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/CodeGen/GlobalISel/Combiner.h"
10 #include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
11 #include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
12 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
13 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
14 #include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
15 #include "llvm/CodeGen/GlobalISel/Utils.h"
16 #include "llvm/CodeGen/MachineDominators.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineInstr.h"
19 #include "llvm/CodeGen/MachineRegisterInfo.h"
20 #include "llvm/CodeGen/TargetInstrInfo.h"
21 #include "llvm/CodeGen/TargetLowering.h"
22 #include "llvm/Support/MathExtras.h"
23 #include "llvm/Target/TargetMachine.h"
24 
25 #define DEBUG_TYPE "gi-combiner"
26 
27 using namespace llvm;
28 using namespace MIPatternMatch;
29 
30 // Option to allow testing of the combiner while no targets know about indexed
31 // addressing.
32 static cl::opt<bool>
33     ForceLegalIndexing("force-legal-indexing", cl::Hidden, cl::init(false),
34                        cl::desc("Force all indexed operations to be "
35                                 "legal for the GlobalISel combiner"));
36 
37 
38 CombinerHelper::CombinerHelper(GISelChangeObserver &Observer,
39                                MachineIRBuilder &B, GISelKnownBits *KB,
40                                MachineDominatorTree *MDT,
41                                const LegalizerInfo *LI)
42     : Builder(B), MRI(Builder.getMF().getRegInfo()), Observer(Observer),
43       KB(KB), MDT(MDT), LI(LI) {
44   (void)this->KB;
45 }
46 
47 void CombinerHelper::replaceRegWith(MachineRegisterInfo &MRI, Register FromReg,
48                                     Register ToReg) const {
49   Observer.changingAllUsesOfReg(MRI, FromReg);
50 
51   if (MRI.constrainRegAttrs(ToReg, FromReg))
52     MRI.replaceRegWith(FromReg, ToReg);
53   else
54     Builder.buildCopy(ToReg, FromReg);
55 
56   Observer.finishedChangingAllUsesOfReg();
57 }
58 
59 void CombinerHelper::replaceRegOpWith(MachineRegisterInfo &MRI,
60                                       MachineOperand &FromRegOp,
61                                       Register ToReg) const {
62   assert(FromRegOp.getParent() && "Expected an operand in an MI");
63   Observer.changingInstr(*FromRegOp.getParent());
64 
65   FromRegOp.setReg(ToReg);
66 
67   Observer.changedInstr(*FromRegOp.getParent());
68 }
69 
70 bool CombinerHelper::tryCombineCopy(MachineInstr &MI) {
71   if (matchCombineCopy(MI)) {
72     applyCombineCopy(MI);
73     return true;
74   }
75   return false;
76 }
77 bool CombinerHelper::matchCombineCopy(MachineInstr &MI) {
78   if (MI.getOpcode() != TargetOpcode::COPY)
79     return false;
80   Register DstReg = MI.getOperand(0).getReg();
81   Register SrcReg = MI.getOperand(1).getReg();
82   return canReplaceReg(DstReg, SrcReg, MRI);
83 }
84 void CombinerHelper::applyCombineCopy(MachineInstr &MI) {
85   Register DstReg = MI.getOperand(0).getReg();
86   Register SrcReg = MI.getOperand(1).getReg();
87   MI.eraseFromParent();
88   replaceRegWith(MRI, DstReg, SrcReg);
89 }
90 
91 bool CombinerHelper::tryCombineConcatVectors(MachineInstr &MI) {
92   bool IsUndef = false;
93   SmallVector<Register, 4> Ops;
94   if (matchCombineConcatVectors(MI, IsUndef, Ops)) {
95     applyCombineConcatVectors(MI, IsUndef, Ops);
96     return true;
97   }
98   return false;
99 }
100 
101 bool CombinerHelper::matchCombineConcatVectors(MachineInstr &MI, bool &IsUndef,
102                                                SmallVectorImpl<Register> &Ops) {
103   assert(MI.getOpcode() == TargetOpcode::G_CONCAT_VECTORS &&
104          "Invalid instruction");
105   IsUndef = true;
106   MachineInstr *Undef = nullptr;
107 
108   // Walk over all the operands of concat vectors and check if they are
109   // build_vector themselves or undef.
110   // Then collect their operands in Ops.
111   for (const MachineOperand &MO : MI.uses()) {
112     Register Reg = MO.getReg();
113     MachineInstr *Def = MRI.getVRegDef(Reg);
114     assert(Def && "Operand not defined");
115     switch (Def->getOpcode()) {
116     case TargetOpcode::G_BUILD_VECTOR:
117       IsUndef = false;
118       // Remember the operands of the build_vector to fold
119       // them into the yet-to-build flattened concat vectors.
120       for (const MachineOperand &BuildVecMO : Def->uses())
121         Ops.push_back(BuildVecMO.getReg());
122       break;
123     case TargetOpcode::G_IMPLICIT_DEF: {
124       LLT OpType = MRI.getType(Reg);
125       // Keep one undef value for all the undef operands.
126       if (!Undef) {
127         Builder.setInsertPt(*MI.getParent(), MI);
128         Undef = Builder.buildUndef(OpType.getScalarType());
129       }
130       assert(MRI.getType(Undef->getOperand(0).getReg()) ==
131                  OpType.getScalarType() &&
132              "All undefs should have the same type");
133       // Break the undef vector in as many scalar elements as needed
134       // for the flattening.
135       for (unsigned EltIdx = 0, EltEnd = OpType.getNumElements();
136            EltIdx != EltEnd; ++EltIdx)
137         Ops.push_back(Undef->getOperand(0).getReg());
138       break;
139     }
140     default:
141       return false;
142     }
143   }
144   return true;
145 }
146 void CombinerHelper::applyCombineConcatVectors(
147     MachineInstr &MI, bool IsUndef, const ArrayRef<Register> Ops) {
148   // We determined that the concat_vectors can be flatten.
149   // Generate the flattened build_vector.
150   Register DstReg = MI.getOperand(0).getReg();
151   Builder.setInsertPt(*MI.getParent(), MI);
152   Register NewDstReg = MRI.cloneVirtualRegister(DstReg);
153 
154   // Note: IsUndef is sort of redundant. We could have determine it by
155   // checking that at all Ops are undef.  Alternatively, we could have
156   // generate a build_vector of undefs and rely on another combine to
157   // clean that up.  For now, given we already gather this information
158   // in tryCombineConcatVectors, just save compile time and issue the
159   // right thing.
160   if (IsUndef)
161     Builder.buildUndef(NewDstReg);
162   else
163     Builder.buildBuildVector(NewDstReg, Ops);
164   MI.eraseFromParent();
165   replaceRegWith(MRI, DstReg, NewDstReg);
166 }
167 
168 bool CombinerHelper::tryCombineShuffleVector(MachineInstr &MI) {
169   SmallVector<Register, 4> Ops;
170   if (matchCombineShuffleVector(MI, Ops)) {
171     applyCombineShuffleVector(MI, Ops);
172     return true;
173   }
174   return false;
175 }
176 
177 bool CombinerHelper::matchCombineShuffleVector(MachineInstr &MI,
178                                                SmallVectorImpl<Register> &Ops) {
179   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR &&
180          "Invalid instruction kind");
181   LLT DstType = MRI.getType(MI.getOperand(0).getReg());
182   Register Src1 = MI.getOperand(1).getReg();
183   LLT SrcType = MRI.getType(Src1);
184   // As bizarre as it may look, shuffle vector can actually produce
185   // scalar! This is because at the IR level a <1 x ty> shuffle
186   // vector is perfectly valid.
187   unsigned DstNumElts = DstType.isVector() ? DstType.getNumElements() : 1;
188   unsigned SrcNumElts = SrcType.isVector() ? SrcType.getNumElements() : 1;
189 
190   // If the resulting vector is smaller than the size of the source
191   // vectors being concatenated, we won't be able to replace the
192   // shuffle vector into a concat_vectors.
193   //
194   // Note: We may still be able to produce a concat_vectors fed by
195   //       extract_vector_elt and so on. It is less clear that would
196   //       be better though, so don't bother for now.
197   //
198   // If the destination is a scalar, the size of the sources doesn't
199   // matter. we will lower the shuffle to a plain copy. This will
200   // work only if the source and destination have the same size. But
201   // that's covered by the next condition.
202   //
203   // TODO: If the size between the source and destination don't match
204   //       we could still emit an extract vector element in that case.
205   if (DstNumElts < 2 * SrcNumElts && DstNumElts != 1)
206     return false;
207 
208   // Check that the shuffle mask can be broken evenly between the
209   // different sources.
210   if (DstNumElts % SrcNumElts != 0)
211     return false;
212 
213   // Mask length is a multiple of the source vector length.
214   // Check if the shuffle is some kind of concatenation of the input
215   // vectors.
216   unsigned NumConcat = DstNumElts / SrcNumElts;
217   SmallVector<int, 8> ConcatSrcs(NumConcat, -1);
218   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
219   for (unsigned i = 0; i != DstNumElts; ++i) {
220     int Idx = Mask[i];
221     // Undef value.
222     if (Idx < 0)
223       continue;
224     // Ensure the indices in each SrcType sized piece are sequential and that
225     // the same source is used for the whole piece.
226     if ((Idx % SrcNumElts != (i % SrcNumElts)) ||
227         (ConcatSrcs[i / SrcNumElts] >= 0 &&
228          ConcatSrcs[i / SrcNumElts] != (int)(Idx / SrcNumElts)))
229       return false;
230     // Remember which source this index came from.
231     ConcatSrcs[i / SrcNumElts] = Idx / SrcNumElts;
232   }
233 
234   // The shuffle is concatenating multiple vectors together.
235   // Collect the different operands for that.
