1 //===- LoadStoreOpt.cpp ----------- Generic memory optimizations -*- C++ -*-==//
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 /// \file
9 /// This file implements the LoadStoreOpt optimization pass.
10 //===----------------------------------------------------------------------===//
11 
12 #include "llvm/CodeGen/GlobalISel/LoadStoreOpt.h"
13 #include "llvm/ADT/STLExtras.h"
14 #include "llvm/ADT/SmallPtrSet.h"
15 #include "llvm/ADT/Statistic.h"
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/Analysis/MemoryLocation.h"
18 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
19 #include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
20 #include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
21 #include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
22 #include "llvm/CodeGen/GlobalISel/Utils.h"
23 #include "llvm/CodeGen/LowLevelTypeUtils.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstr.h"
28 #include "llvm/CodeGen/MachineOptimizationRemarkEmitter.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/Register.h"
31 #include "llvm/CodeGen/TargetLowering.h"
32 #include "llvm/CodeGen/TargetOpcodes.h"
33 #include "llvm/IR/DebugInfoMetadata.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/Support/AtomicOrdering.h"
36 #include "llvm/Support/Casting.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/MathExtras.h"
40 #include <algorithm>
41 
42 #define DEBUG_TYPE "loadstore-opt"
43 
44 using namespace llvm;
45 using namespace ore;
46 using namespace MIPatternMatch;
47 
48 STATISTIC(NumStoresMerged, "Number of stores merged");
49 
50 const unsigned MaxStoreSizeToForm = 128;
51 
52 char LoadStoreOpt::ID = 0;
53 INITIALIZE_PASS_BEGIN(LoadStoreOpt, DEBUG_TYPE, "Generic memory optimizations",
54                       false, false)
55 INITIALIZE_PASS_END(LoadStoreOpt, DEBUG_TYPE, "Generic memory optimizations",
56                     false, false)
57 
58 LoadStoreOpt::LoadStoreOpt(std::function<bool(const MachineFunction &)> F)
59     : MachineFunctionPass(ID), DoNotRunPass(F) {}
60 
61 LoadStoreOpt::LoadStoreOpt()
62     : LoadStoreOpt([](const MachineFunction &) { return false; }) {}
63 
64 void LoadStoreOpt::init(MachineFunction &MF) {
65   this->MF = &MF;
66   MRI = &MF.getRegInfo();
67   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
68   TLI = MF.getSubtarget().getTargetLowering();
69   LI = MF.getSubtarget().getLegalizerInfo();
70   Builder.setMF(MF);
71   IsPreLegalizer = !MF.getProperties().hasProperty(
72       MachineFunctionProperties::Property::Legalized);
73   InstsToErase.clear();
74 }
75 
76 void LoadStoreOpt::getAnalysisUsage(AnalysisUsage &AU) const {
77   AU.addRequired<AAResultsWrapperPass>();
78   AU.setPreservesAll();
79   getSelectionDAGFallbackAnalysisUsage(AU);
80   MachineFunctionPass::getAnalysisUsage(AU);
81 }
82 
83 BaseIndexOffset GISelAddressing::getPointerInfo(Register Ptr,
84                                                 MachineRegisterInfo &MRI) {
85   BaseIndexOffset Info;
86   Register PtrAddRHS;
87   if (!mi_match(Ptr, MRI, m_GPtrAdd(m_Reg(Info.BaseReg), m_Reg(PtrAddRHS)))) {
88     Info.BaseReg = Ptr;
89     Info.IndexReg = Register();
90     Info.IsIndexSignExt = false;
91     return Info;
92   }
93 
94   auto RHSCst = getIConstantVRegValWithLookThrough(PtrAddRHS, MRI);
95   if (RHSCst)
96     Info.Offset = RHSCst->Value.getSExtValue();
97 
98   // Just recognize a simple case for now. In future we'll need to match
99   // indexing patterns for base + index + constant.
100   Info.IndexReg = PtrAddRHS;
101   Info.IsIndexSignExt = false;
102   return Info;
103 }
104 
105 bool GISelAddressing::aliasIsKnownForLoadStore(const MachineInstr &MI1,
106                                                const MachineInstr &MI2,
107                                                bool &IsAlias,
108                                                MachineRegisterInfo &MRI) {
109   auto *LdSt1 = dyn_cast<GLoadStore>(&MI1);
110   auto *LdSt2 = dyn_cast<GLoadStore>(&MI2);
111   if (!LdSt1 || !LdSt2)
112     return false;
113 
114   BaseIndexOffset BasePtr0 = getPointerInfo(LdSt1->getPointerReg(), MRI);
115   BaseIndexOffset BasePtr1 = getPointerInfo(LdSt2->getPointerReg(), MRI);
116 
117   if (!BasePtr0.BaseReg.isValid() || !BasePtr1.BaseReg.isValid())
118     return false;
119 
120   int64_t Size1 = LdSt1->getMemSize();
121   int64_t Size2 = LdSt2->getMemSize();
122 
123   int64_t PtrDiff;
124   if (BasePtr0.BaseReg == BasePtr1.BaseReg) {
125     PtrDiff = BasePtr1.Offset - BasePtr0.Offset;
126     // If the size of memory access is unknown, do not use it to do analysis.
127     // One example of unknown size memory access is to load/store scalable
128     // vector objects on the stack.
