1//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass performs various transformations related to eliminating memcpy
10// calls, or transforming sets of stores into memset's.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/Scalar/MemCpyOptimizer.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/STLExtras.h"
17#include "llvm/ADT/ScopeExit.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/CFG.h"
24#include "llvm/Analysis/CaptureTracking.h"
25#include "llvm/Analysis/GlobalsModRef.h"
26#include "llvm/Analysis/InstructionSimplify.h"
27#include "llvm/Analysis/Loads.h"
28#include "llvm/Analysis/MemoryLocation.h"
29#include "llvm/Analysis/MemorySSA.h"
30#include "llvm/Analysis/MemorySSAUpdater.h"
31#include "llvm/Analysis/PostDominators.h"
32#include "llvm/Analysis/TargetLibraryInfo.h"
33#include "llvm/Analysis/ValueTracking.h"
34#include "llvm/IR/BasicBlock.h"
35#include "llvm/IR/Constants.h"
36#include "llvm/IR/DataLayout.h"
37#include "llvm/IR/DerivedTypes.h"
38#include "llvm/IR/Dominators.h"
39#include "llvm/IR/Function.h"
40#include "llvm/IR/GlobalVariable.h"
41#include "llvm/IR/IRBuilder.h"
42#include "llvm/IR/InstrTypes.h"
43#include "llvm/IR/Instruction.h"
44#include "llvm/IR/Instructions.h"
45#include "llvm/IR/IntrinsicInst.h"
46#include "llvm/IR/Intrinsics.h"
47#include "llvm/IR/LLVMContext.h"
48#include "llvm/IR/Module.h"
49#include "llvm/IR/PassManager.h"
50#include "llvm/IR/ProfDataUtils.h"
51#include "llvm/IR/Type.h"
52#include "llvm/IR/User.h"
53#include "llvm/IR/Value.h"
54#include "llvm/Support/Casting.h"
55#include "llvm/Support/Debug.h"
56#include "llvm/Support/raw_ostream.h"
57#include "llvm/Transforms/Utils/Local.h"
58#include <algorithm>
59#include <cassert>
60#include <cstdint>
61#include <optional>
62
63using namespace llvm;
64
65#define DEBUG_TYPE "memcpyopt"
66
67static cl::opt<bool> EnableMemCpyOptWithoutLibcalls(
68 "enable-memcpyopt-without-libcalls", cl::Hidden,
69 cl::desc("Enable memcpyopt even when libcalls are disabled"));
70
71STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
72STATISTIC(NumMemMoveInstr, "Number of memmove instructions deleted");
73STATISTIC(NumMemSetInfer, "Number of memsets inferred");
74STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy");
75STATISTIC(NumCpyToSet, "Number of memcpys converted to memset");
76STATISTIC(NumCallSlot, "Number of call slot optimizations performed");
77STATISTIC(NumStackMove, "Number of stack-move optimizations performed");
78
79namespace {
80
81/// Represents a range of memset'd bytes with the ByteVal value.
82/// This allows us to analyze stores like:
83/// store 0 -> P+1
84/// store 0 -> P+0
85/// store 0 -> P+3
86/// store 0 -> P+2
87/// which sometimes happens with stores to arrays of structs etc. When we see
88/// the first store, we make a range [1, 2). The second store extends the range
89/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
90/// two ranges into [0, 3) which is memset'able.
91struct MemsetRange {
92 // Start/End - A semi range that describes the span that this range covers.
93 // The range is closed at the start and open at the end: [Start, End).
94 int64_t Start, End;
95
96 /// StartPtr - The getelementptr instruction that points to the start of the
97 /// range.
98 Value *StartPtr;
99
100 /// Alignment - The known alignment of the first store.
101 MaybeAlign Alignment;
102
103 /// TheStores - The actual stores that make up this range.
104 SmallVector<Instruction *, 16> TheStores;
105
106 bool isProfitableToUseMemset(const DataLayout &DL) const;
107};
108
109} // end anonymous namespace
110
111static bool overreadUndefContents(MemorySSA *MSSA, MemCpyInst *MemCpy,
112 MemIntrinsic *MemSrc, BatchAAResults &BAA);
113
114bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
115 // If we found more than 4 stores to merge or 16 bytes, use memset.
116 if (TheStores.size() >= 4 || End - Start >= 16)
117 return true;
118
119 // If there is nothing to merge, don't do anything.
120 if (TheStores.size() < 2)
121 return false;
122
123 // If any of the stores are a memset, then it is always good to extend the
124 // memset.
125 for (Instruction *SI : TheStores)
126 if (!isa<StoreInst>(Val: SI))
127 return true;
128
129 // Assume that the code generator is capable of merging pairs of stores
130 // together if it wants to.
131 if (TheStores.size() == 2)
132 return false;
133
134 // If we have fewer than 8 stores, it can still be worthwhile to do this.
135 // For example, merging 4 i8 stores into an i32 store is useful almost always.
136 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
137 // memset will be split into 2 32-bit stores anyway) and doing so can
138 // pessimize the llvm optimizer.
139 //
140 // Since we don't have perfect knowledge here, make some assumptions: assume
141 // the maximum GPR width is the same size as the largest legal integer
142 // size. If so, check to see whether we will end up actually reducing the
143 // number of stores used.
144 unsigned Bytes = unsigned(End - Start);
145 unsigned MaxIntSize = DL.getLargestLegalIntTypeSizeInBits() / 8;
146 if (MaxIntSize == 0)
147 MaxIntSize = 1;
148 unsigned NumPointerStores = Bytes / MaxIntSize;
149
150 // Assume the remaining bytes if any are done a byte at a time.
151 unsigned NumByteStores = Bytes % MaxIntSize;
152
153 // If we will reduce the # stores (according to this heuristic), do the
154 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
155 // etc.
156 return TheStores.size() > NumPointerStores + NumByteStores;
157}
158
159namespace {
160
161class MemsetRanges {
162 using range_iterator = SmallVectorImpl<MemsetRange>::iterator;
163
164 /// A sorted list of the memset ranges.
165 SmallVector<MemsetRange, 8> Ranges;
166
167 const DataLayout &DL;
168
169public:
170 MemsetRanges(const DataLayout &DL) : DL(DL) {}
171
172 using const_iterator = SmallVectorImpl<MemsetRange>::const_iterator;
173
174 const_iterator begin() const { return Ranges.begin(); }
175 const_iterator end() const { return Ranges.end(); }
176 bool empty() const { return Ranges.empty(); }
177
178 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
179 if (auto *SI = dyn_cast<StoreInst>(Val: Inst))
180 addStore(OffsetFromFirst, SI);
181 else
182 addMemSet(OffsetFromFirst, MSI: cast<MemSetInst>(Val: Inst));
183 }
184
185 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
186 TypeSize StoreSize = DL.getTypeStoreSize(Ty: SI->getOperand(i_nocapture: 0)->getType());
187 assert(!StoreSize.isScalable() && "Can't track scalable-typed stores");
188 addRange(Start: OffsetFromFirst, Size: StoreSize.getFixedValue(),
189 Ptr: SI->getPointerOperand(), Alignment: SI->getAlign(), Inst: SI);
190 }
191
192 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
193 int64_t Size = cast<ConstantInt>(Val: MSI->getLength())->getZExtValue();
194 addRange(Start: OffsetFromFirst, Size, Ptr: MSI->getDest(), Alignment: MSI->getDestAlign(), Inst: MSI);
195 }
196
197 void addRange(int64_t Start, int64_t Size, Value *Ptr, MaybeAlign Alignment,
198 Instruction *Inst);
199};
200
201} // end anonymous namespace
202
203/// Add a new store to the MemsetRanges data structure. This adds a
204/// new range for the specified store at the specified offset, merging into
205/// existing ranges as appropriate.
206void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
207 MaybeAlign Alignment, Instruction *Inst) {
208 int64_t End = Start + Size;
209
210 range_iterator I = partition_point(
211 Range&: Ranges, P: [=](const MemsetRange &O) { return O.End < Start; });
212
213 // We now know that I == E, in which case we didn't find anything to merge
214 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
215 // to insert a new range. Handle this now.
216 if (I == Ranges.end() || End < I->Start) {
217 MemsetRange &R = *Ranges.insert(I, Elt: MemsetRange());
218 R.Start = Start;
219 R.End = End;
220 R.StartPtr = Ptr;
221 R.Alignment = Alignment;
222 R.TheStores.push_back(Elt: Inst);
223 return;
224 }
225
226 // This store overlaps with I, add it.
227 I->TheStores.push_back(Elt: Inst);
228
229 // At this point, we may have an interval that completely contains our store.
230 // If so, just add it to the interval and return.
231 if (I->Start <= Start && I->End >= End)
232 return;
233
234 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
235 // but is not entirely contained within the range.
236
237 // See if the range extends the start of the range. In this case, it couldn't
238 // possibly cause it to join the prior range, because otherwise we would have
239 // stopped on *it*.
240 if (Start < I->Start) {
241 I->Start = Start;
242 I->StartPtr = Ptr;
243 I->Alignment = Alignment;
244 }
245
246 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
247 // is in or right at the end of I), and that End >= I->Start. Extend I out to
248 // End.
249 if (End > I->End) {
250 I->End = End;
251 range_iterator NextI = I;
252 while (++NextI != Ranges.end() && End >= NextI->Start) {
253 // Merge the range in.
254 I->TheStores.append(in_start: NextI->TheStores.begin(), in_end: NextI->TheStores.end());
255 if (NextI->End > I->End)
256 I->End = NextI->End;
257 Ranges.erase(CI: NextI);
258 NextI = I;
259 }
260 }
261}
262
263//===----------------------------------------------------------------------===//
264// MemCpyOptLegacyPass Pass
265//===----------------------------------------------------------------------===//
266
267// Check that V is either not accessible by the caller, or unwinding cannot
268// occur between Start and End.
269static bool mayBeVisibleThroughUnwinding(Value *V, Instruction *Start,
270 Instruction *End) {
271 assert(Start->getParent() == End->getParent() && "Must be in same block");
272 // Function can't unwind, so it also can't be visible through unwinding.
273 if (Start->getFunction()->doesNotThrow())
274 return false;
275
276 // Object is not visible on unwind.
277 // TODO: Support RequiresNoCaptureBeforeUnwind case.
278 bool RequiresNoCaptureBeforeUnwind;
279 if (isNotVisibleOnUnwind(Object: getUnderlyingObject(V),
280 RequiresNoCaptureBeforeUnwind) &&
281 !RequiresNoCaptureBeforeUnwind)
282 return false;
283
284 // Check whether there are any unwinding instructions in the range.
285 return any_of(Range: make_range(x: Start->getIterator(), y: End->getIterator()),
286 P: [](const Instruction &I) { return I.mayThrow(); });
287}
288
289void MemCpyOptPass::eraseInstruction(Instruction *I) {
290 MSSAU->removeMemoryAccess(I);
291 EEA->removeInstruction(I);
292 I->eraseFromParent();
293}
294
295// Check for mod or ref of Loc between Start and End, excluding both boundaries.
296// Start and End must be in the same block.
297// If SkippedLifetimeStart is provided, skip over one clobbering lifetime.start
298// intrinsic and store it inside SkippedLifetimeStart.
299static bool accessedBetween(BatchAAResults &AA, MemoryLocation Loc,
300 const MemoryUseOrDef *Start,
301 const MemoryUseOrDef *End,
302 Instruction **SkippedLifetimeStart = nullptr) {
303 assert(Start->getBlock() == End->getBlock() && "Only local supported");
304 for (const MemoryAccess &MA :
305 make_range(x: ++Start->getIterator(), y: End->getIterator())) {
306 Instruction *I = cast<MemoryUseOrDef>(Val: MA).getMemoryInst();
307 if (isModOrRefSet(MRI: AA.getModRefInfo(I, OptLoc: Loc))) {
308 auto *II = dyn_cast<IntrinsicInst>(Val: I);
309 if (II && II->getIntrinsicID() == Intrinsic::lifetime_start &&
310 SkippedLifetimeStart && !*SkippedLifetimeStart) {
311 *SkippedLifetimeStart = I;
312 continue;
313 }
314
315 return true;
316 }
317 }
318 return false;
319}
320
321// Check for mod of Loc between Start and End, excluding both boundaries.
322// Start and End can be in different blocks.
323static bool writtenBetween(MemorySSA *MSSA, BatchAAResults &AA,
324 MemoryLocation Loc, const MemoryUseOrDef *Start,
325 const MemoryUseOrDef *End) {
326 if (isa<MemoryUse>(Val: End)) {
327 // For MemoryUses, getClobberingMemoryAccess may skip non-clobbering writes.
328 // Manually check read accesses between Start and End, if they are in the
329 // same block, for clobbers. Otherwise assume Loc is clobbered.
330 return Start->getBlock() != End->getBlock() ||
331 any_of(
332 Range: make_range(x: std::next(x: Start->getIterator()), y: End->getIterator()),
333 P: [&AA, Loc](const MemoryAccess &Acc) {
334 if (isa<MemoryUse>(Val: &Acc))
335 return false;
336 Instruction *AccInst =
337 cast<MemoryUseOrDef>(Val: &Acc)->getMemoryInst();
338 return isModSet(MRI: AA.getModRefInfo(I: AccInst, OptLoc: Loc));
339 });
340 }
341
342 // TODO: Only walk until we hit Start.
