Bug Summary

File:llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
Warning:line 1591, column 34
Called C++ object pointer is null

Annotated Source Code

Press '?' to see keyboard shortcuts

clang -cc1 -cc1 -triple x86_64-pc-linux-gnu -analyze -disable-free -disable-llvm-verifier -discard-value-names -main-file-name MemCpyOptimizer.cpp -analyzer-store=region -analyzer-opt-analyze-nested-blocks -analyzer-checker=core -analyzer-checker=apiModeling -analyzer-checker=unix -analyzer-checker=deadcode -analyzer-checker=cplusplus -analyzer-checker=security.insecureAPI.UncheckedReturn -analyzer-checker=security.insecureAPI.getpw -analyzer-checker=security.insecureAPI.gets -analyzer-checker=security.insecureAPI.mktemp -analyzer-checker=security.insecureAPI.mkstemp -analyzer-checker=security.insecureAPI.vfork -analyzer-checker=nullability.NullPassedToNonnull -analyzer-checker=nullability.NullReturnedFromNonnull -analyzer-output plist -w -setup-static-analyzer -analyzer-config-compatibility-mode=true -mrelocation-model pic -pic-level 2 -mframe-pointer=none -fmath-errno -fno-rounding-math -mconstructor-aliases -munwind-tables -target-cpu x86-64 -tune-cpu generic -debugger-tuning=gdb -ffunction-sections -fdata-sections -fcoverage-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -resource-dir /usr/lib/llvm-13/lib/clang/13.0.0 -D _DEBUG -D _GNU_SOURCE -D __STDC_CONSTANT_MACROS -D __STDC_FORMAT_MACROS -D __STDC_LIMIT_MACROS -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/include -I /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/include -D NDEBUG -U NDEBUG -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/x86_64-linux-gnu/c++/10 -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/backward -internal-isystem /usr/lib/llvm-13/lib/clang/13.0.0/include -internal-isystem /usr/local/include -internal-isystem /usr/lib/gcc/x86_64-linux-gnu/10/../../../../x86_64-linux-gnu/include -internal-externc-isystem /usr/include/x86_64-linux-gnu -internal-externc-isystem /include -internal-externc-isystem /usr/include -O2 -Wno-unused-parameter -Wwrite-strings -Wno-missing-field-initializers -Wno-long-long -Wno-maybe-uninitialized -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wno-comment -std=c++14 -fdeprecated-macro -fdebug-compilation-dir=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/build-llvm/lib/Transforms/Scalar -fdebug-prefix-map=/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c=. -ferror-limit 19 -fvisibility-inlines-hidden -stack-protector 2 -fgnuc-version=4.2.1 -vectorize-loops -vectorize-slp -analyzer-output=html -analyzer-config stable-report-filename=true -faddrsig -D__GCC_HAVE_DWARF2_CFI_ASM=1 -o /tmp/scan-build-2021-07-26-235520-9401-1 -x c++ /build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp
1//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This pass performs various transformations related to eliminating memcpy
10// calls, or transforming sets of stores into memset's.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Transforms/Scalar/MemCpyOptimizer.h"
15#include "llvm/ADT/DenseSet.h"
16#include "llvm/ADT/None.h"
17#include "llvm/ADT/STLExtras.h"
18#include "llvm/ADT/SmallVector.h"
19#include "llvm/ADT/Statistic.h"
20#include "llvm/ADT/iterator_range.h"
21#include "llvm/Analysis/AliasAnalysis.h"
22#include "llvm/Analysis/AssumptionCache.h"
23#include "llvm/Analysis/GlobalsModRef.h"
24#include "llvm/Analysis/Loads.h"
25#include "llvm/Analysis/MemoryDependenceAnalysis.h"
26#include "llvm/Analysis/MemoryLocation.h"
27#include "llvm/Analysis/MemorySSA.h"
28#include "llvm/Analysis/MemorySSAUpdater.h"
29#include "llvm/Analysis/TargetLibraryInfo.h"
30#include "llvm/Analysis/ValueTracking.h"
31#include "llvm/IR/Argument.h"
32#include "llvm/IR/BasicBlock.h"
33#include "llvm/IR/Constants.h"
34#include "llvm/IR/DataLayout.h"
35#include "llvm/IR/DerivedTypes.h"
36#include "llvm/IR/Dominators.h"
37#include "llvm/IR/Function.h"
38#include "llvm/IR/GetElementPtrTypeIterator.h"
39#include "llvm/IR/GlobalVariable.h"
40#include "llvm/IR/IRBuilder.h"
41#include "llvm/IR/InstrTypes.h"
42#include "llvm/IR/Instruction.h"
43#include "llvm/IR/Instructions.h"
44#include "llvm/IR/IntrinsicInst.h"
45#include "llvm/IR/Intrinsics.h"
46#include "llvm/IR/LLVMContext.h"
47#include "llvm/IR/Module.h"
48#include "llvm/IR/Operator.h"
49#include "llvm/IR/PassManager.h"
50#include "llvm/IR/Type.h"
51#include "llvm/IR/User.h"
52#include "llvm/IR/Value.h"
53#include "llvm/InitializePasses.h"
54#include "llvm/Pass.h"
55#include "llvm/Support/Casting.h"
56#include "llvm/Support/Debug.h"
57#include "llvm/Support/MathExtras.h"
58#include "llvm/Support/raw_ostream.h"
59#include "llvm/Transforms/Scalar.h"
60#include "llvm/Transforms/Utils/Local.h"
61#include <algorithm>
62#include <cassert>
63#include <cstdint>
64#include <utility>
65
66using namespace llvm;
67
68#define DEBUG_TYPE"memcpyopt" "memcpyopt"
69
70static cl::opt<bool>
71 EnableMemorySSA("enable-memcpyopt-memoryssa", cl::init(true), cl::Hidden,
72 cl::desc("Use MemorySSA-backed MemCpyOpt."));
73
74STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted")static llvm::Statistic NumMemCpyInstr = {"memcpyopt", "NumMemCpyInstr"
, "Number of memcpy instructions deleted"}
;
75STATISTIC(NumMemSetInfer, "Number of memsets inferred")static llvm::Statistic NumMemSetInfer = {"memcpyopt", "NumMemSetInfer"
, "Number of memsets inferred"}
;
76STATISTIC(NumMoveToCpy, "Number of memmoves converted to memcpy")static llvm::Statistic NumMoveToCpy = {"memcpyopt", "NumMoveToCpy"
, "Number of memmoves converted to memcpy"}
;
77STATISTIC(NumCpyToSet, "Number of memcpys converted to memset")static llvm::Statistic NumCpyToSet = {"memcpyopt", "NumCpyToSet"
, "Number of memcpys converted to memset"}
;
78STATISTIC(NumCallSlot, "Number of call slot optimizations performed")static llvm::Statistic NumCallSlot = {"memcpyopt", "NumCallSlot"
, "Number of call slot optimizations performed"}
;
79
80namespace {
81
82/// Represents a range of memset'd bytes with the ByteVal value.