129     // BasePtr1 is PtrDiff away from BasePtr0. They alias if none of the
130     // following situations arise:
131     if (PtrDiff >= 0 &&
132         Size1 != static_cast<int64_t>(MemoryLocation::UnknownSize)) {
133       // [----BasePtr0----]
134       //                         [---BasePtr1--]
135       // ========PtrDiff========>
136       IsAlias = !(Size1 <= PtrDiff);
137       return true;
138     }
139     if (PtrDiff < 0 &&
140         Size2 != static_cast<int64_t>(MemoryLocation::UnknownSize)) {
141       //                     [----BasePtr0----]
142       // [---BasePtr1--]
143       // =====(-PtrDiff)====>
144       IsAlias = !((PtrDiff + Size2) <= 0);
145       return true;
146     }
147     return false;
148   }
149 
150   // If both BasePtr0 and BasePtr1 are FrameIndexes, we will not be
151   // able to calculate their relative offset if at least one arises
152   // from an alloca. However, these allocas cannot overlap and we
153   // can infer there is no alias.
154   auto *Base0Def = getDefIgnoringCopies(BasePtr0.BaseReg, MRI);
155   auto *Base1Def = getDefIgnoringCopies(BasePtr1.BaseReg, MRI);
156   if (!Base0Def || !Base1Def)
157     return false; // Couldn't tell anything.
158 
159 
160   if (Base0Def->getOpcode() != Base1Def->getOpcode())
161     return false;
162 
163   if (Base0Def->getOpcode() == TargetOpcode::G_FRAME_INDEX) {
164     MachineFrameInfo &MFI = Base0Def->getMF()->getFrameInfo();
165     // If the bases have the same frame index but we couldn't find a
166     // constant offset, (indices are different) be conservative.
167     if (Base0Def != Base1Def &&
168         (!MFI.isFixedObjectIndex(Base0Def->getOperand(1).getIndex()) ||
169          !MFI.isFixedObjectIndex(Base1Def->getOperand(1).getIndex()))) {
170       IsAlias = false;
171       return true;
172     }
173   }
174 
175   // This implementation is a lot more primitive than the SDAG one for now.
176   // FIXME: what about constant pools?
177   if (Base0Def->getOpcode() == TargetOpcode::G_GLOBAL_VALUE) {
178     auto GV0 = Base0Def->getOperand(1).getGlobal();
179     auto GV1 = Base1Def->getOperand(1).getGlobal();
180     if (GV0 != GV1) {
181       IsAlias = false;
182       return true;
183     }
184   }
185 
186   // Can't tell anything about aliasing.
187   return false;
188 }
189 
190 bool GISelAddressing::instMayAlias(const MachineInstr &MI,
191                                    const MachineInstr &Other,
192                                    MachineRegisterInfo &MRI,
193                                    AliasAnalysis *AA) {
194   struct MemUseCharacteristics {
195     bool IsVolatile;
196     bool IsAtomic;
197     Register BasePtr;
198     int64_t Offset;
199     uint64_t NumBytes;
200     MachineMemOperand *MMO;
201   };
202 
203   auto getCharacteristics =
204       [&](const MachineInstr *MI) -> MemUseCharacteristics {
205     if (const auto *LS = dyn_cast<GLoadStore>(MI)) {
206       Register BaseReg;
207       int64_t Offset = 0;
208       // No pre/post-inc addressing modes are considered here, unlike in SDAG.
209       if (!mi_match(LS->getPointerReg(), MRI,
210                     m_GPtrAdd(m_Reg(BaseReg), m_ICst(Offset)))) {
211         BaseReg = LS->getPointerReg();
212         Offset = 0;
213       }
214 
215       uint64_t Size = MemoryLocation::getSizeOrUnknown(
216           LS->getMMO().getMemoryType().getSizeInBytes());
217       return {LS->isVolatile(),       LS->isAtomic(),          BaseReg,
218               Offset /*base offset*/, Size, &LS->getMMO()};
219     }
220     // FIXME: support recognizing lifetime instructions.
221     // Default.
222     return {false /*isvolatile*/,
223             /*isAtomic*/ false,          Register(),
224             (int64_t)0 /*offset*/,       0 /*size*/,
225             (MachineMemOperand *)nullptr};
226   };
227   MemUseCharacteristics MUC0 = getCharacteristics(&MI),
228                         MUC1 = getCharacteristics(&Other);
229 
230   // If they are to the same address, then they must be aliases.
231   if (MUC0.BasePtr.isValid() && MUC0.BasePtr == MUC1.BasePtr &&
232       MUC0.Offset == MUC1.Offset)
233     return true;
234 
235   // If they are both volatile then they cannot be reordered.
236   if (MUC0.IsVolatile && MUC1.IsVolatile)
237     return true;
238 
239   // Be conservative about atomics for the moment
240   // TODO: This is way overconservative for unordered atomics (see D66309)
241   if (MUC0.IsAtomic && MUC1.IsAtomic)
242     return true;
243 
244   // If one operation reads from invariant memory, and the other may store, they
245   // cannot alias.
246   if (MUC0.MMO && MUC1.MMO) {
247     if ((MUC0.MMO->isInvariant() && MUC1.MMO->isStore()) ||
248         (MUC1.MMO->isInvariant() && MUC0.MMO->isStore()))
249       return false;
250   }
251 
252   // Try to prove that there is aliasing, or that there is no aliasing. Either
253   // way, we can return now. If nothing can be proved, proceed with more tests.