343 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
344 End->getDefiningAccess(), Loc, AA);
345 return !MSSA->dominates(A: Clobber, B: Start);
346}
347
348/// When scanning forward over instructions, we look for some other patterns to
349/// fold away. In particular, this looks for stores to neighboring locations of
350/// memory. If it sees enough consecutive ones, it attempts to merge them
351/// together into a memcpy/memset.
352Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
353 Value *StartPtr,
354 Value *ByteVal) {
355 const DataLayout &DL = StartInst->getDataLayout();
356
357 // We can't track scalable types
358 if (auto *SI = dyn_cast<StoreInst>(Val: StartInst))
359 if (DL.getTypeStoreSize(Ty: SI->getOperand(i_nocapture: 0)->getType()).isScalable())
360 return nullptr;
361
362 // Okay, so we now have a single store that can be splatable. Scan to find
363 // all subsequent stores of the same value to offset from the same pointer.
364 // Join these together into ranges, so we can decide whether contiguous blocks
365 // are stored.
366 MemsetRanges Ranges(DL);
367
368 BasicBlock::iterator BI(StartInst);
369
370 // Keeps track of the last memory use or def before the insertion point for
371 // the new memset. The new MemoryDef for the inserted memsets will be inserted
372 // after MemInsertPoint.
373 MemoryUseOrDef *MemInsertPoint = nullptr;
374 for (++BI; !BI->isTerminator(); ++BI) {
375 auto *CurrentAcc =
376 cast_or_null<MemoryUseOrDef>(Val: MSSA->getMemoryAccess(I: &*BI));
377 if (CurrentAcc)
378 MemInsertPoint = CurrentAcc;
379
380 // Calls that only access inaccessible memory do not block merging
381 // accessible stores.
382 if (auto *CB = dyn_cast<CallBase>(Val&: BI)) {
383 if (CB->onlyAccessesInaccessibleMemory())
384 continue;
385 }
386
387 if (!isa<StoreInst>(Val: BI) && !isa<MemSetInst>(Val: BI)) {
388 // If the instruction is readnone, ignore it, otherwise bail out. We
389 // don't even allow readonly here because we don't want something like:
390 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
391 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
392 break;
393 continue;
394 }
395
396 if (auto *NextStore = dyn_cast<StoreInst>(Val&: BI)) {
397 // If this is a store, see if we can merge it in.
398 if (!NextStore->isSimple())
399 break;
400
401 Value *StoredVal = NextStore->getValueOperand();
402
403 // Don't convert stores of non-integral pointer types to memsets (which
404 // stores integers).
405 if (DL.isNonIntegralPointerType(Ty: StoredVal->getType()->getScalarType()))
406 break;
407
408 // We can't track ranges involving scalable types.
409 if (DL.getTypeStoreSize(Ty: StoredVal->getType()).isScalable())
410 break;
411
412 // Check to see if this stored value is of the same byte-splattable value.
413 Value *StoredByte = isBytewiseValue(V: StoredVal, DL);
414 // We can blindly merge this store into `StartInst` if it's being filled
415 // with an undef value but we don't because:
416 // 1. `StartInst` can be removed since it's storing an `undef`.
417 // 2. The resulting memset will be much larger than it needs to be.
418 if (ByteVal != StoredByte)
419 break;
420
421 // Check to see if this store is to a constant offset from the start ptr.
422 std::optional<int64_t> Offset =
423 NextStore->getPointerOperand()->getPointerOffsetFrom(Other: StartPtr, DL);
424 if (!Offset)
425 break;
426
427 Ranges.addStore(OffsetFromFirst: *Offset, SI: NextStore);
428 } else {
429 auto *MSI = cast<MemSetInst>(Val&: BI);
430
431 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
432 !isa<ConstantInt>(Val: MSI->getLength()))
433 break;
434
435 // Check to see if this store is to a constant offset from the start ptr.
436 std::optional<int64_t> Offset =
437 MSI->getDest()->getPointerOffsetFrom(Other: StartPtr, DL);
438 if (!Offset)
439 break;
440
441 Ranges.addMemSet(OffsetFromFirst: *Offset, MSI);
442 }
443 }
444
445 // If we have no ranges, then we just had a single store with nothing that
446 // could be merged in. This is a very common case of course.
447 if (Ranges.empty())
448 return nullptr;
449
450 // If we had at least one store that could be merged in, add the starting
451 // store as well. We try to avoid this unless there is at least something
452 // interesting as a small compile-time optimization.
453 Ranges.addInst(OffsetFromFirst: 0, Inst: StartInst);
454
455 // If we create any memsets, we put it right before the first instruction that
456 // isn't part of the memset block. This ensure that the memset is dominated
457 // by any addressing instruction needed by the start of the block.
458 IRBuilder<> Builder(&*BI);
459
460 // Now that we have full information about ranges, loop over the ranges and
461 // emit memset's for anything big enough to be worthwhile.
462 Instruction *AMemSet = nullptr;
463 for (const MemsetRange &Range : Ranges) {
464 if (Range.TheStores.size() == 1)
465 continue;
466
467 // If it is profitable to lower this range to memset, do so now.
468 if (!Range.isProfitableToUseMemset(DL))
469 continue;
470
471 // Otherwise, we do want to transform this! Create a new memset.
472 // Get the starting pointer of the block.
473 StartPtr = Range.StartPtr;
474
475 AMemSet = Builder.CreateMemSet(Ptr: StartPtr, Val: ByteVal, Size: Range.End - Range.Start,
476 Align: Range.Alignment);
477 AMemSet->mergeDIAssignID(SourceInstructions: Range.TheStores);
478
479 LLVM_DEBUG(dbgs() << "Replace stores:\n"; for (Instruction *SI
480 : Range.TheStores) dbgs()
481 << *SI << '\n';
482 dbgs() << "With: " << *AMemSet << '\n');
483 if (!Range.TheStores.empty())
484 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
485
486 auto *NewDef = cast<MemoryDef>(
487 Val: MemInsertPoint->getMemoryInst() == &*BI
488 ? MSSAU->createMemoryAccessBefore(I: AMemSet, Definition: nullptr, InsertPt: MemInsertPoint)
489 : MSSAU->createMemoryAccessAfter(I: AMemSet, Definition: nullptr, InsertPt: MemInsertPoint));
490 MSSAU->insertDef(Def: NewDef, /*RenameUses=*/true);
491 MemInsertPoint = NewDef;
492
493 // Zap all the stores.
494 for (Instruction *SI : Range.TheStores)
495 eraseInstruction(I: SI);
496
497 ++NumMemSetInfer;
498 }
499
500 return AMemSet;
501}
502
503// This method try to lift a store instruction before position P.
504// It will lift the store and its argument + that anything that
505// may alias with these.
506// The method returns true if it was successful.
507bool MemCpyOptPass::moveUp(StoreInst *SI, Instruction *P, const LoadInst *LI) {
508 // If the store alias this position, early bail out.
509 MemoryLocation StoreLoc = MemoryLocation::get(SI);
510 if (isModOrRefSet(MRI: AA->getModRefInfo(I: P, OptLoc: StoreLoc)))
511 return false;
512
513 // Keep track of the arguments of all instruction we plan to lift
514 // so we can make sure to lift them as well if appropriate.
515 DenseSet<Instruction *> Args;
516 auto AddArg = [&](Value *Arg) {
517 auto *I = dyn_cast<Instruction>(Val: Arg);
518 if (I && I->getParent() == SI->getParent()) {
519 // Cannot hoist user of P above P
520 if (I == P)
521 return false;
522 Args.insert(V: I);
523 }
524 return true;
525 };
526 if (!AddArg(SI->getPointerOperand()))
527 return false;
528
529 // Instruction to lift before P.
530 SmallVector<Instruction *, 8> ToLift{SI};
531
532 // Memory locations of lifted instructions.
533 SmallVector<MemoryLocation, 8> MemLocs{StoreLoc};
534
535 // Lifted calls.
536 SmallVector<const CallBase *, 8> Calls;
537
538 const MemoryLocation LoadLoc = MemoryLocation::get(LI);
539
540 for (auto I = --SI->getIterator(), E = P->getIterator(); I != E; --I) {
541 auto *C = &*I;
542
543 // Make sure hoisting does not perform a store that was not guaranteed to
544 // happen.
545 if (!isGuaranteedToTransferExecutionToSuccessor(I: C))
546 return false;
547
548 bool MayAlias = isModOrRefSet(MRI: AA->getModRefInfo(I: C, OptLoc: std::nullopt));
549
550 bool NeedLift = false;
551 if (Args.erase(V: C))
552 NeedLift = true;
553 else if (MayAlias) {
554 NeedLift = llvm::any_of(Range&: MemLocs, P: [C, this](const MemoryLocation &ML) {
555 return isModOrRefSet(MRI: AA->getModRefInfo(I: C, OptLoc: ML));
556 });
557
558 if (!NeedLift)
559 NeedLift = llvm::any_of(Range&: Calls, P: [C, this](const CallBase *Call) {
560 return isModOrRefSet(MRI: AA->getModRefInfo(I: C, Call));
561 });
562 }
563
564 if (!NeedLift)
565 continue;
566
567 if (MayAlias) {
568 // Since LI is implicitly moved downwards past the lifted instructions,
569 // none of them may modify its source.
570 if (isModSet(MRI: AA->getModRefInfo(I: C, OptLoc: LoadLoc)))
571 return false;
572 else if (const auto *Call = dyn_cast<CallBase>(Val: C)) {
573 // If we can't lift this before P, it's game over.
574 if (isModOrRefSet(MRI: AA->getModRefInfo(I: P, Call)))
575 return false;
576
577 Calls.push_back(Elt: Call);
578 } else if (isa<LoadInst>(Val: C) || isa<StoreInst>(Val: C) || isa<VAArgInst>(Val: C)) {
579 // If we can't lift this before P, it's game over.
580 auto ML = MemoryLocation::get(Inst: C);
581 if (isModOrRefSet(MRI: AA->getModRefInfo(I: P, OptLoc: ML)))
582 return false;
583
584 MemLocs.push_back(Elt: ML);
585 } else
586 // We don't know how to lift this instruction.
587 return false;
588 }
589
590 ToLift.push_back(Elt: C);
591 for (Value *Op : C->operands())
592 if (!AddArg(Op))
593 return false;
594 }
595
596 // Find MSSA insertion point. Normally P will always have a corresponding
597 // memory access before which we can insert. However, with non-standard AA
598 // pipelines, there may be a mismatch between AA and MSSA, in which case we
599 // will scan for a memory access before P. In either case, we know for sure
600 // that at least the load will have a memory access.
601 // TODO: Simplify this once P will be determined by MSSA, in which case the
602 // discrepancy can no longer occur.
603 MemoryUseOrDef *MemInsertPoint = nullptr;
604 if (MemoryUseOrDef *MA = MSSA->getMemoryAccess(I: P)) {
605 MemInsertPoint = cast<MemoryUseOrDef>(Val&: --MA->getIterator());
606 } else {
607 const Instruction *ConstP = P;
608 for (const Instruction &I : make_range(x: ++ConstP->getReverseIterator(),
609 y: ++LI->getReverseIterator())) {
610 if (MemoryUseOrDef *MA = MSSA->getMemoryAccess(I: &I)) {
611 MemInsertPoint = MA;
612 break;
613 }
614 }
615 }
616
617 // We made it, we need to lift.
618 for (auto *I : llvm::reverse(C&: ToLift)) {
619 LLVM_DEBUG(dbgs() << "Lifting " << *I << " before " << *P << "\n");
620 I->moveBefore(InsertPos: P->getIterator());
621 assert(MemInsertPoint && "Must have found insert point");
622 if (MemoryUseOrDef *MA = MSSA->getMemoryAccess(I)) {
623 MSSAU->moveAfter(What: MA, Where: MemInsertPoint);
624 MemInsertPoint = MA;
625 }
626 }
627
628 return true;
629}
630
631bool MemCpyOptPass::processStoreOfLoad(StoreInst *SI, LoadInst *LI,
632 const DataLayout &DL,
633 BasicBlock::iterator &BBI) {
634 if (!LI->isSimple() || !LI->hasOneUse() || LI->getParent() != SI->getParent())
635 return false;
636
637 BatchAAResults BAA(*AA, EEA);
638 auto *T = LI->getType();
639 // Don't introduce calls to memcpy/memmove intrinsics out of thin air if
640 // the corresponding libcalls are not available.
641 // TODO: We should really distinguish between libcall availability and
642 // our ability to introduce intrinsics.
643 if (T->isAggregateType() &&
644 (EnableMemCpyOptWithoutLibcalls ||
645 (TLI->has(F: LibFunc_memcpy) && TLI->has(F: LibFunc_memmove)))) {
646 MemoryLocation LoadLoc = MemoryLocation::get(LI);
647
648 // We use alias analysis to check if an instruction may store to
649 // the memory we load from in between the load and the store. If
650 // such an instruction is found, we try to promote there instead
651 // of at the store position.