83/// This allows us to analyze stores like:
84/// store 0 -> P+1
85/// store 0 -> P+0
86/// store 0 -> P+3
87/// store 0 -> P+2
88/// which sometimes happens with stores to arrays of structs etc. When we see
89/// the first store, we make a range [1, 2). The second store extends the range
90/// to [0, 2). The third makes a new range [2, 3). The fourth store joins the
91/// two ranges into [0, 3) which is memset'able.
92struct MemsetRange {
93 // Start/End - A semi range that describes the span that this range covers.
94 // The range is closed at the start and open at the end: [Start, End).
95 int64_t Start, End;
96
97 /// StartPtr - The getelementptr instruction that points to the start of the
98 /// range.
99 Value *StartPtr;
100
101 /// Alignment - The known alignment of the first store.
102 unsigned Alignment;
103
104 /// TheStores - The actual stores that make up this range.
105 SmallVector<Instruction*, 16> TheStores;
106
107 bool isProfitableToUseMemset(const DataLayout &DL) const;
108};
109
110} // end anonymous namespace
111
112bool MemsetRange::isProfitableToUseMemset(const DataLayout &DL) const {
113 // If we found more than 4 stores to merge or 16 bytes, use memset.
114 if (TheStores.size() >= 4 || End-Start >= 16) return true;
115
116 // If there is nothing to merge, don't do anything.
117 if (TheStores.size() < 2) 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>(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) return false;
128
129 // If we have fewer than 8 stores, it can still be worthwhile to do this.
130 // For example, merging 4 i8 stores into an i32 store is useful almost always.
131 // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
132 // memset will be split into 2 32-bit stores anyway) and doing so can
133 // pessimize the llvm optimizer.
134 //
135 // Since we don't have perfect knowledge here, make some assumptions: assume
136 // the maximum GPR width is the same size as the largest legal integer
137 // size. If so, check to see whether we will end up actually reducing the
138 // number of stores used.
139 unsigned Bytes = unsigned(End-Start);
140 unsigned MaxIntSize = DL.getLargestLegalIntTypeSizeInBits() / 8;
141 if (MaxIntSize == 0)
142 MaxIntSize = 1;
143 unsigned NumPointerStores = Bytes / MaxIntSize;
144
145 // Assume the remaining bytes if any are done a byte at a time.
146 unsigned NumByteStores = Bytes % MaxIntSize;
147
148 // If we will reduce the # stores (according to this heuristic), do the
149 // transformation. This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
150 // etc.
151 return TheStores.size() > NumPointerStores+NumByteStores;
152}
153
154namespace {
155
156class MemsetRanges {
157 using range_iterator = SmallVectorImpl<MemsetRange>::iterator;
158
159 /// A sorted list of the memset ranges.
160 SmallVector<MemsetRange, 8> Ranges;
161
162 const DataLayout &DL;
163
164public:
165 MemsetRanges(const DataLayout &DL) : DL(DL) {}
166
167 using const_iterator = SmallVectorImpl<MemsetRange>::const_iterator;
168
169 const_iterator begin() const { return Ranges.begin(); }
170 const_iterator end() const { return Ranges.end(); }
171 bool empty() const { return Ranges.empty(); }
172
173 void addInst(int64_t OffsetFromFirst, Instruction *Inst) {
174 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
175 addStore(OffsetFromFirst, SI);
176 else
177 addMemSet(OffsetFromFirst, cast<MemSetInst>(Inst));
178 }
179
180 void addStore(int64_t OffsetFromFirst, StoreInst *SI) {
181 int64_t StoreSize = DL.getTypeStoreSize(SI->getOperand(0)->getType());
182
183 addRange(OffsetFromFirst, StoreSize, SI->getPointerOperand(),
184 SI->getAlign().value(), SI);
185 }
186
187 void addMemSet(int64_t OffsetFromFirst, MemSetInst *MSI) {
188 int64_t Size = cast<ConstantInt>(MSI->getLength())->getZExtValue();
189 addRange(OffsetFromFirst, Size, MSI->getDest(), MSI->getDestAlignment(), MSI);
190 }
191
192 void addRange(int64_t Start, int64_t Size, Value *Ptr,
193 unsigned Alignment, Instruction *Inst);
194};
195
196} // end anonymous namespace
197
198/// Add a new store to the MemsetRanges data structure. This adds a
199/// new range for the specified store at the specified offset, merging into
200/// existing ranges as appropriate.
