xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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/None.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/iterator_range.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/AssumptionCache.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
25 #include "llvm/Analysis/MemoryLocation.h"
26 #include "llvm/Analysis/TargetLibraryInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/Argument.h"
29 #include "llvm/IR/BasicBlock.h"
30 #include "llvm/IR/Constants.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DerivedTypes.h"
33 #include "llvm/IR/Dominators.h"
34 #include "llvm/IR/Function.h"
35 #include "llvm/IR/GetElementPtrTypeIterator.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/Operator.h"
46 #include "llvm/IR/PassManager.h"
47 #include "llvm/IR/Type.h"
48 #include "llvm/IR/User.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/InitializePasses.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/Debug.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include "llvm/Transforms/Scalar.h"
57 #include "llvm/Transforms/Utils/Local.h"
58 #include <algorithm>
59 #include <cassert>
60 #include <cstdint>
61 #include <utility>
62 
63 using namespace llvm;
64 
65 #define DEBUG_TYPE "memcpyopt"
66 
67 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
68 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
69 STATISTIC(NumMoveToCpy,   "Number of memmoves converted to memcpy");
70 STATISTIC(NumCpyToSet,    "Number of memcpys converted to memset");
71 
72 namespace {
73 
74 /// Represents a range of memset'd bytes with the ByteVal value.
75 /// This allows us to analyze stores like:
76 ///   store 0 -> P+1
77 ///   store 0 -> P+0
78 ///   store 0 -> P+3
79 ///   store 0 -> P+2
80 /// which sometimes happens with stores to arrays of structs etc.  When we see
81 /// the first store, we make a range [1, 2).  The second store extends the range
82 /// to [0, 2).  The third makes a new range [2, 3).  The fourth store joins the
83 /// two ranges into [0, 3) which is memset'able.
84 struct MemsetRange {
85   // Start/End - A semi range that describes the span that this range covers.
86   // The range is closed at the start and open at the end: [Start, End).
87   int64_t Start, End;
88 
89   /// StartPtr - The getelementptr instruction that points to the start of the
90   /// range.
91   Value *StartPtr;
92 
93   /// Alignment - The known alignment of the first store.
94   unsigned Alignment;
95 
96   /// TheStores - The actual stores that make up this range.
97   SmallVector<Instruction*, 16> TheStores;
98 
99   bool isProfitableToUseMemset(const DataLayout &DL) const;
100 };
101 
102 } // end anonymous namespace
103 
104 bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
105   // If we found more than 4 stores to merge or 16 bytes, use memset.
106   if (TheStores.size() >= 4 || End-Start >= 16) return true;
107 
108   // If there is nothing to merge, don't do anything.
109   if (TheStores.size() < 2) return false;
110 
111   // If any of the stores are a memset, then it is always good to extend the
112   // memset.
113   for (Instruction *SI : TheStores)
114     if (!isa<StoreInst>(SI))
115       return true;
116 
117   // Assume that the code generator is capable of merging pairs of stores
118   // together if it wants to.
119   if (TheStores.size() == 2) return false;
120 
121   // If we have fewer than 8 stores, it can still be worthwhile to do this.
122   // For example, merging 4 i8 stores into an i32 store is useful almost always.
123   // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
124   // memset will be split into 2 32-bit stores anyway) and doing so can
125   // pessimize the llvm optimizer.
126   //
127   // Since we don't have perfect knowledge here, make some assumptions: assume
128   // the maximum GPR width is the same size as the largest legal integer
129   // size. If so, check to see whether we will end up actually reducing the
130   // number of stores used.
131   unsigned Bytes = unsigned(End-Start);
132   unsigned MaxIntSize = DL.getLargestLegalIntTypeSizeInBits() / 8;
133   if (MaxIntSize == 0)
134     MaxIntSize = 1;
135   unsigned NumPointerStores = Bytes / MaxIntSize;
136 
137   // Assume the remaining bytes if any are done a byte at a time.
138   unsigned NumByteStores = Bytes % MaxIntSize;
139 
140   // If we will reduce the # stores (according to this heuristic), do the
141   // transformation.  This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
142   // etc.
143   return TheStores.size() > NumPointerStores+NumByteStores;
144 }
145 
146 namespace {
147 
148 class MemsetRanges {
149   using range_iterator = SmallVectorImpl<MemsetRange>::iterator;
150 
151   /// A sorted list of the memset ranges.
152   SmallVector<MemsetRange, 8> Ranges;
153 
154   const DataLayout &DL;
155 
156 public:
157   MemsetRanges(const DataLayout &DL) : DL(DL) {}
158 
159   using const_iterator = SmallVectorImpl<MemsetRange>::const_iterator;
160 
161   const_iterator begin() const { return Ranges.begin(); }
162   const_iterator end() const { return Ranges.end(); }
163   bool empty() const { return Ranges.empty(); }
164 
165   void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
166     if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
167       addStore(OffsetFromFirst, SI);
168     else
169       addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
170   }
171 
172   void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
173     int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
174 
175     addRange(OffsetFromFirst, StoreSize, SI->getPointerOperand(),
176              SI->getAlign().value(), SI);
177   }
178 
179   void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
180     int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
181     addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getDestAlignment(), MSI);
182   }
183 
184   void addRange(int64_t Start, int64_t Size, Value *Ptr,
185                 unsigned Alignment, Instruction *Inst);
186 };
187 
188 } // end anonymous namespace
189 
190 /// Add a new store to the MemsetRanges data structure.  This adds a
191 /// new range for the specified store at the specified offset, merging into
192 /// existing ranges as appropriate.
193 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
194                             unsigned Alignment, Instruction *Inst) {
195   int64_t End = Start+Size;
196 
197   range_iterator I = partition_point(
198       Ranges, [=](const MemsetRange &O) { return O.End < Start; });
199 
200   // We now know that I == E, in which case we didn't find anything to merge
201   // with, or that Start <= I->End.  If End < I->Start or I == E, then we need
202   // to insert a new range.  Handle this now.
203   if (I == Ranges.end() || End < I->Start) {
204     MemsetRange &R = *Ranges.insert(I, MemsetRange());
205     R.Start        = Start;
206     R.End          = End;
207     R.StartPtr     = Ptr;
208     R.Alignment    = Alignment;
209     R.TheStores.push_back(Inst);
210     return;
211   }
212 
213   // This store overlaps with I, add it.
214   I->TheStores.push_back(Inst);
215 
216   // At this point, we may have an interval that completely contains our store.
217   // If so, just add it to the interval and return.
218   if (I->Start <= Start && I->End >= End)
219     return;
220 
221   // Now we know that Start <= I->End and End >= I->Start so the range overlaps
222   // but is not entirely contained within the range.