652 // TODO: Can use MSSA for this.
653 Instruction *P = SI;
654 for (auto &I : make_range(x: ++LI->getIterator(), y: SI->getIterator())) {
655 if (isModSet(MRI: BAA.getModRefInfo(I: &I, OptLoc: LoadLoc))) {
656 P = &I;
657 break;
658 }
659 }
660
661 // If we found an instruction that may write to the loaded memory,
662 // we can try to promote at this position instead of the store
663 // position if nothing aliases the store memory after this and the store
664 // destination is not in the range.
665 if (P == SI || moveUp(SI, P, LI)) {
666 // If we load from memory that may alias the memory we store to,
667 // memmove must be used to preserve semantic. If not, memcpy can
668 // be used. Also, if we load from constant memory, memcpy can be used
669 // as the constant memory won't be modified.
670 bool UseMemMove = false;
671 if (isModSet(MRI: AA->getModRefInfo(I: SI, OptLoc: LoadLoc)))
672 UseMemMove = true;
673
674 IRBuilder<> Builder(P);
675 Value *Size =
676 Builder.CreateTypeSize(Ty: Builder.getInt64Ty(), Size: DL.getTypeStoreSize(Ty: T));
677 Instruction *M;
678 if (UseMemMove)
679 M = Builder.CreateMemMove(Dst: SI->getPointerOperand(), DstAlign: SI->getAlign(),
680 Src: LI->getPointerOperand(), SrcAlign: LI->getAlign(),
681 Size);
682 else
683 M = Builder.CreateMemCpy(Dst: SI->getPointerOperand(), DstAlign: SI->getAlign(),
684 Src: LI->getPointerOperand(), SrcAlign: LI->getAlign(), Size);
685 M->copyMetadata(SrcInst: *SI, WL: LLVMContext::MD_DIAssignID);
686
687 LLVM_DEBUG(dbgs() << "Promoting " << *LI << " to " << *SI << " => " << *M
688 << "\n");
689
690 auto *LastDef = cast<MemoryDef>(Val: MSSA->getMemoryAccess(I: SI));
691 auto *NewAccess = MSSAU->createMemoryAccessAfter(I: M, Definition: nullptr, InsertPt: LastDef);
692 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/true);
693
694 eraseInstruction(I: SI);
695 eraseInstruction(I: LI);
696 ++NumMemCpyInstr;
697
698 // Make sure we do not invalidate the iterator.
699 BBI = M->getIterator();
700 return true;
701 }
702 }
703
704 // Detect cases where we're performing call slot forwarding, but
705 // happen to be using a load-store pair to implement it, rather than
706 // a memcpy.
707 auto GetCall = [&]() -> CallInst * {
708 // We defer this expensive clobber walk until the cheap checks
709 // have been done on the source inside performCallSlotOptzn.
710 if (auto *LoadClobber = dyn_cast<MemoryUseOrDef>(
711 Val: MSSA->getWalker()->getClobberingMemoryAccess(I: LI, AA&: BAA)))
712 return dyn_cast_or_null<CallInst>(Val: LoadClobber->getMemoryInst());
713 return nullptr;
714 };
715
716 bool Changed = performCallSlotOptzn(
717 cpyLoad: LI, cpyStore: SI, cpyDst: SI->getPointerOperand()->stripPointerCasts(),
718 cpySrc: LI->getPointerOperand()->stripPointerCasts(),
719 cpyLen: DL.getTypeStoreSize(Ty: SI->getOperand(i_nocapture: 0)->getType()),
720 cpyAlign: std::min(a: SI->getAlign(), b: LI->getAlign()), BAA, GetC: GetCall);
721 if (Changed) {
722 eraseInstruction(I: SI);
723 eraseInstruction(I: LI);
724 ++NumMemCpyInstr;
725 return true;
726 }
727
728 // If this is a load-store pair from a stack slot to a stack slot, we
729 // might be able to perform the stack-move optimization just as we do for
730 // memcpys from an alloca to an alloca.
731 if (performStackMoveOptzn(Load: LI, Store: SI, DestPtr: SI->getPointerOperand(),
732 SrcPtr: LI->getPointerOperand(), Size: DL.getTypeStoreSize(Ty: T),
733 BAA)) {
734 // Avoid invalidating the iterator.
735 BBI = SI->getNextNode()->getIterator();
736 eraseInstruction(I: SI);
737 eraseInstruction(I: LI);
738 ++NumMemCpyInstr;
739 return true;
740 }
741
742 return false;
743}
744
745bool MemCpyOptPass::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
746 if (!SI->isSimple())
747 return false;
748
749 // Avoid merging nontemporal stores since the resulting
750 // memcpy/memset would not be able to preserve the nontemporal hint.
751 // In theory we could teach how to propagate the !nontemporal metadata to
752 // memset calls. However, that change would force the backend to
753 // conservatively expand !nontemporal memset calls back to sequences of
754 // store instructions (effectively undoing the merging).
755 if (SI->getMetadata(KindID: LLVMContext::MD_nontemporal))
756 return false;
757
758 const DataLayout &DL = SI->getDataLayout();
759
760 Value *StoredVal = SI->getValueOperand();
761
762 // Not all the transforms below are correct for non-integral pointers, bail
763 // until we've audited the individual pieces.
764 if (DL.isNonIntegralPointerType(Ty: StoredVal->getType()->getScalarType()))
765 return false;
766
767 // Load to store forwarding can be interpreted as memcpy.
768 if (auto *LI = dyn_cast<LoadInst>(Val: StoredVal))
769 return processStoreOfLoad(SI, LI, DL, BBI);
770
771 // The following code creates memset intrinsics out of thin air. Don't do
772 // this if the corresponding libfunc is not available.
773 // TODO: We should really distinguish between libcall availability and
774 // our ability to introduce intrinsics.
775 if (!(TLI->has(F: LibFunc_memset) || EnableMemCpyOptWithoutLibcalls))
776 return false;
777
778 // There are two cases that are interesting for this code to handle: memcpy
779 // and memset. Right now we only handle memset.
780
781 // Ensure that the value being stored is something that can be memset'able a
782 // byte at a time like "0" or "-1" or any width, as well as things like
783 // 0xA0A0A0A0 and 0.0.
784 Value *V = SI->getOperand(i_nocapture: 0);
785 Value *ByteVal = isBytewiseValue(V, DL);
786 if (!ByteVal)
787 return false;
788
789 if (Instruction *I =
790 tryMergingIntoMemset(StartInst: SI, StartPtr: SI->getPointerOperand(), ByteVal)) {
791 BBI = I->getIterator(); // Don't invalidate iterator.
792 return true;
793 }
794
795 // If we have an aggregate, we try to promote it to memset regardless
796 // of opportunity for merging as it can expose optimization opportunities
797 // in subsequent passes.
798 auto *T = V->getType();
799 if (!T->isAggregateType())
800 return false;
801
802 TypeSize Size = DL.getTypeStoreSize(Ty: T);
803 if (Size.isScalable())
804 return false;
805
806 IRBuilder<> Builder(SI);
807 auto *M = Builder.CreateMemSet(Ptr: SI->getPointerOperand(), Val: ByteVal, Size,
808 Align: SI->getAlign());
809 M->copyMetadata(SrcInst: *SI, WL: LLVMContext::MD_DIAssignID);
810
811 LLVM_DEBUG(dbgs() << "Promoting " << *SI << " to " << *M << "\n");
812
813 // The newly inserted memset is immediately overwritten by the original
814 // store, so we do not need to rename uses.
815 auto *StoreDef = cast<MemoryDef>(Val: MSSA->getMemoryAccess(I: SI));
816 auto *NewAccess = MSSAU->createMemoryAccessBefore(I: M, Definition: nullptr, InsertPt: StoreDef);
817 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/false);
818
819 eraseInstruction(I: SI);
820 NumMemSetInfer++;
821
822 // Make sure we do not invalidate the iterator.
823 BBI = M->getIterator();
824 return true;
825}
826
827bool MemCpyOptPass::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
828 // See if there is another memset or store neighboring this memset which
829 // allows us to widen out the memset to do a single larger store.
830 if (isa<ConstantInt>(Val: MSI->getLength()) && !MSI->isVolatile())
831 if (Instruction *I =
832 tryMergingIntoMemset(StartInst: MSI, StartPtr: MSI->getDest(), ByteVal: MSI->getValue())) {
833 BBI = I->getIterator(); // Don't invalidate iterator.
834 return true;
835 }
836 return false;
837}
838
839/// Takes a memcpy and a call that it depends on,
840/// and checks for the possibility of a call slot optimization by having
841/// the call write its result directly into the destination of the memcpy.
842bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
843 Instruction *cpyStore, Value *cpyDest,
844 Value *cpySrc, TypeSize cpySize,
845 Align cpyDestAlign,
846 BatchAAResults &BAA,
847 std::function<CallInst *()> GetC) {
848 // The general transformation to keep in mind is
849 //
850 // call @func(..., src, ...)
851 // memcpy(dest, src, ...)
852 //
853 // ->
854 //
855 // memcpy(dest, src, ...)
856 // call @func(..., dest, ...)
857 //
858 // Since moving the memcpy is technically awkward, we additionally check that
859 // src only holds uninitialized values at the moment of the call, meaning that
860 // the memcpy can be discarded rather than moved.
861
862 // We can't optimize scalable types.
863 if (cpySize.isScalable())
864 return false;
865
866 // Require that src be an alloca. This simplifies the reasoning considerably.
867 auto *srcAlloca = dyn_cast<AllocaInst>(Val: cpySrc);
868 if (!srcAlloca)
869 return false;
870
871 const DataLayout &DL = cpyLoad->getDataLayout();
872 // We can't optimize scalable types or variable-length allocas.
873 std::optional<TypeSize> SrcAllocaSize = srcAlloca->getAllocationSize(DL);
874 if (!SrcAllocaSize || SrcAllocaSize->isScalable())
875 return false;
876 uint64_t srcSize = SrcAllocaSize->getFixedValue();
877
878 if (cpySize < srcSize)
879 return false;
880
881 CallInst *C = GetC();
882 if (!C)
883 return false;
884
885 // Lifetime marks shouldn't be operated on.
886 if (Function *F = C->getCalledFunction())
887 if (F->isIntrinsic() && F->getIntrinsicID() == Intrinsic::lifetime_start)
888 return false;
889
890 if (C->getParent() != cpyStore->getParent()) {
891 LLVM_DEBUG(dbgs() << "Call Slot: block local restriction\n");
892 return false;
893 }
894
895 MemoryLocation DestLoc =
896 isa<StoreInst>(Val: cpyStore)
897 ? MemoryLocation::get(Inst: cpyStore)
898 : MemoryLocation::getForDest(MI: cast<MemCpyInst>(Val: cpyStore));
899
900 // Check that nothing touches the dest of the copy between
901 // the call and the store/memcpy.
902 Instruction *SkippedLifetimeStart = nullptr;
903 if (accessedBetween(AA&: BAA, Loc: DestLoc, Start: MSSA->getMemoryAccess(I: C),
904 End: MSSA->getMemoryAccess(I: cpyStore), SkippedLifetimeStart: &SkippedLifetimeStart)) {
905 LLVM_DEBUG(dbgs() << "Call Slot: Dest pointer modified after call\n");
906 return false;
907 }
908
909 // If we need to move a lifetime.start above the call, make sure that we can
910 // actually do so. If the argument is bitcasted for example, we would have to
911 // move the bitcast as well, which we don't handle.
912 if (SkippedLifetimeStart) {
913 auto *LifetimeArg =
914 dyn_cast<Instruction>(Val: SkippedLifetimeStart->getOperand(i: 0));
915 if (LifetimeArg && LifetimeArg->getParent() == C->getParent() &&
916 C->comesBefore(Other: LifetimeArg))
917 return false;
918 }
919
920 // Check that storing to the first srcSize bytes of dest will not cause a
921 // trap or data race.
922 bool ExplicitlyDereferenceableOnly;
923 if (!isWritableObject(Object: getUnderlyingObject(V: cpyDest),
924 ExplicitlyDereferenceableOnly) ||
925 !isDereferenceablePointer(V: cpyDest, Size: APInt(64, cpySize),
926 Q: SimplifyQuery(DL, DT, AC, C))) {
927 // If the call is guaranteed to return normally (willreturn + nounwind),
928 // and there are no instructions between the call and the store that might
929 // trap or throw, execution will reach the store. Since the store would
930 // trap anyway if the pointer was not dereferenceable, we can forward the
931 // pointer to the call. Perform optimization only for non-memcpy/memset
932 // calls, as those are special cased later.
933 if (!isGuaranteedToTransferExecutionToSuccessor(Begin: C->getIterator(),
934 End: cpyStore->getIterator())) {
935 LLVM_DEBUG(dbgs() << "Call Slot: Dest pointer not dereferenceable\n");
936 return false;
937 }
938 }
939
940 // Make sure that nothing can observe cpyDest being written early. There are
941 // a number of cases to consider:
942 // 1. cpyDest cannot be accessed between C and cpyStore as a precondition of
943 // the transform.
944 // 2. C itself may not access cpyDest (prior to the transform). This is
945 // checked further below.