201void MemsetRanges::addRange(int64_t Start, int64_t Size, Value *Ptr,
202 unsigned Alignment, Instruction *Inst) {
203 int64_t End = Start+Size;
204
205 range_iterator I = partition_point(
206 Ranges, [=](const MemsetRange &O) { return O.End < Start; });
207
208 // We now know that I == E, in which case we didn't find anything to merge
209 // with, or that Start <= I->End. If End < I->Start or I == E, then we need
210 // to insert a new range. Handle this now.
211 if (I == Ranges.end() || End < I->Start) {
212 MemsetRange &R = *Ranges.insert(I, MemsetRange());
213 R.Start = Start;
214 R.End = End;
215 R.StartPtr = Ptr;
216 R.Alignment = Alignment;
217 R.TheStores.push_back(Inst);
218 return;
219 }
220
221 // This store overlaps with I, add it.
222 I->TheStores.push_back(Inst);
223
224 // At this point, we may have an interval that completely contains our store.
225 // If so, just add it to the interval and return.
226 if (I->Start <= Start && I->End >= End)
227 return;
228
229 // Now we know that Start <= I->End and End >= I->Start so the range overlaps
230 // but is not entirely contained within the range.
231
232 // See if the range extends the start of the range. In this case, it couldn't
233 // possibly cause it to join the prior range, because otherwise we would have
234 // stopped on *it*.
235 if (Start < I->Start) {
236 I->Start = Start;
237 I->StartPtr = Ptr;
238 I->Alignment = Alignment;
239 }
240
241 // Now we know that Start <= I->End and Start >= I->Start (so the startpoint
242 // is in or right at the end of I), and that End >= I->Start. Extend I out to
243 // End.
244 if (End > I->End) {
245 I->End = End;
246 range_iterator NextI = I;
247 while (++NextI != Ranges.end() && End >= NextI->Start) {
248 // Merge the range in.
249 I->TheStores.append(NextI->TheStores.begin(), NextI->TheStores.end());
250 if (NextI->End > I->End)
251 I->End = NextI->End;
252 Ranges.erase(NextI);
253 NextI = I;
254 }
255 }
256}
257
258//===----------------------------------------------------------------------===//
259// MemCpyOptLegacyPass Pass
260//===----------------------------------------------------------------------===//
261
262namespace {
263
264class MemCpyOptLegacyPass : public FunctionPass {
265 MemCpyOptPass Impl;
266
267public:
268 static char ID; // Pass identification, replacement for typeid
269
270 MemCpyOptLegacyPass() : FunctionPass(ID) {
271 initializeMemCpyOptLegacyPassPass(*PassRegistry::getPassRegistry());
272 }
273
274 bool runOnFunction(Function &F) override;
275
276private:
277 // This transformation requires dominator postdominator info
278 void getAnalysisUsage(AnalysisUsage &AU) const override {
279 AU.setPreservesCFG();
280 AU.addRequired<AssumptionCacheTracker>();
281 AU.addRequired<DominatorTreeWrapperPass>();
282 AU.addPreserved<DominatorTreeWrapperPass>();
283 AU.addPreserved<GlobalsAAWrapperPass>();
284 AU.addRequired<TargetLibraryInfoWrapperPass>();
285 if (!EnableMemorySSA)
286 AU.addRequired<MemoryDependenceWrapperPass>();
287 AU.addPreserved<MemoryDependenceWrapperPass>();
288 AU.addRequired<AAResultsWrapperPass>();
289 AU.addPreserved<AAResultsWrapperPass>();
290 if (EnableMemorySSA)
291 AU.addRequired<MemorySSAWrapperPass>();
292 AU.addPreserved<MemorySSAWrapperPass>();
293 }
294};
295
296} // end anonymous namespace
297
298char MemCpyOptLegacyPass::ID = 0;
299
300/// The public interface to this file...
301FunctionPass *llvm::createMemCpyOptPass() { return new MemCpyOptLegacyPass(); }
302
303INITIALIZE_PASS_BEGIN(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",static void *initializeMemCpyOptLegacyPassPassOnce(PassRegistry
&Registry) {
304 false, false)static void *initializeMemCpyOptLegacyPassPassOnce(PassRegistry
&Registry) {
305INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)initializeAssumptionCacheTrackerPass(Registry);
306INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)initializeDominatorTreeWrapperPassPass(Registry);
307INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)initializeMemoryDependenceWrapperPassPass(Registry);
308INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)initializeTargetLibraryInfoWrapperPassPass(Registry);
309INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)initializeAAResultsWrapperPassPass(Registry);
310INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)initializeGlobalsAAWrapperPassPass(Registry);
311INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass)initializeMemorySSAWrapperPassPass(Registry);
312INITIALIZE_PASS_END(MemCpyOptLegacyPass, "memcpyopt", "MemCpy Optimization",PassInfo *PI = new PassInfo( "MemCpy Optimization", "memcpyopt"
, &MemCpyOptLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<MemCpyOptLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeMemCpyOptLegacyPassPassFlag
; void llvm::initializeMemCpyOptLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeMemCpyOptLegacyPassPassFlag
, initializeMemCpyOptLegacyPassPassOnce, std::ref(Registry));
}
313 false, false)PassInfo *PI = new PassInfo( "MemCpy Optimization", "memcpyopt"
, &MemCpyOptLegacyPass::ID, PassInfo::NormalCtor_t(callDefaultCtor
<MemCpyOptLegacyPass>), false, false); Registry.registerPass
(*PI, true); return PI; } static llvm::once_flag InitializeMemCpyOptLegacyPassPassFlag
; void llvm::initializeMemCpyOptLegacyPassPass(PassRegistry &
Registry) { llvm::call_once(InitializeMemCpyOptLegacyPassPassFlag
, initializeMemCpyOptLegacyPassPassOnce, std::ref(Registry));
}
314
315// Check that V is either not accessible by the caller, or unwinding cannot
316// occur between Start and End.