223 
224   // See if the range extends the start of the range.  In this case, it couldn't
225   // possibly cause it to join the prior range, because otherwise we would have
226   // stopped on *it*.
227   if (Start < I->Start) {
228     I->Start = Start;
229     I->StartPtr = Ptr;
230     I->Alignment = Alignment;
231   }
232 
233   // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
234   // is in or right at the end of I), and that End >= I->Start.  Extend I out to
235   // End.
236   if (End > I->End) {
237     I->End = End;
238     range_iterator NextI = I;
239     while (++NextI != Ranges.end() && End >= NextI->Start) {
240       // Merge the range in.
241       I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
242       if (NextI->End > I->End)
243         I->End = NextI->End;
244       Ranges.erase(NextI);
245       NextI = I;
246     }
247   }
248 }
249 
250 //===----------------------------------------------------------------------===//
251 //                         MemCpyOptLegacyPass Pass
252 //===----------------------------------------------------------------------===//
253 
254 namespace {
255 
256 class MemCpyOptLegacyPass : public FunctionPass {
257   MemCpyOptPass Impl;
258 
259 public:
260   static char ID; // Pass identification, replacement for typeid
261 
262   MemCpyOptLegacyPass() : FunctionPass(ID) {
263     initializeMemCpyOptLegacyPassPass(*PassRegistry::getPassRegistry());
264   }
265 
266   bool runOnFunction(Function &F) override;
267 
268 private:
269   // This transformation requires dominator postdominator info
270   void getAnalysisUsage(AnalysisUsage &AU) const override {
271     AU.setPreservesCFG();
272     AU.addRequired<AssumptionCacheTracker>();
273     AU.addRequired<DominatorTreeWrapperPass>();
274     AU.addRequired<MemoryDependenceWrapperPass>();
275     AU.addRequired<AAResultsWrapperPass>();
276     AU.addRequired<TargetLibraryInfoWrapperPass>();
277     AU.addPreserved<GlobalsAAWrapperPass>();
278     AU.addPreserved<MemoryDependenceWrapperPass>();
279   }
280 };
281 
282 } // end anonymous namespace
283 
284 char MemCpyOptLegacyPass::ID = 0;
285 
286 /// The public interface to this file...
287 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOptLegacyPass(); }
288 
289 INITIALIZE_PASS_BEGIN(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",
290                       false, false)
291 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
292 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
293 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
294 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
295 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
296 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
297 INITIALIZE_PASS_END(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",
298                     false, false)
299 
300 /// When scanning forward over instructions, we look for some other patterns to
301 /// fold away. In particular, this looks for stores to neighboring locations of
302 /// memory. If it sees enough consecutive ones, it attempts to merge them
303 /// together into a memcpy/memset.
304 Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
305                                                  Value *StartPtr,
306                                                  Value *ByteVal) {
307   const DataLayout &DL = StartInst->getModule()->getDataLayout();
308 
309   // Okay, so we now have a single store that can be splatable.  Scan to find
310   // all subsequent stores of the same value to offset from the same pointer.
311   // Join these together into ranges, so we can decide whether contiguous blocks
312   // are stored.
313   MemsetRanges Ranges(DL);
314 
315   BasicBlock::iterator BI(StartInst);
316   for (++BI; !BI->isTerminator(); ++BI) {
317     if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
318       // If the instruction is readnone, ignore it, otherwise bail out.  We
319       // don't even allow readonly here because we don't want something like:
320       // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
321       if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
322         break;
323       continue;
324     }
325 
326     if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
327       // If this is a store, see if we can merge it in.
328       if (!NextStore->isSimple()) break;
329 
330       // Check to see if this stored value is of the same byte-splattable value.
331       Value *StoredByte = isBytewiseValue(NextStore->getOperand(0), DL);
332       if (isa<UndefValue>(ByteVal) && StoredByte)
333         ByteVal = StoredByte;
334       if (ByteVal != StoredByte)
335         break;
336 
337       // Check to see if this store is to a constant offset from the start ptr.
338       Optional<int64_t> Offset =
339           isPointerOffset(StartPtr, NextStore->getPointerOperand(), DL);
340       if (!Offset)
341         break;
342 
343       Ranges.addStore(*Offset, NextStore);
344     } else {
345       MemSetInst *MSI = cast<MemSetInst>(BI);
346 
347       if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
348           !isa<ConstantInt>(MSI->getLength()))
349         break;
350 
351       // Check to see if this store is to a constant offset from the start ptr.
352       Optional<int64_t> Offset = isPointerOffset(StartPtr, MSI->getDest(), DL);
353       if (!Offset)
354         break;
355 
356       Ranges.addMemSet(*Offset, MSI);
357     }
358   }
359 
360   // If we have no ranges, then we just had a single store with nothing that
361   // could be merged in.  This is a very common case of course.
362   if (Ranges.empty())
363     return nullptr;
364 
365   // If we had at least one store that could be merged in, add the starting
366   // store as well.  We try to avoid this unless there is at least something
367   // interesting as a small compile-time optimization.
368   Ranges.addInst(0, StartInst);
369 
370   // If we create any memsets, we put it right before the first instruction that
371   // isn't part of the memset block.  This ensure that the memset is dominated
372   // by any addressing instruction needed by the start of the block.
373   IRBuilder<> Builder(&*BI);
374 
375   // Now that we have full information about ranges, loop over the ranges and
376   // emit memset's for anything big enough to be worthwhile.
377   Instruction *AMemSet = nullptr;
378   for (const MemsetRange &Range : Ranges) {
379     if (Range.TheStores.size() == 1) continue;
380 
381     // If it is profitable to lower this range to memset, do so now.
382     if (!Range.isProfitableToUseMemset(DL))
383       continue;
384 
385     // Otherwise, we do want to transform this!  Create a new memset.
386     // Get the starting pointer of the block.
387     StartPtr = Range.StartPtr;
388 
389     AMemSet = Builder.CreateMemSet(StartPtr, ByteVal, Range.End - Range.Start,
390                                    MaybeAlign(Range.Alignment));
391     LLVM_DEBUG(dbgs() << "Replace stores:\n"; for (Instruction *SI
392                                                    : Range.TheStores) dbgs()
393                                               << *SI << '\n';
394                dbgs() << "With: " << *AMemSet << '\n');
395 
396     if (!Range.TheStores.empty())
397       AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
398 
399     // Zap all the stores.
400     for (Instruction *SI : Range.TheStores) {
401       MD->removeInstruction(SI);
402       SI->eraseFromParent();
403     }
404     ++NumMemSetInfer;
405   }
406 
407   return AMemSet;
408 }
409 
410 // This method try to lift a store instruction before position P.