254   bool IsAlias;
255   if (GISelAddressing::aliasIsKnownForLoadStore(MI, Other, IsAlias, MRI))
256     return IsAlias;
257 
258   // The following all rely on MMO0 and MMO1 being valid.
259   if (!MUC0.MMO || !MUC1.MMO)
260     return true;
261 
262   // FIXME: port the alignment based alias analysis from SDAG's isAlias().
263   int64_t SrcValOffset0 = MUC0.MMO->getOffset();
264   int64_t SrcValOffset1 = MUC1.MMO->getOffset();
265   uint64_t Size0 = MUC0.NumBytes;
266   uint64_t Size1 = MUC1.NumBytes;
267   if (AA && MUC0.MMO->getValue() && MUC1.MMO->getValue() &&
268       Size0 != MemoryLocation::UnknownSize &&
269       Size1 != MemoryLocation::UnknownSize) {
270     // Use alias analysis information.
271     int64_t MinOffset = std::min(SrcValOffset0, SrcValOffset1);
272     int64_t Overlap0 = Size0 + SrcValOffset0 - MinOffset;
273     int64_t Overlap1 = Size1 + SrcValOffset1 - MinOffset;
274     if (AA->isNoAlias(MemoryLocation(MUC0.MMO->getValue(), Overlap0,
275                                      MUC0.MMO->getAAInfo()),
276                       MemoryLocation(MUC1.MMO->getValue(), Overlap1,
277                                      MUC1.MMO->getAAInfo())))
278       return false;
279   }
280 
281   // Otherwise we have to assume they alias.
282   return true;
283 }
284 
285 /// Returns true if the instruction creates an unavoidable hazard that
286 /// forces a boundary between store merge candidates.
287 static bool isInstHardMergeHazard(MachineInstr &MI) {
288   return MI.hasUnmodeledSideEffects() || MI.hasOrderedMemoryRef();
289 }
290 
291 bool LoadStoreOpt::mergeStores(SmallVectorImpl<GStore *> &StoresToMerge) {
292   // Try to merge all the stores in the vector, splitting into separate segments
293   // as necessary.
294   assert(StoresToMerge.size() > 1 && "Expected multiple stores to merge");
295   LLT OrigTy = MRI->getType(StoresToMerge[0]->getValueReg());
296   LLT PtrTy = MRI->getType(StoresToMerge[0]->getPointerReg());
297   unsigned AS = PtrTy.getAddressSpace();
298   // Ensure the legal store info is computed for this address space.
299   initializeStoreMergeTargetInfo(AS);
300   const auto &LegalSizes = LegalStoreSizes[AS];
301 
302 #ifndef NDEBUG
303   for (auto *StoreMI : StoresToMerge)
304     assert(MRI->getType(StoreMI->getValueReg()) == OrigTy);
305 #endif
306 
307   const auto &DL = MF->getFunction().getParent()->getDataLayout();
308   bool AnyMerged = false;
309   do {
310     unsigned NumPow2 = llvm::bit_floor(StoresToMerge.size());
311     unsigned MaxSizeBits = NumPow2 * OrigTy.getSizeInBits().getFixedValue();
312     // Compute the biggest store we can generate to handle the number of stores.
313     unsigned MergeSizeBits;
314     for (MergeSizeBits = MaxSizeBits; MergeSizeBits > 1; MergeSizeBits /= 2) {
315       LLT StoreTy = LLT::scalar(MergeSizeBits);
316       EVT StoreEVT =
317           getApproximateEVTForLLT(StoreTy, DL, MF->getFunction().getContext());
318       if (LegalSizes.size() > MergeSizeBits && LegalSizes[MergeSizeBits] &&
319           TLI->canMergeStoresTo(AS, StoreEVT, *MF) &&
320           (TLI->isTypeLegal(StoreEVT)))
321         break; // We can generate a MergeSize bits store.
322     }
323     if (MergeSizeBits <= OrigTy.getSizeInBits())
324       return AnyMerged; // No greater merge.
325 
326     unsigned NumStoresToMerge = MergeSizeBits / OrigTy.getSizeInBits();
327     // Perform the actual merging.
328     SmallVector<GStore *, 8> SingleMergeStores(
329         StoresToMerge.begin(), StoresToMerge.begin() + NumStoresToMerge);
330     AnyMerged |= doSingleStoreMerge(SingleMergeStores);
331     StoresToMerge.erase(StoresToMerge.begin(),
332                         StoresToMerge.begin() + NumStoresToMerge);
333   } while (StoresToMerge.size() > 1);
334   return AnyMerged;
335 }
336 
337 bool LoadStoreOpt::isLegalOrBeforeLegalizer(const LegalityQuery &Query,
338                                             MachineFunction &MF) const {
339   auto Action = LI->getAction(Query).Action;
340   // If the instruction is unsupported, it can't be legalized at all.
341   if (Action == LegalizeActions::Unsupported)
342     return false;
343   return IsPreLegalizer || Action == LegalizeAction::Legal;
344 }
345 
346 bool LoadStoreOpt::doSingleStoreMerge(SmallVectorImpl<GStore *> &Stores) {
347   assert(Stores.size() > 1);
348   // We know that all the stores are consecutive and there are no aliasing
349   // operations in the range. However, the values that are being stored may be
350   // generated anywhere before each store. To ensure we have the values
351   // available, we materialize the wide value and new store at the place of the
352   // final store in the merge sequence.
