1 //===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
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 //
9 // This pass performs various transformations related to eliminating memcpy
10 // calls, or transforming sets of stores into memset's.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "llvm/Transforms/Scalar/MemCpyOptimizer.h"
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/STLExtras.h"
17 #include "llvm/ADT/SmallVector.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/iterator_range.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/CaptureTracking.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/Loads.h"
25 #include "llvm/Analysis/MemoryLocation.h"
26 #include "llvm/Analysis/MemorySSA.h"
27 #include "llvm/Analysis/MemorySSAUpdater.h"
28 #include "llvm/Analysis/TargetLibraryInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/BasicBlock.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DataLayout.h"
33 #include "llvm/IR/DerivedTypes.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/GlobalVariable.h"
37 #include "llvm/IR/IRBuilder.h"
38 #include "llvm/IR/InstrTypes.h"
39 #include "llvm/IR/Instruction.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/Intrinsics.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/IR/PassManager.h"
46 #include "llvm/IR/Type.h"
47 #include "llvm/IR/User.h"
48 #include "llvm/IR/Value.h"
49 #include "llvm/InitializePasses.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/Scalar.h"
56 #include "llvm/Transforms/Utils/Local.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <cstdint>
60 #include <optional>
61 
62 using namespace llvm;
63 
64 #define DEBUG_TYPE "memcpyopt"
65 
66 static cl::opt<bool> EnableMemCpyOptWithoutLibcalls(
67     "enable-memcpyopt-without-libcalls", cl::Hidden,
68     cl::desc("Enable memcpyopt even when libcalls are disabled"));
69 
70 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
71 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
72 STATISTIC(NumMoveToCpy,   "Number of memmoves converted to memcpy");
73 STATISTIC(NumCpyToSet,    "Number of memcpys converted to memset");
74 STATISTIC(NumCallSlot,    "Number of call slot optimizations performed");
75 
76 namespace {
77 
78 /// Represents a range of memset'd bytes with the ByteVal value.
79 /// This allows us to analyze stores like:
80 ///   store 0 -> P+1
81 ///   store 0 -> P+0
82 ///   store 0 -> P+3
83 ///   store 0 -> P+2
84 /// which sometimes happens with stores to arrays of structs etc.  When we see
85 /// the first store, we make a range [1, 2).  The second store extends the range
86 /// to [0, 2).  The third makes a new range [2, 3).  The fourth store joins the
87 /// two ranges into [0, 3) which is memset'able.
88 struct MemsetRange {
89   // Start/End - A semi range that describes the span that this range covers.
90   // The range is closed at the start and open at the end: [Start, End).
91   int64_t Start, End;
92 
93   /// StartPtr - The getelementptr instruction that points to the start of the
94   /// range.
95   Value *StartPtr;
96 
97   /// Alignment - The known alignment of the first store.
98   MaybeAlign Alignment;
99 
100   /// TheStores - The actual stores that make up this range.
101   SmallVector<Instruction*, 16> TheStores;
102 
103   bool isProfitableToUseMemset(const DataLayout &DL) const;
104 };
105 
106 } // end anonymous namespace
107 
108 bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
109   // If we found more than 4 stores to merge or 16 bytes, use memset.
110   if (TheStores.size() >= 4 || End-Start >= 16) return true;
111 
112   // If there is nothing to merge, don't do anything.
113   if (TheStores.size() < 2) return false;
114 
115   // If any of the stores are a memset, then it is always good to extend the
116   // memset.
117   for (Instruction *SI : TheStores)
118     if (!isa<StoreInst>(SI))
119       return true;
120 
121   // Assume that the code generator is capable of merging pairs of stores
122   // together if it wants to.
123   if (TheStores.size() == 2) return false;
124 
125   // If we have fewer than 8 stores, it can still be worthwhile to do this.
126   // For example, merging 4 i8 stores into an i32 store is useful almost always.
127   // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
128   // memset will be split into 2 32-bit stores anyway) and doing so can
129   // pessimize the llvm optimizer.
130   //
131   // Since we don't have perfect knowledge here, make some assumptions: assume
132   // the maximum GPR width is the same size as the largest legal integer
133   // size. If so, check to see whether we will end up actually reducing the
134   // number of stores used.
135   unsigned Bytes = unsigned(End-Start);
136   unsigned MaxIntSize = DL.getLargestLegalIntTypeSizeInBits() / 8;
137   if (MaxIntSize == 0)
138     MaxIntSize = 1;
139   unsigned NumPointerStores = Bytes / MaxIntSize;
140 
141   // Assume the remaining bytes if any are done a byte at a time.
142   unsigned NumByteStores = Bytes % MaxIntSize;
143 
144   // If we will reduce the # stores (according to this heuristic), do the
145   // transformation.  This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
146   // etc.
147   return TheStores.size() > NumPointerStores+NumByteStores;
148 }
149 
150 namespace {
151 
152 class MemsetRanges {
153   using range_iterator = SmallVectorImpl<MemsetRange>::iterator;
154 
155   /// A sorted list of the memset ranges.
156   SmallVector<MemsetRange, 8> Ranges;
157 
158   const DataLayout &DL;
159 
160 public:
161   MemsetRanges(const DataLayout &DL) : DL(DL) {}
162 
163   using const_iterator = SmallVectorImpl<MemsetRange>::const_iterator;
164 
165   const_iterator begin() const { return Ranges.begin(); }
166   const_iterator end() const { return Ranges.end(); }
167   bool empty() const { return Ranges.empty(); }
168 
169   void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
170     if (auto *SI = dyn_cast<StoreInst>(Inst))
171       addStore(OffsetFromFirst, SI);
172     else
173       addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
174   }
175 
176   void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
177     TypeSize StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
178     assert(!StoreSize.isScalable() && "Can't track scalable-typed stores");
179     addRange(OffsetFromFirst, StoreSize.getFixedValue(),
180              SI->getPointerOperand(), SI->getAlign(), SI);
181   }
182 
183   void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
184     int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
185     addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getDestAlign(), MSI);
186   }
187 
188   void addRange(int64_t Start, int64_t Size, Value *Ptr, MaybeAlign Alignment,
189                 Instruction *Inst);
190 };
191 
192 } // end anonymous namespace
193 
194 /// Add a new store to the MemsetRanges data structure.  This adds a
195 /// new range for the specified store at the specified offset, merging into
196 /// existing ranges as appropriate.
197 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
198                             MaybeAlign Alignment, Instruction *Inst) {
199   int64_t End = Start+Size;
200 
201   range_iterator I = partition_point(
202       Ranges, [=](const MemsetRange &O) { return O.End < Start; });
203 
204   // We now know that I == E, in which case we didn't find anything to merge
205   // with, or that Start <= I->End.  If End < I->Start or I == E, then we need
206   // to insert a new range.  Handle this now.
207   if (I == Ranges.end() || End < I->Start) {
208     MemsetRange &R = *Ranges.insert(I, MemsetRange());
209     R.Start        = Start;
210     R.End          = End;
211     R.StartPtr     = Ptr;
212     R.Alignment    = Alignment;
213     R.TheStores.push_back(Inst);
214     return;
215   }
216 
217   // This store overlaps with I, add it.
218   I->TheStores.push_back(Inst);
219 
220   // At this point, we may have an interval that completely contains our store.
221   // If so, just add it to the interval and return.
222   if (I->Start <= Start && I->End >= End)
223     return;
224 
225   // Now we know that Start <= I->End and End >= I->Start so the range overlaps
226   // but is not entirely contained within the range.
227 
228   // See if the range extends the start of the range.  In this case, it couldn't
229   // possibly cause it to join the prior range, because otherwise we would have
230   // stopped on *it*.
231   if (Start < I->Start) {
232     I->Start = Start;
233     I->StartPtr = Ptr;
234     I->Alignment = Alignment;
235   }
236 
237   // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
238   // is in or right at the end of I), and that End >= I->Start.  Extend I out to
239   // End.
240   if (End > I->End) {
241     I->End = End;
242     range_iterator NextI = I;
243     while (++NextI != Ranges.end() && End >= NextI->Start) {
244       // Merge the range in.
