(*MaybeRHS, Overflow);
7285     MatchInfo = [=](MachineIRBuilder &B) {
7286       B.buildConstant(Dst, Result);
7287       B.buildConstant(Carry, Overflow);
7288     };
7289     return true;
7290   }
7291 
7292   // Fold (addo x, 0) -> x, no carry
7293   if (MaybeRHS && *MaybeRHS == 0 && isConstantLegalOrBeforeLegalizer(CarryTy)) {
7294     MatchInfo = [=](MachineIRBuilder &B) {
7295       B.buildCopy(Dst, LHS);
7296       B.buildConstant(Carry, 0);
7297     };
7298     return true;
7299   }
7300 
7301   // Given 2 constant operands whose sum does not overflow:
7302   // uaddo (X +nuw C0), C1 -> uaddo X, C0 + C1
7303   // saddo (X +nsw C0), C1 -> saddo X, C0 + C1
7304   GAdd *AddLHS = getOpcodeDef<GAdd>(LHS, MRI);
7305   if (MaybeRHS && AddLHS && MRI.hasOneNonDBGUse(Add->getReg(0)) &&
7306       ((IsSigned && AddLHS->getFlag(MachineInstr::MIFlag::NoSWrap)) ||
7307        (!IsSigned && AddLHS->getFlag(MachineInstr::MIFlag::NoUWrap)))) {
7308     std::optional<APInt> MaybeAddRHS =
7309         getConstantOrConstantSplatVector(AddLHS->getRHSReg());
7310     if (MaybeAddRHS) {
7311       bool Overflow;
7312       APInt NewC = IsSigned ? MaybeAddRHS->sadd_ov(*MaybeRHS, Overflow)
7313                             : MaybeAddRHS->uadd_ov(*MaybeRHS, Overflow);
7314       if (!Overflow && isConstantLegalOrBeforeLegalizer(DstTy)) {
7315         if (IsSigned) {
7316           MatchInfo = [=](MachineIRBuilder &B) {
7317             auto ConstRHS = B.buildConstant(DstTy, NewC);
7318             B.buildSAddo(Dst, Carry, AddLHS->getLHSReg(), ConstRHS);
7319           };
7320           return true;
7321         }
7322         // !IsSigned
7323         MatchInfo = [=](MachineIRBuilder &B) {
7324           auto ConstRHS = B.buildConstant(DstTy, NewC);
7325           B.buildUAddo(Dst, Carry, AddLHS->getLHSReg(), ConstRHS);
7326         };
7327         return true;
7328       }
7329     }
7330   };
7331 
7332   // We try to combine addo to non-overflowing add.
7333   if (!isLegalOrBeforeLegalizer({TargetOpcode::G_ADD, {DstTy}}) ||
7334       !isConstantLegalOrBeforeLegalizer(CarryTy))
7335     return false;
7336 
7337   // We try to combine uaddo to non-overflowing add.
7338   if (!IsSigned) {
7339     ConstantRange CRLHS =
7340         ConstantRange::fromKnownBits(KB->getKnownBits(LHS), /*IsSigned=*/false);
7341     ConstantRange CRRHS =
7342         ConstantRange::fromKnownBits(KB->getKnownBits(RHS), /*IsSigned=*/false);
7343 
7344     switch (CRLHS.unsignedAddMayOverflow(CRRHS)) {
7345     case ConstantRange::OverflowResult::MayOverflow:
7346       return false;
7347     case ConstantRange::OverflowResult::NeverOverflows: {
7348       MatchInfo = [=](MachineIRBuilder &B) {
7349         B.buildAdd(Dst, LHS, RHS, MachineInstr::MIFlag::NoUWrap);
7350         B.buildConstant(Carry, 0);
7351       };
7352       return true;
7353     }
7354     case ConstantRange::OverflowResult::AlwaysOverflowsLow:
7355     case ConstantRange::OverflowResult::AlwaysOverflowsHigh: {
7356       MatchInfo = [=](MachineIRBuilder &B) {
7357         B.buildAdd(Dst, LHS, RHS);
7358         B.buildConstant(Carry, 1);
7359       };
7360       return true;
7361     }
7362     }
7363     return false;
7364   }
7365 
7366   // We try to combine saddo to non-overflowing add.
7367 
7368   // If LHS and RHS each have at least two sign bits, then there is no signed
7369   // overflow.