353   GStore *FirstStore = Stores[0];
354   const unsigned NumStores = Stores.size();
355   LLT SmallTy = MRI->getType(FirstStore->getValueReg());
356   LLT WideValueTy =
357       LLT::scalar(NumStores * SmallTy.getSizeInBits().getFixedValue());
358 
359   // For each store, compute pairwise merged debug locs.
360   DebugLoc MergedLoc = Stores.front()->getDebugLoc();
361   for (auto *Store : drop_begin(Stores))
362     MergedLoc = DILocation::getMergedLocation(MergedLoc, Store->getDebugLoc());
363 
364   Builder.setInstr(*Stores.back());
365   Builder.setDebugLoc(MergedLoc);
366 
367   // If all of the store values are constants, then create a wide constant
368   // directly. Otherwise, we need to generate some instructions to merge the
369   // existing values together into a wider type.
370   SmallVector<APInt, 8> ConstantVals;
371   for (auto *Store : Stores) {
372     auto MaybeCst =
373         getIConstantVRegValWithLookThrough(Store->getValueReg(), *MRI);
374     if (!MaybeCst) {
375       ConstantVals.clear();
376       break;
377     }
378     ConstantVals.emplace_back(MaybeCst->Value);
379   }
380 
381   Register WideReg;
382   auto *WideMMO =
383       MF->getMachineMemOperand(&FirstStore->getMMO(), 0, WideValueTy);
384   if (ConstantVals.empty()) {
385     // Mimic the SDAG behaviour here and don't try to do anything for unknown
386     // values. In future, we should also support the cases of loads and
387     // extracted vector elements.
388     return false;
389   }
390 
391   assert(ConstantVals.size() == NumStores);
392   // Check if our wide constant is legal.
393   if (!isLegalOrBeforeLegalizer({TargetOpcode::G_CONSTANT, {WideValueTy}}, *MF))
394     return false;
395   APInt WideConst(WideValueTy.getSizeInBits(), 0);
396   for (unsigned Idx = 0; Idx < ConstantVals.size(); ++Idx) {
397     // Insert the smaller constant into the corresponding position in the
398     // wider one.
399     WideConst.insertBits(ConstantVals[Idx], Idx * SmallTy.getSizeInBits());
400   }
401   WideReg = Builder.buildConstant(WideValueTy, WideConst).getReg(0);
402   auto NewStore =
403       Builder.buildStore(WideReg, FirstStore->getPointerReg(), *WideMMO);
404   (void) NewStore;
405   LLVM_DEBUG(dbgs() << "Merged " << Stores.size()
406                     << " stores into merged store: " << *NewStore);
407   LLVM_DEBUG(for (auto *MI : Stores) dbgs() << "  " << *MI;);
408   NumStoresMerged += Stores.size();
409 
410   MachineOptimizationRemarkEmitter MORE(*MF, nullptr);
411   MORE.emit([&]() {
412     MachineOptimizationRemark R(DEBUG_TYPE, "MergedStore",
413                                 FirstStore->getDebugLoc(),
414                                 FirstStore->getParent());
415     R << "Merged " << NV("NumMerged", Stores.size()) << " stores of "
416       << NV("OrigWidth", SmallTy.getSizeInBytes())
417       << " bytes into a single store of "
418       << NV("NewWidth", WideValueTy.getSizeInBytes()) << " bytes";
419     return R;
420   });
421 
422   for (auto *MI : Stores)
423     InstsToErase.insert(MI);
424   return true;
425 }
426 
427 bool LoadStoreOpt::processMergeCandidate(StoreMergeCandidate &C) {
428   if (C.Stores.size() < 2) {
429     C.reset();
430     return false;
431   }
432 
433   LLVM_DEBUG(dbgs() << "Checking store merge candidate with " << C.Stores.size()
434                     << " stores, starting with " << *C.Stores[0]);
435   // We know that the stores in the candidate are adjacent.
436   // Now we need to check if any potential aliasing instructions recorded
437   // during the search alias with load/stores added to the candidate after.
438   // For example, if we have the candidate:
439   //   C.Stores = [ST1, ST2, ST3, ST4]
440   // and after seeing ST2 we saw a load LD1, which did not alias with ST1 or
441   // ST2, then we would have recorded it into the PotentialAliases structure
442   // with the associated index value of "1". Then we see ST3 and ST4 and add
443   // them to the candidate group. We know that LD1 does not alias with ST1 or
444   // ST2, since we already did that check. However we don't yet know if it
445   // may alias ST3 and ST4, so we perform those checks now.
446   SmallVector<GStore *> StoresToMerge;
447 
448   auto DoesStoreAliasWithPotential = [&](unsigned Idx, GStore &CheckStore) {
449     for (auto AliasInfo : reverse(C.PotentialAliases)) {
450       MachineInstr *PotentialAliasOp = AliasInfo.first;
451       unsigned PreCheckedIdx = AliasInfo.second;
452       if (static_cast<unsigned>(Idx) < PreCheckedIdx) {
453         // Once our store index is lower than the index associated with the
454         // potential alias, we know that we've already checked for this alias
455         // and all of the earlier potential aliases too.
456         return false;
457       }
458       // Need to check this alias.
459       if (GISelAddressing::instMayAlias(CheckStore, *PotentialAliasOp, *MRI,
460                                         AA)) {
461         LLVM_DEBUG(dbgs() << "Potential alias " << *PotentialAliasOp
462                           << " detected\n");
463         return true;
464       }
465     }
466     return false;
467   };
468   // Start from the last store in the group, and check if it aliases with any
469   // of the potential aliasing operations in the list.