245       I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
246       if (NextI->End > I->End)
247         I->End = NextI->End;
248       Ranges.erase(NextI);
249       NextI = I;
250     }
251   }
252 }
253 
254 //===----------------------------------------------------------------------===//
255 //                         MemCpyOptLegacyPass Pass
256 //===----------------------------------------------------------------------===//
257 
258 namespace {
259 
260 class MemCpyOptLegacyPass : public FunctionPass {
261   MemCpyOptPass Impl;
262 
263 public:
264   static char ID; // Pass identification, replacement for typeid
265 
266   MemCpyOptLegacyPass() : FunctionPass(ID) {
267     initializeMemCpyOptLegacyPassPass(*PassRegistry::getPassRegistry());
268   }
269 
270   bool runOnFunction(Function &F) override;
271 
272 private:
273   // This transformation requires dominator postdominator info
274   void getAnalysisUsage(AnalysisUsage &AU) const override {
275     AU.setPreservesCFG();
276     AU.addRequired<AssumptionCacheTracker>();
277     AU.addRequired<DominatorTreeWrapperPass>();
278     AU.addPreserved<DominatorTreeWrapperPass>();
279     AU.addPreserved<GlobalsAAWrapperPass>();
280     AU.addRequired<TargetLibraryInfoWrapperPass>();
281     AU.addRequired<AAResultsWrapperPass>();
282     AU.addPreserved<AAResultsWrapperPass>();
283     AU.addRequired<MemorySSAWrapperPass>();
284     AU.addPreserved<MemorySSAWrapperPass>();
285   }
286 };
287 
288 } // end anonymous namespace
289 
290 char MemCpyOptLegacyPass::ID = 0;
291 
292 /// The public interface to this file...
293 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOptLegacyPass(); }
294 
295 INITIALIZE_PASS_BEGIN(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",
296                       false, false)
297 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
298 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
299 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
300 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
301 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
302 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)
303 INITIALIZE_PASS_END(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",
304                     false, false)
305 
306 // Check that V is either not accessible by the caller, or unwinding cannot
307 // occur between Start and End.
308 static bool mayBeVisibleThroughUnwinding(Value *V, Instruction *Start,
309                                          Instruction *End) {
310   assert(Start->getParent() == End->getParent() && "Must be in same block");
311   // Function can't unwind, so it also can't be visible through unwinding.
312   if (Start->getFunction()->doesNotThrow())
313     return false;
314 
315   // Object is not visible on unwind.
316   // TODO: Support RequiresNoCaptureBeforeUnwind case.
317   bool RequiresNoCaptureBeforeUnwind;
318   if (isNotVisibleOnUnwind(getUnderlyingObject(V),
319                            RequiresNoCaptureBeforeUnwind) &&
320       !RequiresNoCaptureBeforeUnwind)
321     return false;
322 
323   // Check whether there are any unwinding instructions in the range.
324   return any_of(make_range(Start->getIterator(), End->getIterator()),
325                 [](const Instruction &I) { return I.mayThrow(); });
326 }
327 
328 void MemCpyOptPass::eraseInstruction(Instruction *I) {
329   MSSAU->removeMemoryAccess(I);
330   I->eraseFromParent();
331 }
332 
333 // Check for mod or ref of Loc between Start and End, excluding both boundaries.
334 // Start and End must be in the same block.
335 // If SkippedLifetimeStart is provided, skip over one clobbering lifetime.start
336 // intrinsic and store it inside SkippedLifetimeStart.
337 static bool accessedBetween(BatchAAResults &AA, MemoryLocation Loc,
338                             const MemoryUseOrDef *Start,
339                             const MemoryUseOrDef *End,
340                             Instruction **SkippedLifetimeStart = nullptr) {
341   assert(Start->getBlock() == End->getBlock() && "Only local supported");
342   for (const MemoryAccess &MA :
343        make_range(++Start->getIterator(), End->getIterator())) {
344     Instruction *I = cast<MemoryUseOrDef>(MA).getMemoryInst();
345     if (isModOrRefSet(AA.getModRefInfo(I, Loc))) {
346       auto *II = dyn_cast<IntrinsicInst>(I);
347       if (II && II->getIntrinsicID() == Intrinsic::lifetime_start &&
348           SkippedLifetimeStart && !*SkippedLifetimeStart) {
349         *SkippedLifetimeStart = I;
350         continue;
351       }
352 
353       return true;
354     }
355   }
356   return false;
357 }
358 
359 // Check for mod of Loc between Start and End, excluding both boundaries.
360 // Start and End can be in different blocks.
361 static bool writtenBetween(MemorySSA *MSSA, BatchAAResults &AA,
362                            MemoryLocation Loc, const MemoryUseOrDef *Start,
363                            const MemoryUseOrDef *End) {
364   if (isa<MemoryUse>(End)) {
365     // For MemoryUses, getClobberingMemoryAccess may skip non-clobbering writes.
366     // Manually check read accesses between Start and End, if they are in the
367     // same block, for clobbers. Otherwise assume Loc is clobbered.
368     return Start->getBlock() != End->getBlock() ||
369            any_of(
370                make_range(std::next(Start->getIterator()), End->getIterator()),
371                [&AA, Loc](const MemoryAccess &Acc) {
372                  if (isa<MemoryUse>(&Acc))
373                    return false;
374                  Instruction *AccInst =
375                      cast<MemoryUseOrDef>(&Acc)->getMemoryInst();
376                  return isModSet(AA.getModRefInfo(AccInst, Loc));
377                });
378   }
379 
380   // TODO: Only walk until we hit Start.
381   MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
382       End->getDefiningAccess(), Loc, AA);
383   return !MSSA->dominates(Clobber, Start);
384 }
385 
386 /// When scanning forward over instructions, we look for some other patterns to
387 /// fold away. In particular, this looks for stores to neighboring locations of
388 /// memory. If it sees enough consecutive ones, it attempts to merge them
389 /// together into a memcpy/memset.
390 Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
391                                                  Value *StartPtr,
392                                                  Value *ByteVal) {
393   const DataLayout &DL = StartInst->getModule()->getDataLayout();
394 
395   // We can't track scalable types
396   if (auto *SI = dyn_cast<StoreInst>(StartInst))
397     if (DL.getTypeStoreSize(SI->getOperand(0)->getType()).isScalable())
398       return nullptr;
399 
400   // Okay, so we now have a single store that can be splatable.  Scan to find
401   // all subsequent stores of the same value to offset from the same pointer.
402   // Join these together into ranges, so we can decide whether contiguous blocks
403   // are stored.
404   MemsetRanges Ranges(DL);
405 
406   BasicBlock::iterator BI(StartInst);
407 
408   // Keeps track of the last memory use or def before the insertion point for
409   // the new memset. The new MemoryDef for the inserted memsets will be inserted
410   // after MemInsertPoint. It points to either LastMemDef or to the last user
411   // before the insertion point of the memset, if there are any such users.
412   MemoryUseOrDef *MemInsertPoint = nullptr;
413   // Keeps track of the last MemoryDef between StartInst and the insertion point
414   // for the new memset. This will become the defining access of the inserted
415   // memsets.
416   MemoryDef *LastMemDef = nullptr;
417   for (++BI; !BI->isTerminator(); ++BI) {
418     auto *CurrentAcc = cast_or_null<MemoryUseOrDef>(
419         MSSAU->getMemorySSA()->getMemoryAccess(&*BI));
420     if (CurrentAcc) {
421       MemInsertPoint = CurrentAcc;
422       if (auto *CurrentDef = dyn_cast<MemoryDef>(CurrentAcc))
423         LastMemDef = CurrentDef;
424     }
425 
426     // Calls that only access inaccessible memory do not block merging
427     // accessible stores.
428     if (auto *CB = dyn_cast<CallBase>(BI)) {
429       if (CB->onlyAccessesInaccessibleMemory())
430         continue;
431     }
432 
433     if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
434       // If the instruction is readnone, ignore it, otherwise bail out.  We
435       // don't even allow readonly here because we don't want something like:
436       // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
437       if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
438         break;
439       continue;
440     }
441 
442     if (auto *NextStore = dyn_cast<StoreInst>(BI)) {
443       // If this is a store, see if we can merge it in.
444       if (!NextStore->isSimple()) break;
445 
446       Value *StoredVal = NextStore->getValueOperand();
447 
448       // Don't convert stores of non-integral pointer types to memsets (which
449       // stores integers).
450       if (DL.isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
451         break;
452 
453       // We can't track ranges involving scalable types.
454       if (DL.getTypeStoreSize(StoredVal->getType()).isScalable())
455         break;
456 
457       // Check to see if this stored value is of the same byte-splattable value.
458       Value *StoredByte = isBytewiseValue(StoredVal, DL);
459       if (isa<UndefValue>(ByteVal) && StoredByte)
460         ByteVal = StoredByte;
461       if (ByteVal != StoredByte)
462         break;
463 
464       // Check to see if this store is to a constant offset from the start ptr.
465       std::optional<int64_t> Offset =
466           isPointerOffset(StartPtr, NextStore->getPointerOperand(), DL);
467       if (!Offset)
468         break;
469 
470       Ranges.addStore(*Offset, NextStore);
471     } else {
472       auto *MSI = cast<MemSetInst>(BI);
473 
474       if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
475           !isa<ConstantInt>(MSI->getLength()))
476         break;
477 
478       // Check to see if this store is to a constant offset from the start ptr.
479       std::optional<int64_t> Offset =
480           isPointerOffset(StartPtr, MSI->getDest(), DL);
481       if (!Offset)
482         break;
483 
484       Ranges.addMemSet(*Offset, MSI);
485     }
486   }
487 
488   // If we have no ranges, then we just had a single store with nothing that
489   // could be merged in.  This is a very common case of course.