946 // 3. If cpyDest is accessible to the caller of this function (potentially
947 // captured and not based on an alloca), we need to ensure that we cannot
948 // unwind between C and cpyStore. This is checked here.
949 // 4. If cpyDest is potentially captured, there may be accesses to it from
950 // another thread. In this case, we need to check that cpyStore is
951 // guaranteed to be executed if C is. As it is a non-atomic access, it
952 // renders accesses from other threads undefined.
953 // TODO: This is currently not checked.
954 if (mayBeVisibleThroughUnwinding(V: cpyDest, Start: C, End: cpyStore)) {
955 LLVM_DEBUG(dbgs() << "Call Slot: Dest may be visible through unwinding\n");
956 return false;
957 }
958
959 // Check that dest points to memory that is at least as aligned as src.
960 Align srcAlign = srcAlloca->getAlign();
961 bool isDestSufficientlyAligned = srcAlign <= cpyDestAlign;
962 // If dest is not aligned enough and we can't increase its alignment then
963 // bail out.
964 if (!isDestSufficientlyAligned && !isa<AllocaInst>(Val: cpyDest)) {
965 LLVM_DEBUG(dbgs() << "Call Slot: Dest not sufficiently aligned\n");
966 return false;
967 }
968
969 // Check that src is not accessed except via the call and the memcpy. This
970 // guarantees that it holds only undefined values when passed in (so the final
971 // memcpy can be dropped), that it is not read or written between the call and
972 // the memcpy, and that writing beyond the end of it is undefined.
973 SmallVector<User *, 8> srcUseList(srcAlloca->users());
974 while (!srcUseList.empty()) {
975 User *U = srcUseList.pop_back_val();
976
977 if (isa<AddrSpaceCastInst>(Val: U)) {
978 append_range(C&: srcUseList, R: U->users());
979 continue;
980 }
981 if (isa<LifetimeIntrinsic>(Val: U))
982 continue;
983
984 if (U != C && U != cpyLoad) {
985 LLVM_DEBUG(dbgs() << "Call slot: Source accessed by " << *U << "\n");
986 return false;
987 }
988 }
989
990 // Check whether src is captured by the called function, in which case there
991 // may be further indirect uses of src.
992 bool SrcIsCaptured = any_of(Range: C->args(), P: [&](Use &U) {
993 return U->stripPointerCasts() == cpySrc &&
994 !C->doesNotCapture(OpNo: C->getArgOperandNo(U: &U));
995 });
996
997 // If src is captured, then check whether there are any potential uses of
998 // src through the captured pointer before the lifetime of src ends, either
999 // due to a lifetime.end or a return from the function.
1000 if (SrcIsCaptured) {
1001 // Check that dest is not captured before/at the call. We have already
1002 // checked that src is not captured before it. If either had been captured,
1003 // then the call might be comparing the argument against the captured dest
1004 // or src pointer.
1005 Value *DestObj = getUnderlyingObject(V: cpyDest);
1006 if (!isIdentifiedFunctionLocal(V: DestObj) ||
1007 PointerMayBeCapturedBefore(V: DestObj, /* ReturnCaptures */ true, I: C, DT,
1008 /* IncludeI */ true))
1009 return false;
1010
1011 MemoryLocation SrcLoc =
1012 MemoryLocation(srcAlloca, LocationSize::precise(Value: srcSize));
1013 for (Instruction &I :
1014 make_range(x: ++C->getIterator(), y: C->getParent()->end())) {
1015 // Lifetime of srcAlloca ends at lifetime.end.
1016 if (auto *II = dyn_cast<IntrinsicInst>(Val: &I)) {
1017 if (II->getIntrinsicID() == Intrinsic::lifetime_end &&
1018 II->getArgOperand(i: 0) == srcAlloca)
1019 break;
1020 }
1021
1022 // Lifetime of srcAlloca ends at return.
1023 if (isa<ReturnInst>(Val: &I))
1024 break;
1025
1026 // Ignore the direct read of src in the load.
1027 if (&I == cpyLoad)
1028 continue;
1029
1030 // Check whether this instruction may mod/ref src through the captured
1031 // pointer (we have already any direct mod/refs in the loop above).
1032 // Also bail if we hit a terminator, as we don't want to scan into other
1033 // blocks.
1034 if (isModOrRefSet(MRI: BAA.getModRefInfo(I: &I, OptLoc: SrcLoc)) || I.isTerminator())
1035 return false;
1036 }
1037 }
1038
1039 // Since we're changing the parameter to the callsite, we need to make sure
1040 // that what would be the new parameter dominates the callsite.
1041 bool NeedMoveGEP = false;
1042 if (!DT->dominates(Def: cpyDest, User: C)) {
1043 // Support moving a constant index GEP before the call.
1044 auto *GEP = dyn_cast<GetElementPtrInst>(Val: cpyDest);
1045 if (GEP && GEP->hasAllConstantIndices() &&
1046 DT->dominates(Def: GEP->getPointerOperand(), User: C))
1047 NeedMoveGEP = true;
1048 else
1049 return false;
1050 }
1051
1052 // In addition to knowing that the call does not access src in some
1053 // unexpected manner, for example via a global, which we deduce from
1054 // the use analysis, we also need to know that it does not sneakily
1055 // access dest. We rely on AA to figure this out for us.
1056 MemoryLocation DestWithSrcSize(cpyDest, LocationSize::precise(Value: srcSize));
1057 ModRefInfo MR = BAA.getModRefInfo(I: C, OptLoc: DestWithSrcSize);
1058 // If necessary, perform additional analysis.
1059 if (isModOrRefSet(MRI: MR))
1060 MR = BAA.callCapturesBefore(I: C, MemLoc: DestWithSrcSize, DT);
1061 if (isModOrRefSet(MRI: MR))
1062 return false;
1063
1064 // We can't create address space casts here because we don't know if they're
1065 // safe for the target.
1066 if (cpySrc->getType() != cpyDest->getType())
1067 return false;
1068 for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
1069 if (C->getArgOperand(i: ArgI)->stripPointerCasts() == cpySrc &&
1070 cpySrc->getType() != C->getArgOperand(i: ArgI)->getType())
1071 return false;
1072
1073 // All the checks have passed, so do the transformation.
1074 bool changedArgument = false;
1075 for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
1076 if (C->getArgOperand(i: ArgI)->stripPointerCasts() == cpySrc) {
1077 changedArgument = true;
1078 C->setArgOperand(i: ArgI, v: cpyDest);
1079 }
1080
1081 if (!changedArgument)
1082 return false;
1083
1084 // If the destination wasn't sufficiently aligned then increase its alignment.
1085 if (!isDestSufficientlyAligned) {
1086 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!");
1087 AllocaInst *DestAlloca = cast<AllocaInst>(Val: cpyDest);
1088 DestAlloca->setAlignment(std::max(a: DestAlloca->getAlign(), b: srcAlign));
1089 }
1090
1091 if (NeedMoveGEP) {
1092 auto *GEP = dyn_cast<GetElementPtrInst>(Val: cpyDest);
1093 GEP->moveBefore(InsertPos: C->getIterator());
1094 }
1095
1096 if (SkippedLifetimeStart) {
1097 SkippedLifetimeStart->moveBefore(InsertPos: C->getIterator());
1098 MSSAU->moveBefore(What: MSSA->getMemoryAccess(I: SkippedLifetimeStart),
1099 Where: MSSA->getMemoryAccess(I: C));
1100 }
1101
1102 combineAAMetadata(K: C, J: cpyLoad);
1103 if (cpyLoad != cpyStore)
1104 combineAAMetadata(K: C, J: cpyStore);
1105
1106 ++NumCallSlot;
1107 return true;
1108}
1109
1110/// We've found that the (upward scanning) memory dependence of memcpy 'M' is
1111/// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
1112bool MemCpyOptPass::processMemCpyMemCpyDependence(MemCpyInst *M,
1113 MemCpyInst *MDep,
1114 BatchAAResults &BAA) {
1115 // We can only optimize non-volatile memcpy's.
1116 if (MDep->isVolatile())
1117 return false;
1118
1119 // If dep instruction is reading from our current input, then it is a noop
1120 // transfer and substituting the input won't change this instruction. Just
1121 // ignore the input and let someone else zap MDep. This handles cases like:
1122 // memcpy(a <- a)
1123 // memcpy(b <- a)
1124 // This also avoids infinite loops.
1125 if (BAA.isMustAlias(V1: MDep->getDest(), V2: MDep->getSource()))
1126 return false;
1127
1128 int64_t MForwardOffset = 0;
1129 const DataLayout &DL = M->getModule()->getDataLayout();
1130 // We can only transforms memcpy's where the dest of one is the source of the
1131 // other, or they have an offset in a range.
1132 if (M->getSource() != MDep->getDest()) {
1133 std::optional<int64_t> Offset =
1134 M->getSource()->getPointerOffsetFrom(Other: MDep->getDest(), DL);
1135 if (!Offset || *Offset < 0)
1136 return false;
1137 MForwardOffset = *Offset;
1138 }
1139
1140 Value *CopyLength = M->getLength();
1141
1142 // The length of the memcpy's must be the same, or the preceding one must be
1143 // larger than the following one, or the contents of the overread must be
1144 // undefined bytes of a defined size.
1145 if (MForwardOffset != 0 || MDep->getLength() != CopyLength) {
1146 auto *MDepLen = dyn_cast<ConstantInt>(Val: MDep->getLength());
1147 auto *MLen = dyn_cast<ConstantInt>(Val: CopyLength);
1148 // This could be converted to a runtime test (%CopyLength =
1149 // min(max(0, MDepLen - MForwardOffset), MLen)), but it is
1150 // unclear if that is useful
1151 if (!MDepLen || !MLen)
1152 return false;
1153 if (MDepLen->getZExtValue() < MLen->getZExtValue() + MForwardOffset) {
1154 if (!overreadUndefContents(MSSA, MemCpy: M, MemSrc: MDep, BAA))
1155 return false;
1156 if (MDepLen->getZExtValue() <= (uint64_t)MForwardOffset)
1157 return false; // Should not reach here (there is obviously no aliasing
1158 // with MDep), so just bail in case it had incomplete info
1159 // somehow
1160 CopyLength = ConstantInt::get(Ty: CopyLength->getType(),
1161 V: MDepLen->getZExtValue() - MForwardOffset);
1162 }
1163 }
1164
1165 IRBuilder<> Builder(M);
1166 auto *CopySource = MDep->getSource();
1167 Instruction *NewCopySource = nullptr;
1168 llvm::scope_exit CleanupOnRet([&] {
1169 if (NewCopySource && NewCopySource->use_empty())
1170 // Safety: It's safe here because we will only allocate more instructions
1171 // after finishing all BatchAA queries, but we have to be careful if we
1172 // want to do something like this in another place. Then we'd probably
1173 // have to delay instruction removal until all transforms on an
1174 // instruction finished.
1175 eraseInstruction(I: NewCopySource);
1176 });
1177 MaybeAlign CopySourceAlign = MDep->getSourceAlign();
1178 auto MCopyLoc = MemoryLocation::getForSource(MTI: MDep);
1179 // Truncate the size of the MDep access to just the bytes read
1180 if (MDep->getLength() != CopyLength) {
1181 auto *ConstLength = cast<ConstantInt>(Val: CopyLength);
1182 MCopyLoc = MCopyLoc.getWithNewSize(
1183 NewSize: LocationSize::precise(Value: ConstLength->getZExtValue()));
1184 }
1185
1186 // When the forwarding offset is greater than 0, we transform
1187 // memcpy(d1 <- s1)
1188 // memcpy(d2 <- d1+o)
1189 // to
1190 // memcpy(d2 <- s1+o)
1191 if (MForwardOffset > 0) {
1192 // The copy destination of `M` maybe can serve as the source of copying.
1193 std::optional<int64_t> MDestOffset =
1194 M->getRawDest()->getPointerOffsetFrom(Other: MDep->getRawSource(), DL);
1195 if (MDestOffset == MForwardOffset)
1196 CopySource = M->getDest();
1197 else {
1198 CopySource = Builder.CreateInBoundsPtrAdd(
1199 Ptr: CopySource, Offset: Builder.getInt64(C: MForwardOffset));
1200 NewCopySource = dyn_cast<Instruction>(Val: CopySource);
1201 }
1202 // We need to update `MCopyLoc` if an offset exists.
1203 MCopyLoc = MCopyLoc.getWithNewPtr(NewPtr: CopySource);
1204 if (CopySourceAlign)
1205 CopySourceAlign = commonAlignment(A: *CopySourceAlign, Offset: MForwardOffset);
1206 }
1207
1208 // Verify that the copied-from memory doesn't change in between the two
1209 // transfers. For example, in:
1210 // memcpy(a <- b)
1211 // *b = 42;
1212 // memcpy(c <- a)
1213 // It would be invalid to transform the second memcpy into memcpy(c <- b).
1214 //
1215 // TODO: If the code between M and MDep is transparent to the destination "c",
1216 // then we could still perform the xform by moving M up to the first memcpy.
1217 if (writtenBetween(MSSA, AA&: BAA, Loc: MCopyLoc, Start: MSSA->getMemoryAccess(I: MDep),
1218 End: MSSA->getMemoryAccess(I: M)))
1219 return false;
1220
1221 // No need to create `memcpy(a <- a)`.