470   for (int StoreIdx = C.Stores.size() - 1; StoreIdx >= 0; --StoreIdx) {
471     auto *CheckStore = C.Stores[StoreIdx];
472     if (DoesStoreAliasWithPotential(StoreIdx, *CheckStore))
473       continue;
474     StoresToMerge.emplace_back(CheckStore);
475   }
476 
477   LLVM_DEBUG(dbgs() << StoresToMerge.size()
478                     << " stores remaining after alias checks. Merging...\n");
479 
480   // Now we've checked for aliasing hazards, merge any stores left.
481   C.reset();
482   if (StoresToMerge.size() < 2)
483     return false;
484   return mergeStores(StoresToMerge);
485 }
486 
487 bool LoadStoreOpt::operationAliasesWithCandidate(MachineInstr &MI,
488                                                  StoreMergeCandidate &C) {
489   if (C.Stores.empty())
490     return false;
491   return llvm::any_of(C.Stores, [&](MachineInstr *OtherMI) {
492     return instMayAlias(MI, *OtherMI, *MRI, AA);
493   });
494 }
495 
496 void LoadStoreOpt::StoreMergeCandidate::addPotentialAlias(MachineInstr &MI) {
497   PotentialAliases.emplace_back(std::make_pair(&MI, Stores.size() - 1));
498 }
499 
500 bool LoadStoreOpt::addStoreToCandidate(GStore &StoreMI,
501                                        StoreMergeCandidate &C) {
502   // Check if the given store writes to an adjacent address, and other
503   // requirements.
504   LLT ValueTy = MRI->getType(StoreMI.getValueReg());
505   LLT PtrTy = MRI->getType(StoreMI.getPointerReg());
506 
507   // Only handle scalars.
508   if (!ValueTy.isScalar())
509     return false;
510 
511   // Don't allow truncating stores for now.
512   if (StoreMI.getMemSizeInBits() != ValueTy.getSizeInBits())
513     return false;
514 
515   // Avoid adding volatile or ordered stores to the candidate. We already have a
516   // check for this in instMayAlias() but that only get's called later between
517   // potential aliasing hazards.
518   if (!StoreMI.isSimple())
519     return false;
520 
521   Register StoreAddr = StoreMI.getPointerReg();
522   auto BIO = getPointerInfo(StoreAddr, *MRI);
523   Register StoreBase = BIO.BaseReg;
524   uint64_t StoreOffCst = BIO.Offset;
525   if (C.Stores.empty()) {
526     // This is the first store of the candidate.
527     // If the offset can't possibly allow for a lower addressed store with the
528     // same base, don't bother adding it.
529     if (StoreOffCst < ValueTy.getSizeInBytes())
530       return false;
531     C.BasePtr = StoreBase;
532     C.CurrentLowestOffset = StoreOffCst;
533     C.Stores.emplace_back(&StoreMI);
534     LLVM_DEBUG(dbgs() << "Starting a new merge candidate group with: "
535                       << StoreMI);
536     return true;
537   }
538 
539   // Check the store is the same size as the existing ones in the candidate.
540   if (MRI->getType(C.Stores[0]->getValueReg()).getSizeInBits() !=
541       ValueTy.getSizeInBits())
542     return false;
543 
544   if (MRI->getType(C.Stores[0]->getPointerReg()).getAddressSpace() !=
545       PtrTy.getAddressSpace())
546     return false;
547 
548   // There are other stores in the candidate. Check that the store address
549   // writes to the next lowest adjacent address.
550   if (C.BasePtr != StoreBase)
551     return false;
552   if ((C.CurrentLowestOffset - ValueTy.getSizeInBytes()) !=
553       static_cast<uint64_t>(StoreOffCst))
554     return false;
555 
556   // This writes to an adjacent address. Allow it.
557   C.Stores.emplace_back(&StoreMI);
558   C.CurrentLowestOffset = C.CurrentLowestOffset - ValueTy.getSizeInBytes();
559   LLVM_DEBUG(dbgs() << "Candidate added store: " << StoreMI);
560   return true;
561 }
562 
563 bool LoadStoreOpt::mergeBlockStores(MachineBasicBlock &MBB) {
564   bool Changed = false;
565   // Walk through the block bottom-up, looking for merging candidates.
566   StoreMergeCandidate Candidate;
567   for (MachineInstr &MI : llvm::reverse(MBB)) {
568     if (InstsToErase.contains(&MI))
569       continue;
570 
571     if (auto *StoreMI = dyn_cast<GStore>(&MI)) {
572       // We have a G_STORE. Add it to the candidate if it writes to an adjacent
573       // address.
574       if (!addStoreToCandidate(*StoreMI, Candidate)) {
575         // Store wasn't eligible to be added. May need to record it as a
576         // potential alias.
577         if (operationAliasesWithCandidate(*StoreMI, Candidate)) {
578           Changed |= processMergeCandidate(Candidate);
579           continue;
580         }
581         Candidate.addPotentialAlias(*StoreMI);
582       }
583       continue;
584     }
585 
586     // If we don't have any stores yet, this instruction can't pose a problem.
587     if (Candidate.Stores.empty())
588       continue;
589 
590     // We're dealing with some other kind of instruction.