490   if (Ranges.empty())
491     return nullptr;
492 
493   // If we had at least one store that could be merged in, add the starting
494   // store as well.  We try to avoid this unless there is at least something
495   // interesting as a small compile-time optimization.
496   Ranges.addInst(0, StartInst);
497 
498   // If we create any memsets, we put it right before the first instruction that
499   // isn't part of the memset block.  This ensure that the memset is dominated
500   // by any addressing instruction needed by the start of the block.
501   IRBuilder<> Builder(&*BI);
502 
503   // Now that we have full information about ranges, loop over the ranges and
504   // emit memset's for anything big enough to be worthwhile.
505   Instruction *AMemSet = nullptr;
506   for (const MemsetRange &Range : Ranges) {
507     if (Range.TheStores.size() == 1) continue;
508 
509     // If it is profitable to lower this range to memset, do so now.
510     if (!Range.isProfitableToUseMemset(DL))
511       continue;
512 
513     // Otherwise, we do want to transform this!  Create a new memset.
514     // Get the starting pointer of the block.
515     StartPtr = Range.StartPtr;
516 
517     AMemSet = Builder.CreateMemSet(StartPtr, ByteVal, Range.End - Range.Start,
518                                    Range.Alignment);
519     AMemSet->mergeDIAssignID(Range.TheStores);
520 
521     LLVM_DEBUG(dbgs() << "Replace stores:\n"; for (Instruction *SI
522                                                    : Range.TheStores) dbgs()
523                                               << *SI << '\n';
524                dbgs() << "With: " << *AMemSet << '\n');
525     if (!Range.TheStores.empty())
526       AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
527 
528     assert(LastMemDef && MemInsertPoint &&
529            "Both LastMemDef and MemInsertPoint need to be set");
530     auto *NewDef =
531         cast<MemoryDef>(MemInsertPoint->getMemoryInst() == &*BI
532                             ? MSSAU->createMemoryAccessBefore(
533                                   AMemSet, LastMemDef, MemInsertPoint)
534                             : MSSAU->createMemoryAccessAfter(
535                                   AMemSet, LastMemDef, MemInsertPoint));
536     MSSAU->insertDef(NewDef, /*RenameUses=*/true);
537     LastMemDef = NewDef;
538     MemInsertPoint = NewDef;
539 
540     // Zap all the stores.
541     for (Instruction *SI : Range.TheStores)
542       eraseInstruction(SI);
543 
544     ++NumMemSetInfer;
545   }
546 
547   return AMemSet;
548 }
549 
550 // This method try to lift a store instruction before position P.
551 // It will lift the store and its argument + that anything that
552 // may alias with these.
553 // The method returns true if it was successful.
554 bool MemCpyOptPass::moveUp(StoreInst *SI, Instruction *P, const LoadInst *LI) {
555   // If the store alias this position, early bail out.
556   MemoryLocation StoreLoc = MemoryLocation::get(SI);
557   if (isModOrRefSet(AA->getModRefInfo(P, StoreLoc)))
558     return false;
559 
560   // Keep track of the arguments of all instruction we plan to lift
561   // so we can make sure to lift them as well if appropriate.
562   DenseSet<Instruction*> Args;
563   auto AddArg = [&](Value *Arg) {
564     auto *I = dyn_cast<Instruction>(Arg);
565     if (I && I->getParent() == SI->getParent()) {
566       // Cannot hoist user of P above P
567       if (I == P) return false;
568       Args.insert(I);
569     }
570     return true;
571   };
572   if (!AddArg(SI->getPointerOperand()))
573     return false;
574 
575   // Instruction to lift before P.
576   SmallVector<Instruction *, 8> ToLift{SI};
577 
578   // Memory locations of lifted instructions.
579   SmallVector<MemoryLocation, 8> MemLocs{StoreLoc};
580 
581   // Lifted calls.
582   SmallVector<const CallBase *, 8> Calls;
583 
584   const MemoryLocation LoadLoc = MemoryLocation::get(LI);
585 
586   for (auto I = --SI->getIterator(), E = P->getIterator(); I != E; --I) {
587     auto *C = &*I;
588 
589     // Make sure hoisting does not perform a store that was not guaranteed to
590     // happen.
591     if (!isGuaranteedToTransferExecutionToSuccessor(C))
592       return false;
593 
594     bool MayAlias = isModOrRefSet(AA->getModRefInfo(C, std::nullopt));
595 
596     bool NeedLift = false;
597     if (Args.erase(C))
598       NeedLift = true;
599     else if (MayAlias) {
600       NeedLift = llvm::any_of(MemLocs, [C, this](const MemoryLocation &ML) {
601         return isModOrRefSet(AA->getModRefInfo(C, ML));
602       });
603 
604       if (!NeedLift)
605         NeedLift = llvm::any_of(Calls, [C, this](const CallBase *Call) {
606           return isModOrRefSet(AA->getModRefInfo(C, Call));
607         });
608     }
609 
610     if (!NeedLift)
611       continue;
612 
613     if (MayAlias) {
614       // Since LI is implicitly moved downwards past the lifted instructions,
615       // none of them may modify its source.
616       if (isModSet(AA->getModRefInfo(C, LoadLoc)))
617         return false;
618       else if (const auto *Call = dyn_cast<CallBase>(C)) {
619         // If we can't lift this before P, it's game over.
620         if (isModOrRefSet(AA->getModRefInfo(P, Call)))
621           return false;
622 
623         Calls.push_back(Call);
624       } else if (isa<LoadInst>(C) || isa<StoreInst>(C) || isa<VAArgInst>(C)) {
625         // If we can't lift this before P, it's game over.
626         auto ML = MemoryLocation::get(C);
627         if (isModOrRefSet(AA->getModRefInfo(P, ML)))
628           return false;
629 
630         MemLocs.push_back(ML);
631       } else
632         // We don't know how to lift this instruction.
633         return false;
634     }
635 
636     ToLift.push_back(C);
637     for (Value *Op : C->operands())
638       if (!AddArg(Op))
639         return false;
640   }
641 
642   // Find MSSA insertion point. Normally P will always have a corresponding
643   // memory access before which we can insert. However, with non-standard AA
644   // pipelines, there may be a mismatch between AA and MSSA, in which case we
645   // will scan for a memory access before P. In either case, we know for sure
646   // that at least the load will have a memory access.
647   // TODO: Simplify this once P will be determined by MSSA, in which case the
648   // discrepancy can no longer occur.
649   MemoryUseOrDef *MemInsertPoint = nullptr;
650   if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(P)) {
651     MemInsertPoint = cast<MemoryUseOrDef>(--MA->getIterator());
652   } else {
653     const Instruction *ConstP = P;
654     for (const Instruction &I : make_range(++ConstP->getReverseIterator(),
655                                            ++LI->getReverseIterator())) {
656       if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
657         MemInsertPoint = MA;
658         break;
659       }
660     }
661   }
662 
663   // We made it, we need to lift.
664   for (auto *I : llvm::reverse(ToLift)) {
665     LLVM_DEBUG(dbgs() << "Lifting " << *I << " before " << *P << "\n");
666     I->moveBefore(P);
667     assert(MemInsertPoint && "Must have found insert point");
668     if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(I)) {
669       MSSAU->moveAfter(MA, MemInsertPoint);
670       MemInsertPoint = MA;
671     }
672   }
673 
674   return true;
675 }
676 
677 bool MemCpyOptPass::processStoreOfLoad(StoreInst *SI, LoadInst *LI,
678                                        const DataLayout &DL,
679                                        BasicBlock::iterator &BBI) {
680   if (!LI->isSimple() || !LI->hasOneUse() ||
681       LI->getParent() != SI->getParent())
682     return false;
683 
684   auto *T = LI->getType();
685   // Don't introduce calls to memcpy/memmove intrinsics out of thin air if
686   // the corresponding libcalls are not available.
687   // TODO: We should really distinguish between libcall availability and
688   // our ability to introduce intrinsics.
689   if (T->isAggregateType() &&
690       (EnableMemCpyOptWithoutLibcalls ||
691        (TLI->has(LibFunc_memcpy) && TLI->has(LibFunc_memmove)))) {
692     MemoryLocation LoadLoc = MemoryLocation::get(LI);
693 
694     // We use alias analysis to check if an instruction may store to
695     // the memory we load from in between the load and the store. If
696     // such an instruction is found, we try to promote there instead
697     // of at the store position.
698     // TODO: Can use MSSA for this.
699     Instruction *P = SI;
700     for (auto &I : make_range(++LI->getIterator(), SI->getIterator())) {
701       if (isModSet(AA->getModRefInfo(&I, LoadLoc))) {
702         P = &I;
703         break;
704       }
705     }
706 
707     // We found an instruction that may write to the loaded memory.
708     // We can try to promote at this position instead of the store
709     // position if nothing aliases the store memory after this and the store
710     // destination is not in the range.