1222 if (BAA.isMustAlias(V1: M->getDest(), V2: CopySource)) {
1223 // Remove the instruction we're replacing.
1224 eraseInstruction(I: M);
1225 ++NumMemCpyInstr;
1226 return true;
1227 }
1228
1229 // If the dest of the second might alias the source of the first, then the
1230 // source and dest might overlap. In addition, if the source of the first
1231 // points to constant memory, they won't overlap by definition. Otherwise, we
1232 // still want to eliminate the intermediate value, but we have to generate a
1233 // memmove instead of memcpy.
1234 bool UseMemMove = false;
1235 if (isModSet(MRI: BAA.getModRefInfo(I: M, OptLoc: MemoryLocation::getForSource(MTI: MDep)))) {
1236 // Don't convert llvm.memcpy.inline into memmove because memmove can be
1237 // lowered as a call, and that is not allowed for llvm.memcpy.inline (and
1238 // there is no inline version of llvm.memmove)
1239 if (M->isForceInlined())
1240 return false;
1241 UseMemMove = true;
1242 }
1243
1244 // If all checks passed, then we can transform M.
1245 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
1246 << *MDep << '\n'
1247 << *M << '\n');
1248
1249 // TODO: Is this worth it if we're creating a less aligned memcpy? For
1250 // example we could be moving from movaps -> movq on x86.
1251 Instruction *NewM;
1252 if (UseMemMove)
1253 NewM = Builder.CreateMemMove(Dst: M->getDest(), DstAlign: M->getDestAlign(), Src: CopySource,
1254 SrcAlign: CopySourceAlign, Size: CopyLength, isVolatile: M->isVolatile());
1255 else if (M->isForceInlined())
1256 // llvm.memcpy may be promoted to llvm.memcpy.inline, but the converse is
1257 // never allowed since that would allow the latter to be lowered as a call
1258 // to an external function.
1259 NewM = Builder.CreateMemCpyInline(Dst: M->getDest(), DstAlign: M->getDestAlign(),
1260 Src: CopySource, SrcAlign: CopySourceAlign, Size: CopyLength,
1261 isVolatile: M->isVolatile());
1262 else
1263 NewM = Builder.CreateMemCpy(Dst: M->getDest(), DstAlign: M->getDestAlign(), Src: CopySource,
1264 SrcAlign: CopySourceAlign, Size: CopyLength, isVolatile: M->isVolatile());
1265
1266 NewM->copyMetadata(SrcInst: *M, WL: LLVMContext::MD_DIAssignID);
1267
1268 assert(isa<MemoryDef>(MSSA->getMemoryAccess(M)));
1269 auto *LastDef = cast<MemoryDef>(Val: MSSA->getMemoryAccess(I: M));
1270 auto *NewAccess = MSSAU->createMemoryAccessAfter(I: NewM, Definition: nullptr, InsertPt: LastDef);
1271 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/true);
1272
1273 // Remove the instruction we're replacing.
1274 eraseInstruction(I: M);
1275 ++NumMemCpyInstr;
1276 return true;
1277}
1278
1279/// We've found that the (upward scanning) memory dependence of \p MemCpy is
1280/// \p MemSet. Try to simplify \p MemSet to only set the trailing bytes that
1281/// weren't copied over by \p MemCpy.
1282///
1283/// In other words, transform:
1284/// \code
1285/// memset(dst, c, dst_size);
1286/// ...
1287/// memcpy(dst, src, src_size);
1288/// \endcode
1289/// into:
1290/// \code
1291/// ...
1292/// memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
1293/// memcpy(dst, src, src_size);
1294/// \endcode
1295///
1296/// The memset is sunk to just before the memcpy to ensure that src_size is
1297/// present when emitting the simplified memset.
1298bool MemCpyOptPass::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
1299 MemSetInst *MemSet,
1300 BatchAAResults &BAA) {
1301 // We can only transform memset/memcpy with the same destination.
1302 if (!BAA.isMustAlias(V1: MemSet->getDest(), V2: MemCpy->getDest()))
1303 return false;
1304
1305 if (MemSet->isVolatile())
1306 return false;
1307
1308 // Don't perform the transform if src_size may be zero. In that case, the
1309 // transform is essentially a complex no-op and may lead to an infinite
1310 // loop if BasicAA is smart enough to understand that dst and dst + src_size
1311 // are still MustAlias after the transform.
1312 Value *SrcSize = MemCpy->getLength();
1313 if (!isKnownNonZero(V: SrcSize,
1314 Q: SimplifyQuery(MemCpy->getDataLayout(), DT, AC, MemCpy)))
1315 return false;
1316
1317 // Check that src and dst of the memcpy aren't the same. While memcpy
1318 // operands cannot partially overlap, exact equality is allowed.
1319 if (isModSet(MRI: BAA.getModRefInfo(I: MemCpy, OptLoc: MemoryLocation::getForSource(MTI: MemCpy))))
1320 return false;
1321
1322 // We know that dst up to src_size is not written. We now need to make sure
1323 // that dst up to dst_size is not accessed. (If we did not move the memset,
1324 // checking for reads would be sufficient.)
1325 if (accessedBetween(AA&: BAA, Loc: MemoryLocation::getForDest(MI: MemSet),
1326 Start: MSSA->getMemoryAccess(I: MemSet),
1327 End: MSSA->getMemoryAccess(I: MemCpy)))
1328 return false;
1329
1330 // Use the same i8* dest as the memcpy, killing the memset dest if different.
1331 Value *Dest = MemCpy->getRawDest();
1332 Value *DestSize = MemSet->getLength();
1333
1334 if (mayBeVisibleThroughUnwinding(V: Dest, Start: MemSet, End: MemCpy))
1335 return false;
1336
1337 // If the sizes are the same, simply drop the memset instead of generating
1338 // a replacement with zero size.
1339 if (DestSize == SrcSize) {
1340 eraseInstruction(I: MemSet);
1341 return true;
1342 }
1343
1344 // By default, create an unaligned memset.
1345 Align Alignment = Align(1);
1346 // If Dest is aligned, and SrcSize is constant, use the minimum alignment
1347 // of the sum.
1348 const Align DestAlign = std::max(a: MemSet->getDestAlign().valueOrOne(),
1349 b: MemCpy->getDestAlign().valueOrOne());
1350 if (DestAlign > 1)
1351 if (auto *SrcSizeC = dyn_cast<ConstantInt>(Val: SrcSize))
1352 Alignment = commonAlignment(A: DestAlign, Offset: SrcSizeC->getZExtValue());
1353
1354 IRBuilder<> Builder(MemCpy);
1355
1356 // Preserve the debug location of the old memset for the code emitted here
1357 // related to the new memset. This is correct according to the rules in
1358 // https://llvm.org/docs/HowToUpdateDebugInfo.html about "when to preserve an
1359 // instruction location", given that we move the memset within the basic
1360 // block.
1361 assert(MemSet->getParent() == MemCpy->getParent() &&
1362 "Preserving debug location based on moving memset within BB.");
1363 Builder.SetCurrentDebugLocation(MemSet->getDebugLoc());
1364
1365 // If the sizes have different types, zext the smaller one.
1366 if (DestSize->getType() != SrcSize->getType()) {
1367 if (DestSize->getType()->getIntegerBitWidth() >
1368 SrcSize->getType()->getIntegerBitWidth())
1369 SrcSize = Builder.CreateZExt(V: SrcSize, DestTy: DestSize->getType());
1370 else
1371 DestSize = Builder.CreateZExt(V: DestSize, DestTy: SrcSize->getType());
1372 }
1373
1374 Value *Ule = Builder.CreateICmpULE(LHS: DestSize, RHS: SrcSize);
1375 Value *SizeDiff = Builder.CreateSub(LHS: DestSize, RHS: SrcSize);
1376 Value *MemsetLen = Builder.CreateSelect(
1377 C: Ule, True: ConstantInt::getNullValue(Ty: DestSize->getType()), False: SizeDiff);
1378 // FIXME (#167968): we could explore estimating the branch_weights based on
1379 // value profiling data about the 2 sizes.
1380 if (auto *SI = dyn_cast<SelectInst>(Val: MemsetLen))
1381 setExplicitlyUnknownBranchWeightsIfProfiled(I&: *SI, DEBUG_TYPE);
1382 Instruction *NewMemSet =
1383 Builder.CreateMemSet(Ptr: Builder.CreatePtrAdd(Ptr: Dest, Offset: SrcSize),
1384 Val: MemSet->getOperand(i_nocapture: 1), Size: MemsetLen, Align: Alignment);
1385
1386 assert(isa<MemoryDef>(MSSA->getMemoryAccess(MemCpy)) &&
1387 "MemCpy must be a MemoryDef");
1388 // The new memset is inserted before the memcpy, and it is known that the
1389 // memcpy's defining access is the memset about to be removed.
1390 auto *LastDef = cast<MemoryDef>(Val: MSSA->getMemoryAccess(I: MemCpy));
1391 auto *NewAccess =
1392 MSSAU->createMemoryAccessBefore(I: NewMemSet, Definition: nullptr, InsertPt: LastDef);
1393 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/true);
1394
1395 eraseInstruction(I: MemSet);
1396 return true;
1397}
1398
1399/// Determine whether the pointer V had only undefined content (due to Def),
1400/// either because it was freshly alloca'd or started its lifetime.
1401static bool hasUndefContents(MemorySSA *MSSA, BatchAAResults &AA, Value *V,
1402 MemoryDef *Def) {
1403 if (MSSA->isLiveOnEntryDef(MA: Def))
1404 return isa<AllocaInst>(Val: getUnderlyingObject(V));
1405
1406 if (auto *II = dyn_cast_or_null<IntrinsicInst>(Val: Def->getMemoryInst()))
1407 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1408 if (auto *Alloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V)))
1409 return II->getArgOperand(i: 0) == Alloca;
1410
1411 return false;
1412}
1413
1414// If the memcpy is larger than the previous, but the memory was undef prior to
1415// that, we can just ignore the tail. Technically we're only interested in the
1416// bytes from 0..MemSrcOffset and MemSrcLength+MemSrcOffset..CopySize here, but
1417// as we can't easily represent this location (hasUndefContents uses mustAlias
1418// which cannot deal with offsets), we use the full 0..CopySize range.
1419static bool overreadUndefContents(MemorySSA *MSSA, MemCpyInst *MemCpy,
1420 MemIntrinsic *MemSrc, BatchAAResults &BAA) {
1421 MemoryLocation MemCpyLoc = MemoryLocation::getForSource(MTI: MemCpy);
1422 MemoryUseOrDef *MemSrcAccess = MSSA->getMemoryAccess(I: MemSrc);
1423 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
1424 MemSrcAccess->getDefiningAccess(), MemCpyLoc, AA&: BAA);
1425 if (auto *MD = dyn_cast<MemoryDef>(Val: Clobber))
1426 if (hasUndefContents(MSSA, AA&: BAA, V: MemCpy->getSource(), Def: MD))
1427 return true;
1428 return false;
1429}
1430
1431/// Transform memcpy to memset when its source was just memset.
1432/// In other words, turn:
1433/// \code
1434/// memset(dst1, c, dst1_size);
1435/// memcpy(dst2, dst1, dst2_size);
1436/// \endcode
1437/// into:
1438/// \code
1439/// memset(dst1, c, dst1_size);
1440/// memset(dst2, c, dst2_size);
1441/// \endcode
1442bool MemCpyOptPass::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
1443 MemSetInst *MemSet,
1444 BatchAAResults &BAA) {
1445 Value *MemSetSize = MemSet->getLength();
1446 Value *CopySize = MemCpy->getLength();
1447
1448 int64_t MOffset = 0;
1449 const DataLayout &DL = MemCpy->getModule()->getDataLayout();
1450 // We can only transforms memcpy's where the dest of one is the source of the
1451 // other, or they have a known offset.
1452 if (MemCpy->getSource() != MemSet->getDest()) {
1453 std::optional<int64_t> Offset =
1454 MemCpy->getSource()->getPointerOffsetFrom(Other: MemSet->getDest(), DL);
1455 if (!Offset)
1456 return false;
1457 // On positive offsets, the memcpy source is at a offset into the memset'd
1458 // region. On negative offsets, the copy starts at a offset prior to the
1459 // previously memset'd area, namely, we memcpy from a partially initialized
1460 // region.
1461 MOffset = *Offset;
1462 }
1463
1464 if (MOffset != 0 || MemSetSize != CopySize) {
1465 // Make sure the memcpy doesn't read any more than what the memset wrote,
1466 // other than undef. Likewise, the memcpy should not read from an area not
1467 // covered by the memset unless undef bytes. Don't worry about sizes larger
1468 // than i64.
1469 auto *CMemSetSize = dyn_cast<ConstantInt>(Val: MemSetSize);
1470 auto *CCopySize = dyn_cast<ConstantInt>(Val: CopySize);
1471 if (!CMemSetSize || !CCopySize || MOffset < 0 ||
1472 CCopySize->getZExtValue() + MOffset > CMemSetSize->getZExtValue()) {
1473 if (!overreadUndefContents(MSSA, MemCpy, MemSrc: MemSet, BAA))
1474 return false;
1475
1476 if (CMemSetSize && CCopySize) {
1477 uint64_t MemSetSizeVal = CMemSetSize->getZExtValue();
1478 uint64_t MemCpySizeVal = CCopySize->getZExtValue();
1479 uint64_t NewSize;
1480
1481 if (MOffset < 0) {
1482 // Offset from beginning of the initialized region.
