1 //===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass performs various transformations related to eliminating memcpy
11 // calls, or transforming sets of stores into memset's.
12 //
13 //===----------------------------------------------------------------------===//
14
15 #include "llvm/Transforms/Scalar.h"
16 #include "llvm/ADT/SmallVector.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/GlobalsModRef.h"
21 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
22 #include "llvm/Analysis/TargetLibraryInfo.h"
23 #include "llvm/Analysis/ValueTracking.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/GetElementPtrTypeIterator.h"
27 #include "llvm/IR/GlobalVariable.h"
28 #include "llvm/IR/IRBuilder.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/raw_ostream.h"
33 #include "llvm/Transforms/Utils/Local.h"
34 #include <algorithm>
35 using namespace llvm;
36
37 #define DEBUG_TYPE "memcpyopt"
38
39 STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
40 STATISTIC(NumMemSetInfer, "Number of memsets inferred");
41 STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
42 STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
43
GetOffsetFromIndex(const GEPOperator * GEP,unsigned Idx,bool & VariableIdxFound,const DataLayout & DL)44 static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx,
45 bool &VariableIdxFound,
46 const DataLayout &DL) {
47 // Skip over the first indices.
48 gep_type_iterator GTI = gep_type_begin(GEP);
49 for (unsigned i = 1; i != Idx; ++i, ++GTI)
50 /*skip along*/;
51
52 // Compute the offset implied by the rest of the indices.
53 int64_t Offset = 0;
54 for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
55 ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
56 if (!OpC)
57 return VariableIdxFound = true;
58 if (OpC->isZero()) continue; // No offset.
59
60 // Handle struct indices, which add their field offset to the pointer.
61 if (StructType *STy = dyn_cast<StructType>(*GTI)) {
62 Offset += DL.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
63 continue;
64 }
65
66 // Otherwise, we have a sequential type like an array or vector. Multiply
67 // the index by the ElementSize.
68 uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType());
69 Offset += Size*OpC->getSExtValue();
70 }
71
72 return Offset;
73 }
74
75 /// Return true if Ptr1 is provably equal to Ptr2 plus a constant offset, and
76 /// return that constant offset. For example, Ptr1 might be &A[42], and Ptr2
77 /// might be &A[40]. In this case offset would be -8.
IsPointerOffset(Value * Ptr1,Value * Ptr2,int64_t & Offset,const DataLayout & DL)78 static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
79 const DataLayout &DL) {
80 Ptr1 = Ptr1->stripPointerCasts();
81 Ptr2 = Ptr2->stripPointerCasts();
82
83 // Handle the trivial case first.
84 if (Ptr1 == Ptr2) {
85 Offset = 0;
86 return true;
87 }
88
89 GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
90 GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);
91
92 bool VariableIdxFound = false;
93
94 // If one pointer is a GEP and the other isn't, then see if the GEP is a
95 // constant offset from the base, as in "P" and "gep P, 1".
96 if (GEP1 && !GEP2 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
97 Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, DL);
98 return !VariableIdxFound;
99 }
100
101 if (GEP2 && !GEP1 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
102 Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, DL);
103 return !VariableIdxFound;
104 }
105
106 // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
107 // base. After that base, they may have some number of common (and
108 // potentially variable) indices. After that they handle some constant
109 // offset, which determines their offset from each other. At this point, we
110 // handle no other case.
111 if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
112 return false;
113
114 // Skip any common indices and track the GEP types.
115 unsigned Idx = 1;
116 for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
117 if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
118 break;
119
120 int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, DL);
121 int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, DL);
122 if (VariableIdxFound) return false;
123
124 Offset = Offset2-Offset1;
125 return true;
126 }
127
128
129 /// Represents a range of memset'd bytes with the ByteVal value.
130 /// This allows us to analyze stores like:
131 /// store 0 -> P+1
132 /// store 0 -> P+0
133 /// store 0 -> P+3
134 /// store 0 -> P+2
135 /// which sometimes happens with stores to arrays of structs etc. When we see
136 /// the first store, we make a range [1, 2). The second store extends the range
137 /// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
138 /// two ranges into [0, 3) which is memset'able.
139 namespace {
140 struct MemsetRange {
141 // Start/End - A semi range that describes the span that this range covers.
142 // The range is closed at the start and open at the end: [Start, End).
143 int64_t Start, End;
144
145 /// StartPtr - The getelementptr instruction that points to the start of the
146 /// range.
147 Value *StartPtr;
148
149 /// Alignment - The known alignment of the first store.
150 unsigned Alignment;
151
152 /// TheStores - The actual stores that make up this range.
153 SmallVector<Instruction*, 16> TheStores;
154
155 bool isProfitableToUseMemset(const DataLayout &DL) const;
156 };
157 } // end anon namespace
158
isProfitableToUseMemset(const DataLayout & DL) const159 bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
160 // If we found more than 4 stores to merge or 16 bytes, use memset.
161 if (TheStores.size() >= 4 || End-Start >= 16) return true;
162
163 // If there is nothing to merge, don't do anything.
164 if (TheStores.size() < 2) return false;
165
166 // If any of the stores are a memset, then it is always good to extend the
167 // memset.
168 for (Instruction *SI : TheStores)
169 if (!isa<StoreInst>(SI))
170 return true;
171
172 // Assume that the code generator is capable of merging pairs of stores
173 // together if it wants to.
174 if (TheStores.size() == 2) return false;
175
176 // If we have fewer than 8 stores, it can still be worthwhile to do this.