411 // It will lift the store and its argument + that anything that
412 // may alias with these.
413 // The method returns true if it was successful.
414 static bool moveUp(AliasAnalysis &AA, StoreInst *SI, Instruction *P,
415                    const LoadInst *LI) {
416   // If the store alias this position, early bail out.
417   MemoryLocation StoreLoc = MemoryLocation::get(SI);
418   if (isModOrRefSet(AA.getModRefInfo(P, StoreLoc)))
419     return false;
420 
421   // Keep track of the arguments of all instruction we plan to lift
422   // so we can make sure to lift them as well if appropriate.
423   DenseSet<Instruction*> Args;
424   if (auto *Ptr = dyn_cast<Instruction>(SI->getPointerOperand()))
425     if (Ptr->getParent() == SI->getParent())
426       Args.insert(Ptr);
427 
428   // Instruction to lift before P.
429   SmallVector<Instruction*, 8> ToLift;
430 
431   // Memory locations of lifted instructions.
432   SmallVector<MemoryLocation, 8> MemLocs{StoreLoc};
433 
434   // Lifted calls.
435   SmallVector<const CallBase *, 8> Calls;
436 
437   const MemoryLocation LoadLoc = MemoryLocation::get(LI);
438 
439   for (auto I = --SI->getIterator(), E = P->getIterator(); I != E; --I) {
440     auto *C = &*I;
441 
442     bool MayAlias = isModOrRefSet(AA.getModRefInfo(C, None));
443 
444     bool NeedLift = false;
445     if (Args.erase(C))
446       NeedLift = true;
447     else if (MayAlias) {
448       NeedLift = llvm::any_of(MemLocs, [C, &AA](const MemoryLocation &ML) {
449         return isModOrRefSet(AA.getModRefInfo(C, ML));
450       });
451 
452       if (!NeedLift)
453         NeedLift = llvm::any_of(Calls, [C, &AA](const CallBase *Call) {
454           return isModOrRefSet(AA.getModRefInfo(C, Call));
455         });
456     }
457 
458     if (!NeedLift)
459       continue;
460 
461     if (MayAlias) {
462       // Since LI is implicitly moved downwards past the lifted instructions,
463       // none of them may modify its source.
464       if (isModSet(AA.getModRefInfo(C, LoadLoc)))
465         return false;
466       else if (const auto *Call = dyn_cast<CallBase>(C)) {
467         // If we can't lift this before P, it's game over.
468         if (isModOrRefSet(AA.getModRefInfo(P, Call)))
469           return false;
470 
471         Calls.push_back(Call);
472       } else if (isa<LoadInst>(C) || isa<StoreInst>(C) || isa<VAArgInst>(C)) {
473         // If we can't lift this before P, it's game over.
474         auto ML = MemoryLocation::get(C);
475         if (isModOrRefSet(AA.getModRefInfo(P, ML)))
476           return false;
477 
478         MemLocs.push_back(ML);
479       } else
480         // We don't know how to lift this instruction.
481         return false;
482     }
483 
484     ToLift.push_back(C);
485     for (unsigned k = 0, e = C->getNumOperands(); k != e; ++k)
486       if (auto *A = dyn_cast<Instruction>(C->getOperand(k))) {
487         if (A->getParent() == SI->getParent()) {
488           // Cannot hoist user of P above P
489           if(A == P) return false;
490           Args.insert(A);
491         }
492       }
493   }
494 
495   // We made it, we need to lift
496   for (auto *I : llvm::reverse(ToLift)) {
497     LLVM_DEBUG(dbgs() << "Lifting " << *I << " before " << *P << "\n");
498     I->moveBefore(P);
499   }
500 
501   return true;
502 }
503 
504 bool MemCpyOptPass::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
505   if (!SI->isSimple()) return false;
506 
507   // Avoid merging nontemporal stores since the resulting
508   // memcpy/memset would not be able to preserve the nontemporal hint.
509   // In theory we could teach how to propagate the !nontemporal metadata to
510   // memset calls. However, that change would force the backend to
511   // conservatively expand !nontemporal memset calls back to sequences of
512   // store instructions (effectively undoing the merging).
513   if (SI->getMetadata(LLVMContext::MD_nontemporal))
514     return false;
515 
516   const DataLayout &DL = SI->getModule()->getDataLayout();
517 
518   // Load to store forwarding can be interpreted as memcpy.
519   if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
520     if (LI->isSimple() && LI->hasOneUse() &&
521         LI->getParent() == SI->getParent()) {
522 
523       auto *T = LI->getType();
524       if (T->isAggregateType()) {
525         AliasAnalysis &AA = LookupAliasAnalysis();
526         MemoryLocation LoadLoc = MemoryLocation::get(LI);
527 
528         // We use alias analysis to check if an instruction may store to
529         // the memory we load from in between the load and the store. If
530         // such an instruction is found, we try to promote there instead
531         // of at the store position.
532         Instruction *P = SI;
533         for (auto &I : make_range(++LI->getIterator(), SI->getIterator())) {
534           if (isModSet(AA.getModRefInfo(&I, LoadLoc))) {
535             P = &I;
536             break;
537           }
538         }
539 
540         // We found an instruction that may write to the loaded memory.
541         // We can try to promote at this position instead of the store
542         // position if nothing alias the store memory after this and the store
543         // destination is not in the range.
544         if (P && P != SI) {
545           if (!moveUp(AA, SI, P, LI))
546             P = nullptr;
547         }
548 
549         // If a valid insertion position is found, then we can promote
550         // the load/store pair to a memcpy.
551         if (P) {
552           // If we load from memory that may alias the memory we store to,
553           // memmove must be used to preserve semantic. If not, memcpy can
554           // be used.
555           bool UseMemMove = false;
556           if (!AA.isNoAlias(MemoryLocation::get(SI), LoadLoc))
557             UseMemMove = true;
558 
559           uint64_t Size = DL.getTypeStoreSize(T);
560 
561           IRBuilder<> Builder(P);
562           Instruction *M;
563           if (UseMemMove)
564             M = Builder.CreateMemMove(
565                 SI->getPointerOperand(), SI->getAlign(),
566                 LI->getPointerOperand(), LI->getAlign(), Size);
567           else
568             M = Builder.CreateMemCpy(
569                 SI->getPointerOperand(), SI->getAlign(),
570                 LI->getPointerOperand(), LI->getAlign(), Size);
571 
572           LLVM_DEBUG(dbgs() << "Promoting " << *LI << " to " << *SI << " => "
573                             << *M << "\n");
574 
575           MD->removeInstruction(SI);
576           SI->eraseFromParent();
577           MD->removeInstruction(LI);
578           LI->eraseFromParent();
579           ++NumMemCpyInstr;
580 
581           // Make sure we do not invalidate the iterator.