317static bool mayBeVisibleThroughUnwinding(Value *V, Instruction *Start,
318 Instruction *End) {
319 assert(Start->getParent() == End->getParent() && "Must be in same block")(static_cast <bool> (Start->getParent() == End->getParent
() && "Must be in same block") ? void (0) : __assert_fail
("Start->getParent() == End->getParent() && \"Must be in same block\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 319, __extension__ __PRETTY_FUNCTION__))
;
320 if (!Start->getFunction()->doesNotThrow() &&
321 !isa<AllocaInst>(getUnderlyingObject(V))) {
322 for (const Instruction &I :
323 make_range(Start->getIterator(), End->getIterator())) {
324 if (I.mayThrow())
325 return true;
326 }
327 }
328 return false;
329}
330
331void MemCpyOptPass::eraseInstruction(Instruction *I) {
332 if (MSSAU)
333 MSSAU->removeMemoryAccess(I);
334 if (MD)
335 MD->removeInstruction(I);
336 I->eraseFromParent();
337}
338
339// Check for mod or ref of Loc between Start and End, excluding both boundaries.
340// Start and End must be in the same block
341static bool accessedBetween(AliasAnalysis &AA, MemoryLocation Loc,
342 const MemoryUseOrDef *Start,
343 const MemoryUseOrDef *End) {
344 assert(Start->getBlock() == End->getBlock() && "Only local supported")(static_cast <bool> (Start->getBlock() == End->getBlock
() && "Only local supported") ? void (0) : __assert_fail
("Start->getBlock() == End->getBlock() && \"Only local supported\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 344, __extension__ __PRETTY_FUNCTION__))
;
345 for (const MemoryAccess &MA :
346 make_range(++Start->getIterator(), End->getIterator())) {
347 if (isModOrRefSet(AA.getModRefInfo(cast<MemoryUseOrDef>(MA).getMemoryInst(),
348 Loc)))
349 return true;
350 }
351 return false;
352}
353
354// Check for mod of Loc between Start and End, excluding both boundaries.
355// Start and End can be in different blocks.
356static bool writtenBetween(MemorySSA *MSSA, MemoryLocation Loc,
357 const MemoryUseOrDef *Start,
358 const MemoryUseOrDef *End) {
359 // TODO: Only walk until we hit Start.
360 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
361 End->getDefiningAccess(), Loc);
362 return !MSSA->dominates(Clobber, Start);
363}
364
365/// When scanning forward over instructions, we look for some other patterns to
366/// fold away. In particular, this looks for stores to neighboring locations of
367/// memory. If it sees enough consecutive ones, it attempts to merge them
368/// together into a memcpy/memset.
369Instruction *MemCpyOptPass::tryMergingIntoMemset(Instruction *StartInst,
370 Value *StartPtr,
371 Value *ByteVal) {
372 const DataLayout &DL = StartInst->getModule()->getDataLayout();
373
374 // Okay, so we now have a single store that can be splatable. Scan to find
375 // all subsequent stores of the same value to offset from the same pointer.
376 // Join these together into ranges, so we can decide whether contiguous blocks
377 // are stored.
378 MemsetRanges Ranges(DL);
379
380 BasicBlock::iterator BI(StartInst);
381
382 // Keeps track of the last memory use or def before the insertion point for
383 // the new memset. The new MemoryDef for the inserted memsets will be inserted
384 // after MemInsertPoint. It points to either LastMemDef or to the last user
385 // before the insertion point of the memset, if there are any such users.
386 MemoryUseOrDef *MemInsertPoint = nullptr;
387 // Keeps track of the last MemoryDef between StartInst and the insertion point
388 // for the new memset. This will become the defining access of the inserted
389 // memsets.
390 MemoryDef *LastMemDef = nullptr;
391 for (++BI; !BI->isTerminator(); ++BI) {
392 if (MSSAU) {
393 auto *CurrentAcc = cast_or_null<MemoryUseOrDef>(
394 MSSAU->getMemorySSA()->getMemoryAccess(&*BI));
395 if (CurrentAcc) {
396 MemInsertPoint = CurrentAcc;
397 if (auto *CurrentDef = dyn_cast<MemoryDef>(CurrentAcc))
398 LastMemDef = CurrentDef;
399 }
400 }
401
402 // Calls that only access inaccessible memory do not block merging
403 // accessible stores.