177 // For example, merging 4 i8 stores into an i32 store is useful almost always.
178 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
179 // memset will be split into 2 32-bit stores anyway) and doing so can
180 // pessimize the llvm optimizer.
181 //
182 // Since we don't have perfect knowledge here, make some assumptions: assume
183 // the maximum GPR width is the same size as the largest legal integer
184 // size. If so, check to see whether we will end up actually reducing the
185 // number of stores used.
186 unsigned Bytes = unsigned(End-Start);
187 unsigned MaxIntSize = DL.getLargestLegalIntTypeSize();
188 if (MaxIntSize == 0)
189 MaxIntSize = 1;
190 unsigned NumPointerStores = Bytes / MaxIntSize;
191
192 // Assume the remaining bytes if any are done a byte at a time.
193 unsigned NumByteStores = Bytes % MaxIntSize;
194
195 // If we will reduce the # stores (according to this heuristic), do the
196 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
197 // etc.
198 return TheStores.size() > NumPointerStores+NumByteStores;
199 }
200
201
202 namespace {
203 class MemsetRanges {
204 /// A sorted list of the memset ranges.
205 SmallVector<MemsetRange, 8> Ranges;
206 typedef SmallVectorImpl<MemsetRange>::iterator range_iterator;
207 const DataLayout &DL;
208 public:
MemsetRanges(const DataLayout & DL)209 MemsetRanges(const DataLayout &DL) : DL(DL) {}
210
211 typedef SmallVectorImpl<MemsetRange>::const_iterator const_iterator;
begin() const212 const_iterator begin() const { return Ranges.begin(); }
end() const213 const_iterator end() const { return Ranges.end(); }
empty() const214 bool empty() const { return Ranges.empty(); }
215
addInst(int64_t OffsetFromFirst,Instruction * Inst)216 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
217 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
218 addStore(OffsetFromFirst, SI);
219 else
220 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
221 }
222
addStore(int64_t OffsetFromFirst,StoreInst * SI)223 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
224 int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
225
226 addRange(OffsetFromFirst, StoreSize,
227 SI->getPointerOperand(), SI->getAlignment(), SI);
228 }
229
addMemSet(int64_t OffsetFromFirst,MemSetInst * MSI)230 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
231 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
232 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getAlignment(), MSI);
233 }
234
235 void addRange(int64_t Start, int64_t Size, Value *Ptr,
236 unsigned Alignment, Instruction *Inst);
237
238 };
239
240 } // end anon namespace
241
242
243 /// Add a new store to the MemsetRanges data structure. This adds a
244 /// new range for the specified store at the specified offset, merging into
245 /// existing ranges as appropriate.
addRange(int64_t Start,int64_t Size,Value * Ptr,unsigned Alignment,Instruction * Inst)246 void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
247 unsigned Alignment, Instruction *Inst) {
248 int64_t End = Start+Size;
249
250 range_iterator I = std::lower_bound(Ranges.begin(), Ranges.end(), Start,
251 [](const MemsetRange &LHS, int64_t RHS) { return LHS.End < RHS; });
252
253 // We now know that I == E, in which case we didn't find anything to merge
254 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
255 // to insert a new range. Handle this now.
256 if (I == Ranges.end() || End < I->Start) {
257 MemsetRange &R = *Ranges.insert(I, MemsetRange());
258 R.Start = Start;
259 R.End = End;
260 R.StartPtr = Ptr;
261 R.Alignment = Alignment;
262 R.TheStores.push_back(Inst);
263 return;
264 }
265
266 // This store overlaps with I, add it.
267 I->TheStores.push_back(Inst);
268
269 // At this point, we may have an interval that completely contains our store.
270 // If so, just add it to the interval and return.
271 if (I->Start <= Start && I->End >= End)
272 return;
273
274 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
275 // but is not entirely contained within the range.
276
277 // See if the range extends the start of the range. In this case, it couldn't
278 // possibly cause it to join the prior range, because otherwise we would have
279 // stopped on *it*.
280 if (Start < I->Start) {
281 I->Start = Start;
282 I->StartPtr = Ptr;
283 I->Alignment = Alignment;
284 }
285
286 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
287 // is in or right at the end of I), and that End >= I->Start. Extend I out to
288 // End.
289 if (End > I->End) {
290 I->End = End;
291 range_iterator NextI = I;
292 while (++NextI != Ranges.end() && End >= NextI->Start) {
293 // Merge the range in.