582           BBI = M->getIterator();
583           return true;
584         }
585       }
586 
587       // Detect cases where we're performing call slot forwarding, but
588       // happen to be using a load-store pair to implement it, rather than
589       // a memcpy.
590       MemDepResult ldep = MD->getDependency(LI);
591       CallInst *C = nullptr;
592       if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
593         C = dyn_cast<CallInst>(ldep.getInst());
594 
595       if (C) {
596         // Check that nothing touches the dest of the "copy" between
597         // the call and the store.
598         Value *CpyDest = SI->getPointerOperand()->stripPointerCasts();
599         bool CpyDestIsLocal = isa<AllocaInst>(CpyDest);
600         AliasAnalysis &AA = LookupAliasAnalysis();
601         MemoryLocation StoreLoc = MemoryLocation::get(SI);
602         for (BasicBlock::iterator I = --SI->getIterator(), E = C->getIterator();
603              I != E; --I) {
604           if (isModOrRefSet(AA.getModRefInfo(&*I, StoreLoc))) {
605             C = nullptr;
606             break;
607           }
608           // The store to dest may never happen if an exception can be thrown
609           // between the load and the store.
610           if (I->mayThrow() && !CpyDestIsLocal) {
611             C = nullptr;
612             break;
613           }
614         }
615       }
616 
617       if (C) {
618         bool changed = performCallSlotOptzn(
619             LI, SI->getPointerOperand()->stripPointerCasts(),
620             LI->getPointerOperand()->stripPointerCasts(),
621             DL.getTypeStoreSize(SI->getOperand(0)->getType()),
622             commonAlignment(SI->getAlign(), LI->getAlign()), C);
623         if (changed) {
624           MD->removeInstruction(SI);
625           SI->eraseFromParent();
626           MD->removeInstruction(LI);
627           LI->eraseFromParent();
628           ++NumMemCpyInstr;
629           return true;
630         }
631       }
632     }
633   }
634 
635   // There are two cases that are interesting for this code to handle: memcpy
636   // and memset.  Right now we only handle memset.
637 
638   // Ensure that the value being stored is something that can be memset'able a
639   // byte at a time like "0" or "-1" or any width, as well as things like
640   // 0xA0A0A0A0 and 0.0.
641   auto *V = SI->getOperand(0);
642   if (Value *ByteVal = isBytewiseValue(V, DL)) {
643     if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
644                                               ByteVal)) {
645       BBI = I->getIterator(); // Don't invalidate iterator.
646       return true;
647     }
648 
649     // If we have an aggregate, we try to promote it to memset regardless
650     // of opportunity for merging as it can expose optimization opportunities
651     // in subsequent passes.
652     auto *T = V->getType();
653     if (T->isAggregateType()) {
654       uint64_t Size = DL.getTypeStoreSize(T);
655       IRBuilder<> Builder(SI);
656       auto *M = Builder.CreateMemSet(SI->getPointerOperand(), ByteVal, Size,
657                                      SI->getAlign());
658 
659       LLVM_DEBUG(dbgs() << "Promoting " << *SI << " to " << *M << "\n");
660 
661       MD->removeInstruction(SI);
662       SI->eraseFromParent();
663       NumMemSetInfer++;
664 
665       // Make sure we do not invalidate the iterator.
666       BBI = M->getIterator();
667       return true;
668     }
669   }
670 
671   return false;
672 }
673 
674 bool MemCpyOptPass::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
675   // See if there is another memset or store neighboring this memset which
676   // allows us to widen out the memset to do a single larger store.
677   if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
678     if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
679                                               MSI->getValue())) {
680       BBI = I->getIterator(); // Don't invalidate iterator.
681       return true;
682     }
683   return false;
684 }
685 
686 /// Takes a memcpy and a call that it depends on,
687 /// and checks for the possibility of a call slot optimization by having
688 /// the call write its result directly into the destination of the memcpy.
689 bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpy, Value *cpyDest,
690                                          Value *cpySrc, uint64_t cpyLen,
691                                          Align cpyAlign, CallInst *C) {
692   // The general transformation to keep in mind is
693   //
694   //   call @func(..., src, ...)
695   //   memcpy(dest, src, ...)
696   //
697   // ->
698   //
699   //   memcpy(dest, src, ...)
700   //   call @func(..., dest, ...)
701   //
702   // Since moving the memcpy is technically awkward, we additionally check that
703   // src only holds uninitialized values at the moment of the call, meaning that
704   // the memcpy can be discarded rather than moved.
705 
706   // Lifetime marks shouldn't be operated on.
707   if (Function *F = C->getCalledFunction())
708     if (F->isIntrinsic() && F->getIntrinsicID() == Intrinsic::lifetime_start)
709       return false;
710 
711   // Require that src be an alloca.  This simplifies the reasoning considerably.
712   AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
713   if (!srcAlloca)
714     return false;
715 
716   ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
717   if (!srcArraySize)
718     return false;
719 
720   const DataLayout &DL = cpy->getModule()->getDataLayout();
721   uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
722                      srcArraySize->getZExtValue();
723 
724   if (cpyLen < srcSize)
725     return false;
726 
727   // Check that accessing the first srcSize bytes of dest will not cause a
728   // trap.  Otherwise the transform is invalid since it might cause a trap
729   // to occur earlier than it otherwise would.
730   if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
731     // The destination is an alloca.  Check it is larger than srcSize.
732     ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
733     if (!destArraySize)
734       return false;
735 
736     uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) *
737                         destArraySize->getZExtValue();
738 
739     if (destSize < srcSize)
740       return false;
741   } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
742     // The store to dest may never happen if the call can throw.
743     if (C->mayThrow())
744       return false;
745 
746     if (A->getDereferenceableBytes() < srcSize) {
747       // If the destination is an sret parameter then only accesses that are
748       // outside of the returned struct type can trap.
749       if (!A->hasStructRetAttr())
750         return false;
751 
752       Type *StructTy = cast<PointerType>(A->getType())->getElementType();
753       if (!StructTy->isSized()) {
754         // The call may never return and hence the copy-instruction may never
755         // be executed, and therefore it's not safe to say "the destination
756         // has at least <cpyLen> bytes, as implied by the copy-instruction",
757         return false;
758       }
759 
760       uint64_t destSize = DL.getTypeAllocSize(StructTy);
761       if (destSize < srcSize)
762         return false;
763     }
764   } else {
765     return false;
766   }
767 
768   // Check that dest points to memory that is at least as aligned as src.