404 if (auto *CB = dyn_cast<CallBase>(BI)) {
405 if (CB->onlyAccessesInaccessibleMemory())
406 continue;
407 }
408
409 if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
410 // If the instruction is readnone, ignore it, otherwise bail out. We
411 // don't even allow readonly here because we don't want something like:
412 // A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
413 if (BI->mayWriteToMemory() || BI->mayReadFromMemory())
414 break;
415 continue;
416 }
417
418 if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
419 // If this is a store, see if we can merge it in.
420 if (!NextStore->isSimple()) break;
421
422 Value *StoredVal = NextStore->getValueOperand();
423
424 // Don't convert stores of non-integral pointer types to memsets (which
425 // stores integers).
426 if (DL.isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
427 break;
428
429 // Check to see if this stored value is of the same byte-splattable value.
430 Value *StoredByte = isBytewiseValue(StoredVal, DL);
431 if (isa<UndefValue>(ByteVal) && StoredByte)
432 ByteVal = StoredByte;
433 if (ByteVal != StoredByte)
434 break;
435
436 // Check to see if this store is to a constant offset from the start ptr.
437 Optional<int64_t> Offset =
438 isPointerOffset(StartPtr, NextStore->getPointerOperand(), DL);
439 if (!Offset)
440 break;
441
442 Ranges.addStore(*Offset, NextStore);
443 } else {
444 MemSetInst *MSI = cast<MemSetInst>(BI);
445
446 if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
447 !isa<ConstantInt>(MSI->getLength()))
448 break;
449
450 // Check to see if this store is to a constant offset from the start ptr.
451 Optional<int64_t> Offset = isPointerOffset(StartPtr, MSI->getDest(), DL);
452 if (!Offset)
453 break;
454
455 Ranges.addMemSet(*Offset, MSI);
456 }
457 }
458
459 // If we have no ranges, then we just had a single store with nothing that
460 // could be merged in. This is a very common case of course.
461 if (Ranges.empty())
462 return nullptr;
463
464 // If we had at least one store that could be merged in, add the starting
465 // store as well. We try to avoid this unless there is at least something
466 // interesting as a small compile-time optimization.
467 Ranges.addInst(0, StartInst);
468
469 // If we create any memsets, we put it right before the first instruction that
470 // isn't part of the memset block. This ensure that the memset is dominated
471 // by any addressing instruction needed by the start of the block.
472 IRBuilder<> Builder(&*BI);
473
474 // Now that we have full information about ranges, loop over the ranges and
475 // emit memset's for anything big enough to be worthwhile.
476 Instruction *AMemSet = nullptr;
477 for (const MemsetRange &Range : Ranges) {
478 if (Range.TheStores.size() == 1) continue;
479
480 // If it is profitable to lower this range to memset, do so now.
481 if (!Range.isProfitableToUseMemset(DL))
482 continue;
483
484 // Otherwise, we do want to transform this! Create a new memset.
485 // Get the starting pointer of the block.
486 StartPtr = Range.StartPtr;
487
488 AMemSet = Builder.CreateMemSet(StartPtr, ByteVal, Range.End - Range.Start,
489 MaybeAlign(Range.Alignment));
490 LLVM_DEBUG(dbgs() << "Replace stores:\n"; for (Instruction *SIdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Replace stores:\n"; for (Instruction
*SI : Range.TheStores) dbgs() << *SI << '\n'; dbgs
() << "With: " << *AMemSet << '\n'; } } while
(false)
491 : Range.TheStores) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Replace stores:\n"; for (Instruction
*SI : Range.TheStores) dbgs() << *SI << '\n'; dbgs
() << "With: " << *AMemSet << '\n'; } } while
(false)
492 << *SI << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Replace stores:\n"; for (Instruction
*SI : Range.TheStores) dbgs() << *SI << '\n'; dbgs
() << "With: " << *AMemSet << '\n'; } } while
(false)
493 dbgs() << "With: " << *AMemSet << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Replace stores:\n"; for (Instruction
*SI : Range.TheStores) dbgs() << *SI << '\n'; dbgs
() << "With: " << *AMemSet << '\n'; } } while
(false)
;
494 if (!Range.TheStores.empty())
495 AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
496
497 if (MSSAU) {
498 assert(LastMemDef && MemInsertPoint &&(static_cast <bool> (LastMemDef && MemInsertPoint
&& "Both LastMemDef and MemInsertPoint need to be set"
) ? void (0) : __assert_fail ("LastMemDef && MemInsertPoint && \"Both LastMemDef and MemInsertPoint need to be set\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 499, __extension__ __PRETTY_FUNCTION__))
499 "Both LastMemDef and MemInsertPoint need to be set")(static_cast <bool> (LastMemDef && MemInsertPoint
&& "Both LastMemDef and MemInsertPoint need to be set"
) ? void (0) : __assert_fail ("LastMemDef && MemInsertPoint && \"Both LastMemDef and MemInsertPoint need to be set\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 499, __extension__ __PRETTY_FUNCTION__))
;
500 auto *NewDef =
501 cast<MemoryDef>(MemInsertPoint->getMemoryInst() == &*BI
502 ? MSSAU->createMemoryAccessBefore(
503 AMemSet, LastMemDef, MemInsertPoint)
504 : MSSAU->createMemoryAccessAfter(
505 AMemSet, LastMemDef, MemInsertPoint));
506 MSSAU->insertDef(NewDef, /*RenameUses=*/true);
507 LastMemDef = NewDef;
508 MemInsertPoint = NewDef;
509 }
510
511 // Zap all the stores.