294 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
295 if (NextI->End > I->End)
296 I->End = NextI->End;
297 Ranges.erase(NextI);
298 NextI = I;
299 }
300 }
301 }
302
303 //===----------------------------------------------------------------------===//
304 // MemCpyOpt Pass
305 //===----------------------------------------------------------------------===//
306
307 namespace {
308 class MemCpyOpt : public FunctionPass {
309 MemoryDependenceAnalysis *MD;
310 TargetLibraryInfo *TLI;
311 public:
312 static char ID; // Pass identification, replacement for typeid
MemCpyOpt()313 MemCpyOpt() : FunctionPass(ID) {
314 initializeMemCpyOptPass(*PassRegistry::getPassRegistry());
315 MD = nullptr;
316 TLI = nullptr;
317 }
318
319 bool runOnFunction(Function &F) override;
320
321 private:
322 // This transformation requires dominator postdominator info
getAnalysisUsage(AnalysisUsage & AU) const323 void getAnalysisUsage(AnalysisUsage &AU) const override {
324 AU.setPreservesCFG();
325 AU.addRequired<AssumptionCacheTracker>();
326 AU.addRequired<DominatorTreeWrapperPass>();
327 AU.addRequired<MemoryDependenceAnalysis>();
328 AU.addRequired<AAResultsWrapperPass>();
329 AU.addRequired<TargetLibraryInfoWrapperPass>();
330 AU.addPreserved<GlobalsAAWrapperPass>();
331 AU.addPreserved<MemoryDependenceAnalysis>();
332 }
333
334 // Helper functions
335 bool processStore(StoreInst *SI, BasicBlock::iterator &BBI);
336 bool processMemSet(MemSetInst *SI, BasicBlock::iterator &BBI);
337 bool processMemCpy(MemCpyInst *M);
338 bool processMemMove(MemMoveInst *M);
339 bool performCallSlotOptzn(Instruction *cpy, Value *cpyDst, Value *cpySrc,
340 uint64_t cpyLen, unsigned cpyAlign, CallInst *C);
341 bool processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep);
342 bool processMemSetMemCpyDependence(MemCpyInst *M, MemSetInst *MDep);
343 bool performMemCpyToMemSetOptzn(MemCpyInst *M, MemSetInst *MDep);
344 bool processByValArgument(CallSite CS, unsigned ArgNo);
345 Instruction *tryMergingIntoMemset(Instruction *I, Value *StartPtr,
346 Value *ByteVal);
347
348 bool iterateOnFunction(Function &F);
349 };
350
351 char MemCpyOpt::ID = 0;
352 }
353
354 /// The public interface to this file...
createMemCpyOptPass()355 FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOpt(); }
356
357 INITIALIZE_PASS_BEGIN(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
358 false, false)
INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)359 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
360 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
361 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
362 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
363 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
364 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
365 INITIALIZE_PASS_END(MemCpyOpt, "memcpyopt", "MemCpy Optimization",
366 false, false)
367
368 /// When scanning forward over instructions, we look for some other patterns to
369 /// fold away. In particular, this looks for stores to neighboring locations of
370 /// memory. If it sees enough consecutive ones, it attempts to merge them
371 /// together into a memcpy/memset.
372 Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
373 Value *StartPtr, Value *ByteVal) {
374 const DataLayout &DL = StartInst->getModule()->getDataLayout();
375
376 // Okay, so we now have a single store that can be splatable. Scan to find
377 // all subsequent stores of the same value to offset from the same pointer.
378 // Join these together into ranges, so we can decide whether contiguous blocks
379 // are stored.
380 MemsetRanges Ranges(DL);
381
382 BasicBlock::iterator BI(StartInst);
383 for (++BI; !isa<TerminatorInst>(BI); ++BI) {
384 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
385 // If the instruction is readnone, ignore it, otherwise bail out. We
386 // don't even allow readonly here because we don't want something like:
387 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
388 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
389 break;
390 continue;
391 }
392
393 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
394 // If this is a store, see if we can merge it in.
395 if (!NextStore->isSimple()) break;
396
397 // Check to see if this stored value is of the same byte-splattable value.
398 if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
399 break;
400
401 // Check to see if this store is to a constant offset from the start ptr.
402 int64_t Offset;
403 if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset,
404 DL))
405 break;
406
407 Ranges.addStore(Offset, NextStore);
408 } else {
409 MemSetInst *MSI = cast<MemSetInst>(BI);
410
411 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
412 !isa<ConstantInt>(MSI->getLength()))
413 break;
414
415 // Check to see if this store is to a constant offset from the start ptr.
416 int64_t Offset;
417 if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, DL))
418 break;
419
420 Ranges.addMemSet(Offset, MSI);
421 }
422 }
423
424 // If we have no ranges, then we just had a single store with nothing that
425 // could be merged in. This is a very common case of course.
426 if (Ranges.empty())
427 return nullptr;
428
429 // If we had at least one store that could be merged in, add the starting
430 // store as well. We try to avoid this unless there is at least something
431 // interesting as a small compile-time optimization.
432 Ranges.addInst(0, StartInst);
433
434 // If we create any memsets, we put it right before the first instruction that
435 // isn't part of the memset block. This ensure that the memset is dominated
436 // by any addressing instruction needed by the start of the block.
437 IRBuilder<> Builder(&*BI);
438
439 // Now that we have full information about ranges, loop over the ranges and
440 // emit memset's for anything big enough to be worthwhile.
441 Instruction *AMemSet = nullptr;
442 for (const MemsetRange &Range : Ranges) {
443
444 if (Range.TheStores.size() == 1) continue;
445
446 // If it is profitable to lower this range to memset, do so now.
447 if (!Range.isProfitableToUseMemset(DL))
448 continue;
449
450 // Otherwise, we do want to transform this! Create a new memset.
451 // Get the starting pointer of the block.
452 StartPtr = Range.StartPtr;
453
454 // Determine alignment
455 unsigned Alignment = Range.Alignment;
456 if (Alignment == 0) {
457 Type *EltType =
458 cast<PointerType>(StartPtr->getType())->getElementType();
459 Alignment = DL.getABITypeAlignment(EltType);
460 }
461
462 AMemSet =
463 Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
464
465 DEBUG(dbgs() << "Replace stores:\n";
466 for (Instruction *SI : Range.TheStores)
467 dbgs() << *SI << '\n';
468 dbgs() << "With: " << *AMemSet << '\n');
469
470 if (!Range.TheStores.empty())
471 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
472
473 // Zap all the stores.