769   Align srcAlign = srcAlloca->getAlign();
770   bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
771   // If dest is not aligned enough and we can't increase its alignment then
772   // bail out.
773   if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
774     return false;
775 
776   // Check that src is not accessed except via the call and the memcpy.  This
777   // guarantees that it holds only undefined values when passed in (so the final
778   // memcpy can be dropped), that it is not read or written between the call and
779   // the memcpy, and that writing beyond the end of it is undefined.
780   SmallVector<User*, 8> srcUseList(srcAlloca->user_begin(),
781                                    srcAlloca->user_end());
782   while (!srcUseList.empty()) {
783     User *U = srcUseList.pop_back_val();
784 
785     if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
786       for (User *UU : U->users())
787         srcUseList.push_back(UU);
788       continue;
789     }
790     if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) {
791       if (!G->hasAllZeroIndices())
792         return false;
793 
794       for (User *UU : U->users())
795         srcUseList.push_back(UU);
796       continue;
797     }
798     if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U))
799       if (IT->isLifetimeStartOrEnd())
800         continue;
801 
802     if (U != C && U != cpy)
803       return false;
804   }
805 
806   // Check that src isn't captured by the called function since the
807   // transformation can cause aliasing issues in that case.
808   for (unsigned ArgI = 0, E = C->arg_size(); ArgI != E; ++ArgI)
809     if (C->getArgOperand(ArgI) == cpySrc && !C->doesNotCapture(ArgI))
810       return false;
811 
812   // Since we're changing the parameter to the callsite, we need to make sure
813   // that what would be the new parameter dominates the callsite.
814   DominatorTree &DT = LookupDomTree();
815   if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
816     if (!DT.dominates(cpyDestInst, C))
817       return false;
818 
819   // In addition to knowing that the call does not access src in some
820   // unexpected manner, for example via a global, which we deduce from
821   // the use analysis, we also need to know that it does not sneakily
822   // access dest.  We rely on AA to figure this out for us.
823   AliasAnalysis &AA = LookupAliasAnalysis();
824   ModRefInfo MR = AA.getModRefInfo(C, cpyDest, LocationSize::precise(srcSize));
825   // If necessary, perform additional analysis.
826   if (isModOrRefSet(MR))
827     MR = AA.callCapturesBefore(C, cpyDest, LocationSize::precise(srcSize), &DT);
828   if (isModOrRefSet(MR))
829     return false;
830 
831   // We can't create address space casts here because we don't know if they're
832   // safe for the target.
833   if (cpySrc->getType()->getPointerAddressSpace() !=
834       cpyDest->getType()->getPointerAddressSpace())
835     return false;
836   for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
837     if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc &&
838         cpySrc->getType()->getPointerAddressSpace() !=
839             C->getArgOperand(ArgI)->getType()->getPointerAddressSpace())
840       return false;
841 
842   // All the checks have passed, so do the transformation.
843   bool changedArgument = false;
844   for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
845     if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc) {
846       Value *Dest = cpySrc->getType() == cpyDest->getType() ?  cpyDest
847         : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
848                                       cpyDest->getName(), C);
849       changedArgument = true;
850       if (C->getArgOperand(ArgI)->getType() == Dest->getType())
851         C->setArgOperand(ArgI, Dest);
852       else
853         C->setArgOperand(ArgI, CastInst::CreatePointerCast(
854                                    Dest, C->getArgOperand(ArgI)->getType(),
855                                    Dest->getName(), C));
856     }
857 
858   if (!changedArgument)
859     return false;
860 
861   // If the destination wasn't sufficiently aligned then increase its alignment.
862   if (!isDestSufficientlyAligned) {
863     assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
864     cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
865   }
866 
867   // Drop any cached information about the call, because we may have changed
868   // its dependence information by changing its parameter.
869   MD->removeInstruction(C);
870 
871   // Update AA metadata
872   // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
873   // handled here, but combineMetadata doesn't support them yet
874   unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
875                          LLVMContext::MD_noalias,
876                          LLVMContext::MD_invariant_group,
877                          LLVMContext::MD_access_group};
878   combineMetadata(C, cpy, KnownIDs, true);
879 
880   // Remove the memcpy.
881   MD->removeInstruction(cpy);
882   ++NumMemCpyInstr;
883 
884   return true;
885 }
886 
887 /// We've found that the (upward scanning) memory dependence of memcpy 'M' is
888 /// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
889 bool MemCpyOptPass::processMemCpyMemCpyDependence(MemCpyInst *M,
890                                                   MemCpyInst *MDep) {
891   // We can only transforms memcpy's where the dest of one is the source of the
892   // other.
893   if (M->getSource() != MDep->getDest() || MDep->isVolatile())
894     return false;
895 
896   // If dep instruction is reading from our current input, then it is a noop
897   // transfer and substituting the input won't change this instruction.  Just
898   // ignore the input and let someone else zap MDep.  This handles cases like:
899   //    memcpy(a <- a)
900   //    memcpy(b <- a)
901   if (M->getSource() == MDep->getSource())
902     return false;
903 
904   // Second, the length of the memcpy's must be the same, or the preceding one
905   // must be larger than the following one.
906   ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
907   ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
908   if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
909     return false;
910 
911   AliasAnalysis &AA = LookupAliasAnalysis();
912 
913   // Verify that the copied-from memory doesn't change in between the two
914   // transfers.  For example, in:
915   //    memcpy(a <- b)
916   //    *b = 42;
917   //    memcpy(c <- a)
918   // It would be invalid to transform the second memcpy into memcpy(c <- b).
919   //
920   // TODO: If the code between M and MDep is transparent to the destination "c",
921   // then we could still perform the xform by moving M up to the first memcpy.
922   //
923   // NOTE: This is conservative, it will stop on any read from the source loc,
924   // not just the defining memcpy.
925   MemDepResult SourceDep =
926       MD->getPointerDependencyFrom(MemoryLocation::getForSource(MDep), false,
927                                    M->getIterator(), M->getParent());
928   if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
929     return false;
930 
931   // If the dest of the second might alias the source of the first, then the
932   // source and dest might overlap.  We still want to eliminate the intermediate
933   // value, but we have to generate a memmove instead of memcpy.
934   bool UseMemMove = false;
935   if (!AA.isNoAlias(MemoryLocation::getForDest(M),
936                     MemoryLocation::getForSource(MDep)))
937     UseMemMove = true;
938 
939   // If all checks passed, then we can transform M.