512 for (Instruction *SI : Range.TheStores)
513 eraseInstruction(SI);
514
515 ++NumMemSetInfer;
516 }
517
518 return AMemSet;
519}
520
521// This method try to lift a store instruction before position P.
522// It will lift the store and its argument + that anything that
523// may alias with these.
524// The method returns true if it was successful.
525bool MemCpyOptPass::moveUp(StoreInst *SI, Instruction *P, const LoadInst *LI) {
526 // If the store alias this position, early bail out.
527 MemoryLocation StoreLoc = MemoryLocation::get(SI);
528 if (isModOrRefSet(AA->getModRefInfo(P, StoreLoc)))
529 return false;
530
531 // Keep track of the arguments of all instruction we plan to lift
532 // so we can make sure to lift them as well if appropriate.
533 DenseSet<Instruction*> Args;
534 if (auto *Ptr = dyn_cast<Instruction>(SI->getPointerOperand()))
535 if (Ptr->getParent() == SI->getParent())
536 Args.insert(Ptr);
537
538 // Instruction to lift before P.
539 SmallVector<Instruction *, 8> ToLift{SI};
540
541 // Memory locations of lifted instructions.
542 SmallVector<MemoryLocation, 8> MemLocs{StoreLoc};
543
544 // Lifted calls.
545 SmallVector<const CallBase *, 8> Calls;
546
547 const MemoryLocation LoadLoc = MemoryLocation::get(LI);
548
549 for (auto I = --SI->getIterator(), E = P->getIterator(); I != E; --I) {
550 auto *C = &*I;
551
552 // Make sure hoisting does not perform a store that was not guaranteed to
553 // happen.
554 if (!isGuaranteedToTransferExecutionToSuccessor(C))
555 return false;
556
557 bool MayAlias = isModOrRefSet(AA->getModRefInfo(C, None));
558
559 bool NeedLift = false;
560 if (Args.erase(C))
561 NeedLift = true;
562 else if (MayAlias) {
563 NeedLift = llvm::any_of(MemLocs, [C, this](const MemoryLocation &ML) {
564 return isModOrRefSet(AA->getModRefInfo(C, ML));
565 });
566
567 if (!NeedLift)
568 NeedLift = llvm::any_of(Calls, [C, this](const CallBase *Call) {
569 return isModOrRefSet(AA->getModRefInfo(C, Call));
570 });
571 }
572
573 if (!NeedLift)
574 continue;
575
576 if (MayAlias) {
577 // Since LI is implicitly moved downwards past the lifted instructions,
578 // none of them may modify its source.
579 if (isModSet(AA->getModRefInfo(C, LoadLoc)))
580 return false;
581 else if (const auto *Call = dyn_cast<CallBase>(C)) {
582 // If we can't lift this before P, it's game over.
583 if (isModOrRefSet(AA->getModRefInfo(P, Call)))
584 return false;
585
586 Calls.push_back(Call);
587 } else if (isa<LoadInst>(C) || isa<StoreInst>(C) || isa<VAArgInst>(C)) {
588 // If we can't lift this before P, it's game over.
589 auto ML = MemoryLocation::get(C);
590 if (isModOrRefSet(AA->getModRefInfo(P, ML)))
591 return false;
592
593 MemLocs.push_back(ML);
594 } else
595 // We don't know how to lift this instruction.
596 return false;
597 }
598
599 ToLift.push_back(C);
600 for (unsigned k = 0, e = C->getNumOperands(); k != e; ++k)
601 if (auto *A = dyn_cast<Instruction>(C->getOperand(k))) {
602 if (A->getParent() == SI->getParent()) {
603 // Cannot hoist user of P above P
604 if(A == P) return false;
605 Args.insert(A);
606 }
607 }
608 }
609
610 // Find MSSA insertion point. Normally P will always have a corresponding
611 // memory access before which we can insert. However, with non-standard AA
612 // pipelines, there may be a mismatch between AA and MSSA, in which case we
613 // will scan for a memory access before P. In either case, we know for sure
614 // that at least the load will have a memory access.
615 // TODO: Simplify this once P will be determined by MSSA, in which case the
616 // discrepancy can no longer occur.
617 MemoryUseOrDef *MemInsertPoint = nullptr;
618 if (MSSAU) {
619 if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(P)) {
620 MemInsertPoint = cast<MemoryUseOrDef>(--MA->getIterator());
621 } else {
622 const Instruction *ConstP = P;
623 for (const Instruction &I : make_range(++ConstP->getReverseIterator(),
624 ++LI->getReverseIterator())) {
625 if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(&I)) {
626 MemInsertPoint = MA;
627 break;
628 }
629 }
630 }
631 }
632
633 // We made it, we need to lift.
634 for (auto *I : llvm::reverse(ToLift)) {
635 LLVM_DEBUG(dbgs() << "Lifting " << *I << " before " << *P << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Lifting " << *I <<
" before " << *P << "\n"; } } while (false)
;
636 I->moveBefore(P);
637 if (MSSAU) {
638 assert(MemInsertPoint && "Must have found insert point")(static_cast <bool> (MemInsertPoint && "Must have found insert point"
) ? void (0) : __assert_fail ("MemInsertPoint && \"Must have found insert point\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 638, __extension__ __PRETTY_FUNCTION__))
;
639 if (MemoryUseOrDef *MA = MSSAU->getMemorySSA()->getMemoryAccess(I)) {
640 MSSAU->moveAfter(MA, MemInsertPoint);
641 MemInsertPoint = MA;
642 }
643 }
644 }
645
646 return true;
647}
648
649bool MemCpyOptPass::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
650 if (!SI->isSimple()) return false;
651
652 // Avoid merging nontemporal stores since the resulting
653 // memcpy/memset would not be able to preserve the nontemporal hint.