1483 uint64_t Offset = -MOffset;
1484 NewSize = MemCpySizeVal <= Offset ? 0 : MemCpySizeVal - Offset;
1485 } else if (MOffset == 0) {
1486 NewSize = MemSetSizeVal;
1487 } else {
1488 NewSize =
1489 MemSetSizeVal <= (uint64_t)MOffset ? 0 : MemSetSizeVal - MOffset;
1490 }
1491 CopySize = ConstantInt::get(Ty: CopySize->getType(), V: NewSize);
1492 } else {
1493 if (MOffset < 0)
1494 return false;
1495 }
1496 }
1497 }
1498
1499 IRBuilder<> Builder(MemCpy);
1500 Value *DestPtr = MemCpy->getRawDest();
1501 MaybeAlign Align = MemCpy->getDestAlign();
1502 if (MOffset < 0) {
1503 DestPtr = Builder.CreatePtrAdd(Ptr: DestPtr, Offset: Builder.getInt64(C: -MOffset));
1504 if (Align)
1505 Align = commonAlignment(A: *Align, Offset: -MOffset);
1506 }
1507
1508 Instruction *NewM =
1509 Builder.CreateMemSet(Ptr: DestPtr, Val: MemSet->getOperand(i_nocapture: 1), Size: CopySize, Align);
1510 auto *LastDef = cast<MemoryDef>(Val: MSSA->getMemoryAccess(I: MemCpy));
1511 auto *NewAccess = MSSAU->createMemoryAccessAfter(I: NewM, Definition: nullptr, InsertPt: LastDef);
1512 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/true);
1513
1514 return true;
1515}
1516
1517// Attempts to optimize the pattern whereby memory is copied from an alloca to
1518// another alloca, where the two allocas don't have conflicting mod/ref. If
1519// successful, the two allocas can be merged into one and the transfer can be
1520// deleted. This pattern is generated frequently in Rust, due to the ubiquity of
1521// move operations in that language.
1522//
1523// Once we determine that the optimization is safe to perform, we replace all
1524// uses of the destination alloca with the source alloca. We also "shrink wrap"
1525// the lifetime markers of the single merged alloca to before the first use
1526// and after the last use. Note that the "shrink wrapping" procedure is a safe
1527// transformation only because we restrict the scope of this optimization to
1528// allocas that aren't captured.
1529bool MemCpyOptPass::performStackMoveOptzn(Instruction *Load, Instruction *Store,
1530 Value *DestPtr, Value *SrcPtr,
1531 TypeSize Size, BatchAAResults &BAA) {
1532 LLVM_DEBUG(dbgs() << "Stack Move: Attempting to optimize:\n"
1533 << *Store << "\n");
1534
1535 AllocaInst *DestAlloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: DestPtr));
1536 if (!DestAlloca)
1537 return false;
1538
1539 AllocaInst *SrcAlloca = dyn_cast<AllocaInst>(Val: getUnderlyingObject(V: SrcPtr));
1540 if (!SrcAlloca)
1541 return false;
1542
1543 // Explicitly don't handle degenerate case of a partial copy within one
1544 // alloca. It would always fail the dominator check later anyways, and
1545 // possibly the modref checks also.
1546 if (SrcAlloca == DestAlloca)
1547 return false;
1548
1549 // Make sure the two allocas are in the same address space.
1550 if (SrcAlloca->getAddressSpace() != DestAlloca->getAddressSpace()) {
1551 LLVM_DEBUG(dbgs() << "Stack Move: Address space mismatch\n");
1552 return false;
1553 }
1554
1555 if (!SrcAlloca->isStaticAlloca() || !DestAlloca->isStaticAlloca())
1556 return false;
1557
1558 // Check that copy is full with static size.
1559 const DataLayout &DL = DestAlloca->getDataLayout();
1560
1561 auto DestOffset = DestPtr->getPointerOffsetFrom(Other: DestAlloca, DL);
1562 if (!DestOffset)
1563 return false;
1564
1565 auto SrcOffset = SrcPtr->getPointerOffsetFrom(Other: SrcAlloca, DL);
1566 if (!SrcOffset || *SrcOffset < *DestOffset || *SrcOffset < 0)
1567 return false;
1568 // Offset difference must preserve dest alloca's alignment.
1569 if ((*SrcOffset - *DestOffset) % DestAlloca->getAlign().value() != 0)
1570 return false;
1571 std::optional<TypeSize> SrcSize = SrcAlloca->getAllocationSize(DL);
1572 std::optional<TypeSize> DestSize = DestAlloca->getAllocationSize(DL);
1573 if (!SrcSize || !DestSize)
1574 return false;
1575 if (*SrcSize != *DestSize)
1576 if (!SrcSize->isFixed() || !DestSize->isFixed())
1577 return false;
1578 // Check that copy covers entirety of dest alloca.
1579 if (Size != *DestSize || *DestOffset != 0) {
1580 LLVM_DEBUG(dbgs() << "Stack Move: Destination alloca size mismatch\n");
1581 return false;
1582 }
1583
1584 if (*SrcOffset) {
1585 // Make sure that the copied offset is actually part of the alloca. There
1586 // might be an out-of-bounds copy in dead code.
1587 if (!Size.isFixed() || *SrcOffset + Size > *SrcSize)
1588 return false;
1589 }
1590
1591 // Check if it will be legal to combine allocas without breaking dominator.
1592 bool MoveSrc = !DT->dominates(Def: SrcAlloca, User: DestAlloca);
1593 if (MoveSrc) {
1594 if (!DT->dominates(Def: DestAlloca, User: SrcAlloca))
1595 return false;
1596 }
1597
1598 // Check that src and dest are never captured, unescaped allocas. Also
1599 // find the nearest common dominator and postdominator for all users in
1600 // order to shrink wrap the lifetimes, and instructions with noalias metadata
1601 // to remove them.
1602
1603 SmallVector<Instruction *, 4> LifetimeMarkers;
1604 SmallPtrSet<Instruction *, 4> AAMetadataInstrs;
1605
1606 auto CaptureTrackingWithModRef =
1607 [&](Instruction *AI, function_ref<bool(Instruction *)> ModRefCallback,
1608 bool &AddressCaptured) -> bool {
1609 SmallVector<Instruction *, 8> Worklist;
1610 Worklist.push_back(Elt: AI);
1611 unsigned MaxUsesToExplore = getDefaultMaxUsesToExploreForCaptureTracking();
1612 Worklist.reserve(N: MaxUsesToExplore);
1613 SmallPtrSet<const Use *, 20> Visited;
1614 while (!Worklist.empty()) {
1615 Instruction *I = Worklist.pop_back_val();
1616 for (const Use &U : I->uses()) {
1617 auto *UI = cast<Instruction>(Val: U.getUser());
1618
1619 if (Visited.size() >= MaxUsesToExplore) {
1620 LLVM_DEBUG(
1621 dbgs()
1622 << "Stack Move: Exceeded max uses to see ModRef, bailing\n");
1623 return false;
1624 }
1625 if (!Visited.insert(Ptr: &U).second)
1626 continue;
1627 UseCaptureInfo CI = DetermineUseCaptureKind(U, Base: AI);
1628 if (capturesAnyProvenance(CC: CI.UseCC))
1629 return false;
1630 AddressCaptured |= capturesAddress(CC: CI.UseCC);
1631
1632 if (UI->mayReadOrWriteMemory()) {
1633 if (UI->isLifetimeStartOrEnd()) {
1634 // We note the locations of these intrinsic calls so that we can
1635 // delete them later if the optimization succeeds, this is safe
1636 // since both llvm.lifetime.start and llvm.lifetime.end intrinsics
1637 // practically fill all the bytes of the alloca with an undefined
1638 // value, although conceptually marked as alive/dead.
1639 LifetimeMarkers.push_back(Elt: UI);
1640 continue;
1641 }
1642 AAMetadataInstrs.insert(Ptr: UI);
1643
1644 if (!ModRefCallback(UI))
1645 return false;
1646 }
1647
1648 if (capturesAnything(CC: CI.ResultCC)) {
1649 Worklist.push_back(Elt: UI);
1650 continue;
1651 }
1652 }
1653 }
1654 return true;
1655 };
1656
1657 // Check that dest alloca has no Mod/Ref, from the alloca to the Store. And
1658 // collect modref inst for the reachability check.
1659 ModRefInfo DestModRef = ModRefInfo::NoModRef;
1660 MemoryLocation DestLoc(DestAlloca, LocationSize::precise(Value: *DestSize));
1661 SmallVector<BasicBlock *, 8> ReachabilityWorklist;
1662 auto DestModRefCallback = [&](Instruction *UI) -> bool {
1663 // We don't care about the store itself.
1664 if (UI == Store)
1665 return true;
1666 ModRefInfo Res = BAA.getModRefInfo(I: UI, OptLoc: DestLoc);
1667 DestModRef |= Res;
1668 if (isModOrRefSet(MRI: Res)) {
1669 // Instructions reachability checks.
1670 // FIXME: adding the Instruction version isPotentiallyReachableFromMany on
1671 // lib/Analysis/CFG.cpp (currently only for BasicBlocks) might be helpful.
1672 if (UI->getParent() == Store->getParent()) {
1673 // The same block case is special because it's the only time we're
1674 // looking within a single block to see which instruction comes first.
1675 // Once we start looking at multiple blocks, the first instruction of
1676 // the block is reachable, so we only need to determine reachability
1677 // between whole blocks.
1678 BasicBlock *BB = UI->getParent();
1679
1680 // If A comes before B, then B is definitively reachable from A.
1681 if (UI->comesBefore(Other: Store))
1682 return false;
1683
1684 // If the user's parent block is entry, no predecessor exists.
1685 if (BB->isEntryBlock())
1686 return true;
1687
1688 // Otherwise, continue doing the normal per-BB CFG walk.
1689 ReachabilityWorklist.append(in_start: succ_begin(BB), in_end: succ_end(BB));
1690 } else {
1691 ReachabilityWorklist.push_back(Elt: UI->getParent());
1692 }
1693 }
1694 return true;
1695 };
1696
1697 bool DestAddressCaptured = false;
1698 if (!CaptureTrackingWithModRef(DestAlloca, DestModRefCallback,
1699 DestAddressCaptured))
1700 return false;
1701 // Bailout if Dest may have any ModRef before Store.
1702 if (!ReachabilityWorklist.empty() &&
1703 isPotentiallyReachableFromMany(Worklist&: ReachabilityWorklist, StopBB: Store->getParent(),
1704 ExclusionSet: nullptr, DT, LI: nullptr))
1705 return false;
1706
1707 // Check that, from after the Load to the end of the BB,
1708 // - if the dest has any Mod, src has no Ref, and
1709 // - if the dest has any Ref, src has no Mod except full-sized lifetimes
1710 // Where:
1711 // - src is defined as the memory from max(SrcAlloca, SrcPtr minus
1712 // dest_offset) to min(dest_size, SrcSize minus SrcOffset)
1713 // - dest_offset and dest_size could be computed by DestModRefCallback
1714 // to be the bounds of the first and last mod region, and which is at
1715 // least as large as DestOffset to DestSize, and at most as large as
1716 // SrcAlloca to SrcSize.
1717 // - Currently DestOffset==0 and DestSize==Size, so this math is simplified.
1718 MemoryLocation SrcLoc(SrcPtr, LocationSize::precise(Value: Size));
1719
1720 auto SrcModRefCallback = [&](Instruction *UI) -> bool {
1721 // Any ModRef post-dominated by Load doesn't matter, also Load and Store
1722 // themselves can be ignored.
1723 if (PDT->dominates(I1: Load, I2: UI) || UI == Load || UI == Store)
1724 return true;
1725 ModRefInfo Res = BAA.getModRefInfo(I: UI, OptLoc: SrcLoc);
1726 if ((isModSet(MRI: DestModRef) && isRefSet(MRI: Res)) ||
1727 (isRefSet(MRI: DestModRef) && isModSet(MRI: Res)))
1728 return false;
1729
1730 return true;
1731 };
1732
1733 bool SrcAddressCaptured = false;
1734 if (!CaptureTrackingWithModRef(SrcAlloca, SrcModRefCallback,
1735 SrcAddressCaptured))
1736 return false;
1737
1738 // If both the source and destination address are captured, the fact that they
1739 // are no longer two separate allocations may be observed.
1740 if (DestAddressCaptured && SrcAddressCaptured)
1741 return false;
1742
1743 // We can now do the transformation. First move the Src if it was after Dest.
1744 if (MoveSrc)
1745 SrcAlloca->moveBefore(InsertPos: DestAlloca->getIterator());
1746
1747 // Align the allocas appropriately.
1748 SrcAlloca->setAlignment(
1749 std::max(a: SrcAlloca->getAlign(), b: DestAlloca->getAlign()));
1750
1751 // Size the allocas appropriately.