940   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
941                     << *MDep << '\n' << *M << '\n');
942 
943   // TODO: Is this worth it if we're creating a less aligned memcpy? For
944   // example we could be moving from movaps -> movq on x86.
945   IRBuilder<> Builder(M);
946   if (UseMemMove)
947     Builder.CreateMemMove(M->getRawDest(), M->getDestAlign(),
948                           MDep->getRawSource(), MDep->getSourceAlign(),
949                           M->getLength(), M->isVolatile());
950   else
951     Builder.CreateMemCpy(M->getRawDest(), M->getDestAlign(),
952                          MDep->getRawSource(), MDep->getSourceAlign(),
953                          M->getLength(), M->isVolatile());
954 
955   // Remove the instruction we're replacing.
956   MD->removeInstruction(M);
957   M->eraseFromParent();
958   ++NumMemCpyInstr;
959   return true;
960 }
961 
962 /// We've found that the (upward scanning) memory dependence of \p MemCpy is
963 /// \p MemSet.  Try to simplify \p MemSet to only set the trailing bytes that
964 /// weren't copied over by \p MemCpy.
965 ///
966 /// In other words, transform:
967 /// \code
968 ///   memset(dst, c, dst_size);
969 ///   memcpy(dst, src, src_size);
970 /// \endcode
971 /// into:
972 /// \code
973 ///   memcpy(dst, src, src_size);
974 ///   memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
975 /// \endcode
976 bool MemCpyOptPass::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
977                                                   MemSetInst *MemSet) {
978   // We can only transform memset/memcpy with the same destination.
979   if (MemSet->getDest() != MemCpy->getDest())
980     return false;
981 
982   // Check that there are no other dependencies on the memset destination.
983   MemDepResult DstDepInfo =
984       MD->getPointerDependencyFrom(MemoryLocation::getForDest(MemSet), false,
985                                    MemCpy->getIterator(), MemCpy->getParent());
986   if (DstDepInfo.getInst() != MemSet)
987     return false;
988 
989   // Use the same i8* dest as the memcpy, killing the memset dest if different.
990   Value *Dest = MemCpy->getRawDest();
991   Value *DestSize = MemSet->getLength();
992   Value *SrcSize = MemCpy->getLength();
993 
994   // By default, create an unaligned memset.
995   unsigned Align = 1;
996   // If Dest is aligned, and SrcSize is constant, use the minimum alignment
997   // of the sum.
998   const unsigned DestAlign =
999       std::max(MemSet->getDestAlignment(), MemCpy->getDestAlignment());
1000   if (DestAlign > 1)
1001     if (ConstantInt *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
1002       Align = MinAlign(SrcSizeC->getZExtValue(), DestAlign);
1003 
1004   IRBuilder<> Builder(MemCpy);
1005 
1006   // If the sizes have different types, zext the smaller one.
1007   if (DestSize->getType() != SrcSize->getType()) {
1008     if (DestSize->getType()->getIntegerBitWidth() >
1009         SrcSize->getType()->getIntegerBitWidth())
1010       SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
1011     else
1012       DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
1013   }
1014 
1015   Value *Ule = Builder.CreateICmpULE(DestSize, SrcSize);
1016   Value *SizeDiff = Builder.CreateSub(DestSize, SrcSize);
1017   Value *MemsetLen = Builder.CreateSelect(
1018       Ule, ConstantInt::getNullValue(DestSize->getType()), SizeDiff);
1019   Builder.CreateMemSet(
1020       Builder.CreateGEP(Dest->getType()->getPointerElementType(), Dest,
1021                         SrcSize),
1022       MemSet->getOperand(1), MemsetLen, MaybeAlign(Align));
1023 
1024   MD->removeInstruction(MemSet);
1025   MemSet->eraseFromParent();
1026   return true;
1027 }
1028 
1029 /// Determine whether the instruction has undefined content for the given Size,
1030 /// either because it was freshly alloca'd or started its lifetime.
1031 static bool hasUndefContents(Instruction *I, ConstantInt *Size) {
1032   if (isa<AllocaInst>(I))
1033     return true;
1034 
1035   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1036     if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1037       if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
1038         if (LTSize->getZExtValue() >= Size->getZExtValue())
1039           return true;
1040 
1041   return false;
1042 }
1043 
1044 /// Transform memcpy to memset when its source was just memset.
1045 /// In other words, turn:
1046 /// \code
1047 ///   memset(dst1, c, dst1_size);
1048 ///   memcpy(dst2, dst1, dst2_size);
1049 /// \endcode
1050 /// into:
1051 /// \code
1052 ///   memset(dst1, c, dst1_size);
1053 ///   memset(dst2, c, dst2_size);
1054 /// \endcode
1055 /// When dst2_size <= dst1_size.
1056 ///
1057 /// The \p MemCpy must have a Constant length.
1058 bool MemCpyOptPass::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
1059                                                MemSetInst *MemSet) {
1060   AliasAnalysis &AA = LookupAliasAnalysis();
1061 
1062   // Make sure that memcpy(..., memset(...), ...), that is we are memsetting and
1063   // memcpying from the same address. Otherwise it is hard to reason about.
1064   if (!AA.isMustAlias(MemSet->getRawDest(), MemCpy->getRawSource()))
1065     return false;
1066 
1067   // A known memset size is required.
1068   ConstantInt *MemSetSize = dyn_cast<ConstantInt>(MemSet->getLength());
1069   if (!MemSetSize)
1070     return false;
1071 
1072   // Make sure the memcpy doesn't read any more than what the memset wrote.
1073   // Don't worry about sizes larger than i64.
1074   ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());
1075   if (CopySize->getZExtValue() > MemSetSize->getZExtValue()) {
1076     // If the memcpy is larger than the memset, but the memory was undef prior
1077     // to the memset, we can just ignore the tail. Technically we're only
1078     // interested in the bytes from MemSetSize..CopySize here, but as we can't
1079     // easily represent this location, we use the full 0..CopySize range.
1080     MemoryLocation MemCpyLoc = MemoryLocation::getForSource(MemCpy);
1081     MemDepResult DepInfo = MD->getPointerDependencyFrom(
1082         MemCpyLoc, true, MemSet->getIterator(), MemSet->getParent());
1083     if (DepInfo.isDef() && hasUndefContents(DepInfo.getInst(), CopySize))
1084       CopySize = MemSetSize;
1085     else
1086       return false;
1087   }
1088 
1089   IRBuilder<> Builder(MemCpy);
1090   Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1), CopySize,
1091                        MaybeAlign(MemCpy->getDestAlignment()));
1092   return true;
1093 }
1094 
1095 /// Perform simplification of memcpy's.  If we have memcpy A
1096 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
1097 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
1098 /// circumstances). This allows later passes to remove the first memcpy
1099 /// altogether.