236   Register UndefReg;
237   Register Src2 = MI.getOperand(2).getReg();
238   for (auto Src : ConcatSrcs) {
239     if (Src < 0) {
240       if (!UndefReg) {
241         Builder.setInsertPt(*MI.getParent(), MI);
242         UndefReg = Builder.buildUndef(SrcType).getReg(0);
243       }
244       Ops.push_back(UndefReg);
245     } else if (Src == 0)
246       Ops.push_back(Src1);
247     else
248       Ops.push_back(Src2);
249   }
250   return true;
251 }
252 
253 void CombinerHelper::applyCombineShuffleVector(MachineInstr &MI,
254                                                const ArrayRef<Register> Ops) {
255   Register DstReg = MI.getOperand(0).getReg();
256   Builder.setInsertPt(*MI.getParent(), MI);
257   Register NewDstReg = MRI.cloneVirtualRegister(DstReg);
258 
259   if (Ops.size() == 1)
260     Builder.buildCopy(NewDstReg, Ops[0]);
261   else
262     Builder.buildMerge(NewDstReg, Ops);
263 
264   MI.eraseFromParent();
265   replaceRegWith(MRI, DstReg, NewDstReg);
266 }
267 
268 namespace {
269 
270 /// Select a preference between two uses. CurrentUse is the current preference
271 /// while *ForCandidate is attributes of the candidate under consideration.
272 PreferredTuple ChoosePreferredUse(PreferredTuple &CurrentUse,
273                                   const LLT TyForCandidate,
274                                   unsigned OpcodeForCandidate,
275                                   MachineInstr *MIForCandidate) {
276   if (!CurrentUse.Ty.isValid()) {
277     if (CurrentUse.ExtendOpcode == OpcodeForCandidate ||
278         CurrentUse.ExtendOpcode == TargetOpcode::G_ANYEXT)
279       return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
280     return CurrentUse;
281   }
282 
283   // We permit the extend to hoist through basic blocks but this is only
284   // sensible if the target has extending loads. If you end up lowering back
285   // into a load and extend during the legalizer then the end result is
286   // hoisting the extend up to the load.
287 
288   // Prefer defined extensions to undefined extensions as these are more
289   // likely to reduce the number of instructions.
290   if (OpcodeForCandidate == TargetOpcode::G_ANYEXT &&
291       CurrentUse.ExtendOpcode != TargetOpcode::G_ANYEXT)
292     return CurrentUse;
293   else if (CurrentUse.ExtendOpcode == TargetOpcode::G_ANYEXT &&
294            OpcodeForCandidate != TargetOpcode::G_ANYEXT)
295     return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
296 
297   // Prefer sign extensions to zero extensions as sign-extensions tend to be
298   // more expensive.
299   if (CurrentUse.Ty == TyForCandidate) {
300     if (CurrentUse.ExtendOpcode == TargetOpcode::G_SEXT &&
301         OpcodeForCandidate == TargetOpcode::G_ZEXT)
302       return CurrentUse;
303     else if (CurrentUse.ExtendOpcode == TargetOpcode::G_ZEXT &&
304              OpcodeForCandidate == TargetOpcode::G_SEXT)
305       return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
306   }
307 
308   // This is potentially target specific. We've chosen the largest type
309   // because G_TRUNC is usually free. One potential catch with this is that
310   // some targets have a reduced number of larger registers than smaller
311   // registers and this choice potentially increases the live-range for the
312   // larger value.
313   if (TyForCandidate.getSizeInBits() > CurrentUse.Ty.getSizeInBits()) {
314     return {TyForCandidate, OpcodeForCandidate, MIForCandidate};
315   }
316   return CurrentUse;
317 }
318 
319 /// Find a suitable place to insert some instructions and insert them. This
320 /// function accounts for special cases like inserting before a PHI node.
321 /// The current strategy for inserting before PHI's is to duplicate the
322 /// instructions for each predecessor. However, while that's ok for G_TRUNC
323 /// on most targets since it generally requires no code, other targets/cases may
324 /// want to try harder to find a dominating block.
325 static void InsertInsnsWithoutSideEffectsBeforeUse(
326     MachineIRBuilder &Builder, MachineInstr &DefMI, MachineOperand &UseMO,
327     std::function<void(MachineBasicBlock *, MachineBasicBlock::iterator,
328                        MachineOperand &UseMO)>
329         Inserter) {
330   MachineInstr &UseMI = *UseMO.getParent();
331 
332   MachineBasicBlock *InsertBB = UseMI.getParent();
333 
334   // If the use is a PHI then we want the predecessor block instead.
335   if (UseMI.isPHI()) {
336     MachineOperand *PredBB = std::next(&UseMO);
337     InsertBB = PredBB->getMBB();
338   }
339 
340   // If the block is the same block as the def then we want to insert just after
341   // the def instead of at the start of the block.
342   if (InsertBB == DefMI.getParent()) {
343     MachineBasicBlock::iterator InsertPt = &DefMI;
344     Inserter(InsertBB, std::next(InsertPt), UseMO);
345     return;
346   }
347 
348   // Otherwise we want the start of the BB
349   Inserter(InsertBB, InsertBB->getFirstNonPHI(), UseMO);
350 }
351 } // end anonymous namespace
352 
353 bool CombinerHelper::tryCombineExtendingLoads(MachineInstr &MI) {
354   PreferredTuple Preferred;
355   if (matchCombineExtendingLoads(MI, Preferred)) {
356     applyCombineExtendingLoads(MI, Preferred);
357     return true;
358   }
359   return false;
360 }
361 
362 bool CombinerHelper::matchCombineExtendingLoads(MachineInstr &MI,
363                                                 PreferredTuple &Preferred) {
364   // We match the loads and follow the uses to the extend instead of matching
365   // the extends and following the def to the load. This is because the load
366   // must remain in the same position for correctness (unless we also add code
367   // to find a safe place to sink it) whereas the extend is freely movable.
368   // It also prevents us from duplicating the load for the volatile case or just
369   // for performance.
370 
371   if (MI.getOpcode() != TargetOpcode::G_LOAD &&
372       MI.getOpcode() != TargetOpcode::G_SEXTLOAD &&
373       MI.getOpcode() != TargetOpcode::G_ZEXTLOAD)
374     return false;
375 
376   auto &LoadValue = MI.getOperand(0);
377   assert(LoadValue.isReg() && "Result wasn't a register?");
378 
379   LLT LoadValueTy = MRI.getType(LoadValue.getReg());
380   if (!LoadValueTy.isScalar())
381     return false;
382 
383   // Most architectures are going to legalize <s8 loads into at least a 1 byte
384   // load, and the MMOs can only describe memory accesses in multiples of bytes.
385   // If we try to perform extload combining on those, we can end up with
386   // %a(s8) = extload %ptr (load 1 byte from %ptr)
387   // ... which is an illegal extload instruction.
388   if (LoadValueTy.getSizeInBits() < 8)
389     return false;
390 
391   // For non power-of-2 types, they will very likely be legalized into multiple
392   // loads. Don't bother trying to match them into extending loads.
393   if (!isPowerOf2_32(LoadValueTy.getSizeInBits()))
394     return false;
395 
396   // Find the preferred type aside from the any-extends (unless it's the only
397   // one) and non-extending ops. We'll emit an extending load to that type and
398   // and emit a variant of (extend (trunc X)) for the others according to the
399   // relative type sizes. At the same time, pick an extend to use based on the
400   // extend involved in the chosen type.
401   unsigned PreferredOpcode = MI.getOpcode() == TargetOpcode::G_LOAD
402                                  ? TargetOpcode::G_ANYEXT
403                                  : MI.getOpcode() == TargetOpcode::G_SEXTLOAD
404                                        ? TargetOpcode::G_SEXT
405                                        : TargetOpcode::G_ZEXT;
406   Preferred = {LLT(), PreferredOpcode, nullptr};
407   for (auto &UseMI : MRI.use_nodbg_instructions(LoadValue.getReg())) {
408     if (UseMI.getOpcode() == TargetOpcode::G_SEXT ||
409         UseMI.getOpcode() == TargetOpcode::G_ZEXT ||
410         (UseMI.getOpcode() == TargetOpcode::G_ANYEXT)) {
411       // Check for legality.
412       if (LI) {
413         LegalityQuery::MemDesc MMDesc;
414         const auto &MMO = **MI.memoperands_begin();
415         MMDesc.SizeInBits = MMO.getSizeInBits();
416         MMDesc.AlignInBits = MMO.getAlign().value() * 8;
417         MMDesc.Ordering = MMO.getOrdering();
418         LLT UseTy = MRI.getType(UseMI.getOperand(0).getReg());
419         LLT SrcTy = MRI.getType(MI.getOperand(1).getReg());
420         if (LI->getAction({MI.getOpcode(), {UseTy, SrcTy}, {MMDesc}}).Action !=
421             LegalizeActions::Legal)
422           continue;
423       }
424       Preferred = ChoosePreferredUse(Preferred,
425                                      MRI.getType(UseMI.getOperand(0).getReg()),
426                                      UseMI.getOpcode(), &UseMI);
427     }
428   }
429 
430   // There were no extends
431   if (!Preferred.MI)
432     return false;
433   // It should be impossible to chose an extend without selecting a different
434   // type since by definition the result of an extend is larger.
435   assert(Preferred.Ty != LoadValueTy && "Extending to same type?");
436 
437   LLVM_DEBUG(dbgs() << "Preferred use is: " << *Preferred.MI);
438   return true;
439 }
440 
441 void CombinerHelper::applyCombineExtendingLoads(MachineInstr &MI,
442                                                 PreferredTuple &Preferred) {
443   // Rewrite the load to the chosen extending load.
444   Register ChosenDstReg = Preferred.MI->getOperand(0).getReg();
445 
446   // Inserter to insert a truncate back to the original type at a given point
447   // with some basic CSE to limit truncate duplication to one per BB.