474 for (Instruction *SI : Range.TheStores) {
475 MD->removeInstruction(SI);
476 SI->eraseFromParent();
477 }
478 ++NumMemSetInfer;
479 }
480
481 return AMemSet;
482 }
483
484
processStore(StoreInst * SI,BasicBlock::iterator & BBI)485 bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
486 if (!SI->isSimple()) return false;
487
488 // Avoid merging nontemporal stores since the resulting
489 // memcpy/memset would not be able to preserve the nontemporal hint.
490 // In theory we could teach how to propagate the !nontemporal metadata to
491 // memset calls. However, that change would force the backend to
492 // conservatively expand !nontemporal memset calls back to sequences of
493 // store instructions (effectively undoing the merging).
494 if (SI->getMetadata(LLVMContext::MD_nontemporal))
495 return false;
496
497 const DataLayout &DL = SI->getModule()->getDataLayout();
498
499 // Detect cases where we're performing call slot forwarding, but
500 // happen to be using a load-store pair to implement it, rather than
501 // a memcpy.
502 if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
503 if (LI->isSimple() && LI->hasOneUse() &&
504 LI->getParent() == SI->getParent()) {
505 MemDepResult ldep = MD->getDependency(LI);
506 CallInst *C = nullptr;
507 if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
508 C = dyn_cast<CallInst>(ldep.getInst());
509
510 if (C) {
511 // Check that nothing touches the dest of the "copy" between
512 // the call and the store.
513 AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
514 MemoryLocation StoreLoc = MemoryLocation::get(SI);
515 for (BasicBlock::iterator I = --SI->getIterator(), E = C->getIterator();
516 I != E; --I) {
517 if (AA.getModRefInfo(&*I, StoreLoc) != MRI_NoModRef) {
518 C = nullptr;
519 break;
520 }
521 }
522 }
523
524 if (C) {
525 unsigned storeAlign = SI->getAlignment();
526 if (!storeAlign)
527 storeAlign = DL.getABITypeAlignment(SI->getOperand(0)->getType());
528 unsigned loadAlign = LI->getAlignment();
529 if (!loadAlign)
530 loadAlign = DL.getABITypeAlignment(LI->getType());
531
532 bool changed = performCallSlotOptzn(
533 LI, SI->getPointerOperand()->stripPointerCasts(),
534 LI->getPointerOperand()->stripPointerCasts(),
535 DL.getTypeStoreSize(SI->getOperand(0)->getType()),
536 std::min(storeAlign, loadAlign), C);
537 if (changed) {
538 MD->removeInstruction(SI);
539 SI->eraseFromParent();
540 MD->removeInstruction(LI);
541 LI->eraseFromParent();
542 ++NumMemCpyInstr;
543 return true;
544 }
545 }
546 }
547 }
548
549 // There are two cases that are interesting for this code to handle: memcpy
550 // and memset. Right now we only handle memset.
551
552 // Ensure that the value being stored is something that can be memset'able a
553 // byte at a time like "0" or "-1" or any width, as well as things like
554 // 0xA0A0A0A0 and 0.0.
555 if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
556 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
557 ByteVal)) {
558 BBI = I->getIterator(); // Don't invalidate iterator.
559 return true;
560 }
561
562 return false;
563 }
564
processMemSet(MemSetInst * MSI,BasicBlock::iterator & BBI)565 bool MemCpyOpt::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
566 // See if there is another memset or store neighboring this memset which
567 // allows us to widen out the memset to do a single larger store.
568 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
569 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
570 MSI->getValue())) {
571 BBI = I->getIterator(); // Don't invalidate iterator.
572 return true;
573 }
574 return false;
575 }
576
577
578 /// Takes a memcpy and a call that it depends on,
579 /// and checks for the possibility of a call slot optimization by having
580 /// the call write its result directly into the destination of the memcpy.
performCallSlotOptzn(Instruction * cpy,Value * cpyDest,Value * cpySrc,uint64_t cpyLen,unsigned cpyAlign,CallInst * C)581 bool MemCpyOpt::performCallSlotOptzn(Instruction *cpy,
582 Value *cpyDest, Value *cpySrc,
583 uint64_t cpyLen, unsigned cpyAlign,
584 CallInst *C) {
585 // The general transformation to keep in mind is
586 //
587 // call @func(..., src, ...)
588 // memcpy(dest, src, ...)
589 //
590 // ->
591 //
592 // memcpy(dest, src, ...)
593 // call @func(..., dest, ...)
594 //
595 // Since moving the memcpy is technically awkward, we additionally check that
596 // src only holds uninitialized values at the moment of the call, meaning that
597 // the memcpy can be discarded rather than moved.
598
599 // Deliberately get the source and destination with bitcasts stripped away,
600 // because we'll need to do type comparisons based on the underlying type.
601 CallSite CS(C);
602
603 // Require that src be an alloca. This simplifies the reasoning considerably.
604 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
605 if (!srcAlloca)
606 return false;
607
608 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
609 if (!srcArraySize)
610 return false;
611
612 const DataLayout &DL = cpy->getModule()->getDataLayout();
613 uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
614 srcArraySize->getZExtValue();
615
616 if (cpyLen < srcSize)
617 return false;
618
619 // Check that accessing the first srcSize bytes of dest will not cause a
620 // trap. Otherwise the transform is invalid since it might cause a trap
621 // to occur earlier than it otherwise would.
622 if (AllocaInst *A = dyn_cast<AllocaInst>(cpyDest)) {
623 // The destination is an alloca. Check it is larger than srcSize.