654 // In theory we could teach how to propagate the !nontemporal metadata to
655 // memset calls. However, that change would force the backend to
656 // conservatively expand !nontemporal memset calls back to sequences of
657 // store instructions (effectively undoing the merging).
658 if (SI->getMetadata(LLVMContext::MD_nontemporal))
659 return false;
660
661 const DataLayout &DL = SI->getModule()->getDataLayout();
662
663 Value *StoredVal = SI->getValueOperand();
664
665 // Not all the transforms below are correct for non-integral pointers, bail
666 // until we've audited the individual pieces.
667 if (DL.isNonIntegralPointerType(StoredVal->getType()->getScalarType()))
668 return false;
669
670 // Load to store forwarding can be interpreted as memcpy.
671 if (LoadInst *LI = dyn_cast<LoadInst>(StoredVal)) {
672 if (LI->isSimple() && LI->hasOneUse() &&
673 LI->getParent() == SI->getParent()) {
674
675 auto *T = LI->getType();
676 if (T->isAggregateType()) {
677 MemoryLocation LoadLoc = MemoryLocation::get(LI);
678
679 // We use alias analysis to check if an instruction may store to
680 // the memory we load from in between the load and the store. If
681 // such an instruction is found, we try to promote there instead
682 // of at the store position.
683 // TODO: Can use MSSA for this.
684 Instruction *P = SI;
685 for (auto &I : make_range(++LI->getIterator(), SI->getIterator())) {
686 if (isModSet(AA->getModRefInfo(&I, LoadLoc))) {
687 P = &I;
688 break;
689 }
690 }
691
692 // We found an instruction that may write to the loaded memory.
693 // We can try to promote at this position instead of the store
694 // position if nothing aliases the store memory after this and the store
695 // destination is not in the range.
696 if (P && P != SI) {
697 if (!moveUp(SI, P, LI))
698 P = nullptr;
699 }
700
701 // If a valid insertion position is found, then we can promote
702 // the load/store pair to a memcpy.
703 if (P) {
704 // If we load from memory that may alias the memory we store to,
705 // memmove must be used to preserve semantic. If not, memcpy can
706 // be used.
707 bool UseMemMove = false;
708 if (!AA->isNoAlias(MemoryLocation::get(SI), LoadLoc))
709 UseMemMove = true;
710
711 uint64_t Size = DL.getTypeStoreSize(T);
712
713 IRBuilder<> Builder(P);
714 Instruction *M;
715 if (UseMemMove)
716 M = Builder.CreateMemMove(
717 SI->getPointerOperand(), SI->getAlign(),
718 LI->getPointerOperand(), LI->getAlign(), Size);
719 else
720 M = Builder.CreateMemCpy(
721 SI->getPointerOperand(), SI->getAlign(),
722 LI->getPointerOperand(), LI->getAlign(), Size);
723
724 LLVM_DEBUG(dbgs() << "Promoting " << *LI << " to " << *SI << " => "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Promoting " << *LI <<
" to " << *SI << " => " << *M << "\n"
; } } while (false)
725 << *M << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Promoting " << *LI <<
" to " << *SI << " => " << *M << "\n"
; } } while (false)
;
726
727 if (MSSAU) {
728 auto *LastDef =
729 cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(SI));
730 auto *NewAccess =
731 MSSAU->createMemoryAccessAfter(M, LastDef, LastDef);
732 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
733 }
734
735 eraseInstruction(SI);
736 eraseInstruction(LI);
737 ++NumMemCpyInstr;
738
739 // Make sure we do not invalidate the iterator.
740 BBI = M->getIterator();
741 return true;
742 }
743 }
744
745 // Detect cases where we're performing call slot forwarding, but
746 // happen to be using a load-store pair to implement it, rather than
747 // a memcpy.
748 CallInst *C = nullptr;
749 if (EnableMemorySSA) {
750 if (auto *LoadClobber = dyn_cast<MemoryUseOrDef>(
751 MSSA->getWalker()->getClobberingMemoryAccess(LI))) {
752 // The load most post-dom the call. Limit to the same block for now.
753 // TODO: Support non-local call-slot optimization?
754 if (LoadClobber->getBlock() == SI->getParent())
755 C = dyn_cast_or_null<CallInst>(LoadClobber->getMemoryInst());
756 }
757 } else {
758 MemDepResult ldep = MD->getDependency(LI);
759 if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
760 C = dyn_cast<CallInst>(ldep.getInst());
761 }
762
763 if (C) {
764 // Check that nothing touches the dest of the "copy" between
765 // the call and the store.
766 MemoryLocation StoreLoc = MemoryLocation::get(SI);
767 if (EnableMemorySSA) {
768 if (accessedBetween(*AA, StoreLoc, MSSA->getMemoryAccess(C),
769 MSSA->getMemoryAccess(SI)))
770 C = nullptr;
771 } else {
772 for (BasicBlock::iterator I = --SI->getIterator(),
773 E = C->getIterator();
774 I != E; --I) {
775 if (isModOrRefSet(AA->getModRefInfo(&*I, StoreLoc))) {
776 C = nullptr;
777 break;
778 }
779 }
780 }
781 }
782
783 if (C) {
784 bool changed = performCallSlotOptzn(
785 LI, SI, SI->getPointerOperand()->stripPointerCasts(),
786 LI->getPointerOperand()->stripPointerCasts(),
787 DL.getTypeStoreSize(SI->getOperand(0)->getType()),
788 commonAlignment(SI->getAlign(), LI->getAlign()), C);
789 if (changed) {
790 eraseInstruction(SI);
791 eraseInstruction(LI);
792 ++NumMemCpyInstr;
793 return true;
794 }
795 }
796 }
797 }
798
799 // There are two cases that are interesting for this code to handle: memcpy
800 // and memset. Right now we only handle memset.