711     if (P && P != SI) {
712       if (!moveUp(SI, P, LI))
713         P = nullptr;
714     }
715 
716     // If a valid insertion position is found, then we can promote
717     // the load/store pair to a memcpy.
718     if (P) {
719       // If we load from memory that may alias the memory we store to,
720       // memmove must be used to preserve semantic. If not, memcpy can
721       // be used. Also, if we load from constant memory, memcpy can be used
722       // as the constant memory won't be modified.
723       bool UseMemMove = false;
724       if (isModSet(AA->getModRefInfo(SI, LoadLoc)))
725         UseMemMove = true;
726 
727       uint64_t Size = DL.getTypeStoreSize(T);
728 
729       IRBuilder<> Builder(P);
730       Instruction *M;
731       if (UseMemMove)
732         M = Builder.CreateMemMove(
733             SI->getPointerOperand(), SI->getAlign(),
734             LI->getPointerOperand(), LI->getAlign(), Size);
735       else
736         M = Builder.CreateMemCpy(
737             SI->getPointerOperand(), SI->getAlign(),
738             LI->getPointerOperand(), LI->getAlign(), Size);
739       M->copyMetadata(*SI, LLVMContext::MD_DIAssignID);
740 
741       LLVM_DEBUG(dbgs() << "Promoting " << *LI << " to " << *SI << " => "
742                         << *M << "\n");
743 
744       auto *LastDef =
745           cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(SI));
746       auto *NewAccess = MSSAU->createMemoryAccessAfter(M, LastDef, LastDef);
747       MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
748 
749       eraseInstruction(SI);
750       eraseInstruction(LI);
751       ++NumMemCpyInstr;
752 
753       // Make sure we do not invalidate the iterator.
754       BBI = M->getIterator();
755       return true;
756     }
757   }
758 
759   // Detect cases where we're performing call slot forwarding, but
760   // happen to be using a load-store pair to implement it, rather than
761   // a memcpy.
762   BatchAAResults BAA(*AA);
763   auto GetCall = [&]() -> CallInst * {
764     // We defer this expensive clobber walk until the cheap checks
765     // have been done on the source inside performCallSlotOptzn.
766     if (auto *LoadClobber = dyn_cast<MemoryUseOrDef>(
767             MSSA->getWalker()->getClobberingMemoryAccess(LI, BAA)))
768       return dyn_cast_or_null<CallInst>(LoadClobber->getMemoryInst());
769     return nullptr;
770   };
771 
772   bool Changed = performCallSlotOptzn(
773       LI, SI, SI->getPointerOperand()->stripPointerCasts(),
774       LI->getPointerOperand()->stripPointerCasts(),
775       DL.getTypeStoreSize(SI->getOperand(0)->getType()),
776       std::min(SI->getAlign(), LI->getAlign()), BAA, GetCall);
777   if (Changed) {
778     eraseInstruction(SI);
779     eraseInstruction(LI);
780     ++NumMemCpyInstr;
781     return true;
782   }
783 
784   return false;
785 }
786 
787 bool MemCpyOptPass::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
788   if (!SI->isSimple()) return false;
789 
790   // Avoid merging nontemporal stores since the resulting
791   // memcpy/memset would not be able to preserve the nontemporal hint.
792   // In theory we could teach how to propagate the !nontemporal metadata to
793   // memset calls. However, that change would force the backend to
794   // conservatively expand !nontemporal memset calls back to sequences of
795   // store instructions (effectively undoing the merging).
796   if (SI->getMetadata(LLVMContext::MD_nontemporal))
797     return false;
798 
799   const DataLayout &DL = SI->getModule()->getDataLayout();
800 
801   Value *StoredVal = SI->getValueOperand();
802 
803   // Not all the transforms below are correct for non-integral pointers, bail
804   // until we've audited the individual pieces.
805   if (DL.isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
806     return false;
807 
808   // Load to store forwarding can be interpreted as memcpy.
809   if (auto *LI = dyn_cast<LoadInst>(StoredVal))
810     return processStoreOfLoad(SI, LI, DL, BBI);
811 
812   // The following code creates memset intrinsics out of thin air. Don't do
813   // this if the corresponding libfunc is not available.
814   // TODO: We should really distinguish between libcall availability and
815   // our ability to introduce intrinsics.
816   if (!(TLI->has(LibFunc_memset) || EnableMemCpyOptWithoutLibcalls))
817     return false;
818 
819   // There are two cases that are interesting for this code to handle: memcpy
820   // and memset.  Right now we only handle memset.
821 
822   // Ensure that the value being stored is something that can be memset'able a
823   // byte at a time like "0" or "-1" or any width, as well as things like
824   // 0xA0A0A0A0 and 0.0.
825   auto *V = SI->getOperand(0);
826   if (Value *ByteVal = isBytewiseValue(V, DL)) {
827     if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
828                                               ByteVal)) {
829       BBI = I->getIterator(); // Don't invalidate iterator.
830       return true;
831     }
832 
833     // If we have an aggregate, we try to promote it to memset regardless
834     // of opportunity for merging as it can expose optimization opportunities
835     // in subsequent passes.
836     auto *T = V->getType();
837     if (T->isAggregateType()) {
838       uint64_t Size = DL.getTypeStoreSize(T);
839       IRBuilder<> Builder(SI);
840       auto *M = Builder.CreateMemSet(SI->getPointerOperand(), ByteVal, Size,
841                                      SI->getAlign());
842       M->copyMetadata(*SI, LLVMContext::MD_DIAssignID);
843 
844       LLVM_DEBUG(dbgs() << "Promoting " << *SI << " to " << *M << "\n");
845 
846       // The newly inserted memset is immediately overwritten by the original
847       // store, so we do not need to rename uses.
848       auto *StoreDef = cast<MemoryDef>(MSSA->getMemoryAccess(SI));
849       auto *NewAccess = MSSAU->createMemoryAccessBefore(
850           M, StoreDef->getDefiningAccess(), StoreDef);
851       MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/false);
852 
853       eraseInstruction(SI);
854       NumMemSetInfer++;
855 
856       // Make sure we do not invalidate the iterator.
857       BBI = M->getIterator();
858       return true;
859     }
860   }
861 
862   return false;
863 }
864 
865 bool MemCpyOptPass::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
866   // See if there is another memset or store neighboring this memset which
867   // allows us to widen out the memset to do a single larger store.
868   if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
869     if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
870                                               MSI->getValue())) {
871       BBI = I->getIterator(); // Don't invalidate iterator.
872       return true;
873     }
874   return false;
875 }
876 
877 /// Takes a memcpy and a call that it depends on,
878 /// and checks for the possibility of a call slot optimization by having
879 /// the call write its result directly into the destination of the memcpy.
880 bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
881                                          Instruction *cpyStore, Value *cpyDest,
882                                          Value *cpySrc, TypeSize cpySize,
883                                          Align cpyDestAlign, BatchAAResults &BAA,
884                                          std::function<CallInst *()> GetC) {
885   // The general transformation to keep in mind is
886   //
887   //   call @func(..., src, ...)
888   //   memcpy(dest, src, ...)
889   //
890   // ->
891   //
892   //   memcpy(dest, src, ...)
893   //   call @func(..., dest, ...)
894   //
895   // Since moving the memcpy is technically awkward, we additionally check that
896   // src only holds uninitialized values at the moment of the call, meaning that
897   // the memcpy can be discarded rather than moved.
898 
899   // We can't optimize scalable types.
900   if (cpySize.isScalable())
901     return false;
902 
903   // Require that src be an alloca.  This simplifies the reasoning considerably.
904   auto *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
905   if (!srcAlloca)
906     return false;
907 
908   ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
909   if (!srcArraySize)
910     return false;
911 
912   const DataLayout &DL = cpyLoad->getModule()->getDataLayout();
913   uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
914                      srcArraySize->getZExtValue();
915 
916   if (cpySize < srcSize)
917     return false;
918 
919   CallInst *C = GetC();
920   if (!C)
921     return false;
922 
923   // Lifetime marks shouldn't be operated on.
924   if (Function *F = C->getCalledFunction())
925     if (F->isIntrinsic() && F->getIntrinsicID() == Intrinsic::lifetime_start)
926       return false;
927 
928 
929   if (C->getParent() != cpyStore->getParent()) {
930     LLVM_DEBUG(dbgs() << "Call Slot: block local restriction\n");
931     return false;
932   }
933 
934   MemoryLocation DestLoc = isa<StoreInst>(cpyStore) ?
935     MemoryLocation::get(cpyStore) :
936     MemoryLocation::getForDest(cast<MemCpyInst>(cpyStore));
937 
938   // Check that nothing touches the dest of the copy between
939   // the call and the store/memcpy.
940   Instruction *SkippedLifetimeStart = nullptr;
941   if (accessedBetween(BAA, DestLoc, MSSA->getMemoryAccess(C),
942                       MSSA->getMemoryAccess(cpyStore), &SkippedLifetimeStart)) {
943     LLVM_DEBUG(dbgs() << "Call Slot: Dest pointer modified after call\n");
944     return false;
945   }
946 
947   // If we need to move a lifetime.start above the call, make sure that we can
948   // actually do so. If the argument is bitcasted for example, we would have to
949   // move the bitcast as well, which we don't handle.