624 ConstantInt *destArraySize = dyn_cast<ConstantInt>(A->getArraySize());
625 if (!destArraySize)
626 return false;
627
628 uint64_t destSize = DL.getTypeAllocSize(A->getAllocatedType()) *
629 destArraySize->getZExtValue();
630
631 if (destSize < srcSize)
632 return false;
633 } else if (Argument *A = dyn_cast<Argument>(cpyDest)) {
634 if (A->getDereferenceableBytes() < srcSize) {
635 // If the destination is an sret parameter then only accesses that are
636 // outside of the returned struct type can trap.
637 if (!A->hasStructRetAttr())
638 return false;
639
640 Type *StructTy = cast<PointerType>(A->getType())->getElementType();
641 if (!StructTy->isSized()) {
642 // The call may never return and hence the copy-instruction may never
643 // be executed, and therefore it's not safe to say "the destination
644 // has at least <cpyLen> bytes, as implied by the copy-instruction",
645 return false;
646 }
647
648 uint64_t destSize = DL.getTypeAllocSize(StructTy);
649 if (destSize < srcSize)
650 return false;
651 }
652 } else {
653 return false;
654 }
655
656 // Check that dest points to memory that is at least as aligned as src.
657 unsigned srcAlign = srcAlloca->getAlignment();
658 if (!srcAlign)
659 srcAlign = DL.getABITypeAlignment(srcAlloca->getAllocatedType());
660 bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
661 // If dest is not aligned enough and we can't increase its alignment then
662 // bail out.
663 if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
664 return false;
665
666 // Check that src is not accessed except via the call and the memcpy. This
667 // guarantees that it holds only undefined values when passed in (so the final
668 // memcpy can be dropped), that it is not read or written between the call and
669 // the memcpy, and that writing beyond the end of it is undefined.
670 SmallVector<User*, 8> srcUseList(srcAlloca->user_begin(),
671 srcAlloca->user_end());
672 while (!srcUseList.empty()) {
673 User *U = srcUseList.pop_back_val();
674
675 if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
676 for (User *UU : U->users())
677 srcUseList.push_back(UU);
678 continue;
679 }
680 if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) {
681 if (!G->hasAllZeroIndices())
682 return false;
683
684 for (User *UU : U->users())
685 srcUseList.push_back(UU);
686 continue;
687 }
688 if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U))
689 if (IT->getIntrinsicID() == Intrinsic::lifetime_start ||
690 IT->getIntrinsicID() == Intrinsic::lifetime_end)
691 continue;
692
693 if (U != C && U != cpy)
694 return false;
695 }
696
697 // Check that src isn't captured by the called function since the
698 // transformation can cause aliasing issues in that case.
699 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
700 if (CS.getArgument(i) == cpySrc && !CS.doesNotCapture(i))
701 return false;
702
703 // Since we're changing the parameter to the callsite, we need to make sure
704 // that what would be the new parameter dominates the callsite.
705 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
706 if (Instruction *cpyDestInst = dyn_cast<Instruction>(cpyDest))
707 if (!DT.dominates(cpyDestInst, C))
708 return false;
709
710 // In addition to knowing that the call does not access src in some
711 // unexpected manner, for example via a global, which we deduce from
712 // the use analysis, we also need to know that it does not sneakily
713 // access dest. We rely on AA to figure this out for us.
714 AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
715 ModRefInfo MR = AA.getModRefInfo(C, cpyDest, srcSize);
716 // If necessary, perform additional analysis.
717 if (MR != MRI_NoModRef)
718 MR = AA.callCapturesBefore(C, cpyDest, srcSize, &DT);
719 if (MR != MRI_NoModRef)
720 return false;
721
722 // All the checks have passed, so do the transformation.
723 bool changedArgument = false;
724 for (unsigned i = 0; i < CS.arg_size(); ++i)
725 if (CS.getArgument(i)->stripPointerCasts() == cpySrc) {
726 Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
727 : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
728 cpyDest->getName(), C);
729 changedArgument = true;
730 if (CS.getArgument(i)->getType() == Dest->getType())
731 CS.setArgument(i, Dest);
732 else
733 CS.setArgument(i, CastInst::CreatePointerCast(Dest,
734 CS.getArgument(i)->getType(), Dest->getName(), C));
735 }
736
737 if (!changedArgument)
738 return false;
739
740 // If the destination wasn't sufficiently aligned then increase its alignment.
741 if (!isDestSufficientlyAligned) {
742 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
743 cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
744 }
745
746 // Drop any cached information about the call, because we may have changed
747 // its dependence information by changing its parameter.
748 MD->removeInstruction(C);
749
750 // Update AA metadata
751 // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
752 // handled here, but combineMetadata doesn't support them yet
753 unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
754 LLVMContext::MD_noalias,
755 LLVMContext::MD_invariant_group};
756 combineMetadata(C, cpy, KnownIDs);
757
758 // Remove the memcpy.
759 MD->removeInstruction(cpy);
760 ++NumMemCpyInstr;
761
762 return true;
763 }
764
765 /// We've found that the (upward scanning) memory dependence of memcpy 'M' is
766 /// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
processMemCpyMemCpyDependence(MemCpyInst * M,MemCpyInst * MDep)767 bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep) {
768 // We can only transforms memcpy's where the dest of one is the source of the
769 // other.