1752 if (*SrcSize != *DestSize) {
1753 // Only possible if both sizes are fixed (due to earlier check)
1754 // Set Src to the type and array size of Dest if Dest was larger
1755 if (DestSize->getFixedValue() > SrcSize->getFixedValue()) {
1756 SrcAlloca->setAllocatedType(DestAlloca->getAllocatedType());
1757 SrcAlloca->setOperand(i_nocapture: 0, Val_nocapture: DestAlloca->getArraySize());
1758 }
1759 }
1760
1761 // Merge the two allocas.
1762 Value *NewDestPtr = SrcAlloca;
1763 if (*SrcOffset != *DestOffset) {
1764 IRBuilder<> Builder(DestAlloca);
1765 NewDestPtr = Builder.CreateInBoundsPtrAdd(
1766 Ptr: SrcAlloca, Offset: Builder.getInt64(C: *SrcOffset - *DestOffset));
1767 }
1768 DestAlloca->replaceAllUsesWith(V: NewDestPtr);
1769 eraseInstruction(I: DestAlloca);
1770
1771 // Drop metadata on the source alloca.
1772 SrcAlloca->dropUnknownNonDebugMetadata();
1773
1774 // TODO: Reconstruct merged lifetime markers.
1775 // Remove all other lifetime markers. if the original lifetime intrinsics
1776 // exists.
1777 if (!LifetimeMarkers.empty()) {
1778 for (Instruction *I : LifetimeMarkers)
1779 eraseInstruction(I);
1780 }
1781
1782 // As this transformation can cause memory accesses that didn't previously
1783 // alias to begin to alias one another, we remove !alias.scope, !noalias,
1784 // !tbaa and !tbaa_struct metadata from any uses of either alloca.
1785 // This is conservative, but more precision doesn't seem worthwhile
1786 // right now.
1787 for (Instruction *I : AAMetadataInstrs) {
1788 I->setMetadata(KindID: LLVMContext::MD_alias_scope, Node: nullptr);
1789 I->setMetadata(KindID: LLVMContext::MD_noalias, Node: nullptr);
1790 I->setMetadata(KindID: LLVMContext::MD_tbaa, Node: nullptr);
1791 I->setMetadata(KindID: LLVMContext::MD_tbaa_struct, Node: nullptr);
1792 }
1793
1794 LLVM_DEBUG(dbgs() << "Stack Move: Performed stack-move optimization\n");
1795 NumStackMove++;
1796 return true;
1797}
1798
1799static bool isZeroSize(Value *Size) {
1800 if (auto *I = dyn_cast<Instruction>(Val: Size))
1801 if (auto *Res = simplifyInstruction(I, Q: I->getDataLayout()))
1802 Size = Res;
1803 // Treat undef/poison size like zero.
1804 if (auto *C = dyn_cast<Constant>(Val: Size))
1805 return isa<UndefValue>(Val: C) || C->isNullValue();
1806 return false;
1807}
1808
1809/// Perform simplification of memcpy's. If we have memcpy A
1810/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
1811/// B to be a memcpy from X to Z (or potentially a memmove, depending on
1812/// circumstances). This allows later passes to remove the first memcpy
1813/// altogether.
1814bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
1815 // We can only optimize non-volatile memcpy's.
1816 if (M->isVolatile())
1817 return false;
1818
1819 // If the source and destination of the memcpy are the same, then zap it.
1820 if (M->getSource() == M->getDest()) {
1821 ++BBI;
1822 eraseInstruction(I: M);
1823 return true;
1824 }
1825
1826 // If the size is zero, remove the memcpy.
1827 if (isZeroSize(Size: M->getLength())) {
1828 ++BBI;
1829 eraseInstruction(I: M);
1830 return true;
1831 }
1832
1833 MemoryUseOrDef *MA = MSSA->getMemoryAccess(I: M);
1834 if (!MA)
1835 // Degenerate case: memcpy marked as not accessing memory.
1836 return false;
1837
1838 // If copying from a constant, try to turn the memcpy into a memset.
1839 if (auto *GV = dyn_cast<GlobalVariable>(Val: getUnderlyingObject(V: M->getSource())))
1840 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1841 if (Value *ByteVal = isBytewiseValue(V: GV->getInitializer(),
1842 DL: M->getDataLayout())) {
1843 IRBuilder<> Builder(M);
1844 Instruction *NewM = Builder.CreateMemSet(
1845 Ptr: M->getRawDest(), Val: ByteVal, Size: M->getLength(), Align: M->getDestAlign(), isVolatile: false);
1846 auto *LastDef = cast<MemoryDef>(Val: MA);
1847 auto *NewAccess =
1848 MSSAU->createMemoryAccessAfter(I: NewM, Definition: nullptr, InsertPt: LastDef);
1849 MSSAU->insertDef(Def: cast<MemoryDef>(Val: NewAccess), /*RenameUses=*/true);
1850
1851 eraseInstruction(I: M);
1852 ++NumCpyToSet;
1853 return true;
1854 }
1855
1856 BatchAAResults BAA(*AA, EEA);
1857 // FIXME: Not using getClobberingMemoryAccess() here due to PR54682.
1858 MemoryAccess *AnyClobber = MA->getDefiningAccess();
1859 MemoryLocation DestLoc = MemoryLocation::getForDest(MI: M);
1860 const MemoryAccess *DestClobber =
1861 MSSA->getWalker()->getClobberingMemoryAccess(AnyClobber, DestLoc, AA&: BAA);
1862
1863 // Try to turn a partially redundant memset + memcpy into
1864 // smaller memset + memcpy. We don't need the memcpy size for this.
1865 // The memcpy must post-dom the memset, so limit this to the same basic
1866 // block. A non-local generalization is likely not worthwhile.
1867 if (auto *MD = dyn_cast<MemoryDef>(Val: DestClobber))
1868 if (auto *MDep = dyn_cast_or_null<MemSetInst>(Val: MD->getMemoryInst()))
1869 if (DestClobber->getBlock() == M->getParent())
1870 if (processMemSetMemCpyDependence(MemCpy: M, MemSet: MDep, BAA))
1871 return true;
1872
1873 MemoryAccess *SrcClobber = MSSA->getWalker()->getClobberingMemoryAccess(
1874 AnyClobber, MemoryLocation::getForSource(MTI: M), AA&: BAA);
1875
1876 // There are five possible optimizations we can do for memcpy:
1877 // a) memcpy-memcpy xform which exposes redundance for DSE.
1878 // b) call-memcpy xform for return slot optimization.
1879 // c) memcpy from freshly alloca'd space or space that has just started
1880 // its lifetime copies undefined data, and we can therefore eliminate
1881 // the memcpy in favor of the data that was already at the destination.
1882 // d) memcpy from a just-memset'd source can be turned into memset.
1883 // e) elimination of memcpy via stack-move optimization.
1884 if (auto *MD = dyn_cast<MemoryDef>(Val: SrcClobber)) {
1885 if (Instruction *MI = MD->getMemoryInst()) {
1886 if (auto *CopySize = dyn_cast<ConstantInt>(Val: M->getLength())) {
1887 if (auto *C = dyn_cast<CallInst>(Val: MI)) {
1888 if (performCallSlotOptzn(cpyLoad: M, cpyStore: M, cpyDest: M->getDest(), cpySrc: M->getSource(),
1889 cpySize: TypeSize::getFixed(ExactSize: CopySize->getZExtValue()),
1890 cpyDestAlign: M->getDestAlign().valueOrOne(), BAA,
1891 GetC: [C]() -> CallInst * { return C; })) {
1892 LLVM_DEBUG(dbgs() << "Performed call slot optimization:\n"
1893 << " call: " << *C << "\n"
1894 << " memcpy: " << *M << "\n");
1895 eraseInstruction(I: M);
1896 ++NumMemCpyInstr;
1897 return true;
1898 }
1899 }
1900 }
1901 if (auto *MDep = dyn_cast<MemCpyInst>(Val: MI))
1902 if (processMemCpyMemCpyDependence(M, MDep, BAA))
1903 return true;
1904 if (auto *MDep = dyn_cast<MemSetInst>(Val: MI)) {
1905 if (performMemCpyToMemSetOptzn(MemCpy: M, MemSet: MDep, BAA)) {
1906 LLVM_DEBUG(dbgs() << "Converted memcpy to memset\n");
1907 eraseInstruction(I: M);
1908 ++NumCpyToSet;
1909 return true;
1910 }
1911 }
1912 }
1913
1914 if (hasUndefContents(MSSA, AA&: BAA, V: M->getSource(), Def: MD)) {
1915 LLVM_DEBUG(dbgs() << "Removed memcpy from undef\n");
1916 eraseInstruction(I: M);
1917 ++NumMemCpyInstr;
1918 return true;
1919 }
1920 }
1921
1922 // If the transfer is from a stack slot to a stack slot, then we may be able
1923 // to perform the stack-move optimization. See the comments in
1924 // performStackMoveOptzn() for more details.
1925 ConstantInt *Len = dyn_cast<ConstantInt>(Val: M->getLength());
1926 if (Len == nullptr)
1927 return false;
1928 if (performStackMoveOptzn(Load: M, Store: M, DestPtr: M->getDest(), SrcPtr: M->getSource(),
1929 Size: TypeSize::getFixed(ExactSize: Len->getZExtValue()), BAA)) {
1930 // Avoid invalidating the iterator.
1931 BBI = M->getNextNode()->getIterator();
1932 eraseInstruction(I: M);
1933 ++NumMemCpyInstr;
1934 return true;
1935 }
1936
1937 return false;
1938}
1939
1940/// Memmove calls with overlapping src/dest buffers that come after a memset may
1941/// be removed.
1942bool MemCpyOptPass::isMemMoveMemSetDependency(MemMoveInst *M) {
1943 const auto &DL = M->getDataLayout();
1944 MemoryUseOrDef *MemMoveAccess = MSSA->getMemoryAccess(I: M);
1945 if (!MemMoveAccess)
1946 return false;
1947
1948 // The memmove is of form memmove(x, x + A, B).
1949 MemoryLocation SourceLoc = MemoryLocation::getForSource(MTI: M);
1950 auto *MemMoveSourceOp = M->getSource();
1951 auto *Source = dyn_cast<GEPOperator>(Val: MemMoveSourceOp);
1952 if (!Source)
1953 return false;
1954
1955 APInt Offset(DL.getIndexTypeSizeInBits(Ty: Source->getType()), 0);
1956 LocationSize MemMoveLocSize = SourceLoc.Size;
1957 if (Source->getPointerOperand() != M->getDest() ||
1958 !MemMoveLocSize.hasValue() ||
1959 !Source->accumulateConstantOffset(DL, Offset) || Offset.isNegative()) {
1960 return false;
1961 }
1962
1963 uint64_t MemMoveSize = MemMoveLocSize.getValue();
1964 LocationSize TotalSize =
1965 LocationSize::precise(Value: Offset.getZExtValue() + MemMoveSize);
1966 MemoryLocation CombinedLoc(M->getDest(), TotalSize);
1967
1968 // The first dominating clobbering MemoryAccess for the combined location
1969 // needs to be a memset.
1970 BatchAAResults BAA(*AA);
1971 MemoryAccess *FirstDef = MemMoveAccess->getDefiningAccess();
1972 auto *DestClobber = dyn_cast<MemoryDef>(
1973 Val: MSSA->getWalker()->getClobberingMemoryAccess(FirstDef, CombinedLoc, AA&: BAA));
1974 if (!DestClobber)
1975 return false;
1976
1977 auto *MS = dyn_cast_or_null<MemSetInst>(Val: DestClobber->getMemoryInst());
1978 if (!MS)
1979 return false;
1980
1981 // Memset length must be sufficiently large.
1982 auto *MemSetLength = dyn_cast<ConstantInt>(Val: MS->getLength());
1983 if (!MemSetLength ||
1984 MemSetLength->getZExtValue() < Offset.getZExtValue() + MemMoveSize)
1985 return false;
1986
1987 // The destination buffer must have been memset'd.
1988 if (!BAA.isMustAlias(V1: MS->getDest(), V2: M->getDest()))
1989 return false;
1990
1991 return true;
1992}
1993
1994/// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1995/// not to alias.
1996bool MemCpyOptPass::processMemMove(MemMoveInst *M, BasicBlock::iterator &BBI) {
1997 // See if the source could be modified by this memmove potentially.
1998 if (isModSet(MRI: AA->getModRefInfo(I: M, OptLoc: MemoryLocation::getForSource(MTI: M)))) {
1999 // On the off-chance the memmove clobbers src with previously memset'd
2000 // bytes, the memmove may be redundant.
2001 if (!M->isVolatile() && isMemMoveMemSetDependency(M)) {
2002 LLVM_DEBUG(dbgs() << "Removed redundant memmove.\n");
2003 ++BBI;
2004 eraseInstruction(I: M);
2005 ++NumMemMoveInstr;
2006 return true;
2007 }
2008 return false;
2009 }
2010
2011 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: " << *M
2012 << "\n");
2013
2014 // If not, then we know we can transform this.