950   if (SkippedLifetimeStart) {
951     auto *LifetimeArg =
952         dyn_cast<Instruction>(SkippedLifetimeStart->getOperand(1));
953     if (LifetimeArg && LifetimeArg->getParent() == C->getParent() &&
954         C->comesBefore(LifetimeArg))
955       return false;
956   }
957 
958   // Check that accessing the first srcSize bytes of dest will not cause a
959   // trap.  Otherwise the transform is invalid since it might cause a trap
960   // to occur earlier than it otherwise would.
961   if (!isDereferenceableAndAlignedPointer(cpyDest, Align(1), APInt(64, cpySize),
962                                           DL, C, AC, DT)) {
963     LLVM_DEBUG(dbgs() << "Call Slot: Dest pointer not dereferenceable\n");
964     return false;
965   }
966 
967   // Make sure that nothing can observe cpyDest being written early. There are
968   // a number of cases to consider:
969   //  1. cpyDest cannot be accessed between C and cpyStore as a precondition of
970   //     the transform.
971   //  2. C itself may not access cpyDest (prior to the transform). This is
972   //     checked further below.
973   //  3. If cpyDest is accessible to the caller of this function (potentially
974   //     captured and not based on an alloca), we need to ensure that we cannot
975   //     unwind between C and cpyStore. This is checked here.
976   //  4. If cpyDest is potentially captured, there may be accesses to it from
977   //     another thread. In this case, we need to check that cpyStore is
978   //     guaranteed to be executed if C is. As it is a non-atomic access, it
979   //     renders accesses from other threads undefined.
980   //     TODO: This is currently not checked.
981   if (mayBeVisibleThroughUnwinding(cpyDest, C, cpyStore)) {
982     LLVM_DEBUG(dbgs() << "Call Slot: Dest may be visible through unwinding\n");
983     return false;
984   }
985 
986   // Check that dest points to memory that is at least as aligned as src.
987   Align srcAlign = srcAlloca->getAlign();
988   bool isDestSufficientlyAligned = srcAlign <= cpyDestAlign;
989   // If dest is not aligned enough and we can't increase its alignment then
990   // bail out.
991   if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest)) {
992     LLVM_DEBUG(dbgs() << "Call Slot: Dest not sufficiently aligned\n");
993     return false;
994   }
995 
996   // Check that src is not accessed except via the call and the memcpy.  This
997   // guarantees that it holds only undefined values when passed in (so the final
998   // memcpy can be dropped), that it is not read or written between the call and
999   // the memcpy, and that writing beyond the end of it is undefined.
1000   SmallVector<User *, 8> srcUseList(srcAlloca->users());
1001   while (!srcUseList.empty()) {
1002     User *U = srcUseList.pop_back_val();
1003 
1004     if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
1005       append_range(srcUseList, U->users());
1006       continue;
1007     }
1008     if (const auto *G = dyn_cast<GetElementPtrInst>(U)) {
1009       if (!G->hasAllZeroIndices())
1010         return false;
1011 
1012       append_range(srcUseList, U->users());
1013       continue;
1014     }
1015     if (const auto *IT = dyn_cast<IntrinsicInst>(U))
1016       if (IT->isLifetimeStartOrEnd())
1017         continue;
1018 
1019     if (U != C && U != cpyLoad)
1020       return false;
1021   }
1022 
1023   // Check whether src is captured by the called function, in which case there
1024   // may be further indirect uses of src.
1025   bool SrcIsCaptured = any_of(C->args(), [&](Use &U) {
1026     return U->stripPointerCasts() == cpySrc &&
1027            !C->doesNotCapture(C->getArgOperandNo(&U));
1028   });
1029 
1030   // If src is captured, then check whether there are any potential uses of
1031   // src through the captured pointer before the lifetime of src ends, either
1032   // due to a lifetime.end or a return from the function.
1033   if (SrcIsCaptured) {
1034     // Check that dest is not captured before/at the call. We have already
1035     // checked that src is not captured before it. If either had been captured,
1036     // then the call might be comparing the argument against the captured dest
1037     // or src pointer.
1038     Value *DestObj = getUnderlyingObject(cpyDest);
1039     if (!isIdentifiedFunctionLocal(DestObj) ||
1040         PointerMayBeCapturedBefore(DestObj, /* ReturnCaptures */ true,
1041                                    /* StoreCaptures */ true, C, DT,
1042                                    /* IncludeI */ true))
1043       return false;
1044 
1045     MemoryLocation SrcLoc =
1046         MemoryLocation(srcAlloca, LocationSize::precise(srcSize));
1047     for (Instruction &I :
1048          make_range(++C->getIterator(), C->getParent()->end())) {
1049       // Lifetime of srcAlloca ends at lifetime.end.
1050       if (auto *II = dyn_cast<IntrinsicInst>(&I)) {
1051         if (II->getIntrinsicID() == Intrinsic::lifetime_end &&
1052             II->getArgOperand(1)->stripPointerCasts() == srcAlloca &&
1053             cast<ConstantInt>(II->getArgOperand(0))->uge(srcSize))
1054           break;
1055       }
1056 
1057       // Lifetime of srcAlloca ends at return.
1058       if (isa<ReturnInst>(&I))
1059         break;
1060 
1061       // Ignore the direct read of src in the load.
1062       if (&I == cpyLoad)
1063         continue;
1064 
1065       // Check whether this instruction may mod/ref src through the captured
1066       // pointer (we have already any direct mod/refs in the loop above).
1067       // Also bail if we hit a terminator, as we don't want to scan into other
1068       // blocks.
1069       if (isModOrRefSet(BAA.getModRefInfo(&I, SrcLoc)) || I.isTerminator())
1070         return false;
1071     }
1072   }
1073 
1074   // Since we're changing the parameter to the callsite, we need to make sure
1075   // that what would be the new parameter dominates the callsite.
1076   if (!DT->dominates(cpyDest, C)) {
1077     // Support moving a constant index GEP before the call.
1078     auto *GEP = dyn_cast<GetElementPtrInst>(cpyDest);
1079     if (GEP && GEP->hasAllConstantIndices() &&
1080         DT->dominates(GEP->getPointerOperand(), C))
1081       GEP->moveBefore(C);
1082     else
1083       return false;
1084   }
1085 
1086   // In addition to knowing that the call does not access src in some
1087   // unexpected manner, for example via a global, which we deduce from
1088   // the use analysis, we also need to know that it does not sneakily
1089   // access dest.  We rely on AA to figure this out for us.
1090   MemoryLocation DestWithSrcSize(cpyDest, LocationSize::precise(srcSize));
1091   ModRefInfo MR = BAA.getModRefInfo(C, DestWithSrcSize);
1092   // If necessary, perform additional analysis.
1093   if (isModOrRefSet(MR))
1094     MR = BAA.callCapturesBefore(C, DestWithSrcSize, DT);
1095   if (isModOrRefSet(MR))
1096     return false;
1097 
1098   // We can't create address space casts here because we don't know if they're
1099   // safe for the target.
1100   if (cpySrc->getType()->getPointerAddressSpace() !=
1101       cpyDest->getType()->getPointerAddressSpace())
1102     return false;
1103   for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
1104     if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc &&
1105         cpySrc->getType()->getPointerAddressSpace() !=
1106             C->getArgOperand(ArgI)->getType()->getPointerAddressSpace())
1107       return false;
1108 
1109   // All the checks have passed, so do the transformation.
1110   bool changedArgument = false;
1111   for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
1112     if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc) {
1113       Value *Dest = cpySrc->getType() == cpyDest->getType() ?  cpyDest
1114         : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
1115                                       cpyDest->getName(), C);
1116       changedArgument = true;
1117       if (C->getArgOperand(ArgI)->getType() == Dest->getType())
1118         C->setArgOperand(ArgI, Dest);
1119       else
1120         C->setArgOperand(ArgI, CastInst::CreatePointerCast(
1121                                    Dest, C->getArgOperand(ArgI)->getType(),
1122                                    Dest->getName(), C));
1123     }
1124 
1125   if (!changedArgument)
1126     return false;
1127 
1128   // If the destination wasn't sufficiently aligned then increase its alignment.