7370   if (KB->computeNumSignBits(RHS) > 1 && KB->computeNumSignBits(LHS) > 1) {
7371     MatchInfo = [=](MachineIRBuilder &B) {
7372       B.buildAdd(Dst, LHS, RHS, MachineInstr::MIFlag::NoSWrap);
7373       B.buildConstant(Carry, 0);
7374     };
7375     return true;
7376   }
7377 
7378   ConstantRange CRLHS =
7379       ConstantRange::fromKnownBits(KB->getKnownBits(LHS), /*IsSigned=*/true);
7380   ConstantRange CRRHS =
7381       ConstantRange::fromKnownBits(KB->getKnownBits(RHS), /*IsSigned=*/true);
7382 
7383   switch (CRLHS.signedAddMayOverflow(CRRHS)) {
7384   case ConstantRange::OverflowResult::MayOverflow:
7385     return false;
7386   case ConstantRange::OverflowResult::NeverOverflows: {
7387     MatchInfo = [=](MachineIRBuilder &B) {
7388       B.buildAdd(Dst, LHS, RHS, MachineInstr::MIFlag::NoSWrap);
7389       B.buildConstant(Carry, 0);
7390     };
7391     return true;
7392   }
7393   case ConstantRange::OverflowResult::AlwaysOverflowsLow:
7394   case ConstantRange::OverflowResult::AlwaysOverflowsHigh: {
7395     MatchInfo = [=](MachineIRBuilder &B) {
7396       B.buildAdd(Dst, LHS, RHS);
7397       B.buildConstant(Carry, 1);
7398     };
7399     return true;
7400   }
7401   }
7402 
7403   return false;
7404 }
7405 
applyBuildFnMO(const MachineOperand & MO,BuildFnTy & MatchInfo)7406 void CombinerHelper::applyBuildFnMO(const MachineOperand &MO,
7407                                     BuildFnTy &MatchInfo) {
7408   MachineInstr *Root = getDefIgnoringCopies(MO.getReg(), MRI);
7409   MatchInfo(Builder);
7410   Root->eraseFromParent();
7411 }
7412 
matchFPowIExpansion(MachineInstr & MI,int64_t Exponent)7413 bool CombinerHelper::matchFPowIExpansion(MachineInstr &MI, int64_t Exponent) {
7414   bool OptForSize = MI.getMF()->getFunction().hasOptSize();
7415   return getTargetLowering().isBeneficialToExpandPowI(Exponent, OptForSize);
7416 }
7417 
applyExpandFPowI(MachineInstr & MI,int64_t Exponent)7418 void CombinerHelper::applyExpandFPowI(MachineInstr &MI, int64_t Exponent) {
7419   auto [Dst, Base] = MI.getFirst2Regs();
7420   LLT Ty = MRI.getType(Dst);
7421   int64_t ExpVal = Exponent;
7422 
7423   if (ExpVal == 0) {
7424     Builder.buildFConstant(Dst, 1.0);
7425     MI.removeFromParent();
7426     return;
7427   }
7428 
7429   if (ExpVal < 0)
7430     ExpVal = -ExpVal;
7431 
7432   // We use the simple binary decomposition method from SelectionDAG ExpandPowI
7433   // to generate the multiply sequence. There are more optimal ways to do this
7434   // (for example, powi(x,15) generates one more multiply than it should), but
7435   // this has the benefit of being both really simple and much better than a
7436   // libcall.
7437   std::optional<SrcOp> Res;
7438   SrcOp CurSquare = Base;
7439   while (ExpVal > 0) {
7440     if (ExpVal & 1) {
7441       if (!Res)
7442         Res = CurSquare;
7443       else
7444         Res = Builder.buildFMul(Ty, *Res, CurSquare);
7445     }
7446 
7447     CurSquare = Builder.buildFMul(Ty, CurSquare, CurSquare);
7448     ExpVal >>= 1;
7449   }
7450 
7451   // If the original exponent was negative, invert the result, producing
7452   // 1/(x*x*x).