448   DenseMap<MachineBasicBlock *, MachineInstr *> EmittedInsns;
449   auto InsertTruncAt = [&](MachineBasicBlock *InsertIntoBB,
450                            MachineBasicBlock::iterator InsertBefore,
451                            MachineOperand &UseMO) {
452     MachineInstr *PreviouslyEmitted = EmittedInsns.lookup(InsertIntoBB);
453     if (PreviouslyEmitted) {
454       Observer.changingInstr(*UseMO.getParent());
455       UseMO.setReg(PreviouslyEmitted->getOperand(0).getReg());
456       Observer.changedInstr(*UseMO.getParent());
457       return;
458     }
459 
460     Builder.setInsertPt(*InsertIntoBB, InsertBefore);
461     Register NewDstReg = MRI.cloneVirtualRegister(MI.getOperand(0).getReg());
462     MachineInstr *NewMI = Builder.buildTrunc(NewDstReg, ChosenDstReg);
463     EmittedInsns[InsertIntoBB] = NewMI;
464     replaceRegOpWith(MRI, UseMO, NewDstReg);
465   };
466 
467   Observer.changingInstr(MI);
468   MI.setDesc(
469       Builder.getTII().get(Preferred.ExtendOpcode == TargetOpcode::G_SEXT
470                                ? TargetOpcode::G_SEXTLOAD
471                                : Preferred.ExtendOpcode == TargetOpcode::G_ZEXT
472                                      ? TargetOpcode::G_ZEXTLOAD
473                                      : TargetOpcode::G_LOAD));
474 
475   // Rewrite all the uses to fix up the types.
476   auto &LoadValue = MI.getOperand(0);
477   SmallVector<MachineOperand *, 4> Uses;
478   for (auto &UseMO : MRI.use_operands(LoadValue.getReg()))
479     Uses.push_back(&UseMO);
480 
481   for (auto *UseMO : Uses) {
482     MachineInstr *UseMI = UseMO->getParent();
483 
484     // If the extend is compatible with the preferred extend then we should fix
485     // up the type and extend so that it uses the preferred use.
486     if (UseMI->getOpcode() == Preferred.ExtendOpcode ||
487         UseMI->getOpcode() == TargetOpcode::G_ANYEXT) {
488       Register UseDstReg = UseMI->getOperand(0).getReg();
489       MachineOperand &UseSrcMO = UseMI->getOperand(1);
490       const LLT UseDstTy = MRI.getType(UseDstReg);
491       if (UseDstReg != ChosenDstReg) {
492         if (Preferred.Ty == UseDstTy) {
493           // If the use has the same type as the preferred use, then merge
494           // the vregs and erase the extend. For example:
495           //    %1:_(s8) = G_LOAD ...
496           //    %2:_(s32) = G_SEXT %1(s8)
497           //    %3:_(s32) = G_ANYEXT %1(s8)
498           //    ... = ... %3(s32)
499           // rewrites to:
500           //    %2:_(s32) = G_SEXTLOAD ...
501           //    ... = ... %2(s32)
502           replaceRegWith(MRI, UseDstReg, ChosenDstReg);
503           Observer.erasingInstr(*UseMO->getParent());
504           UseMO->getParent()->eraseFromParent();
505         } else if (Preferred.Ty.getSizeInBits() < UseDstTy.getSizeInBits()) {
506           // If the preferred size is smaller, then keep the extend but extend
507           // from the result of the extending load. For example:
508           //    %1:_(s8) = G_LOAD ...
509           //    %2:_(s32) = G_SEXT %1(s8)
510           //    %3:_(s64) = G_ANYEXT %1(s8)
511           //    ... = ... %3(s64)
512           /// rewrites to:
513           //    %2:_(s32) = G_SEXTLOAD ...
514           //    %3:_(s64) = G_ANYEXT %2:_(s32)
515           //    ... = ... %3(s64)
516           replaceRegOpWith(MRI, UseSrcMO, ChosenDstReg);
517         } else {
518           // If the preferred size is large, then insert a truncate. For
519           // example:
520           //    %1:_(s8) = G_LOAD ...
521           //    %2:_(s64) = G_SEXT %1(s8)
522           //    %3:_(s32) = G_ZEXT %1(s8)
523           //    ... = ... %3(s32)
524           /// rewrites to:
525           //    %2:_(s64) = G_SEXTLOAD ...
526           //    %4:_(s8) = G_TRUNC %2:_(s32)
527           //    %3:_(s64) = G_ZEXT %2:_(s8)
528           //    ... = ... %3(s64)
529           InsertInsnsWithoutSideEffectsBeforeUse(Builder, MI, *UseMO,
530                                                  InsertTruncAt);
531         }
532         continue;
533       }
534       // The use is (one of) the uses of the preferred use we chose earlier.
535       // We're going to update the load to def this value later so just erase
536       // the old extend.
537       Observer.erasingInstr(*UseMO->getParent());
538       UseMO->getParent()->eraseFromParent();
539       continue;
540     }
541 
542     // The use isn't an extend. Truncate back to the type we originally loaded.
543     // This is free on many targets.
544     InsertInsnsWithoutSideEffectsBeforeUse(Builder, MI, *UseMO, InsertTruncAt);
545   }
546 
547   MI.getOperand(0).setReg(ChosenDstReg);
548   Observer.changedInstr(MI);
549 }
550 
551 bool CombinerHelper::isPredecessor(const MachineInstr &DefMI,
552                                    const MachineInstr &UseMI) {
553   assert(!DefMI.isDebugInstr() && !UseMI.isDebugInstr() &&
554          "shouldn't consider debug uses");
555   assert(DefMI.getParent() == UseMI.getParent());
556   if (&DefMI == &UseMI)
557     return false;
558 
559   // Loop through the basic block until we find one of the instructions.
560   MachineBasicBlock::const_iterator I = DefMI.getParent()->begin();
561   for (; &*I != &DefMI && &*I != &UseMI; ++I)
562     return &*I == &DefMI;
563 
564   llvm_unreachable("Block must contain instructions");
565 }
566 
567 bool CombinerHelper::dominates(const MachineInstr &DefMI,
568                                const MachineInstr &UseMI) {
569   assert(!DefMI.isDebugInstr() && !UseMI.isDebugInstr() &&
570          "shouldn't consider debug uses");
571   if (MDT)
572     return MDT->dominates(&DefMI, &UseMI);
573   else if (DefMI.getParent() != UseMI.getParent())
574     return false;
575 
576   return isPredecessor(DefMI, UseMI);
577 }
578 
579 bool CombinerHelper::matchSextAlreadyExtended(MachineInstr &MI) {
580   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
581   Register SrcReg = MI.getOperand(1).getReg();
582   unsigned SrcSignBits = KB->computeNumSignBits(SrcReg);
583   unsigned NumSextBits =
584       MRI.getType(MI.getOperand(0).getReg()).getScalarSizeInBits() -
585       MI.getOperand(2).getImm();
586   return SrcSignBits >= NumSextBits;
587 }
588 
589 bool CombinerHelper::applySextAlreadyExtended(MachineInstr &MI) {
590   assert(MI.getOpcode() == TargetOpcode::G_SEXT_INREG);
591   MachineIRBuilder MIB(MI);
592   MIB.buildCopy(MI.getOperand(0).getReg(), MI.getOperand(1).getReg());
593   MI.eraseFromParent();
594   return true;
595 }
596 
597 bool CombinerHelper::findPostIndexCandidate(MachineInstr &MI, Register &Addr,
598                                             Register &Base, Register &Offset) {
599   auto &MF = *MI.getParent()->getParent();
600   const auto &TLI = *MF.getSubtarget().getTargetLowering();
601 
602 #ifndef NDEBUG
603   unsigned Opcode = MI.getOpcode();
604   assert(Opcode == TargetOpcode::G_LOAD || Opcode == TargetOpcode::G_SEXTLOAD ||
605          Opcode == TargetOpcode::G_ZEXTLOAD || Opcode == TargetOpcode::G_STORE);
606 #endif
607 
608   Base = MI.getOperand(1).getReg();
609   MachineInstr *BaseDef = MRI.getUniqueVRegDef(Base);
610   if (BaseDef && BaseDef->getOpcode() == TargetOpcode::G_FRAME_INDEX)
611     return false;
612 
613   LLVM_DEBUG(dbgs() << "Searching for post-indexing opportunity for: " << MI);
614 
615   for (auto &Use : MRI.use_nodbg_instructions(Base)) {
616     if (Use.getOpcode() != TargetOpcode::G_PTR_ADD)
617       continue;
618 
619     Offset = Use.getOperand(2).getReg();
620     if (!ForceLegalIndexing &&
621         !TLI.isIndexingLegal(MI, Base, Offset, /*IsPre*/ false, MRI)) {
622       LLVM_DEBUG(dbgs() << "    Ignoring candidate with illegal addrmode: "
623                         << Use);
624       continue;
625     }
626 
627     // Make sure the offset calculation is before the potentially indexed op.
628     // FIXME: we really care about dependency here. The offset calculation might
629     // be movable.
630     MachineInstr *OffsetDef = MRI.getUniqueVRegDef(Offset);
631     if (!OffsetDef || !dominates(*OffsetDef, MI)) {
632       LLVM_DEBUG(dbgs() << "    Ignoring candidate with offset after mem-op: "
633                         << Use);
634       continue;
635     }
636 
637     // FIXME: check whether all uses of Base are load/store with foldable
638     // addressing modes. If so, using the normal addr-modes is better than
639     // forming an indexed one.