1129   if (!isDestSufficientlyAligned) {
1130     assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
1131     cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
1132   }
1133 
1134   if (SkippedLifetimeStart) {
1135     SkippedLifetimeStart->moveBefore(C);
1136     MSSAU->moveBefore(MSSA->getMemoryAccess(SkippedLifetimeStart),
1137                       MSSA->getMemoryAccess(C));
1138   }
1139 
1140   // Update AA metadata
1141   // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
1142   // handled here, but combineMetadata doesn't support them yet
1143   unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1144                          LLVMContext::MD_noalias,
1145                          LLVMContext::MD_invariant_group,
1146                          LLVMContext::MD_access_group};
1147   combineMetadata(C, cpyLoad, KnownIDs, true);
1148   if (cpyLoad != cpyStore)
1149     combineMetadata(C, cpyStore, KnownIDs, true);
1150 
1151   ++NumCallSlot;
1152   return true;
1153 }
1154 
1155 /// We've found that the (upward scanning) memory dependence of memcpy 'M' is
1156 /// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
1157 bool MemCpyOptPass::processMemCpyMemCpyDependence(MemCpyInst *M,
1158                                                   MemCpyInst *MDep,
1159                                                   BatchAAResults &BAA) {
1160   // We can only transforms memcpy's where the dest of one is the source of the
1161   // other.
1162   if (M->getSource() != MDep->getDest() || MDep->isVolatile())
1163     return false;
1164 
1165   // If dep instruction is reading from our current input, then it is a noop
1166   // transfer and substituting the input won't change this instruction.  Just
1167   // ignore the input and let someone else zap MDep.  This handles cases like:
1168   //    memcpy(a <- a)
1169   //    memcpy(b <- a)
1170   if (M->getSource() == MDep->getSource())
1171     return false;
1172 
1173   // Second, the length of the memcpy's must be the same, or the preceding one
1174   // must be larger than the following one.
1175   if (MDep->getLength() != M->getLength()) {
1176     auto *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
1177     auto *MLen = dyn_cast<ConstantInt>(M->getLength());
1178     if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
1179       return false;
1180   }
1181 
1182   // Verify that the copied-from memory doesn't change in between the two
1183   // transfers.  For example, in:
1184   //    memcpy(a <- b)
1185   //    *b = 42;
1186   //    memcpy(c <- a)
1187   // It would be invalid to transform the second memcpy into memcpy(c <- b).
1188   //
1189   // TODO: If the code between M and MDep is transparent to the destination "c",
1190   // then we could still perform the xform by moving M up to the first memcpy.
1191   // TODO: It would be sufficient to check the MDep source up to the memcpy
1192   // size of M, rather than MDep.
1193   if (writtenBetween(MSSA, BAA, MemoryLocation::getForSource(MDep),
1194                      MSSA->getMemoryAccess(MDep), MSSA->getMemoryAccess(M)))
1195     return false;
1196 
1197   // If the dest of the second might alias the source of the first, then the
1198   // source and dest might overlap. In addition, if the source of the first
1199   // points to constant memory, they won't overlap by definition. Otherwise, we
1200   // still want to eliminate the intermediate value, but we have to generate a
1201   // memmove instead of memcpy.
1202   bool UseMemMove = false;
1203   if (isModSet(BAA.getModRefInfo(M, MemoryLocation::getForSource(MDep))))
1204     UseMemMove = true;
1205 
1206   // If all checks passed, then we can transform M.
1207   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
1208                     << *MDep << '\n' << *M << '\n');
1209 
1210   // TODO: Is this worth it if we're creating a less aligned memcpy? For
1211   // example we could be moving from movaps -> movq on x86.
1212   IRBuilder<> Builder(M);
1213   Instruction *NewM;
1214   if (UseMemMove)
1215     NewM = Builder.CreateMemMove(M->getRawDest(), M->getDestAlign(),
1216                                  MDep->getRawSource(), MDep->getSourceAlign(),
1217                                  M->getLength(), M->isVolatile());
1218   else if (isa<MemCpyInlineInst>(M)) {
1219     // llvm.memcpy may be promoted to llvm.memcpy.inline, but the converse is
1220     // never allowed since that would allow the latter to be lowered as a call
1221     // to an external function.
1222     NewM = Builder.CreateMemCpyInline(
1223         M->getRawDest(), M->getDestAlign(), MDep->getRawSource(),
1224         MDep->getSourceAlign(), M->getLength(), M->isVolatile());
1225   } else
1226     NewM = Builder.CreateMemCpy(M->getRawDest(), M->getDestAlign(),
1227                                 MDep->getRawSource(), MDep->getSourceAlign(),
1228                                 M->getLength(), M->isVolatile());
1229   NewM->copyMetadata(*M, LLVMContext::MD_DIAssignID);
1230 
1231   assert(isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M)));
1232   auto *LastDef = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M));
1233   auto *NewAccess = MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1234   MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1235 
1236   // Remove the instruction we're replacing.
1237   eraseInstruction(M);
1238   ++NumMemCpyInstr;
1239   return true;
1240 }
1241 
1242 /// We've found that the (upward scanning) memory dependence of \p MemCpy is
1243 /// \p MemSet.  Try to simplify \p MemSet to only set the trailing bytes that
1244 /// weren't copied over by \p MemCpy.
1245 ///
1246 /// In other words, transform:
1247 /// \code
1248 ///   memset(dst, c, dst_size);
1249 ///   memcpy(dst, src, src_size);
1250 /// \endcode
1251 /// into:
1252 /// \code
1253 ///   memcpy(dst, src, src_size);
1254 ///   memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
1255 /// \endcode
1256 bool MemCpyOptPass::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
1257                                                   MemSetInst *MemSet,
1258                                                   BatchAAResults &BAA) {
1259   // We can only transform memset/memcpy with the same destination.
1260   if (!BAA.isMustAlias(MemSet->getDest(), MemCpy->getDest()))
1261     return false;
1262 
1263   // Check that src and dst of the memcpy aren't the same. While memcpy
1264   // operands cannot partially overlap, exact equality is allowed.
1265   if (isModSet(BAA.getModRefInfo(MemCpy, MemoryLocation::getForSource(MemCpy))))
1266     return false;
1267 
1268   // We know that dst up to src_size is not written. We now need to make sure
1269   // that dst up to dst_size is not accessed. (If we did not move the memset,
1270   // checking for reads would be sufficient.)
1271   if (accessedBetween(BAA, MemoryLocation::getForDest(MemSet),
1272                       MSSA->getMemoryAccess(MemSet),
1273                       MSSA->getMemoryAccess(MemCpy)))
1274     return false;
1275 
1276   // Use the same i8* dest as the memcpy, killing the memset dest if different.
1277   Value *Dest = MemCpy->getRawDest();
1278   Value *DestSize = MemSet->getLength();
1279   Value *SrcSize = MemCpy->getLength();
1280 
1281   if (mayBeVisibleThroughUnwinding(Dest, MemSet, MemCpy))
1282     return false;
1283 
1284   // If the sizes are the same, simply drop the memset instead of generating
1285   // a replacement with zero size.
1286   if (DestSize == SrcSize) {
1287     eraseInstruction(MemSet);
1288     return true;
1289   }
1290 
1291   // By default, create an unaligned memset.
1292   Align Alignment = Align(1);
1293   // If Dest is aligned, and SrcSize is constant, use the minimum alignment
1294   // of the sum.
1295   const Align DestAlign = std::max(MemSet->getDestAlign().valueOrOne(),
1296                                    MemCpy->getDestAlign().valueOrOne());
1297   if (DestAlign > 1)
1298     if (auto *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
1299       Alignment = commonAlignment(DestAlign, SrcSizeC->getZExtValue());
1300 
1301   IRBuilder<> Builder(MemCpy);
1302 
1303   // If the sizes have different types, zext the smaller one.
1304   if (DestSize->getType() != SrcSize->getType()) {
1305     if (DestSize->getType()->getIntegerBitWidth() >
1306         SrcSize->getType()->getIntegerBitWidth())
1307       SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
1308     else
1309       DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
1310   }
1311 
1312   Value *Ule = Builder.CreateICmpULE(DestSize, SrcSize);
1313   Value *SizeDiff = Builder.CreateSub(DestSize, SrcSize);
1314   Value *MemsetLen = Builder.CreateSelect(
1315       Ule, ConstantInt::getNullValue(DestSize->getType()), SizeDiff);
1316   unsigned DestAS = Dest->getType()->getPointerAddressSpace();
1317   Instruction *NewMemSet = Builder.CreateMemSet(
1318       Builder.CreateGEP(
1319           Builder.getInt8Ty(),
1320           Builder.CreatePointerCast(Dest, Builder.getInt8PtrTy(DestAS)),
1321           SrcSize),
1322       MemSet->getOperand(1), MemsetLen, Alignment);
1323 
1324   assert(isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy)) &&
1325          "MemCpy must be a MemoryDef");
1326   // The new memset is inserted after the memcpy, but it is known that its
1327   // defining access is the memset about to be removed which immediately
1328   // precedes the memcpy.
1329   auto *LastDef =
1330       cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy));
1331   auto *NewAccess = MSSAU->createMemoryAccessBefore(
1332       NewMemSet, LastDef->getDefiningAccess(), LastDef);
1333   MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1334 
1335   eraseInstruction(MemSet);
1336   return true;
1337 }
1338 
1339 /// Determine whether the instruction has undefined content for the given Size,
1340 /// either because it was freshly alloca'd or started its lifetime.