2015 Type *ArgTys[3] = {M->getRawDest()->getType(), M->getRawSource()->getType(),
2016 M->getLength()->getType()};
2017 M->setCalledFunction(Intrinsic::getOrInsertDeclaration(
2018 M: M->getModule(), id: Intrinsic::memcpy, OverloadTys: ArgTys));
2019
2020 // For MemorySSA nothing really changes (except that memcpy may imply stricter
2021 // aliasing guarantees).
2022
2023 ++NumMoveToCpy;
2024 return true;
2025}
2026
2027/// This is called on every byval argument in call sites.
2028bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
2029 const DataLayout &DL = CB.getDataLayout();
2030 // Find out what feeds this byval argument.
2031 Value *ByValArg = CB.getArgOperand(i: ArgNo);
2032 Type *ByValTy = CB.getParamByValType(ArgNo);
2033 TypeSize ByValSize = DL.getTypeAllocSize(Ty: ByValTy);
2034 MemoryLocation Loc(ByValArg, LocationSize::precise(Value: ByValSize));
2035 MemoryUseOrDef *CallAccess = MSSA->getMemoryAccess(I: &CB);
2036 if (!CallAccess)
2037 return false;
2038 MemCpyInst *MDep = nullptr;
2039 BatchAAResults BAA(*AA, EEA);
2040 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
2041 CallAccess->getDefiningAccess(), Loc, AA&: BAA);
2042 if (auto *MD = dyn_cast<MemoryDef>(Val: Clobber))
2043 MDep = dyn_cast_or_null<MemCpyInst>(Val: MD->getMemoryInst());
2044
2045 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
2046 // a memcpy, see if we can byval from the source of the memcpy instead of the
2047 // result.
2048 if (!MDep || MDep->isVolatile() ||
2049 ByValArg->stripPointerCasts() != MDep->getDest())
2050 return false;
2051
2052 // The length of the memcpy must be larger or equal to the size of the byval.
2053 auto *C1 = dyn_cast<ConstantInt>(Val: MDep->getLength());
2054 if (!C1 || !TypeSize::isKnownGE(
2055 LHS: TypeSize::getFixed(ExactSize: C1->getValue().getZExtValue()), RHS: ByValSize))
2056 return false;
2057
2058 // Get the alignment of the byval. If the call doesn't specify the alignment,
2059 // then it is some target specific value that we can't know.
2060 MaybeAlign ByValAlign = CB.getParamAlign(ArgNo);
2061 if (!ByValAlign)
2062 return false;
2063
2064 // If it is greater than the memcpy, then we check to see if we can force the
2065 // source of the memcpy to the alignment we need. If we fail, we bail out.
2066 MaybeAlign MemDepAlign = MDep->getSourceAlign();
2067 if ((!MemDepAlign || *MemDepAlign < *ByValAlign) &&
2068 getOrEnforceKnownAlignment(V: MDep->getSource(), PrefAlign: ByValAlign, DL, CxtI: &CB, AC,
2069 DT) < *ByValAlign)
2070 return false;
2071
2072 // The type of the memcpy source must match the byval argument
2073 if (MDep->getSource()->getType() != ByValArg->getType())
2074 return false;
2075
2076 // Verify that the copied-from memory doesn't change in between the memcpy and
2077 // the byval call.
2078 // memcpy(a <- b)
2079 // *b = 42;
2080 // foo(*a)
2081 // It would be invalid to transform the second memcpy into foo(*b).
2082 if (writtenBetween(MSSA, AA&: BAA, Loc: MemoryLocation::getForSource(MTI: MDep),
2083 Start: MSSA->getMemoryAccess(I: MDep), End: CallAccess))
2084 return false;
2085
2086 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
2087 << " " << *MDep << "\n"
2088 << " " << CB << "\n");
2089
2090 // Otherwise we're good! Update the byval argument.
2091 combineAAMetadata(K: &CB, J: MDep);
2092 CB.setArgOperand(i: ArgNo, v: MDep->getSource());
2093 ++NumMemCpyInstr;
2094 return true;
2095}
2096
2097/// This is called on memcpy dest pointer arguments attributed as immutable
2098/// during call. Try to use memcpy source directly if all of the following
2099/// conditions are satisfied.
2100/// 1. The memcpy dst is neither modified during the call nor captured by the
2101/// call.
2102/// 2. The memcpy dst is an alloca with known alignment & size.
2103/// 2-1. The memcpy length == the alloca size which ensures that the new
2104/// pointer is dereferenceable for the required range
2105/// 2-2. The src pointer has alignment >= the alloca alignment or can be
2106/// enforced so.
2107/// 3. The memcpy dst and src is not modified between the memcpy and the call.
2108/// (if MSSA clobber check is safe.)
2109/// 4. The memcpy src is not modified during the call. (ModRef check shows no
2110/// Mod.)
2111bool MemCpyOptPass::processImmutArgument(CallBase &CB, unsigned ArgNo) {
2112 BatchAAResults BAA(*AA, EEA);
2113 Value *ImmutArg = CB.getArgOperand(i: ArgNo);
2114
2115 // 1. Ensure passed argument is immutable during call.
2116 if (!CB.doesNotCapture(OpNo: ArgNo))
2117 return false;
2118
2119 // We know that the argument is readonly at this point, but the function
2120 // might still modify the same memory through a different pointer. Exclude
2121 // this either via noalias, or alias analysis.
2122 if (!CB.paramHasAttr(ArgNo, Kind: Attribute::NoAlias) &&
2123 isModSet(
2124 MRI: BAA.getModRefInfo(I: &CB, OptLoc: MemoryLocation::getBeforeOrAfter(Ptr: ImmutArg))))
2125 return false;
2126
2127 const DataLayout &DL = CB.getDataLayout();
2128
2129 // 2. Check that arg is alloca
2130 // TODO: Even if the arg gets back to branches, we can remove memcpy if all
2131 // the alloca alignments can be enforced to source alignment.
2132 auto *AI = dyn_cast<AllocaInst>(Val: ImmutArg->stripPointerCasts());
2133 if (!AI)
2134 return false;
2135
2136 std::optional<TypeSize> AllocaSize = AI->getAllocationSize(DL);
2137 // Can't handle unknown size alloca.
2138 // (e.g. Variable Length Array, Scalable Vector)
2139 if (!AllocaSize || AllocaSize->isScalable())
2140 return false;
2141 MemoryLocation Loc(ImmutArg, LocationSize::precise(Value: *AllocaSize));
2142 MemoryUseOrDef *CallAccess = MSSA->getMemoryAccess(I: &CB);
2143 if (!CallAccess)
2144 return false;
2145
2146 MemCpyInst *MDep = nullptr;
2147 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
2148 CallAccess->getDefiningAccess(), Loc, AA&: BAA);
2149 if (auto *MD = dyn_cast<MemoryDef>(Val: Clobber))
2150 MDep = dyn_cast_or_null<MemCpyInst>(Val: MD->getMemoryInst());
2151
2152 // If the immut argument isn't fed by a memcpy, ignore it. If it is fed by
2153 // a memcpy, check that the arg equals the memcpy dest.
2154 if (!MDep || MDep->isVolatile() || AI != MDep->getDest())
2155 return false;
2156
2157 // The type of the memcpy source must match the immut argument
2158 if (MDep->getSource()->getType() != ImmutArg->getType())
2159 return false;
2160
2161 // 2-1. The length of the memcpy must be equal to the size of the alloca.
2162 auto *MDepLen = dyn_cast<ConstantInt>(Val: MDep->getLength());
2163 if (!MDepLen || AllocaSize != MDepLen->getValue())
2164 return false;
2165
2166 // 2-2. the memcpy source align must be larger than or equal the alloca's
2167 // align. If not so, we check to see if we can force the source of the memcpy
2168 // to the alignment we need. If we fail, we bail out.
2169 Align MemDepAlign = MDep->getSourceAlign().valueOrOne();
2170 Align AllocaAlign = AI->getAlign();
2171 if (MemDepAlign < AllocaAlign &&
2172 getOrEnforceKnownAlignment(V: MDep->getSource(), PrefAlign: AllocaAlign, DL, CxtI: &CB, AC,
2173 DT) < AllocaAlign)
2174 return false;
2175
2176 // 3. Verify that the source doesn't change in between the memcpy and
2177 // the call.
2178 // memcpy(a <- b)
2179 // *b = 42;
2180 // foo(*a)
2181 // It would be invalid to transform the second memcpy into foo(*b).
2182 if (writtenBetween(MSSA, AA&: BAA, Loc: MemoryLocation::getForSource(MTI: MDep),
2183 Start: MSSA->getMemoryAccess(I: MDep), End: CallAccess))
2184 return false;
2185
2186 // 4. The memcpy src must not be modified during the call.
2187 if (isModSet(MRI: BAA.getModRefInfo(I: &CB, OptLoc: MemoryLocation::getForSource(MTI: MDep))))
2188 return false;
2189
2190 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy to Immut src:\n"
2191 << " " << *MDep << "\n"
2192 << " " << CB << "\n");
2193
2194 // Otherwise we're good! Update the immut argument.
2195 combineAAMetadata(K: &CB, J: MDep);
2196 CB.setArgOperand(i: ArgNo, v: MDep->getSource());
2197 ++NumMemCpyInstr;
2198 return true;
2199}
2200
2201/// Executes one iteration of MemCpyOptPass.
2202bool MemCpyOptPass::iterateOnFunction(Function &F) {
2203 bool MadeChange = false;
2204
2205 // Walk all instruction in the function.
2206 for (BasicBlock &BB : F) {
2207 // Skip unreachable blocks. For example processStore assumes that an
2208 // instruction in a BB can't be dominated by a later instruction in the
2209 // same BB (which is a scenario that can happen for an unreachable BB that
2210 // has itself as a predecessor).
2211 if (!DT->isReachableFromEntry(A: &BB))
2212 continue;
2213
2214 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
2215 // Avoid invalidating the iterator.
2216 Instruction *I = &*BI++;
2217
2218 bool RepeatInstruction = false;
2219
2220 if (auto *SI = dyn_cast<StoreInst>(Val: I))
2221 MadeChange |= processStore(SI, BBI&: BI);
2222 else if (auto *M = dyn_cast<MemSetInst>(Val: I))
2223 RepeatInstruction = processMemSet(MSI: M, BBI&: BI);
2224 else if (auto *M = dyn_cast<MemCpyInst>(Val: I))
2225 RepeatInstruction = processMemCpy(M, BBI&: BI);
2226 else if (auto *M = dyn_cast<MemMoveInst>(Val: I))
2227 RepeatInstruction = processMemMove(M, BBI&: BI);
2228 else if (auto *CB = dyn_cast<CallBase>(Val: I)) {
2229 for (unsigned i = 0, e = CB->arg_size(); i != e; ++i) {
2230 if (CB->isByValArgument(ArgNo: i))
2231 MadeChange |= processByValArgument(CB&: *CB, ArgNo: i);
2232 else if (CB->onlyReadsMemory(OpNo: i))
2233 MadeChange |= processImmutArgument(CB&: *CB, ArgNo: i);
2234 }
2235 }
2236
2237 // Reprocess the instruction if desired.
2238 if (RepeatInstruction) {
2239 if (BI != BB.begin())
2240 --BI;
2241 MadeChange = true;
2242 }
2243 }
2244 }
2245
2246 return MadeChange;
2247}
2248
2249PreservedAnalyses MemCpyOptPass::run(Function &F, FunctionAnalysisManager &AM) {
2250 auto &TLI = AM.getResult<TargetLibraryAnalysis>(IR&: F);
2251 auto *AA = &AM.getResult<AAManager>(IR&: F);
2252 auto *AC = &AM.getResult<AssumptionAnalysis>(IR&: F);
2253 auto *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F);
2254 auto *PDT = &AM.getResult<PostDominatorTreeAnalysis>(IR&: F);
2255 auto *MSSA = &AM.getResult<MemorySSAAnalysis>(IR&: F);
2256
2257 bool MadeChange = runImpl(F, TLI: &TLI, AA, AC, DT, PDT, MSSA: &MSSA->getMSSA());
2258 if (!MadeChange)
2259 return PreservedAnalyses::all();
2260
2261 PreservedAnalyses PA;
2262 PA.preserveSet<CFGAnalyses>();
2263 PA.preserve<MemorySSAAnalysis>();
2264 return PA;
2265}
2266
2267bool MemCpyOptPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
2268 AliasAnalysis *AA_, AssumptionCache *AC_,
2269 DominatorTree *DT_, PostDominatorTree *PDT_,
2270 MemorySSA *MSSA_) {
2271 bool MadeChange = false;
2272 TLI = TLI_;
2273 AA = AA_;
2274 AC = AC_;
2275 DT = DT_;
2276 PDT = PDT_;
2277 MSSA = MSSA_;
2278 MemorySSAUpdater MSSAU_(MSSA_);
2279 MSSAU = &MSSAU_;
2280 EarliestEscapeAnalysis EEA_(*DT);
2281 EEA = &EEA_;
2282
2283 while (true) {
2284 if (!iterateOnFunction(F))
2285 break;
2286 MadeChange = true;
2287 }
2288
2289 if (VerifyMemorySSA)
2290 MSSA_->verifyMemorySSA();
2291
2292 return MadeChange;
2293}
2294