640 
641     bool MemOpDominatesAddrUses = true;
642     for (auto &PtrAddUse :
643          MRI.use_nodbg_instructions(Use.getOperand(0).getReg())) {
644       if (!dominates(MI, PtrAddUse)) {
645         MemOpDominatesAddrUses = false;
646         break;
647       }
648     }
649 
650     if (!MemOpDominatesAddrUses) {
651       LLVM_DEBUG(
652           dbgs() << "    Ignoring candidate as memop does not dominate uses: "
653                  << Use);
654       continue;
655     }
656 
657     LLVM_DEBUG(dbgs() << "    Found match: " << Use);
658     Addr = Use.getOperand(0).getReg();
659     return true;
660   }
661 
662   return false;
663 }
664 
665 bool CombinerHelper::findPreIndexCandidate(MachineInstr &MI, Register &Addr,
666                                            Register &Base, Register &Offset) {
667   auto &MF = *MI.getParent()->getParent();
668   const auto &TLI = *MF.getSubtarget().getTargetLowering();
669 
670 #ifndef NDEBUG
671   unsigned Opcode = MI.getOpcode();
672   assert(Opcode == TargetOpcode::G_LOAD || Opcode == TargetOpcode::G_SEXTLOAD ||
673          Opcode == TargetOpcode::G_ZEXTLOAD || Opcode == TargetOpcode::G_STORE);
674 #endif
675 
676   Addr = MI.getOperand(1).getReg();
677   MachineInstr *AddrDef = getOpcodeDef(TargetOpcode::G_PTR_ADD, Addr, MRI);
678   if (!AddrDef || MRI.hasOneNonDBGUse(Addr))
679     return false;
680 
681   Base = AddrDef->getOperand(1).getReg();
682   Offset = AddrDef->getOperand(2).getReg();
683 
684   LLVM_DEBUG(dbgs() << "Found potential pre-indexed load_store: " << MI);
685 
686   if (!ForceLegalIndexing &&
687       !TLI.isIndexingLegal(MI, Base, Offset, /*IsPre*/ true, MRI)) {
688     LLVM_DEBUG(dbgs() << "    Skipping, not legal for target");
689     return false;
690   }
691 
692   MachineInstr *BaseDef = getDefIgnoringCopies(Base, MRI);
693   if (BaseDef->getOpcode() == TargetOpcode::G_FRAME_INDEX) {
694     LLVM_DEBUG(dbgs() << "    Skipping, frame index would need copy anyway.");
695     return false;
696   }
697 
698   if (MI.getOpcode() == TargetOpcode::G_STORE) {
699     // Would require a copy.
700     if (Base == MI.getOperand(0).getReg()) {
701       LLVM_DEBUG(dbgs() << "    Skipping, storing base so need copy anyway.");
702       return false;
703     }
704 
705     // We're expecting one use of Addr in MI, but it could also be the
706     // value stored, which isn't actually dominated by the instruction.
707     if (MI.getOperand(0).getReg() == Addr) {
708       LLVM_DEBUG(dbgs() << "    Skipping, does not dominate all addr uses");
709       return false;
710     }
711   }
712 
713   // FIXME: check whether all uses of the base pointer are constant PtrAdds.
714   // That might allow us to end base's liveness here by adjusting the constant.
715 
716   for (auto &UseMI : MRI.use_nodbg_instructions(Addr)) {
717     if (!dominates(MI, UseMI)) {
718       LLVM_DEBUG(dbgs() << "    Skipping, does not dominate all addr uses.");
719       return false;
720     }
721   }
722 
723   return true;
724 }
725 
726 bool CombinerHelper::tryCombineIndexedLoadStore(MachineInstr &MI) {
727   IndexedLoadStoreMatchInfo MatchInfo;
728   if (matchCombineIndexedLoadStore(MI, MatchInfo)) {
729     applyCombineIndexedLoadStore(MI, MatchInfo);
730     return true;
731   }
732   return false;
733 }
734 
735 bool CombinerHelper::matchCombineIndexedLoadStore(MachineInstr &MI, IndexedLoadStoreMatchInfo &MatchInfo) {
736   unsigned Opcode = MI.getOpcode();
737   if (Opcode != TargetOpcode::G_LOAD && Opcode != TargetOpcode::G_SEXTLOAD &&
738       Opcode != TargetOpcode::G_ZEXTLOAD && Opcode != TargetOpcode::G_STORE)
739     return false;
740 
741   MatchInfo.IsPre = findPreIndexCandidate(MI, MatchInfo.Addr, MatchInfo.Base,
742                                           MatchInfo.Offset);
743   if (!MatchInfo.IsPre &&
744       !findPostIndexCandidate(MI, MatchInfo.Addr, MatchInfo.Base,
745                               MatchInfo.Offset))
746     return false;
747 
748   return true;
749 }
750 
751 void CombinerHelper::applyCombineIndexedLoadStore(
752     MachineInstr &MI, IndexedLoadStoreMatchInfo &MatchInfo) {
753   MachineInstr &AddrDef = *MRI.getUniqueVRegDef(MatchInfo.Addr);
754   MachineIRBuilder MIRBuilder(MI);
755   unsigned Opcode = MI.getOpcode();
756   bool IsStore = Opcode == TargetOpcode::G_STORE;
757   unsigned NewOpcode;
758   switch (Opcode) {
759   case TargetOpcode::G_LOAD:
760     NewOpcode = TargetOpcode::G_INDEXED_LOAD;
761     break;
762   case TargetOpcode::G_SEXTLOAD:
763     NewOpcode = TargetOpcode::G_INDEXED_SEXTLOAD;
764     break;
765   case TargetOpcode::G_ZEXTLOAD:
766     NewOpcode = TargetOpcode::G_INDEXED_ZEXTLOAD;
767     break;
768   case TargetOpcode::G_STORE:
769     NewOpcode = TargetOpcode::G_INDEXED_STORE;
770     break;
771   default:
772     llvm_unreachable("Unknown load/store opcode");
773   }
774 
775   auto MIB = MIRBuilder.buildInstr(NewOpcode);
776   if (IsStore) {
777     MIB.addDef(MatchInfo.Addr);
778     MIB.addUse(MI.getOperand(0).getReg());
779   } else {
780     MIB.addDef(MI.getOperand(0).getReg());
781     MIB.addDef(MatchInfo.Addr);
782   }
783 
784   MIB.addUse(MatchInfo.Base);
785   MIB.addUse(MatchInfo.Offset);
786   MIB.addImm(MatchInfo.IsPre);
787   MI.eraseFromParent();
788   AddrDef.eraseFromParent();
789 
790   LLVM_DEBUG(dbgs() << "    Combinined to indexed operation");
791 }
792 
793 bool CombinerHelper::matchElideBrByInvertingCond(MachineInstr &MI) {
794   if (MI.getOpcode() != TargetOpcode::G_BR)
795     return false;
796 
797   // Try to match the following:
798   // bb1:
799   //   %c(s32) = G_ICMP pred, %a, %b
800   //   %c1(s1) = G_TRUNC %c(s32)
801   //   G_BRCOND %c1, %bb2
802   //   G_BR %bb3
803   // bb2:
804   // ...
805   // bb3:
806 
807   // The above pattern does not have a fall through to the successor bb2, always
808   // resulting in a branch no matter which path is taken. Here we try to find
809   // and replace that pattern with conditional branch to bb3 and otherwise
810   // fallthrough to bb2.
811 
812   MachineBasicBlock *MBB = MI.getParent();
813   MachineBasicBlock::iterator BrIt(MI);
814   if (BrIt == MBB->begin())
815     return false;
816   assert(std::next(BrIt) == MBB->end() && "expected G_BR to be a terminator");
817 
818   MachineInstr *BrCond = &*std::prev(BrIt);
819   if (BrCond->getOpcode() != TargetOpcode::G_BRCOND)
820     return false;
821 
822   // Check that the next block is the conditional branch target.
823   if (!MBB->isLayoutSuccessor(BrCond->getOperand(1).getMBB()))
824     return false;
825 
826   MachineInstr *CmpMI = MRI.getVRegDef(BrCond->getOperand(0).getReg());
827   if (!CmpMI || CmpMI->getOpcode() != TargetOpcode::G_ICMP ||
828       !MRI.hasOneNonDBGUse(CmpMI->getOperand(0).getReg()))
829     return false;
830   return true;
831 }
832 
833 bool CombinerHelper::tryElideBrByInvertingCond(MachineInstr &MI) {
834   if (!matchElideBrByInvertingCond(MI))
835     return false;
836   applyElideBrByInvertingCond(MI);
837   return true;
838 }
839 
840 void CombinerHelper::applyElideBrByInvertingCond(MachineInstr &MI) {
841   MachineBasicBlock *BrTarget = MI.getOperand(0).getMBB();
842   MachineBasicBlock::iterator BrIt(MI);
843   MachineInstr *BrCond = &*std::prev(BrIt);
844   MachineInstr *CmpMI = MRI.getVRegDef(BrCond->getOperand(0).getReg());
845 
846   CmpInst::Predicate InversePred = CmpInst::getInversePredicate(
847       (CmpInst::Predicate)CmpMI->getOperand(1).getPredicate());
848 
849   // Invert the G_ICMP condition.
850   Observer.changingInstr(*CmpMI);
851   CmpMI->getOperand(1).setPredicate(InversePred);
852   Observer.changedInstr(*CmpMI);
853 
854   // Change the conditional branch target.
855   Observer.changingInstr(*BrCond);
856   BrCond->getOperand(1).setMBB(BrTarget);
857   Observer.changedInstr(*BrCond);
858   MI.eraseFromParent();
859 }
860 
861 static bool shouldLowerMemFuncForSize(const MachineFunction &MF) {
862   // On Darwin, -Os means optimize for size without hurting performance, so
863   // only really optimize for size when -Oz (MinSize) is used.
864   if (MF.getTarget().getTargetTriple().isOSDarwin())
865     return MF.getFunction().hasMinSize();
866   return MF.getFunction().hasOptSize();
867 }
868 
869 // Returns a list of types to use for memory op lowering in MemOps. A partial
870 // port of findOptimalMemOpLowering in TargetLowering.
871 static bool findGISelOptimalMemOpLowering(std::vector<LLT> &MemOps,
872                                           unsigned Limit, const MemOp &Op,
873                                           unsigned DstAS, unsigned SrcAS,
874                                           const AttributeList &FuncAttributes,
875                                           const TargetLowering &TLI) {
876   if (Op.isMemcpyWithFixedDstAlign() && Op.getSrcAlign() < Op.getDstAlign())
877     return false;
878 
879   LLT Ty = TLI.getOptimalMemOpLLT(Op, FuncAttributes);
880 
881   if (Ty == LLT()) {
882     // Use the largest scalar type whose alignment constraints are satisfied.
883     // We only need to check DstAlign here as SrcAlign is always greater or
884     // equal to DstAlign (or zero).