1341 static bool hasUndefContents(MemorySSA *MSSA, BatchAAResults &AA, Value *V,
1342                              MemoryDef *Def, Value *Size) {
1343   if (MSSA->isLiveOnEntryDef(Def))
1344     return isa<AllocaInst>(getUnderlyingObject(V));
1345 
1346   if (auto *II = dyn_cast_or_null<IntrinsicInst>(Def->getMemoryInst())) {
1347     if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1348       auto *LTSize = cast<ConstantInt>(II->getArgOperand(0));
1349 
1350       if (auto *CSize = dyn_cast<ConstantInt>(Size)) {
1351         if (AA.isMustAlias(V, II->getArgOperand(1)) &&
1352             LTSize->getZExtValue() >= CSize->getZExtValue())
1353           return true;
1354       }
1355 
1356       // If the lifetime.start covers a whole alloca (as it almost always
1357       // does) and we're querying a pointer based on that alloca, then we know
1358       // the memory is definitely undef, regardless of how exactly we alias.
1359       // The size also doesn't matter, as an out-of-bounds access would be UB.
1360       if (auto *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(V))) {
1361         if (getUnderlyingObject(II->getArgOperand(1)) == Alloca) {
1362           const DataLayout &DL = Alloca->getModule()->getDataLayout();
1363           if (std::optional<TypeSize> AllocaSize =
1364                   Alloca->getAllocationSize(DL))
1365             if (*AllocaSize == LTSize->getValue())
1366               return true;
1367         }
1368       }
1369     }
1370   }
1371 
1372   return false;
1373 }
1374 
1375 /// Transform memcpy to memset when its source was just memset.
1376 /// In other words, turn:
1377 /// \code
1378 ///   memset(dst1, c, dst1_size);
1379 ///   memcpy(dst2, dst1, dst2_size);
1380 /// \endcode
1381 /// into:
1382 /// \code
1383 ///   memset(dst1, c, dst1_size);
1384 ///   memset(dst2, c, dst2_size);
1385 /// \endcode
1386 /// When dst2_size <= dst1_size.
1387 bool MemCpyOptPass::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
1388                                                MemSetInst *MemSet,
1389                                                BatchAAResults &BAA) {
1390   // Make sure that memcpy(..., memset(...), ...), that is we are memsetting and
1391   // memcpying from the same address. Otherwise it is hard to reason about.
1392   if (!BAA.isMustAlias(MemSet->getRawDest(), MemCpy->getRawSource()))
1393     return false;
1394 
1395   Value *MemSetSize = MemSet->getLength();
1396   Value *CopySize = MemCpy->getLength();
1397 
1398   if (MemSetSize != CopySize) {
1399     // Make sure the memcpy doesn't read any more than what the memset wrote.
1400     // Don't worry about sizes larger than i64.
1401 
1402     // A known memset size is required.
1403     auto *CMemSetSize = dyn_cast<ConstantInt>(MemSetSize);
1404     if (!CMemSetSize)
1405       return false;
1406 
1407     // A known memcpy size is also required.
1408     auto  *CCopySize = dyn_cast<ConstantInt>(CopySize);
1409     if (!CCopySize)
1410       return false;
1411     if (CCopySize->getZExtValue() > CMemSetSize->getZExtValue()) {
1412       // If the memcpy is larger than the memset, but the memory was undef prior
1413       // to the memset, we can just ignore the tail. Technically we're only
1414       // interested in the bytes from MemSetSize..CopySize here, but as we can't
1415       // easily represent this location, we use the full 0..CopySize range.
1416       MemoryLocation MemCpyLoc = MemoryLocation::getForSource(MemCpy);
1417       bool CanReduceSize = false;
1418       MemoryUseOrDef *MemSetAccess = MSSA->getMemoryAccess(MemSet);
1419       MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
1420           MemSetAccess->getDefiningAccess(), MemCpyLoc, BAA);
1421       if (auto *MD = dyn_cast<MemoryDef>(Clobber))
1422         if (hasUndefContents(MSSA, BAA, MemCpy->getSource(), MD, CopySize))
1423           CanReduceSize = true;
1424 
1425       if (!CanReduceSize)
1426         return false;
1427       CopySize = MemSetSize;
1428     }
1429   }
1430 
1431   IRBuilder<> Builder(MemCpy);
1432   Instruction *NewM =
1433       Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),
1434                            CopySize, MemCpy->getDestAlign());
1435   auto *LastDef =
1436       cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy));
1437   auto *NewAccess = MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1438   MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1439 
1440   return true;
1441 }
1442 
1443 /// Perform simplification of memcpy's.  If we have memcpy A
1444 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
1445 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
1446 /// circumstances). This allows later passes to remove the first memcpy
1447 /// altogether.
1448 bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
1449   // We can only optimize non-volatile memcpy's.
1450   if (M->isVolatile()) return false;
1451 
1452   // If the source and destination of the memcpy are the same, then zap it.
1453   if (M->getSource() == M->getDest()) {
1454     ++BBI;
1455     eraseInstruction(M);
1456     return true;
1457   }
1458 
1459   // If copying from a constant, try to turn the memcpy into a memset.
1460   if (auto *GV = dyn_cast<GlobalVariable>(M->getSource()))
1461     if (GV->isConstant() && GV->hasDefinitiveInitializer())
1462       if (Value *ByteVal = isBytewiseValue(GV->getInitializer(),
1463                                            M->getModule()->getDataLayout())) {
1464         IRBuilder<> Builder(M);
1465         Instruction *NewM = Builder.CreateMemSet(
1466             M->getRawDest(), ByteVal, M->getLength(), M->getDestAlign(), false);
1467         auto *LastDef =
1468             cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M));
1469         auto *NewAccess =
1470             MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1471         MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1472 
1473         eraseInstruction(M);
1474         ++NumCpyToSet;
1475         return true;
1476       }
1477 
1478   BatchAAResults BAA(*AA);
1479   MemoryUseOrDef *MA = MSSA->getMemoryAccess(M);
1480   // FIXME: Not using getClobberingMemoryAccess() here due to PR54682.
1481   MemoryAccess *AnyClobber = MA->getDefiningAccess();
1482   MemoryLocation DestLoc = MemoryLocation::getForDest(M);
1483   const MemoryAccess *DestClobber =
1484       MSSA->getWalker()->getClobberingMemoryAccess(AnyClobber, DestLoc, BAA);
1485 
1486   // Try to turn a partially redundant memset + memcpy into
1487   // memcpy + smaller memset.  We don't need the memcpy size for this.
1488   // The memcpy most post-dom the memset, so limit this to the same basic
1489   // block. A non-local generalization is likely not worthwhile.
1490   if (auto *MD = dyn_cast<MemoryDef>(DestClobber))
1491     if (auto *MDep = dyn_cast_or_null<MemSetInst>(MD->getMemoryInst()))
1492       if (DestClobber->getBlock() == M->getParent())
1493         if (processMemSetMemCpyDependence(M, MDep, BAA))
1494           return true;
1495 
1496   MemoryAccess *SrcClobber = MSSA->getWalker()->getClobberingMemoryAccess(
1497       AnyClobber, MemoryLocation::getForSource(M), BAA);
1498 
1499   // There are four possible optimizations we can do for memcpy:
1500   //   a) memcpy-memcpy xform which exposes redundance for DSE.
1501   //   b) call-memcpy xform for return slot optimization.
1502   //   c) memcpy from freshly alloca'd space or space that has just started
1503   //      its lifetime copies undefined data, and we can therefore eliminate
1504   //      the memcpy in favor of the data that was already at the destination.
1505   //   d) memcpy from a just-memset'd source can be turned into memset.
1506   if (auto *MD = dyn_cast<MemoryDef>(SrcClobber)) {
1507     if (Instruction *MI = MD->getMemoryInst()) {
1508       if (auto *CopySize = dyn_cast<ConstantInt>(M->getLength())) {
1509         if (auto *C = dyn_cast<CallInst>(MI)) {
1510           if (performCallSlotOptzn(M, M, M->getDest(), M->getSource(),
1511                                    TypeSize::getFixed(CopySize->getZExtValue()),
1512                                    M->getDestAlign().valueOrOne(), BAA,
1513                                    [C]() -> CallInst * { return C; })) {
1514             LLVM_DEBUG(dbgs() << "Performed call slot optimization:\n"
1515                               << "    call: " << *C << "\n"
1516                               << "    memcpy: " << *M << "\n");
1517             eraseInstruction(M);
1518             ++NumMemCpyInstr;
1519             return true;
1520           }
1521         }
1522       }
1523       if (auto *MDep = dyn_cast<MemCpyInst>(MI))
1524         return processMemCpyMemCpyDependence(M, MDep, BAA);
1525       if (auto *MDep = dyn_cast<MemSetInst>(MI)) {
1526         if (performMemCpyToMemSetOptzn(M, MDep, BAA)) {
1527           LLVM_DEBUG(dbgs() << "Converted memcpy to memset\n");
1528           eraseInstruction(M);
1529           ++NumCpyToSet;
1530           return true;
1531         }
1532       }
1533     }
1534 
1535     if (hasUndefContents(MSSA, BAA, M->getSource(), MD, M->getLength())) {
1536       LLVM_DEBUG(dbgs() << "Removed memcpy from undef\n");
1537       eraseInstruction(M);
1538       ++NumMemCpyInstr;
1539       return true;
1540     }
1541   }
1542 
1543   return false;
1544 }
1545 
1546 /// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1547 /// not to alias.