770 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
771 return false;
772
773 // If dep instruction is reading from our current input, then it is a noop
774 // transfer and substituting the input won't change this instruction. Just
775 // ignore the input and let someone else zap MDep. This handles cases like:
776 // memcpy(a <- a)
777 // memcpy(b <- a)
778 if (M->getSource() == MDep->getSource())
779 return false;
780
781 // Second, the length of the memcpy's must be the same, or the preceding one
782 // must be larger than the following one.
783 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
784 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
785 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
786 return false;
787
788 AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
789
790 // Verify that the copied-from memory doesn't change in between the two
791 // transfers. For example, in:
792 // memcpy(a <- b)
793 // *b = 42;
794 // memcpy(c <- a)
795 // It would be invalid to transform the second memcpy into memcpy(c <- b).
796 //
797 // TODO: If the code between M and MDep is transparent to the destination "c",
798 // then we could still perform the xform by moving M up to the first memcpy.
799 //
800 // NOTE: This is conservative, it will stop on any read from the source loc,
801 // not just the defining memcpy.
802 MemDepResult SourceDep =
803 MD->getPointerDependencyFrom(MemoryLocation::getForSource(MDep), false,
804 M->getIterator(), M->getParent());
805 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
806 return false;
807
808 // If the dest of the second might alias the source of the first, then the
809 // source and dest might overlap. We still want to eliminate the intermediate
810 // value, but we have to generate a memmove instead of memcpy.
811 bool UseMemMove = false;
812 if (!AA.isNoAlias(MemoryLocation::getForDest(M),
813 MemoryLocation::getForSource(MDep)))
814 UseMemMove = true;
815
816 // If all checks passed, then we can transform M.
817
818 // Make sure to use the lesser of the alignment of the source and the dest
819 // since we're changing where we're reading from, but don't want to increase
820 // the alignment past what can be read from or written to.
821 // TODO: Is this worth it if we're creating a less aligned memcpy? For
822 // example we could be moving from movaps -> movq on x86.
823 unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
824
825 IRBuilder<> Builder(M);
826 if (UseMemMove)
827 Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
828 Align, M->isVolatile());
829 else
830 Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
831 Align, M->isVolatile());
832
833 // Remove the instruction we're replacing.
834 MD->removeInstruction(M);
835 M->eraseFromParent();
836 ++NumMemCpyInstr;
837 return true;
838 }
839
840 /// We've found that the (upward scanning) memory dependence of \p MemCpy is
841 /// \p MemSet. Try to simplify \p MemSet to only set the trailing bytes that
842 /// weren't copied over by \p MemCpy.
843 ///
844 /// In other words, transform:
845 /// \code
846 /// memset(dst, c, dst_size);
847 /// memcpy(dst, src, src_size);
848 /// \endcode
849 /// into:
850 /// \code
851 /// memcpy(dst, src, src_size);
852 /// memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
853 /// \endcode
processMemSetMemCpyDependence(MemCpyInst * MemCpy,MemSetInst * MemSet)854 bool MemCpyOpt::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
855 MemSetInst *MemSet) {
856 // We can only transform memset/memcpy with the same destination.
857 if (MemSet->getDest() != MemCpy->getDest())
858 return false;
859
860 // Check that there are no other dependencies on the memset destination.
861 MemDepResult DstDepInfo =
862 MD->getPointerDependencyFrom(MemoryLocation::getForDest(MemSet), false,
863 MemCpy->getIterator(), MemCpy->getParent());
864 if (DstDepInfo.getInst() != MemSet)
865 return false;
866
867 // Use the same i8* dest as the memcpy, killing the memset dest if different.
868 Value *Dest = MemCpy->getRawDest();
869 Value *DestSize = MemSet->getLength();
870 Value *SrcSize = MemCpy->getLength();
871
872 // By default, create an unaligned memset.
873 unsigned Align = 1;
874 // If Dest is aligned, and SrcSize is constant, use the minimum alignment
875 // of the sum.
876 const unsigned DestAlign =
877 std::max(MemSet->getAlignment(), MemCpy->getAlignment());
878 if (DestAlign > 1)
879 if (ConstantInt *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
880 Align = MinAlign(SrcSizeC->getZExtValue(), DestAlign);
881
882 IRBuilder<> Builder(MemCpy);
883
884 // If the sizes have different types, zext the smaller one.
885 if (DestSize->getType() != SrcSize->getType()) {
886 if (DestSize->getType()->getIntegerBitWidth() >
887 SrcSize->getType()->getIntegerBitWidth())
888 SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
889 else
890 DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
891 }
892
893 Value *MemsetLen =
894 Builder.CreateSelect(Builder.CreateICmpULE(DestSize, SrcSize),
895 ConstantInt::getNullValue(DestSize->getType()),
896 Builder.CreateSub(DestSize, SrcSize));
897 Builder.CreateMemSet(Builder.CreateGEP(Dest, SrcSize), MemSet->getOperand(1),
898 MemsetLen, Align);
899
900 MD->removeInstruction(MemSet);
901 MemSet->eraseFromParent();
902 return true;
903 }
904
905 /// Transform memcpy to memset when its source was just memset.
906 /// In other words, turn:
907 /// \code
908 /// memset(dst1, c, dst1_size);
909 /// memcpy(dst2, dst1, dst2_size);
910 /// \endcode
911 /// into:
912 /// \code
913 /// memset(dst1, c, dst1_size);
914 /// memset(dst2, c, dst2_size);
915 /// \endcode
916 /// When dst2_size <= dst1_size.