885     Ty = LLT::scalar(64);
886     if (Op.isFixedDstAlign())
887       while (Op.getDstAlign() < Ty.getSizeInBytes() &&
888              !TLI.allowsMisalignedMemoryAccesses(Ty, DstAS, Op.getDstAlign()))
889         Ty = LLT::scalar(Ty.getSizeInBytes());
890     assert(Ty.getSizeInBits() > 0 && "Could not find valid type");
891     // FIXME: check for the largest legal type we can load/store to.
892   }
893 
894   unsigned NumMemOps = 0;
895   uint64_t Size = Op.size();
896   while (Size) {
897     unsigned TySize = Ty.getSizeInBytes();
898     while (TySize > Size) {
899       // For now, only use non-vector load / store's for the left-over pieces.
900       LLT NewTy = Ty;
901       // FIXME: check for mem op safety and legality of the types. Not all of
902       // SDAGisms map cleanly to GISel concepts.
903       if (NewTy.isVector())
904         NewTy = NewTy.getSizeInBits() > 64 ? LLT::scalar(64) : LLT::scalar(32);
905       NewTy = LLT::scalar(PowerOf2Floor(NewTy.getSizeInBits() - 1));
906       unsigned NewTySize = NewTy.getSizeInBytes();
907       assert(NewTySize > 0 && "Could not find appropriate type");
908 
909       // If the new LLT cannot cover all of the remaining bits, then consider
910       // issuing a (or a pair of) unaligned and overlapping load / store.
911       bool Fast;
912       // Need to get a VT equivalent for allowMisalignedMemoryAccesses().
913       MVT VT = getMVTForLLT(Ty);
914       if (NumMemOps && Op.allowOverlap() && NewTySize < Size &&
915           TLI.allowsMisalignedMemoryAccesses(
916               VT, DstAS, Op.isFixedDstAlign() ? Op.getDstAlign().value() : 0,
917               MachineMemOperand::MONone, &Fast) &&
918           Fast)
919         TySize = Size;
920       else {
921         Ty = NewTy;
922         TySize = NewTySize;
923       }
924     }
925 
926     if (++NumMemOps > Limit)
927       return false;
928 
929     MemOps.push_back(Ty);
930     Size -= TySize;
931   }
932 
933   return true;
934 }
935 
936 static Type *getTypeForLLT(LLT Ty, LLVMContext &C) {
937   if (Ty.isVector())
938     return FixedVectorType::get(IntegerType::get(C, Ty.getScalarSizeInBits()),
939                                 Ty.getNumElements());
940   return IntegerType::get(C, Ty.getSizeInBits());
941 }
942 
943 // Get a vectorized representation of the memset value operand, GISel edition.
944 static Register getMemsetValue(Register Val, LLT Ty, MachineIRBuilder &MIB) {
945   MachineRegisterInfo &MRI = *MIB.getMRI();
946   unsigned NumBits = Ty.getScalarSizeInBits();
947   auto ValVRegAndVal = getConstantVRegValWithLookThrough(Val, MRI);
948   if (!Ty.isVector() && ValVRegAndVal) {
949     unsigned KnownVal = ValVRegAndVal->Value;
950     APInt Scalar = APInt(8, KnownVal);
951     APInt SplatVal = APInt::getSplat(NumBits, Scalar);
952     return MIB.buildConstant(Ty, SplatVal).getReg(0);
953   }
954 
955   // Extend the byte value to the larger type, and then multiply by a magic
956   // value 0x010101... in order to replicate it across every byte.
957   // Unless it's zero, in which case just emit a larger G_CONSTANT 0.
958   if (ValVRegAndVal && ValVRegAndVal->Value == 0) {
959     return MIB.buildConstant(Ty, 0).getReg(0);
960   }
961 
962   LLT ExtType = Ty.getScalarType();
963   auto ZExt = MIB.buildZExtOrTrunc(ExtType, Val);
964   if (NumBits > 8) {
965     APInt Magic = APInt::getSplat(NumBits, APInt(8, 0x01));
966     auto MagicMI = MIB.buildConstant(ExtType, Magic);
967     Val = MIB.buildMul(ExtType, ZExt, MagicMI).getReg(0);
968   }
969 
970   // For vector types create a G_BUILD_VECTOR.
971   if (Ty.isVector())
972     Val = MIB.buildSplatVector(Ty, Val).getReg(0);
973 
974   return Val;
975 }
976 
977 bool CombinerHelper::optimizeMemset(MachineInstr &MI, Register Dst,
978                                     Register Val, unsigned KnownLen,
979                                     Align Alignment, bool IsVolatile) {
980   auto &MF = *MI.getParent()->getParent();
981   const auto &TLI = *MF.getSubtarget().getTargetLowering();
982   auto &DL = MF.getDataLayout();
983   LLVMContext &C = MF.getFunction().getContext();
984 
985   assert(KnownLen != 0 && "Have a zero length memset length!");
986 
987   bool DstAlignCanChange = false;
988   MachineFrameInfo &MFI = MF.getFrameInfo();
989   bool OptSize = shouldLowerMemFuncForSize(MF);
990 
991   MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI);
992   if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex()))
993     DstAlignCanChange = true;
994 
995   unsigned Limit = TLI.getMaxStoresPerMemset(OptSize);
996   std::vector<LLT> MemOps;
997 
998   const auto &DstMMO = **MI.memoperands_begin();
999   MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo();
1000 
1001   auto ValVRegAndVal = getConstantVRegValWithLookThrough(Val, MRI);
1002   bool IsZeroVal = ValVRegAndVal && ValVRegAndVal->Value == 0;
1003 
1004   if (!findGISelOptimalMemOpLowering(MemOps, Limit,
1005                                      MemOp::Set(KnownLen, DstAlignCanChange,
1006                                                 Alignment,
1007                                                 /*IsZeroMemset=*/IsZeroVal,
1008                                                 /*IsVolatile=*/IsVolatile),
1009                                      DstPtrInfo.getAddrSpace(), ~0u,
1010                                      MF.getFunction().getAttributes(), TLI))
1011     return false;
1012 
1013   if (DstAlignCanChange) {
1014     // Get an estimate of the type from the LLT.
1015     Type *IRTy = getTypeForLLT(MemOps[0], C);
1016     Align NewAlign = DL.getABITypeAlign(IRTy);
1017     if (NewAlign > Alignment) {
1018       Alignment = NewAlign;
1019       unsigned FI = FIDef->getOperand(1).getIndex();
1020       // Give the stack frame object a larger alignment if needed.
1021       if (MFI.getObjectAlign(FI) < Alignment)
1022         MFI.setObjectAlignment(FI, Alignment);
1023     }
1024   }
1025 
1026   MachineIRBuilder MIB(MI);
1027   // Find the largest store and generate the bit pattern for it.
1028   LLT LargestTy = MemOps[0];
1029   for (unsigned i = 1; i < MemOps.size(); i++)
1030     if (MemOps[i].getSizeInBits() > LargestTy.getSizeInBits())
1031       LargestTy = MemOps[i];
1032 
1033   // The memset stored value is always defined as an s8, so in order to make it
1034   // work with larger store types we need to repeat the bit pattern across the
1035   // wider type.
1036   Register MemSetValue = getMemsetValue(Val, LargestTy, MIB);
1037 
1038   if (!MemSetValue)
1039     return false;
1040 
1041   // Generate the stores. For each store type in the list, we generate the
1042   // matching store of that type to the destination address.
1043   LLT PtrTy = MRI.getType(Dst);
1044   unsigned DstOff = 0;
1045   unsigned Size = KnownLen;
1046   for (unsigned I = 0; I < MemOps.size(); I++) {
1047     LLT Ty = MemOps[I];
1048     unsigned TySize = Ty.getSizeInBytes();
1049     if (TySize > Size) {
1050       // Issuing an unaligned load / store pair that overlaps with the previous
1051       // pair. Adjust the offset accordingly.
1052       assert(I == MemOps.size() - 1 && I != 0);
1053       DstOff -= TySize - Size;
1054     }
1055 
1056     // If this store is smaller than the largest store see whether we can get
1057     // the smaller value for free with a truncate.
1058     Register Value = MemSetValue;
1059     if (Ty.getSizeInBits() < LargestTy.getSizeInBits()) {
1060       MVT VT = getMVTForLLT(Ty);
1061       MVT LargestVT = getMVTForLLT(LargestTy);
1062       if (!LargestTy.isVector() && !Ty.isVector() &&
1063           TLI.isTruncateFree(LargestVT, VT))
1064         Value = MIB.buildTrunc(Ty, MemSetValue).getReg(0);
1065       else
1066         Value = getMemsetValue(Val, Ty, MIB);
1067       if (!Value)
1068         return false;
1069     }
1070 
1071     auto *StoreMMO =
1072         MF.getMachineMemOperand(&DstMMO, DstOff, Ty.getSizeInBytes());
1073 
1074     Register Ptr = Dst;
1075     if (DstOff != 0) {
1076       auto Offset =
1077           MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), DstOff);
1078       Ptr = MIB.buildPtrAdd(PtrTy, Dst, Offset).getReg(0);
1079     }
1080 
1081     MIB.buildStore(Value, Ptr, *StoreMMO);
1082     DstOff += Ty.getSizeInBytes();
1083     Size -= TySize;
1084   }
1085 
1086   MI.eraseFromParent();
1087   return true;
1088 }
1089 
1090 bool CombinerHelper::optimizeMemcpy(MachineInstr &MI, Register Dst,
1091                                     Register Src, unsigned KnownLen,
1092                                     Align DstAlign, Align SrcAlign,
1093                                     bool IsVolatile) {
1094   auto &MF = *MI.getParent()->getParent();
1095   const auto &TLI = *MF.getSubtarget().getTargetLowering();
1096   auto &DL = MF.getDataLayout();
1097   LLVMContext &C = MF.getFunction().getContext();
1098 
1099   assert(KnownLen != 0 && "Have a zero length memcpy length!");
1100 
1101   bool DstAlignCanChange = false;
1102   MachineFrameInfo &MFI = MF.getFrameInfo();
1103   bool OptSize = shouldLowerMemFuncForSize(MF);
1104   Align Alignment = commonAlignment(DstAlign, SrcAlign);
1105 
1106   MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI);
1107   if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex()))
1108     DstAlignCanChange = true;
1109 
1110   // FIXME: infer better src pointer alignment like SelectionDAG does here.