1548 bool MemCpyOptPass::processMemMove(MemMoveInst *M) {
1549   // See if the source could be modified by this memmove potentially.
1550   if (isModSet(AA->getModRefInfo(M, MemoryLocation::getForSource(M))))
1551     return false;
1552 
1553   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: " << *M
1554                     << "\n");
1555 
1556   // If not, then we know we can transform this.
1557   Type *ArgTys[3] = { M->getRawDest()->getType(),
1558                       M->getRawSource()->getType(),
1559                       M->getLength()->getType() };
1560   M->setCalledFunction(Intrinsic::getDeclaration(M->getModule(),
1561                                                  Intrinsic::memcpy, ArgTys));
1562 
1563   // For MemorySSA nothing really changes (except that memcpy may imply stricter
1564   // aliasing guarantees).
1565 
1566   ++NumMoveToCpy;
1567   return true;
1568 }
1569 
1570 /// This is called on every byval argument in call sites.
1571 bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
1572   const DataLayout &DL = CB.getCaller()->getParent()->getDataLayout();
1573   // Find out what feeds this byval argument.
1574   Value *ByValArg = CB.getArgOperand(ArgNo);
1575   Type *ByValTy = CB.getParamByValType(ArgNo);
1576   TypeSize ByValSize = DL.getTypeAllocSize(ByValTy);
1577   MemoryLocation Loc(ByValArg, LocationSize::precise(ByValSize));
1578   MemoryUseOrDef *CallAccess = MSSA->getMemoryAccess(&CB);
1579   if (!CallAccess)
1580     return false;
1581   MemCpyInst *MDep = nullptr;
1582   BatchAAResults BAA(*AA);
1583   MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
1584       CallAccess->getDefiningAccess(), Loc, BAA);
1585   if (auto *MD = dyn_cast<MemoryDef>(Clobber))
1586     MDep = dyn_cast_or_null<MemCpyInst>(MD->getMemoryInst());
1587 
1588   // If the byval argument isn't fed by a memcpy, ignore it.  If it is fed by
1589   // a memcpy, see if we can byval from the source of the memcpy instead of the
1590   // result.
1591   if (!MDep || MDep->isVolatile() ||
1592       ByValArg->stripPointerCasts() != MDep->getDest())
1593     return false;
1594 
1595   // The length of the memcpy must be larger or equal to the size of the byval.
1596   auto *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1597   if (!C1 || !TypeSize::isKnownGE(
1598                  TypeSize::getFixed(C1->getValue().getZExtValue()), ByValSize))
1599     return false;
1600 
1601   // Get the alignment of the byval.  If the call doesn't specify the alignment,
1602   // then it is some target specific value that we can't know.
1603   MaybeAlign ByValAlign = CB.getParamAlign(ArgNo);
1604   if (!ByValAlign) return false;
1605 
1606   // If it is greater than the memcpy, then we check to see if we can force the
1607   // source of the memcpy to the alignment we need.  If we fail, we bail out.
1608   MaybeAlign MemDepAlign = MDep->getSourceAlign();
1609   if ((!MemDepAlign || *MemDepAlign < *ByValAlign) &&
1610       getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL, &CB, AC,
1611                                  DT) < *ByValAlign)
1612     return false;
1613 
1614   // The address space of the memcpy source must match the byval argument
1615   if (MDep->getSource()->getType()->getPointerAddressSpace() !=
1616       ByValArg->getType()->getPointerAddressSpace())
1617     return false;
1618 
1619   // Verify that the copied-from memory doesn't change in between the memcpy and
1620   // the byval call.
1621   //    memcpy(a <- b)
1622   //    *b = 42;
1623   //    foo(*a)
1624   // It would be invalid to transform the second memcpy into foo(*b).
1625   if (writtenBetween(MSSA, BAA, MemoryLocation::getForSource(MDep),
1626                      MSSA->getMemoryAccess(MDep), MSSA->getMemoryAccess(&CB)))
1627     return false;
1628 
1629   Value *TmpCast = MDep->getSource();
1630   if (MDep->getSource()->getType() != ByValArg->getType()) {
1631     BitCastInst *TmpBitCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
1632                                               "tmpcast", &CB);
1633     // Set the tmpcast's DebugLoc to MDep's
1634     TmpBitCast->setDebugLoc(MDep->getDebugLoc());
1635     TmpCast = TmpBitCast;
1636   }
1637 
1638   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
1639                     << "  " << *MDep << "\n"
1640                     << "  " << CB << "\n");
1641 
1642   // Otherwise we're good!  Update the byval argument.
1643   CB.setArgOperand(ArgNo, TmpCast);
1644   ++NumMemCpyInstr;
1645   return true;
1646 }
1647 
1648 /// Executes one iteration of MemCpyOptPass.
1649 bool MemCpyOptPass::iterateOnFunction(Function &F) {
1650   bool MadeChange = false;
1651 
1652   // Walk all instruction in the function.
1653   for (BasicBlock &BB : F) {
1654     // Skip unreachable blocks. For example processStore assumes that an
1655     // instruction in a BB can't be dominated by a later instruction in the
1656     // same BB (which is a scenario that can happen for an unreachable BB that
1657     // has itself as a predecessor).
1658     if (!DT->isReachableFromEntry(&BB))
1659       continue;
1660 
1661     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
1662         // Avoid invalidating the iterator.
1663       Instruction *I = &*BI++;
1664 
1665       bool RepeatInstruction = false;
1666 
1667       if (auto *SI = dyn_cast<StoreInst>(I))
1668         MadeChange |= processStore(SI, BI);
1669       else if (auto *M = dyn_cast<MemSetInst>(I))
1670         RepeatInstruction = processMemSet(M, BI);
1671       else if (auto *M = dyn_cast<MemCpyInst>(I))
1672         RepeatInstruction = processMemCpy(M, BI);
1673       else if (auto *M = dyn_cast<MemMoveInst>(I))
1674         RepeatInstruction = processMemMove(M);
1675       else if (auto *CB = dyn_cast<CallBase>(I)) {
1676         for (unsigned i = 0, e = CB->arg_size(); i != e; ++i)
1677           if (CB->isByValArgument(i))
1678             MadeChange |= processByValArgument(*CB, i);
1679       }
1680 
1681       // Reprocess the instruction if desired.
1682       if (RepeatInstruction) {
1683         if (BI != BB.begin())
1684           --BI;
1685         MadeChange = true;
1686       }
1687     }
1688   }
1689 
1690   return MadeChange;
1691 }
1692 
1693 PreservedAnalyses MemCpyOptPass::run(Function &F, FunctionAnalysisManager &AM) {
1694   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1695   auto *AA = &AM.getResult<AAManager>(F);
1696   auto *AC = &AM.getResult<AssumptionAnalysis>(F);
1697   auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
1698   auto *MSSA = &AM.getResult<MemorySSAAnalysis>(F);
1699 
1700   bool MadeChange = runImpl(F, &TLI, AA, AC, DT, &MSSA->getMSSA());
1701   if (!MadeChange)
1702     return PreservedAnalyses::all();
1703 
1704   PreservedAnalyses PA;
1705   PA.preserveSet<CFGAnalyses>();
1706   PA.preserve<MemorySSAAnalysis>();
1707   return PA;
1708 }
1709 
1710 bool MemCpyOptPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
1711                             AliasAnalysis *AA_, AssumptionCache *AC_,
1712                             DominatorTree *DT_, MemorySSA *MSSA_) {
1713   bool MadeChange = false;
1714   TLI = TLI_;
1715   AA = AA_;
1716   AC = AC_;
1717   DT = DT_;
1718   MSSA = MSSA_;
1719   MemorySSAUpdater MSSAU_(MSSA_);
1720   MSSAU = &MSSAU_;
1721 
1722   while (true) {
1723     if (!iterateOnFunction(F))
1724       break;
1725     MadeChange = true;
1726   }
1727 
1728   if (VerifyMemorySSA)
1729     MSSA_->verifyMemorySSA();
1730 
1731   return MadeChange;
1732 }
1733 
1734 /// This is the main transformation entry point for a function.
1735 bool MemCpyOptLegacyPass::runOnFunction(Function &F) {
1736   if (skipFunction(F))
1737     return false;
1738 
1739   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1740   auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1741   auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1742   auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1743   auto *MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA();
1744 
1745   return Impl.runImpl(F, TLI, AA, AC, DT, MSSA);
1746 }
1747