591     if (isInstHardMergeHazard(MI)) {
592       Changed |= processMergeCandidate(Candidate);
593       Candidate.Stores.clear();
594       continue;
595     }
596 
597     if (!MI.mayLoadOrStore())
598       continue;
599 
600     if (operationAliasesWithCandidate(MI, Candidate)) {
601       // We have a potential alias, so process the current candidate if we can
602       // and then continue looking for a new candidate.
603       Changed |= processMergeCandidate(Candidate);
604       continue;
605     }
606 
607     // Record this instruction as a potential alias for future stores that are
608     // added to the candidate.
609     Candidate.addPotentialAlias(MI);
610   }
611 
612   // Process any candidate left after finishing searching the entire block.
613   Changed |= processMergeCandidate(Candidate);
614 
615   // Erase instructions now that we're no longer iterating over the block.
616   for (auto *MI : InstsToErase)
617     MI->eraseFromParent();
618   InstsToErase.clear();
619   return Changed;
620 }
621 
622 /// Check if the store \p Store is a truncstore that can be merged. That is,
623 /// it's a store of a shifted value of \p SrcVal. If \p SrcVal is an empty
624 /// Register then it does not need to match and SrcVal is set to the source
625 /// value found.
626 /// On match, returns the start byte offset of the \p SrcVal that is being
627 /// stored.
628 static std::optional<int64_t>
629 getTruncStoreByteOffset(GStore &Store, Register &SrcVal,
630                         MachineRegisterInfo &MRI) {
631   Register TruncVal;
632   if (!mi_match(Store.getValueReg(), MRI, m_GTrunc(m_Reg(TruncVal))))
633     return std::nullopt;
634 
635   // The shift amount must be a constant multiple of the narrow type.
636   // It is translated to the offset address in the wide source value "y".
637   //
638   // x = G_LSHR y, ShiftAmtC
639   // s8 z = G_TRUNC x
640   // store z, ...
641   Register FoundSrcVal;
642   int64_t ShiftAmt;
643   if (!mi_match(TruncVal, MRI,
644                 m_any_of(m_GLShr(m_Reg(FoundSrcVal), m_ICst(ShiftAmt)),
645                          m_GAShr(m_Reg(FoundSrcVal), m_ICst(ShiftAmt))))) {
646     if (!SrcVal.isValid() || TruncVal == SrcVal) {
647       if (!SrcVal.isValid())
648         SrcVal = TruncVal;
649       return 0; // If it's the lowest index store.
650     }
651     return std::nullopt;
652   }
653 
654   unsigned NarrowBits = Store.getMMO().getMemoryType().getScalarSizeInBits();
655   if (ShiftAmt % NarrowBits != 0)
656     return std::nullopt;
657   const unsigned Offset = ShiftAmt / NarrowBits;
658 
659   if (SrcVal.isValid() && FoundSrcVal != SrcVal)
660     return std::nullopt;
661 
662   if (!SrcVal.isValid())
663     SrcVal = FoundSrcVal;
664   else if (MRI.getType(SrcVal) != MRI.getType(FoundSrcVal))
665     return std::nullopt;
666   return Offset;
667 }
668 
669 /// Match a pattern where a wide type scalar value is stored by several narrow
670 /// stores. Fold it into a single store or a BSWAP and a store if the targets
671 /// supports it.
672 ///
673 /// Assuming little endian target:
674 ///  i8 *p = ...
675 ///  i32 val = ...
676 ///  p[0] = (val >> 0) & 0xFF;
677 ///  p[1] = (val >> 8) & 0xFF;
678 ///  p[2] = (val >> 16) & 0xFF;
679 ///  p[3] = (val >> 24) & 0xFF;
680 /// =>
681 ///  *((i32)p) = val;
682 ///
683 ///  i8 *p = ...
684 ///  i32 val = ...
685 ///  p[0] = (val >> 24) & 0xFF;
686 ///  p[1] = (val >> 16) & 0xFF;
687 ///  p[2] = (val >> 8) & 0xFF;
688 ///  p[3] = (val >> 0) & 0xFF;
689 /// =>
690 ///  *((i32)p) = BSWAP(val);
691 bool LoadStoreOpt::mergeTruncStore(GStore &StoreMI,
692                                    SmallPtrSetImpl<GStore *> &DeletedStores) {
693   LLT MemTy = StoreMI.getMMO().getMemoryType();
694 
695   // We only handle merging simple stores of 1-4 bytes.
696   if (!MemTy.isScalar())
697     return false;
698   switch (MemTy.getSizeInBits()) {
699   case 8:
700   case 16:
701   case 32:
702     break;
703   default:
704     return false;
705   }
706   if (!StoreMI.isSimple())
707     return false;
708 
709   // We do a simple search for mergeable stores prior to this one.
710   // Any potential alias hazard along the way terminates the search.
711   SmallVector<GStore *> FoundStores;
712 
713   // We're looking for:
714   // 1) a (store(trunc(...)))
715   // 2) of an LSHR/ASHR of a single wide value, by the appropriate shift to get
716   //    the partial value stored.
717   // 3) where the offsets form either a little or big-endian sequence.
718 
719   auto &LastStore = StoreMI;
720 
721   // The single base pointer that all stores must use.
722   Register BaseReg;
723   int64_t LastOffset;
724   if (!mi_match(LastStore.getPointerReg(), *MRI,
725                 m_GPtrAdd(m_Reg(BaseReg), m_ICst(LastOffset)))) {
726     BaseReg = LastStore.getPointerReg();
727     LastOffset = 0;
728   }
729 
730   GStore *LowestIdxStore = &LastStore;
731   int64_t LowestIdxOffset = LastOffset;
732 
733   Register WideSrcVal;
734   auto LowestShiftAmt = getTruncStoreByteOffset(LastStore, WideSrcVal, *MRI);
735   if (!LowestShiftAmt)
736     return false; // Didn't match a trunc.