1111   // FIXME: also use the equivalent of isMemSrcFromConstant and alwaysinlining
1112   // if the memcpy is in a tail call position.
1113 
1114   unsigned Limit = TLI.getMaxStoresPerMemcpy(OptSize);
1115   std::vector<LLT> MemOps;
1116 
1117   const auto &DstMMO = **MI.memoperands_begin();
1118   const auto &SrcMMO = **std::next(MI.memoperands_begin());
1119   MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo();
1120   MachinePointerInfo SrcPtrInfo = SrcMMO.getPointerInfo();
1121 
1122   if (!findGISelOptimalMemOpLowering(
1123           MemOps, Limit,
1124           MemOp::Copy(KnownLen, DstAlignCanChange, Alignment, SrcAlign,
1125                       IsVolatile),
1126           DstPtrInfo.getAddrSpace(), SrcPtrInfo.getAddrSpace(),
1127           MF.getFunction().getAttributes(), TLI))
1128     return false;
1129 
1130   if (DstAlignCanChange) {
1131     // Get an estimate of the type from the LLT.
1132     Type *IRTy = getTypeForLLT(MemOps[0], C);
1133     Align NewAlign = DL.getABITypeAlign(IRTy);
1134 
1135     // Don't promote to an alignment that would require dynamic stack
1136     // realignment.
1137     const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1138     if (!TRI->needsStackRealignment(MF))
1139       while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign))
1140         NewAlign = NewAlign / 2;
1141 
1142     if (NewAlign > Alignment) {
1143       Alignment = NewAlign;
1144       unsigned FI = FIDef->getOperand(1).getIndex();
1145       // Give the stack frame object a larger alignment if needed.
1146       if (MFI.getObjectAlign(FI) < Alignment)
1147         MFI.setObjectAlignment(FI, Alignment);
1148     }
1149   }
1150 
1151   LLVM_DEBUG(dbgs() << "Inlining memcpy: " << MI << " into loads & stores\n");
1152 
1153   MachineIRBuilder MIB(MI);
1154   // Now we need to emit a pair of load and stores for each of the types we've
1155   // collected. I.e. for each type, generate a load from the source pointer of
1156   // that type width, and then generate a corresponding store to the dest buffer
1157   // of that value loaded. This can result in a sequence of loads and stores
1158   // mixed types, depending on what the target specifies as good types to use.
1159   unsigned CurrOffset = 0;
1160   LLT PtrTy = MRI.getType(Src);
1161   unsigned Size = KnownLen;
1162   for (auto CopyTy : MemOps) {
1163     // Issuing an unaligned load / store pair  that overlaps with the previous
1164     // pair. Adjust the offset accordingly.
1165     if (CopyTy.getSizeInBytes() > Size)
1166       CurrOffset -= CopyTy.getSizeInBytes() - Size;
1167 
1168     // Construct MMOs for the accesses.
1169     auto *LoadMMO =
1170         MF.getMachineMemOperand(&SrcMMO, CurrOffset, CopyTy.getSizeInBytes());
1171     auto *StoreMMO =
1172         MF.getMachineMemOperand(&DstMMO, CurrOffset, CopyTy.getSizeInBytes());
1173 
1174     // Create the load.
1175     Register LoadPtr = Src;
1176     Register Offset;
1177     if (CurrOffset != 0) {
1178       Offset = MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), CurrOffset)
1179                    .getReg(0);
1180       LoadPtr = MIB.buildPtrAdd(PtrTy, Src, Offset).getReg(0);
1181     }
1182     auto LdVal = MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO);
1183 
1184     // Create the store.
1185     Register StorePtr =
1186         CurrOffset == 0 ? Dst : MIB.buildPtrAdd(PtrTy, Dst, Offset).getReg(0);
1187     MIB.buildStore(LdVal, StorePtr, *StoreMMO);
1188     CurrOffset += CopyTy.getSizeInBytes();
1189     Size -= CopyTy.getSizeInBytes();
1190   }
1191 
1192   MI.eraseFromParent();
1193   return true;
1194 }
1195 
1196 bool CombinerHelper::optimizeMemmove(MachineInstr &MI, Register Dst,
1197                                      Register Src, unsigned KnownLen,
1198                                      Align DstAlign, Align SrcAlign,
1199                                      bool IsVolatile) {
1200   auto &MF = *MI.getParent()->getParent();
1201   const auto &TLI = *MF.getSubtarget().getTargetLowering();
1202   auto &DL = MF.getDataLayout();
1203   LLVMContext &C = MF.getFunction().getContext();
1204 
1205   assert(KnownLen != 0 && "Have a zero length memmove length!");
1206 
1207   bool DstAlignCanChange = false;
1208   MachineFrameInfo &MFI = MF.getFrameInfo();
1209   bool OptSize = shouldLowerMemFuncForSize(MF);
1210   Align Alignment = commonAlignment(DstAlign, SrcAlign);
1211 
1212   MachineInstr *FIDef = getOpcodeDef(TargetOpcode::G_FRAME_INDEX, Dst, MRI);
1213   if (FIDef && !MFI.isFixedObjectIndex(FIDef->getOperand(1).getIndex()))
1214     DstAlignCanChange = true;
1215 
1216   unsigned Limit = TLI.getMaxStoresPerMemmove(OptSize);
1217   std::vector<LLT> MemOps;
1218 
1219   const auto &DstMMO = **MI.memoperands_begin();
1220   const auto &SrcMMO = **std::next(MI.memoperands_begin());
1221   MachinePointerInfo DstPtrInfo = DstMMO.getPointerInfo();
1222   MachinePointerInfo SrcPtrInfo = SrcMMO.getPointerInfo();
1223 
1224   // FIXME: SelectionDAG always passes false for 'AllowOverlap', apparently due
1225   // to a bug in it's findOptimalMemOpLowering implementation. For now do the
1226   // same thing here.
1227   if (!findGISelOptimalMemOpLowering(
1228           MemOps, Limit,
1229           MemOp::Copy(KnownLen, DstAlignCanChange, Alignment, SrcAlign,
1230                       /*IsVolatile*/ true),
1231           DstPtrInfo.getAddrSpace(), SrcPtrInfo.getAddrSpace(),
1232           MF.getFunction().getAttributes(), TLI))
1233     return false;
1234 
1235   if (DstAlignCanChange) {
1236     // Get an estimate of the type from the LLT.
1237     Type *IRTy = getTypeForLLT(MemOps[0], C);
1238     Align NewAlign = DL.getABITypeAlign(IRTy);
1239 
1240     // Don't promote to an alignment that would require dynamic stack
1241     // realignment.
1242     const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1243     if (!TRI->needsStackRealignment(MF))
1244       while (NewAlign > Alignment && DL.exceedsNaturalStackAlignment(NewAlign))
1245         NewAlign = NewAlign / 2;
1246 
1247     if (NewAlign > Alignment) {
1248       Alignment = NewAlign;
1249       unsigned FI = FIDef->getOperand(1).getIndex();
1250       // Give the stack frame object a larger alignment if needed.
1251       if (MFI.getObjectAlign(FI) < Alignment)
1252         MFI.setObjectAlignment(FI, Alignment);
1253     }
1254   }
1255 
1256   LLVM_DEBUG(dbgs() << "Inlining memmove: " << MI << " into loads & stores\n");
1257 
1258   MachineIRBuilder MIB(MI);
1259   // Memmove requires that we perform the loads first before issuing the stores.
1260   // Apart from that, this loop is pretty much doing the same thing as the
1261   // memcpy codegen function.
1262   unsigned CurrOffset = 0;
1263   LLT PtrTy = MRI.getType(Src);
1264   SmallVector<Register, 16> LoadVals;
1265   for (auto CopyTy : MemOps) {
1266     // Construct MMO for the load.
1267     auto *LoadMMO =
1268         MF.getMachineMemOperand(&SrcMMO, CurrOffset, CopyTy.getSizeInBytes());
1269 
1270     // Create the load.
1271     Register LoadPtr = Src;
1272     if (CurrOffset != 0) {
1273       auto Offset =
1274           MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), CurrOffset);
1275       LoadPtr = MIB.buildPtrAdd(PtrTy, Src, Offset).getReg(0);
1276     }
1277     LoadVals.push_back(MIB.buildLoad(CopyTy, LoadPtr, *LoadMMO).getReg(0));
1278     CurrOffset += CopyTy.getSizeInBytes();
1279   }
1280 
1281   CurrOffset = 0;
1282   for (unsigned I = 0; I < MemOps.size(); ++I) {
1283     LLT CopyTy = MemOps[I];
1284     // Now store the values loaded.
1285     auto *StoreMMO =
1286         MF.getMachineMemOperand(&DstMMO, CurrOffset, CopyTy.getSizeInBytes());
1287 
1288     Register StorePtr = Dst;
1289     if (CurrOffset != 0) {
1290       auto Offset =
1291           MIB.buildConstant(LLT::scalar(PtrTy.getSizeInBits()), CurrOffset);
1292       StorePtr = MIB.buildPtrAdd(PtrTy, Dst, Offset).getReg(0);
1293     }
1294     MIB.buildStore(LoadVals[I], StorePtr, *StoreMMO);
1295     CurrOffset += CopyTy.getSizeInBytes();
1296   }
1297   MI.eraseFromParent();
1298   return true;
1299 }
1300 
1301 bool CombinerHelper::tryCombineMemCpyFamily(MachineInstr &MI, unsigned MaxLen) {
1302   // This combine is fairly complex so it's not written with a separate
1303   // matcher function.
1304   assert(MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
1305   Intrinsic::ID ID = (Intrinsic::ID)MI.getIntrinsicID();
1306   assert((ID == Intrinsic::memcpy || ID == Intrinsic::memmove ||
1307           ID == Intrinsic::memset) &&
1308          "Expected a memcpy like intrinsic");
1309 
1310   auto MMOIt = MI.memoperands_begin();
1311   const MachineMemOperand *MemOp = *MMOIt;
1312   bool IsVolatile = MemOp->isVolatile();
1313   // Don't try to optimize volatile.