917 ///
918 /// The \p MemCpy must have a Constant length.
performMemCpyToMemSetOptzn(MemCpyInst * MemCpy,MemSetInst * MemSet)919 bool MemCpyOpt::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
920 MemSetInst *MemSet) {
921 // This only makes sense on memcpy(..., memset(...), ...).
922 if (MemSet->getRawDest() != MemCpy->getRawSource())
923 return false;
924
925 ConstantInt *CopySize = cast<ConstantInt>(MemCpy->getLength());
926 ConstantInt *MemSetSize = dyn_cast<ConstantInt>(MemSet->getLength());
927 // Make sure the memcpy doesn't read any more than what the memset wrote.
928 // Don't worry about sizes larger than i64.
929 if (!MemSetSize || CopySize->getZExtValue() > MemSetSize->getZExtValue())
930 return false;
931
932 IRBuilder<> Builder(MemCpy);
933 Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),
934 CopySize, MemCpy->getAlignment());
935 return true;
936 }
937
938 /// Perform simplification of memcpy's. If we have memcpy A
939 /// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
940 /// B to be a memcpy from X to Z (or potentially a memmove, depending on
941 /// circumstances). This allows later passes to remove the first memcpy
942 /// altogether.
processMemCpy(MemCpyInst * M)943 bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
944 // We can only optimize non-volatile memcpy's.
945 if (M->isVolatile()) return false;
946
947 // If the source and destination of the memcpy are the same, then zap it.
948 if (M->getSource() == M->getDest()) {
949 MD->removeInstruction(M);
950 M->eraseFromParent();
951 return false;
952 }
953
954 // If copying from a constant, try to turn the memcpy into a memset.
955 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
956 if (GV->isConstant() && GV->hasDefinitiveInitializer())
957 if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
958 IRBuilder<> Builder(M);
959 Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
960 M->getAlignment(), false);
961 MD->removeInstruction(M);
962 M->eraseFromParent();
963 ++NumCpyToSet;
964 return true;
965 }
966
967 MemDepResult DepInfo = MD->getDependency(M);
968
969 // Try to turn a partially redundant memset + memcpy into
970 // memcpy + smaller memset. We don't need the memcpy size for this.
971 if (DepInfo.isClobber())
972 if (MemSetInst *MDep = dyn_cast<MemSetInst>(DepInfo.getInst()))
973 if (processMemSetMemCpyDependence(M, MDep))
974 return true;
975
976 // The optimizations after this point require the memcpy size.
977 ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
978 if (!CopySize) return false;
979
980 // There are four possible optimizations we can do for memcpy:
981 // a) memcpy-memcpy xform which exposes redundance for DSE.
982 // b) call-memcpy xform for return slot optimization.
983 // c) memcpy from freshly alloca'd space or space that has just started its
984 // lifetime copies undefined data, and we can therefore eliminate the
985 // memcpy in favor of the data that was already at the destination.
986 // d) memcpy from a just-memset'd source can be turned into memset.
987 if (DepInfo.isClobber()) {
988 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
989 if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
990 CopySize->getZExtValue(), M->getAlignment(),
991 C)) {
992 MD->removeInstruction(M);
993 M->eraseFromParent();
994 return true;
995 }
996 }
997 }
998
999 MemoryLocation SrcLoc = MemoryLocation::getForSource(M);
1000 MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(
1001 SrcLoc, true, M->getIterator(), M->getParent());
1002
1003 if (SrcDepInfo.isClobber()) {
1004 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
1005 return processMemCpyMemCpyDependence(M, MDep);
1006 } else if (SrcDepInfo.isDef()) {
1007 Instruction *I = SrcDepInfo.getInst();
1008 bool hasUndefContents = false;
1009
1010 if (isa<AllocaInst>(I)) {
1011 hasUndefContents = true;
1012 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1013 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1014 if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
1015 if (LTSize->getZExtValue() >= CopySize->getZExtValue())
1016 hasUndefContents = true;
1017 }
1018
1019 if (hasUndefContents) {
1020 MD->removeInstruction(M);
1021 M->eraseFromParent();
1022 ++NumMemCpyInstr;
1023 return true;
1024 }
1025 }
1026
1027 if (SrcDepInfo.isClobber())
1028 if (MemSetInst *MDep = dyn_cast<MemSetInst>(SrcDepInfo.getInst()))
1029 if (performMemCpyToMemSetOptzn(M, MDep)) {
1030 MD->removeInstruction(M);
1031 M->eraseFromParent();
1032 ++NumCpyToSet;
1033 return true;
1034 }
1035
1036 return false;
1037 }
1038
1039 /// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1040 /// not to alias.
processMemMove(MemMoveInst * M)1041 bool MemCpyOpt::processMemMove(MemMoveInst *M) {
1042 AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1043
1044 if (!TLI->has(LibFunc::memmove))
1045 return false;
1046
1047 // See if the pointers alias.
1048 if (!AA.isNoAlias(MemoryLocation::getForDest(M),
1049 MemoryLocation::getForSource(M)))
1050 return false;
1051
1052 DEBUG(dbgs() << "MemCpyOpt: Optimizing memmove -> memcpy: " << *M << "\n");
1053
1054 // If not, then we know we can transform this.
1055 Type *ArgTys[3] = { M->getRawDest()->getType(),
1056 M->getRawSource()->getType(),
1057 M->getLength()->getType() };
1058 M->setCalledFunction(Intrinsic::getDeclaration(M->getModule(),
1059 Intrinsic::memcpy, ArgTys));
1060
1061 // MemDep may have over conservative information about this instruction, just
1062 // conservatively flush it from the cache.