1100 bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
1101   // We can only optimize non-volatile memcpy's.
1102   if (M->isVolatile()) return false;
1103 
1104   // If the source and destination of the memcpy are the same, then zap it.
1105   if (M->getSource() == M->getDest()) {
1106     ++BBI;
1107     MD->removeInstruction(M);
1108     M->eraseFromParent();
1109     return true;
1110   }
1111 
1112   // If copying from a constant, try to turn the memcpy into a memset.
1113   if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
1114     if (GV->isConstant() && GV->hasDefinitiveInitializer())
1115       if (Value *ByteVal = isBytewiseValue(GV->getInitializer(),
1116                                            M->getModule()->getDataLayout())) {
1117         IRBuilder<> Builder(M);
1118         Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
1119                              MaybeAlign(M->getDestAlignment()), false);
1120         MD->removeInstruction(M);
1121         M->eraseFromParent();
1122         ++NumCpyToSet;
1123         return true;
1124       }
1125 
1126   MemDepResult DepInfo = MD->getDependency(M);
1127 
1128   // Try to turn a partially redundant memset + memcpy into
1129   // memcpy + smaller memset.  We don't need the memcpy size for this.
1130   if (DepInfo.isClobber())
1131     if (MemSetInst *MDep = dyn_cast<MemSetInst>(DepInfo.getInst()))
1132       if (processMemSetMemCpyDependence(M, MDep))
1133         return true;
1134 
1135   // The optimizations after this point require the memcpy size.
1136   ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
1137   if (!CopySize) return false;
1138 
1139   // There are four possible optimizations we can do for memcpy:
1140   //   a) memcpy-memcpy xform which exposes redundance for DSE.
1141   //   b) call-memcpy xform for return slot optimization.
1142   //   c) memcpy from freshly alloca'd space or space that has just started its
1143   //      lifetime copies undefined data, and we can therefore eliminate the
1144   //      memcpy in favor of the data that was already at the destination.
1145   //   d) memcpy from a just-memset'd source can be turned into memset.
1146   if (DepInfo.isClobber()) {
1147     if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
1148       // FIXME: Can we pass in either of dest/src alignment here instead
1149       // of conservatively taking the minimum?
1150       Align Alignment = std::min(M->getDestAlign().valueOrOne(),
1151                                  M->getSourceAlign().valueOrOne());
1152       if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
1153                                CopySize->getZExtValue(), Alignment, C)) {
1154         MD->removeInstruction(M);
1155         M->eraseFromParent();
1156         return true;
1157       }
1158     }
1159   }
1160 
1161   MemoryLocation SrcLoc = MemoryLocation::getForSource(M);
1162   MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(
1163       SrcLoc, true, M->getIterator(), M->getParent());
1164 
1165   if (SrcDepInfo.isClobber()) {
1166     if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
1167       return processMemCpyMemCpyDependence(M, MDep);
1168   } else if (SrcDepInfo.isDef()) {
1169     if (hasUndefContents(SrcDepInfo.getInst(), CopySize)) {
1170       MD->removeInstruction(M);
1171       M->eraseFromParent();
1172       ++NumMemCpyInstr;
1173       return true;
1174     }
1175   }
1176 
1177   if (SrcDepInfo.isClobber())
1178     if (MemSetInst *MDep = dyn_cast<MemSetInst>(SrcDepInfo.getInst()))
1179       if (performMemCpyToMemSetOptzn(M, MDep)) {
1180         MD->removeInstruction(M);
1181         M->eraseFromParent();
1182         ++NumCpyToSet;
1183         return true;
1184       }
1185 
1186   return false;
1187 }
1188 
1189 /// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1190 /// not to alias.
1191 bool MemCpyOptPass::processMemMove(MemMoveInst *M) {
1192   AliasAnalysis &AA = LookupAliasAnalysis();
1193 
1194   if (!TLI->has(LibFunc_memmove))
1195     return false;
1196 
1197   // See if the pointers alias.
1198   if (!AA.isNoAlias(MemoryLocation::getForDest(M),
1199                     MemoryLocation::getForSource(M)))
1200     return false;
1201 
1202   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: " << *M
1203                     << "\n");
1204 
1205   // If not, then we know we can transform this.
1206   Type *ArgTys[3] = { M->getRawDest()->getType(),
1207                       M->getRawSource()->getType(),
1208                       M->getLength()->getType() };
1209   M->setCalledFunction(Intrinsic::getDeclaration(M->getModule(),
1210                                                  Intrinsic::memcpy, ArgTys));
1211 
1212   // MemDep may have over conservative information about this instruction, just
1213   // conservatively flush it from the cache.
1214   MD->removeInstruction(M);
1215 
1216   ++NumMoveToCpy;
1217   return true;
1218 }
1219 
1220 /// This is called on every byval argument in call sites.
1221 bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
1222   const DataLayout &DL = CB.getCaller()->getParent()->getDataLayout();
1223   // Find out what feeds this byval argument.
1224   Value *ByValArg = CB.getArgOperand(ArgNo);
1225   Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
1226   uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
1227   MemDepResult DepInfo = MD->getPointerDependencyFrom(
1228       MemoryLocation(ByValArg, LocationSize::precise(ByValSize)), true,
1229       CB.getIterator(), CB.getParent());
1230   if (!DepInfo.isClobber())
1231     return false;
1232 
1233   // If the byval argument isn't fed by a memcpy, ignore it.  If it is fed by
1234   // a memcpy, see if we can byval from the source of the memcpy instead of the
1235   // result.
1236   MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
1237   if (!MDep || MDep->isVolatile() ||
1238       ByValArg->stripPointerCasts() != MDep->getDest())
1239     return false;
1240 
1241   // The length of the memcpy must be larger or equal to the size of the byval.
1242   ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1243   if (!C1 || C1->getValue().getZExtValue() < ByValSize)
1244     return false;
1245 
1246   // Get the alignment of the byval.  If the call doesn't specify the alignment,
1247   // then it is some target specific value that we can't know.
1248   MaybeAlign ByValAlign = CB.getParamAlign(ArgNo);
1249   if (!ByValAlign) return false;
1250 
1251   // If it is greater than the memcpy, then we check to see if we can force the
1252   // source of the memcpy to the alignment we need.  If we fail, we bail out.