1314   if (IsVolatile)
1315     return false;
1316 
1317   Align DstAlign = MemOp->getBaseAlign();
1318   Align SrcAlign;
1319   Register Dst = MI.getOperand(1).getReg();
1320   Register Src = MI.getOperand(2).getReg();
1321   Register Len = MI.getOperand(3).getReg();
1322 
1323   if (ID != Intrinsic::memset) {
1324     assert(MMOIt != MI.memoperands_end() && "Expected a second MMO on MI");
1325     MemOp = *(++MMOIt);
1326     SrcAlign = MemOp->getBaseAlign();
1327   }
1328 
1329   // See if this is a constant length copy
1330   auto LenVRegAndVal = getConstantVRegValWithLookThrough(Len, MRI);
1331   if (!LenVRegAndVal)
1332     return false; // Leave it to the legalizer to lower it to a libcall.
1333   unsigned KnownLen = LenVRegAndVal->Value;
1334 
1335   if (KnownLen == 0) {
1336     MI.eraseFromParent();
1337     return true;
1338   }
1339 
1340   if (MaxLen && KnownLen > MaxLen)
1341     return false;
1342 
1343   if (ID == Intrinsic::memcpy)
1344     return optimizeMemcpy(MI, Dst, Src, KnownLen, DstAlign, SrcAlign, IsVolatile);
1345   if (ID == Intrinsic::memmove)
1346     return optimizeMemmove(MI, Dst, Src, KnownLen, DstAlign, SrcAlign, IsVolatile);
1347   if (ID == Intrinsic::memset)
1348     return optimizeMemset(MI, Dst, Src, KnownLen, DstAlign, IsVolatile);
1349   return false;
1350 }
1351 
1352 bool CombinerHelper::matchPtrAddImmedChain(MachineInstr &MI,
1353                                            PtrAddChain &MatchInfo) {
1354   // We're trying to match the following pattern:
1355   //   %t1 = G_PTR_ADD %base, G_CONSTANT imm1
1356   //   %root = G_PTR_ADD %t1, G_CONSTANT imm2
1357   // -->
1358   //   %root = G_PTR_ADD %base, G_CONSTANT (imm1 + imm2)
1359 
1360   if (MI.getOpcode() != TargetOpcode::G_PTR_ADD)
1361     return false;
1362 
1363   Register Add2 = MI.getOperand(1).getReg();
1364   Register Imm1 = MI.getOperand(2).getReg();
1365   auto MaybeImmVal = getConstantVRegValWithLookThrough(Imm1, MRI);
1366   if (!MaybeImmVal)
1367     return false;
1368 
1369   MachineInstr *Add2Def = MRI.getUniqueVRegDef(Add2);
1370   if (!Add2Def || Add2Def->getOpcode() != TargetOpcode::G_PTR_ADD)
1371     return false;
1372 
1373   Register Base = Add2Def->getOperand(1).getReg();
1374   Register Imm2 = Add2Def->getOperand(2).getReg();
1375   auto MaybeImm2Val = getConstantVRegValWithLookThrough(Imm2, MRI);
1376   if (!MaybeImm2Val)
1377     return false;
1378 
1379   // Pass the combined immediate to the apply function.
1380   MatchInfo.Imm = MaybeImmVal->Value + MaybeImm2Val->Value;
1381   MatchInfo.Base = Base;
1382   return true;
1383 }
1384 
1385 bool CombinerHelper::applyPtrAddImmedChain(MachineInstr &MI,
1386                                            PtrAddChain &MatchInfo) {
1387   assert(MI.getOpcode() == TargetOpcode::G_PTR_ADD && "Expected G_PTR_ADD");
1388   MachineIRBuilder MIB(MI);
1389   LLT OffsetTy = MRI.getType(MI.getOperand(2).getReg());
1390   auto NewOffset = MIB.buildConstant(OffsetTy, MatchInfo.Imm);
1391   Observer.changingInstr(MI);
1392   MI.getOperand(1).setReg(MatchInfo.Base);
1393   MI.getOperand(2).setReg(NewOffset.getReg(0));
1394   Observer.changedInstr(MI);
1395   return true;
1396 }
1397 
1398 bool CombinerHelper::matchCombineMulToShl(MachineInstr &MI,
1399                                           unsigned &ShiftVal) {
1400   assert(MI.getOpcode() == TargetOpcode::G_MUL && "Expected a G_MUL");
1401   auto MaybeImmVal =
1402       getConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
1403   if (!MaybeImmVal || !isPowerOf2_64(MaybeImmVal->Value))
1404     return false;
1405   ShiftVal = Log2_64(MaybeImmVal->Value);
1406   return true;
1407 }
1408 
1409 bool CombinerHelper::applyCombineMulToShl(MachineInstr &MI,
1410                                           unsigned &ShiftVal) {
1411   assert(MI.getOpcode() == TargetOpcode::G_MUL && "Expected a G_MUL");
1412   MachineIRBuilder MIB(MI);
1413   LLT ShiftTy = MRI.getType(MI.getOperand(0).getReg());
1414   auto ShiftCst = MIB.buildConstant(ShiftTy, ShiftVal);
1415   Observer.changingInstr(MI);
1416   MI.setDesc(MIB.getTII().get(TargetOpcode::G_SHL));
1417   MI.getOperand(2).setReg(ShiftCst.getReg(0));
1418   Observer.changedInstr(MI);
1419   return true;
1420 }
1421 
1422 bool CombinerHelper::matchCombineShiftToUnmerge(MachineInstr &MI,
1423                                                 unsigned TargetShiftSize,
1424                                                 unsigned &ShiftVal) {
1425   assert((MI.getOpcode() == TargetOpcode::G_SHL ||
1426           MI.getOpcode() == TargetOpcode::G_LSHR ||
1427           MI.getOpcode() == TargetOpcode::G_ASHR) && "Expected a shift");
1428 
1429   LLT Ty = MRI.getType(MI.getOperand(0).getReg());
1430   if (Ty.isVector()) // TODO:
1431     return false;
1432 
1433   // Don't narrow further than the requested size.
1434   unsigned Size = Ty.getSizeInBits();
1435   if (Size <= TargetShiftSize)
1436     return false;
1437 
1438   auto MaybeImmVal =
1439     getConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
1440   if (!MaybeImmVal)
1441     return false;
1442 
1443   ShiftVal = MaybeImmVal->Value;
1444   return ShiftVal >= Size / 2 && ShiftVal < Size;
1445 }
1446 
1447 bool CombinerHelper::applyCombineShiftToUnmerge(MachineInstr &MI,
1448                                                 const unsigned &ShiftVal) {
1449   Register DstReg = MI.getOperand(0).getReg();
1450   Register SrcReg = MI.getOperand(1).getReg();
1451   LLT Ty = MRI.getType(SrcReg);
1452   unsigned Size = Ty.getSizeInBits();
1453   unsigned HalfSize = Size / 2;
1454   assert(ShiftVal >= HalfSize);
1455 
1456   LLT HalfTy = LLT::scalar(HalfSize);
1457 
1458   Builder.setInstr(MI);
1459   auto Unmerge = Builder.buildUnmerge(HalfTy, SrcReg);
1460   unsigned NarrowShiftAmt = ShiftVal - HalfSize;
1461 
1462   if (MI.getOpcode() == TargetOpcode::G_LSHR) {
1463     Register Narrowed = Unmerge.getReg(1);
1464 
1465     //  dst = G_LSHR s64:x, C for C >= 32
1466     // =>
1467     //   lo, hi = G_UNMERGE_VALUES x
1468     //   dst = G_MERGE_VALUES (G_LSHR hi, C - 32), 0
1469 
1470     if (NarrowShiftAmt != 0) {
1471       Narrowed = Builder.buildLShr(HalfTy, Narrowed,
1472         Builder.buildConstant(HalfTy, NarrowShiftAmt)).getReg(0);
1473     }
1474 
1475     auto Zero = Builder.buildConstant(HalfTy, 0);
1476     Builder.buildMerge(DstReg, { Narrowed, Zero });
1477   } else if (MI.getOpcode() == TargetOpcode::G_SHL) {
1478     Register Narrowed = Unmerge.getReg(0);
1479     //  dst = G_SHL s64:x, C for C >= 32
1480     // =>
1481     //   lo, hi = G_UNMERGE_VALUES x
1482     //   dst = G_MERGE_VALUES 0, (G_SHL hi, C - 32)
1483     if (NarrowShiftAmt != 0) {
1484       Narrowed = Builder.buildShl(HalfTy, Narrowed,
1485         Builder.buildConstant(HalfTy, NarrowShiftAmt)).getReg(0);
1486     }
1487 
1488     auto Zero = Builder.buildConstant(HalfTy, 0);
1489     Builder.buildMerge(DstReg, { Zero, Narrowed });
1490   } else {
1491     assert(MI.getOpcode() == TargetOpcode::G_ASHR);
1492     auto Hi = Builder.buildAShr(
1493       HalfTy, Unmerge.getReg(1),
1494       Builder.buildConstant(HalfTy, HalfSize - 1));
1495 
1496     if (ShiftVal == HalfSize) {
1497       // (G_ASHR i64:x, 32) ->
1498       //   G_MERGE_VALUES hi_32(x), (G_ASHR hi_32(x), 31)
1499       Builder.buildMerge(DstReg, { Unmerge.getReg(1), Hi });
1500     } else if (ShiftVal == Size - 1) {
1501       // Don't need a second shift.