7453   if (Exponent < 0)
7454     Res = Builder.buildFDiv(Ty, Builder.buildFConstant(Ty, 1.0), *Res,
7455                             MI.getFlags());
7456 
7457   Builder.buildCopy(Dst, *Res);
7458   MI.eraseFromParent();
7459 }
7460 
matchSextOfTrunc(const MachineOperand & MO,BuildFnTy & MatchInfo)7461 bool CombinerHelper::matchSextOfTrunc(const MachineOperand &MO,
7462                                       BuildFnTy &MatchInfo) {
7463   GSext *Sext = cast<GSext>(getDefIgnoringCopies(MO.getReg(), MRI));
7464   GTrunc *Trunc = cast<GTrunc>(getDefIgnoringCopies(Sext->getSrcReg(), MRI));
7465 
7466   Register Dst = Sext->getReg(0);
7467   Register Src = Trunc->getSrcReg();
7468 
7469   LLT DstTy = MRI.getType(Dst);
7470   LLT SrcTy = MRI.getType(Src);
7471 
7472   if (DstTy == SrcTy) {
7473     MatchInfo = [=](MachineIRBuilder &B) { B.buildCopy(Dst, Src); };
7474     return true;
7475   }
7476 
7477   if (DstTy.getScalarSizeInBits() < SrcTy.getScalarSizeInBits() &&
7478       isLegalOrBeforeLegalizer({TargetOpcode::G_TRUNC, {DstTy, SrcTy}})) {
7479     MatchInfo = [=](MachineIRBuilder &B) {
7480       B.buildTrunc(Dst, Src, MachineInstr::MIFlag::NoSWrap);
7481     };
7482     return true;
7483   }
7484 
7485   if (DstTy.getScalarSizeInBits() > SrcTy.getScalarSizeInBits() &&
7486       isLegalOrBeforeLegalizer({TargetOpcode::G_SEXT, {DstTy, SrcTy}})) {
7487     MatchInfo = [=](MachineIRBuilder &B) { B.buildSExt(Dst, Src); };
7488     return true;
7489   }
7490 
7491   return false;
7492 }
7493 
matchZextOfTrunc(const MachineOperand & MO,BuildFnTy & MatchInfo)7494 bool CombinerHelper::matchZextOfTrunc(const MachineOperand &MO,
7495                                       BuildFnTy &MatchInfo) {
7496   GZext *Zext = cast<GZext>(getDefIgnoringCopies(MO.getReg(), MRI));
7497   GTrunc *Trunc = cast<GTrunc>(getDefIgnoringCopies(Zext->getSrcReg(), MRI));
7498 
7499   Register Dst = Zext->getReg(0);
7500   Register Src = Trunc->getSrcReg();
7501 
7502   LLT DstTy = MRI.getType(Dst);
7503   LLT SrcTy = MRI.getType(Src);
7504 
7505   if (DstTy == SrcTy) {
7506     MatchInfo = [=](MachineIRBuilder &B) { B.buildCopy(Dst, Src); };
7507     return true;
7508   }
7509 
7510   if (DstTy.getScalarSizeInBits() < SrcTy.getScalarSizeInBits() &&
7511       isLegalOrBeforeLegalizer({TargetOpcode::G_TRUNC, {DstTy, SrcTy}})) {
7512     MatchInfo = [=](MachineIRBuilder &B) {
7513       B.buildTrunc(Dst, Src, MachineInstr::MIFlag::NoUWrap);
7514     };
7515     return true;
7516   }
7517 
7518   if (DstTy.getScalarSizeInBits() > SrcTy.getScalarSizeInBits() &&
7519       isLegalOrBeforeLegalizer({TargetOpcode::G_ZEXT, {DstTy, SrcTy}})) {
7520     MatchInfo = [=](MachineIRBuilder &B) {
7521       B.buildZExt(Dst, Src, MachineInstr::MIFlag::NonNeg);
7522     };
7523     return true;
7524   }
7525 
7526   return false;
7527 }
7528 
matchNonNegZext(const MachineOperand & MO,BuildFnTy & MatchInfo)7529 bool CombinerHelper::matchNonNegZext(const MachineOperand &MO,
7530                                      BuildFnTy &MatchInfo) {
7531   GZext *Zext = cast<GZext>(MRI.getVRegDef(MO.getReg()));
7532 
7533   Register Dst = Zext->getReg(0);
7534   Register Src = Zext->getSrcReg();
7535 
7536   LLT DstTy = MRI.getType(Dst);
7537   LLT SrcTy = MRI.getType(Src);
7538   const auto &TLI = getTargetLowering();
7539 
7540   // Convert zext nneg to sext if sext is the preferred form for the target.
7541   if (isLegalOrBeforeLegalizer({TargetOpcode::G_SEXT, {DstTy, SrcTy}}) &&
7542       TLI.isSExtCheaperThanZExt(getMVTForLLT(SrcTy), getMVTForLLT(DstTy))) {
7543     MatchInfo = [=](MachineIRBuilder &B) { B.buildSExt(Dst, Src); };
7544     return true;
7545   }
7546 
7547   return false;
7548 }
7549