737   assert(WideSrcVal.isValid());
738 
739   LLT WideStoreTy = MRI->getType(WideSrcVal);
740   // The wide type might not be a multiple of the memory type, e.g. s48 and s32.
741   if (WideStoreTy.getSizeInBits() % MemTy.getSizeInBits() != 0)
742     return false;
743   const unsigned NumStoresRequired =
744       WideStoreTy.getSizeInBits() / MemTy.getSizeInBits();
745 
746   SmallVector<int64_t, 8> OffsetMap(NumStoresRequired, INT64_MAX);
747   OffsetMap[*LowestShiftAmt] = LastOffset;
748   FoundStores.emplace_back(&LastStore);
749 
750   const int MaxInstsToCheck = 10;
751   int NumInstsChecked = 0;
752   for (auto II = ++LastStore.getReverseIterator();
753        II != LastStore.getParent()->rend() && NumInstsChecked < MaxInstsToCheck;
754        ++II) {
755     NumInstsChecked++;
756     GStore *NewStore;
757     if ((NewStore = dyn_cast<GStore>(&*II))) {
758       if (NewStore->getMMO().getMemoryType() != MemTy || !NewStore->isSimple())
759         break;
760     } else if (II->isLoadFoldBarrier() || II->mayLoad()) {
761       break;
762     } else {
763       continue; // This is a safe instruction we can look past.
764     }
765 
766     Register NewBaseReg;
767     int64_t MemOffset;
768     // Check we're storing to the same base + some offset.
769     if (!mi_match(NewStore->getPointerReg(), *MRI,
770                   m_GPtrAdd(m_Reg(NewBaseReg), m_ICst(MemOffset)))) {
771       NewBaseReg = NewStore->getPointerReg();
772       MemOffset = 0;
773     }
774     if (BaseReg != NewBaseReg)
775       break;
776 
777     auto ShiftByteOffset = getTruncStoreByteOffset(*NewStore, WideSrcVal, *MRI);
778     if (!ShiftByteOffset)
779       break;
780     if (MemOffset < LowestIdxOffset) {
781       LowestIdxOffset = MemOffset;
782       LowestIdxStore = NewStore;
783     }
784 
785     // Map the offset in the store and the offset in the combined value, and
786     // early return if it has been set before.
787     if (*ShiftByteOffset < 0 || *ShiftByteOffset >= NumStoresRequired ||
788         OffsetMap[*ShiftByteOffset] != INT64_MAX)
789       break;
790     OffsetMap[*ShiftByteOffset] = MemOffset;
791 
792     FoundStores.emplace_back(NewStore);
793     // Reset counter since we've found a matching inst.
794     NumInstsChecked = 0;
795     if (FoundStores.size() == NumStoresRequired)
796       break;
797   }
798 
799   if (FoundStores.size() != NumStoresRequired) {
800     if (FoundStores.size() == 1)
801       return false;
802     // We didn't find enough stores to merge into the size of the original
803     // source value, but we may be able to generate a smaller store if we
804     // truncate the source value.
805     WideStoreTy = LLT::scalar(FoundStores.size() * MemTy.getScalarSizeInBits());
806   }
807 
808   unsigned NumStoresFound = FoundStores.size();
809 
810   const auto &DL = LastStore.getMF()->getDataLayout();
811   auto &C = LastStore.getMF()->getFunction().getContext();
812   // Check that a store of the wide type is both allowed and fast on the target
813   unsigned Fast = 0;
814   bool Allowed = TLI->allowsMemoryAccess(
815       C, DL, WideStoreTy, LowestIdxStore->getMMO(), &Fast);
816   if (!Allowed || !Fast)
817     return false;
818 
819   // Check if the pieces of the value are going to the expected places in memory
820   // to merge the stores.
821   unsigned NarrowBits = MemTy.getScalarSizeInBits();
822   auto checkOffsets = [&](bool MatchLittleEndian) {
823     if (MatchLittleEndian) {
824       for (unsigned i = 0; i != NumStoresFound; ++i)
825         if (OffsetMap[i] != i * (NarrowBits / 8) + LowestIdxOffset)
826           return false;
827     } else { // MatchBigEndian by reversing loop counter.
828       for (unsigned i = 0, j = NumStoresFound - 1; i != NumStoresFound;
829            ++i, --j)
830         if (OffsetMap[j] != i * (NarrowBits / 8) + LowestIdxOffset)
831           return false;
832     }
833     return true;
834   };
835 
836   // Check if the offsets line up for the native data layout of this target.
837   bool NeedBswap = false;
838   bool NeedRotate = false;
839   if (!checkOffsets(DL.isLittleEndian())) {
840     // Special-case: check if byte offsets line up for the opposite endian.