1502       // (G_ASHR i64:x, 63) ->
1503       //   %narrowed = (G_ASHR hi_32(x), 31)
1504       //   G_MERGE_VALUES %narrowed, %narrowed
1505       Builder.buildMerge(DstReg, { Hi, Hi });
1506     } else {
1507       auto Lo = Builder.buildAShr(
1508         HalfTy, Unmerge.getReg(1),
1509         Builder.buildConstant(HalfTy, ShiftVal - HalfSize));
1510 
1511       // (G_ASHR i64:x, C) ->, for C >= 32
1512       //   G_MERGE_VALUES (G_ASHR hi_32(x), C - 32), (G_ASHR hi_32(x), 31)
1513       Builder.buildMerge(DstReg, { Lo, Hi });
1514     }
1515   }
1516 
1517   MI.eraseFromParent();
1518   return true;
1519 }
1520 
1521 bool CombinerHelper::tryCombineShiftToUnmerge(MachineInstr &MI,
1522                                               unsigned TargetShiftAmount) {
1523   unsigned ShiftAmt;
1524   if (matchCombineShiftToUnmerge(MI, TargetShiftAmount, ShiftAmt)) {
1525     applyCombineShiftToUnmerge(MI, ShiftAmt);
1526     return true;
1527   }
1528 
1529   return false;
1530 }
1531 
1532 bool CombinerHelper::matchAnyExplicitUseIsUndef(MachineInstr &MI) {
1533   return any_of(MI.explicit_uses(), [this](const MachineOperand &MO) {
1534     return MO.isReg() &&
1535            getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MO.getReg(), MRI);
1536   });
1537 }
1538 
1539 bool CombinerHelper::matchAllExplicitUsesAreUndef(MachineInstr &MI) {
1540   return all_of(MI.explicit_uses(), [this](const MachineOperand &MO) {
1541     return !MO.isReg() ||
1542            getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MO.getReg(), MRI);
1543   });
1544 }
1545 
1546 bool CombinerHelper::matchUndefShuffleVectorMask(MachineInstr &MI) {
1547   assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
1548   ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
1549   return all_of(Mask, [](int Elt) { return Elt < 0; });
1550 }
1551 
1552 bool CombinerHelper::matchUndefStore(MachineInstr &MI) {
1553   assert(MI.getOpcode() == TargetOpcode::G_STORE);
1554   return getOpcodeDef(TargetOpcode::G_IMPLICIT_DEF, MI.getOperand(0).getReg(),
1555                       MRI);
1556 }
1557 
1558 bool CombinerHelper::eraseInst(MachineInstr &MI) {
1559   MI.eraseFromParent();
1560   return true;
1561 }
1562 
1563 bool CombinerHelper::matchEqualDefs(const MachineOperand &MOP1,
1564                                     const MachineOperand &MOP2) {
1565   if (!MOP1.isReg() || !MOP2.isReg())
1566     return false;
1567   MachineInstr *I1 = getDefIgnoringCopies(MOP1.getReg(), MRI);
1568   if (!I1)
1569     return false;
1570   MachineInstr *I2 = getDefIgnoringCopies(MOP2.getReg(), MRI);
1571   if (!I2)
1572     return false;
1573 
1574   // Handle a case like this:
1575   //
1576   // %0:_(s64), %1:_(s64) = G_UNMERGE_VALUES %2:_(<2 x s64>)
1577   //
1578   // Even though %0 and %1 are produced by the same instruction they are not
1579   // the same values.
1580   if (I1 == I2)
1581     return MOP1.getReg() == MOP2.getReg();
1582 
1583   // If we have an instruction which loads or stores, we can't guarantee that
1584   // it is identical.
1585   //
1586   // For example, we may have
1587   //
1588   // %x1 = G_LOAD %addr (load N from @somewhere)
1589   // ...
1590   // call @foo
1591   // ...
1592   // %x2 = G_LOAD %addr (load N from @somewhere)
1593   // ...
1594   // %or = G_OR %x1, %x2
1595   //
1596   // It's possible that @foo will modify whatever lives at the address we're
1597   // loading from. To be safe, let's just assume that all loads and stores
1598   // are different (unless we have something which is guaranteed to not
1599   // change.)
1600   if (I1->mayLoadOrStore() && !I1->isDereferenceableInvariantLoad(nullptr))
1601     return false;
1602 
1603   // Check for physical registers on the instructions first to avoid cases
1604   // like this:
1605   //
1606   // %a = COPY $physreg
1607   // ...
1608   // SOMETHING implicit-def $physreg
1609   // ...
1610   // %b = COPY $physreg
1611   //
1612   // These copies are not equivalent.
1613   if (any_of(I1->uses(), [](const MachineOperand &MO) {
1614         return MO.isReg() && MO.getReg().isPhysical();
1615       })) {
1616     // Check if we have a case like this:
1617     //
1618     // %a = COPY $physreg
1619     // %b = COPY %a
1620     //
1621     // In this case, I1 and I2 will both be equal to %a = COPY $physreg.
1622     // From that, we know that they must have the same value, since they must
1623     // have come from the same COPY.
1624     return I1->isIdenticalTo(*I2);
1625   }
1626 
1627   // We don't have any physical registers, so we don't necessarily need the
1628   // same vreg defs.
1629   //
1630   // On the off-chance that there's some target instruction feeding into the
1631   // instruction, let's use produceSameValue instead of isIdenticalTo.
1632   return Builder.getTII().produceSameValue(*I1, *I2, &MRI);
1633 }
1634 
1635 bool CombinerHelper::matchConstantOp(const MachineOperand &MOP, int64_t C) {
1636   if (!MOP.isReg())
1637     return false;
1638   // MIPatternMatch doesn't let us look through G_ZEXT etc.
1639   auto ValAndVReg = getConstantVRegValWithLookThrough(MOP.getReg(), MRI);
1640   return ValAndVReg && ValAndVReg->Value == C;
1641 }
1642 
1643 bool CombinerHelper::replaceSingleDefInstWithOperand(MachineInstr &MI,
1644                                                      unsigned OpIdx) {
1645   assert(MI.getNumExplicitDefs() == 1 && "Expected one explicit def?");
1646   Register OldReg = MI.getOperand(0).getReg();
1647   Register Replacement = MI.getOperand(OpIdx).getReg();
1648   assert(canReplaceReg(OldReg, Replacement, MRI) && "Cannot replace register?");
1649   MI.eraseFromParent();
1650   replaceRegWith(MRI, OldReg, Replacement);
1651   return true;
1652 }
1653 
1654 bool CombinerHelper::matchSelectSameVal(MachineInstr &MI) {
1655   assert(MI.getOpcode() == TargetOpcode::G_SELECT);
1656   // Match (cond ? x : x)
1657   return matchEqualDefs(MI.getOperand(2), MI.getOperand(3)) &&
1658          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(2).getReg(),
1659                        MRI);
1660 }
1661 
1662 bool CombinerHelper::matchBinOpSameVal(MachineInstr &MI) {
1663   return matchEqualDefs(MI.getOperand(1), MI.getOperand(2)) &&
1664          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(1).getReg(),
1665                        MRI);
1666 }
1667 
1668 bool CombinerHelper::matchOperandIsZero(MachineInstr &MI, unsigned OpIdx) {
1669   return matchConstantOp(MI.getOperand(OpIdx), 0) &&
1670          canReplaceReg(MI.getOperand(0).getReg(), MI.getOperand(OpIdx).getReg(),
1671                        MRI);
1672 }
1673 
1674 bool CombinerHelper::replaceInstWithFConstant(MachineInstr &MI, double C) {
1675   assert(MI.getNumDefs() == 1 && "Expected only one def?");
1676   Builder.setInstr(MI);
1677   Builder.buildFConstant(MI.getOperand(0), C);
1678   MI.eraseFromParent();
1679   return true;
1680 }
1681 
1682 bool CombinerHelper::replaceInstWithConstant(MachineInstr &MI, int64_t C) {
1683   assert(MI.getNumDefs() == 1 && "Expected only one def?");
1684   Builder.setInstr(MI);
1685   Builder.buildConstant(MI.getOperand(0), C);
1686   MI.eraseFromParent();
1687   return true;
1688 }
1689 
1690 bool CombinerHelper::replaceInstWithUndef(MachineInstr &MI) {
1691   assert(MI.getNumDefs() == 1 && "Expected only one def?");
1692   Builder.setInstr(MI);
1693   Builder.buildUndef(MI.getOperand(0));
1694   MI.eraseFromParent();
1695   return true;
1696 }
1697 
1698 bool CombinerHelper::matchSimplifyAddToSub(
1699     MachineInstr &MI, std::tuple<Register, Register> &MatchInfo) {
1700   Register LHS = MI.getOperand(1).getReg();
1701   Register RHS = MI.getOperand(2).getReg();
1702   Register &NewLHS = std::get<0>(MatchInfo);
1703   Register &NewRHS = std::get<1>(MatchInfo);
1704 
1705   // Helper lambda to check for opportunities for
1706   // ((0-A) + B) -> B - A
1707   // (A + (0-B)) -> A - B
1708   auto CheckFold = [&](Register &MaybeSub, Register &MaybeNewLHS) {
1709     int64_t Cst;
1710     if (!mi_match(MaybeSub, MRI, m_GSub(m_ICst(Cst), m_Reg(NewRHS))) ||
1711         Cst != 0)
1712       return false;
1713     NewLHS = MaybeNewLHS;
1714     return true;
1715   };
1716 
1717   return CheckFold(LHS, RHS) || CheckFold(RHS, LHS);
1718 }
1719 
1720 bool CombinerHelper::applySimplifyAddToSub(
1721     MachineInstr &MI, std::tuple<Register, Register> &MatchInfo) {
1722   Builder.setInstr(MI);
1723   Register SubLHS, SubRHS;
1724   std::tie(SubLHS, SubRHS) = MatchInfo;
1725   Builder.buildSub(MI.getOperand(0).getReg(), SubLHS, SubRHS);
1726   MI.eraseFromParent();
1727   return true;
1728 }
1729 
1730 bool CombinerHelper::tryCombine(MachineInstr &MI) {
1731   if (tryCombineCopy(MI))
1732     return true;
1733   if (tryCombineExtendingLoads(MI))
1734     return true;
1735   if (tryCombineIndexedLoadStore(MI))
1736     return true;
1737   return false;
1738 }
1739