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