841     if (NarrowBits == 8 && checkOffsets(DL.isBigEndian()))
842       NeedBswap = true;
843     else if (NumStoresFound == 2 && checkOffsets(DL.isBigEndian()))
844       NeedRotate = true;
845     else
846       return false;
847   }
848 
849   if (NeedBswap &&
850       !isLegalOrBeforeLegalizer({TargetOpcode::G_BSWAP, {WideStoreTy}}, *MF))
851     return false;
852   if (NeedRotate &&
853       !isLegalOrBeforeLegalizer(
854           {TargetOpcode::G_ROTR, {WideStoreTy, WideStoreTy}}, *MF))
855     return false;
856 
857   Builder.setInstrAndDebugLoc(StoreMI);
858 
859   if (WideStoreTy != MRI->getType(WideSrcVal))
860     WideSrcVal = Builder.buildTrunc(WideStoreTy, WideSrcVal).getReg(0);
861 
862   if (NeedBswap) {
863     WideSrcVal = Builder.buildBSwap(WideStoreTy, WideSrcVal).getReg(0);
864   } else if (NeedRotate) {
865     assert(WideStoreTy.getSizeInBits() % 2 == 0 &&
866            "Unexpected type for rotate");
867     auto RotAmt =
868         Builder.buildConstant(WideStoreTy, WideStoreTy.getSizeInBits() / 2);
869     WideSrcVal =
870         Builder.buildRotateRight(WideStoreTy, WideSrcVal, RotAmt).getReg(0);
871   }
872 
873   Builder.buildStore(WideSrcVal, LowestIdxStore->getPointerReg(),
874                      LowestIdxStore->getMMO().getPointerInfo(),
875                      LowestIdxStore->getMMO().getAlign());
876 
877   // Erase the old stores.
878   for (auto *ST : FoundStores) {
879     ST->eraseFromParent();
880     DeletedStores.insert(ST);
881   }
882   return true;
883 }
884 
885 bool LoadStoreOpt::mergeTruncStoresBlock(MachineBasicBlock &BB) {
886   bool Changed = false;
887   SmallVector<GStore *, 16> Stores;
888   SmallPtrSet<GStore *, 8> DeletedStores;
889   // Walk up the block so we can see the most eligible stores.
890   for (MachineInstr &MI : llvm::reverse(BB))
891     if (auto *StoreMI = dyn_cast<GStore>(&MI))
892       Stores.emplace_back(StoreMI);
893 
894   for (auto *StoreMI : Stores) {
895     if (DeletedStores.count(StoreMI))
896       continue;
897     if (mergeTruncStore(*StoreMI, DeletedStores))
898       Changed = true;
899   }
900   return Changed;
901 }
902 
903 bool LoadStoreOpt::mergeFunctionStores(MachineFunction &MF) {
904   bool Changed = false;
905   for (auto &BB : MF){
906     Changed |= mergeBlockStores(BB);
907     Changed |= mergeTruncStoresBlock(BB);
908   }
909 
910   // Erase all dead instructions left over by the merging.
911   if (Changed) {
912     for (auto &BB : MF) {
913       for (auto &I : make_early_inc_range(make_range(BB.rbegin(), BB.rend()))) {
914         if (isTriviallyDead(I, *MRI))
915           I.eraseFromParent();
916       }
917     }
918   }
919 
920   return Changed;
921 }
922 
923 void LoadStoreOpt::initializeStoreMergeTargetInfo(unsigned AddrSpace) {
924   // Query the legalizer info to record what store types are legal.
925   // We record this because we don't want to bother trying to merge stores into
926   // illegal ones, which would just result in being split again.
927 
928   if (LegalStoreSizes.count(AddrSpace)) {
929     assert(LegalStoreSizes[AddrSpace].any());
930     return; // Already cached sizes for this address space.
931   }
932 
933   // Need to reserve at least MaxStoreSizeToForm + 1 bits.
934   BitVector LegalSizes(MaxStoreSizeToForm * 2);
935   const auto &LI = *MF->getSubtarget().getLegalizerInfo();
936   const auto &DL = MF->getFunction().getParent()->getDataLayout();
937   Type *IntPtrIRTy =
938       DL.getIntPtrType(MF->getFunction().getContext(), AddrSpace);
939   LLT PtrTy = getLLTForType(*IntPtrIRTy->getPointerTo(AddrSpace), DL);
940   // We assume that we're not going to be generating any stores wider than
941   // MaxStoreSizeToForm bits for now.
942   for (unsigned Size = 2; Size <= MaxStoreSizeToForm; Size *= 2) {
943     LLT Ty = LLT::scalar(Size);
944     SmallVector<LegalityQuery::MemDesc, 2> MemDescrs(
945         {{Ty, Ty.getSizeInBits(), AtomicOrdering::NotAtomic}});
946     SmallVector<LLT> StoreTys({Ty, PtrTy});
947     LegalityQuery Q(TargetOpcode::G_STORE, StoreTys, MemDescrs);
948     LegalizeActionStep ActionStep = LI.getAction(Q);
949     if (ActionStep.Action == LegalizeActions::Legal)
950       LegalSizes.set(Size);
951   }
952   assert(LegalSizes.any() && "Expected some store sizes to be legal!");
953   LegalStoreSizes[AddrSpace] = LegalSizes;
954 }
955 
956 bool LoadStoreOpt::runOnMachineFunction(MachineFunction &MF) {
957   // If the ISel pipeline failed, do not bother running that pass.
958   if (MF.getProperties().hasProperty(
959           MachineFunctionProperties::Property::FailedISel))
960     return false;
961 
962   LLVM_DEBUG(dbgs() << "Begin memory optimizations for: " << MF.getName()
963                     << '\n');
964 
965   init(MF);
966   bool Changed = false;
967   Changed |= mergeFunctionStores(MF);
968 
969   LegalStoreSizes.clear();
970   return Changed;
971 }
972