1063 MD->removeInstruction(M);
1064
1065 ++NumMoveToCpy;
1066 return true;
1067 }
1068
1069 /// This is called on every byval argument in call sites.
processByValArgument(CallSite CS,unsigned ArgNo)1070 bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
1071 const DataLayout &DL = CS.getCaller()->getParent()->getDataLayout();
1072 // Find out what feeds this byval argument.
1073 Value *ByValArg = CS.getArgument(ArgNo);
1074 Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
1075 uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
1076 MemDepResult DepInfo = MD->getPointerDependencyFrom(
1077 MemoryLocation(ByValArg, ByValSize), true,
1078 CS.getInstruction()->getIterator(), CS.getInstruction()->getParent());
1079 if (!DepInfo.isClobber())
1080 return false;
1081
1082 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
1083 // a memcpy, see if we can byval from the source of the memcpy instead of the
1084 // result.
1085 MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
1086 if (!MDep || MDep->isVolatile() ||
1087 ByValArg->stripPointerCasts() != MDep->getDest())
1088 return false;
1089
1090 // The length of the memcpy must be larger or equal to the size of the byval.
1091 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1092 if (!C1 || C1->getValue().getZExtValue() < ByValSize)
1093 return false;
1094
1095 // Get the alignment of the byval. If the call doesn't specify the alignment,
1096 // then it is some target specific value that we can't know.
1097 unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
1098 if (ByValAlign == 0) return false;
1099
1100 // If it is greater than the memcpy, then we check to see if we can force the
1101 // source of the memcpy to the alignment we need. If we fail, we bail out.
1102 AssumptionCache &AC =
1103 getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
1104 *CS->getParent()->getParent());
1105 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1106 if (MDep->getAlignment() < ByValAlign &&
1107 getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL,
1108 CS.getInstruction(), &AC, &DT) < ByValAlign)
1109 return false;
1110
1111 // Verify that the copied-from memory doesn't change in between the memcpy and
1112 // the byval call.
1113 // memcpy(a <- b)
1114 // *b = 42;
1115 // foo(*a)
1116 // It would be invalid to transform the second memcpy into foo(*b).
1117 //
1118 // NOTE: This is conservative, it will stop on any read from the source loc,
1119 // not just the defining memcpy.
1120 MemDepResult SourceDep = MD->getPointerDependencyFrom(
1121 MemoryLocation::getForSource(MDep), false,
1122 CS.getInstruction()->getIterator(), MDep->getParent());
1123 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
1124 return false;
1125
1126 Value *TmpCast = MDep->getSource();
1127 if (MDep->getSource()->getType() != ByValArg->getType())
1128 TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
1129 "tmpcast", CS.getInstruction());
1130
1131 DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
1132 << " " << *MDep << "\n"
1133 << " " << *CS.getInstruction() << "\n");
1134
1135 // Otherwise we're good! Update the byval argument.
1136 CS.setArgument(ArgNo, TmpCast);
1137 ++NumMemCpyInstr;
1138 return true;
1139 }
1140
1141 /// Executes one iteration of MemCpyOpt.
iterateOnFunction(Function & F)1142 bool MemCpyOpt::iterateOnFunction(Function &F) {
1143 bool MadeChange = false;
1144
1145 // Walk all instruction in the function.
1146 for (Function::iterator BB = F.begin(), BBE = F.end(); BB != BBE; ++BB) {
1147 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
1148 // Avoid invalidating the iterator.
1149 Instruction *I = &*BI++;
1150
1151 bool RepeatInstruction = false;
1152
1153 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1154 MadeChange |= processStore(SI, BI);
1155 else if (MemSetInst *M = dyn_cast<MemSetInst>(I))
1156 RepeatInstruction = processMemSet(M, BI);
1157 else if (MemCpyInst *M = dyn_cast<MemCpyInst>(I))
1158 RepeatInstruction = processMemCpy(M);
1159 else if (MemMoveInst *M = dyn_cast<MemMoveInst>(I))
1160 RepeatInstruction = processMemMove(M);
1161 else if (auto CS = CallSite(I)) {
1162 for (unsigned i = 0, e = CS.arg_size(); i != e; ++i)
1163 if (CS.isByValArgument(i))
1164 MadeChange |= processByValArgument(CS, i);
1165 }
1166
1167 // Reprocess the instruction if desired.
1168 if (RepeatInstruction) {
1169 if (BI != BB->begin()) --BI;
1170 MadeChange = true;
1171 }
1172 }
1173 }
1174
1175 return MadeChange;
1176 }
1177
1178 /// This is the main transformation entry point for a function.
runOnFunction(Function & F)1179 bool MemCpyOpt::runOnFunction(Function &F) {
1180 if (skipOptnoneFunction(F))
1181 return false;
1182
1183 bool MadeChange = false;
1184 MD = &getAnalysis<MemoryDependenceAnalysis>();
1185 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
1186
1187 // If we don't have at least memset and memcpy, there is little point of doing
1188 // anything here. These are required by a freestanding implementation, so if
1189 // even they are disabled, there is no point in trying hard.
1190 if (!TLI->has(LibFunc::memset) || !TLI->has(LibFunc::memcpy))
1191 return false;
1192
1193 while (1) {
1194 if (!iterateOnFunction(F))
1195 break;
1196 MadeChange = true;
1197 }
1198
1199 MD = nullptr;
1200 return MadeChange;
1201 }
1202