1253   AssumptionCache &AC = LookupAssumptionCache();
1254   DominatorTree &DT = LookupDomTree();
1255   MaybeAlign MemDepAlign = MDep->getSourceAlign();
1256   if ((!MemDepAlign || *MemDepAlign < *ByValAlign) &&
1257       getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL, &CB, &AC,
1258                                  &DT) < *ByValAlign)
1259     return false;
1260 
1261   // The address space of the memcpy source must match the byval argument
1262   if (MDep->getSource()->getType()->getPointerAddressSpace() !=
1263       ByValArg->getType()->getPointerAddressSpace())
1264     return false;
1265 
1266   // Verify that the copied-from memory doesn't change in between the memcpy and
1267   // the byval call.
1268   //    memcpy(a <- b)
1269   //    *b = 42;
1270   //    foo(*a)
1271   // It would be invalid to transform the second memcpy into foo(*b).
1272   //
1273   // NOTE: This is conservative, it will stop on any read from the source loc,
1274   // not just the defining memcpy.
1275   MemDepResult SourceDep = MD->getPointerDependencyFrom(
1276       MemoryLocation::getForSource(MDep), false,
1277       CB.getIterator(), MDep->getParent());
1278   if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
1279     return false;
1280 
1281   Value *TmpCast = MDep->getSource();
1282   if (MDep->getSource()->getType() != ByValArg->getType()) {
1283     BitCastInst *TmpBitCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
1284                                               "tmpcast", &CB);
1285     // Set the tmpcast's DebugLoc to MDep's
1286     TmpBitCast->setDebugLoc(MDep->getDebugLoc());
1287     TmpCast = TmpBitCast;
1288   }
1289 
1290   LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
1291                     << "  " << *MDep << "\n"
1292                     << "  " << CB << "\n");
1293 
1294   // Otherwise we're good!  Update the byval argument.
1295   CB.setArgOperand(ArgNo, TmpCast);
1296   ++NumMemCpyInstr;
1297   return true;
1298 }
1299 
1300 /// Executes one iteration of MemCpyOptPass.
1301 bool MemCpyOptPass::iterateOnFunction(Function &F) {
1302   bool MadeChange = false;
1303 
1304   DominatorTree &DT = LookupDomTree();
1305 
1306   // Walk all instruction in the function.
1307   for (BasicBlock &BB : F) {
1308     // Skip unreachable blocks. For example processStore assumes that an
1309     // instruction in a BB can't be dominated by a later instruction in the
1310     // same BB (which is a scenario that can happen for an unreachable BB that
1311     // has itself as a predecessor).
1312     if (!DT.isReachableFromEntry(&BB))
1313       continue;
1314 
1315     for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
1316         // Avoid invalidating the iterator.
1317       Instruction *I = &*BI++;
1318 
1319       bool RepeatInstruction = false;
1320 
1321       if (StoreInst *SI = dyn_cast<StoreInst>(I))
1322         MadeChange |= processStore(SI, BI);
1323       else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
1324         RepeatInstruction = processMemSet(M, BI);
1325       else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
1326         RepeatInstruction = processMemCpy(M, BI);
1327       else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
1328         RepeatInstruction = processMemMove(M);
1329       else if (auto *CB = dyn_cast<CallBase>(I)) {
1330         for (unsigned i = 0, e = CB->arg_size(); i != e; ++i)
1331           if (CB->isByValArgument(i))
1332             MadeChange |= processByValArgument(*CB, i);
1333       }
1334 
1335       // Reprocess the instruction if desired.
1336       if (RepeatInstruction) {
1337         if (BI != BB.begin())
1338           --BI;
1339         MadeChange = true;
1340       }
1341     }
1342   }
1343 
1344   return MadeChange;
1345 }
1346 
1347 PreservedAnalyses MemCpyOptPass::run(Function &F, FunctionAnalysisManager &AM) {
1348   auto &MD = AM.getResult<MemoryDependenceAnalysis>(F);
1349   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1350 
1351   auto LookupAliasAnalysis = [&]() -> AliasAnalysis & {
1352     return AM.getResult<AAManager>(F);
1353   };
1354   auto LookupAssumptionCache = [&]() -> AssumptionCache & {
1355     return AM.getResult<AssumptionAnalysis>(F);
1356   };
1357   auto LookupDomTree = [&]() -> DominatorTree & {
1358     return AM.getResult<DominatorTreeAnalysis>(F);
1359   };
1360 
1361   bool MadeChange = runImpl(F, &MD, &TLI, LookupAliasAnalysis,
1362                             LookupAssumptionCache, LookupDomTree);
1363   if (!MadeChange)
1364     return PreservedAnalyses::all();
1365 
1366   PreservedAnalyses PA;
1367   PA.preserveSet<CFGAnalyses>();
1368   PA.preserve<GlobalsAA>();
1369   PA.preserve<MemoryDependenceAnalysis>();
1370   return PA;
1371 }
1372 
1373 bool MemCpyOptPass::runImpl(
1374     Function &F, MemoryDependenceResults *MD_, TargetLibraryInfo *TLI_,
1375     std::function<AliasAnalysis &()> LookupAliasAnalysis_,
1376     std::function<AssumptionCache &()> LookupAssumptionCache_,
1377     std::function<DominatorTree &()> LookupDomTree_) {
1378   bool MadeChange = false;
1379   MD = MD_;
1380   TLI = TLI_;
1381   LookupAliasAnalysis = std::move(LookupAliasAnalysis_);
1382   LookupAssumptionCache = std::move(LookupAssumptionCache_);
1383   LookupDomTree = std::move(LookupDomTree_);
1384 
1385   // If we don't have at least memset and memcpy, there is little point of doing
1386   // anything here.  These are required by a freestanding implementation, so if
1387   // even they are disabled, there is no point in trying hard.
1388   if (!TLI->has(LibFunc_memset) || !TLI->has(LibFunc_memcpy))
1389     return false;
1390 
1391   while (true) {
1392     if (!iterateOnFunction(F))
1393       break;
1394     MadeChange = true;
1395   }
1396 
1397   MD = nullptr;
1398   return MadeChange;
1399 }
1400 
1401 /// This is the main transformation entry point for a function.
1402 bool MemCpyOptLegacyPass::runOnFunction(Function &F) {
1403   if (skipFunction(F))
1404     return false;
1405 
1406   auto *MD = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
1407   auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1408 
1409   auto LookupAliasAnalysis = [this]() -> AliasAnalysis & {
1410     return getAnalysis<AAResultsWrapperPass>().getAAResults();
1411   };
1412   auto LookupAssumptionCache = [this, &F]() -> AssumptionCache & {
1413     return getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1414   };
1415   auto LookupDomTree = [this]() -> DominatorTree & {
1416     return getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1417   };
1418 
1419   return Impl.runImpl(F, MD, TLI, LookupAliasAnalysis, LookupAssumptionCache,
1420                       LookupDomTree);
1421 }
1422