801
802 // Ensure that the value being stored is something that can be memset'able a
803 // byte at a time like "0" or "-1" or any width, as well as things like
804 // 0xA0A0A0A0 and 0.0.
805 auto *V = SI->getOperand(0);
806 if (Value *ByteVal = isBytewiseValue(V, DL)) {
807 if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
808 ByteVal)) {
809 BBI = I->getIterator(); // Don't invalidate iterator.
810 return true;
811 }
812
813 // If we have an aggregate, we try to promote it to memset regardless
814 // of opportunity for merging as it can expose optimization opportunities
815 // in subsequent passes.
816 auto *T = V->getType();
817 if (T->isAggregateType()) {
818 uint64_t Size = DL.getTypeStoreSize(T);
819 IRBuilder<> Builder(SI);
820 auto *M = Builder.CreateMemSet(SI->getPointerOperand(), ByteVal, Size,
821 SI->getAlign());
822
823 LLVM_DEBUG(dbgs() << "Promoting " << *SI << " to " << *M << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Promoting " << *SI <<
" to " << *M << "\n"; } } while (false)
;
824
825 if (MSSAU) {
826 assert(isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(SI)))(static_cast <bool> (isa<MemoryDef>(MSSAU->getMemorySSA
()->getMemoryAccess(SI))) ? void (0) : __assert_fail ("isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(SI))"
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 826, __extension__ __PRETTY_FUNCTION__))
;
827 auto *LastDef =
828 cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(SI));
829 auto *NewAccess = MSSAU->createMemoryAccessAfter(M, LastDef, LastDef);
830 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
831 }
832
833 eraseInstruction(SI);
834 NumMemSetInfer++;
835
836 // Make sure we do not invalidate the iterator.
837 BBI = M->getIterator();
838 return true;
839 }
840 }
841
842 return false;
843}
844
845bool MemCpyOptPass::processMemSet(MemSetInst *MSI, BasicBlock::iterator &BBI) {
846 // See if there is another memset or store neighboring this memset which
847 // allows us to widen out the memset to do a single larger store.
848 if (isa<ConstantInt>(MSI->getLength()) && !MSI->isVolatile())
849 if (Instruction *I = tryMergingIntoMemset(MSI, MSI->getDest(),
850 MSI->getValue())) {
851 BBI = I->getIterator(); // Don't invalidate iterator.
852 return true;
853 }
854 return false;
855}
856
857/// Takes a memcpy and a call that it depends on,
858/// and checks for the possibility of a call slot optimization by having
859/// the call write its result directly into the destination of the memcpy.
860bool MemCpyOptPass::performCallSlotOptzn(Instruction *cpyLoad,
861 Instruction *cpyStore, Value *cpyDest,
862 Value *cpySrc, uint64_t cpyLen,
863 Align cpyAlign, CallInst *C) {
864 // The general transformation to keep in mind is
865 //
866 // call @func(..., src, ...)
867 // memcpy(dest, src, ...)
868 //
869 // ->
870 //
871 // memcpy(dest, src, ...)
872 // call @func(..., dest, ...)
873 //
874 // Since moving the memcpy is technically awkward, we additionally check that
875 // src only holds uninitialized values at the moment of the call, meaning that
876 // the memcpy can be discarded rather than moved.
877
878 // Lifetime marks shouldn't be operated on.
879 if (Function *F = C->getCalledFunction())
880 if (F->isIntrinsic() && F->getIntrinsicID() == Intrinsic::lifetime_start)
881 return false;
882
883 // Require that src be an alloca. This simplifies the reasoning considerably.
884 AllocaInst *srcAlloca = dyn_cast<AllocaInst>(cpySrc);
885 if (!srcAlloca)
886 return false;
887
888 ConstantInt *srcArraySize = dyn_cast<ConstantInt>(srcAlloca->getArraySize());
889 if (!srcArraySize)
890 return false;
891
892 const DataLayout &DL = cpyLoad->getModule()->getDataLayout();
893 uint64_t srcSize = DL.getTypeAllocSize(srcAlloca->getAllocatedType()) *
894 srcArraySize->getZExtValue();
895
896 if (cpyLen < srcSize)
897 return false;
898
899 // Check that accessing the first srcSize bytes of dest will not cause a
900 // trap. Otherwise the transform is invalid since it might cause a trap
901 // to occur earlier than it otherwise would.
902 if (!isDereferenceableAndAlignedPointer(cpyDest, Align(1), APInt(64, cpyLen),
903 DL, C, DT))
904 return false;
905
906 // Make sure that nothing can observe cpyDest being written early. There are
907 // a number of cases to consider:
908 // 1. cpyDest cannot be accessed between C and cpyStore as a precondition of
909 // the transform.
910 // 2. C itself may not access cpyDest (prior to the transform). This is
911 // checked further below.
912 // 3. If cpyDest is accessible to the caller of this function (potentially
913 // captured and not based on an alloca), we need to ensure that we cannot
914 // unwind between C and cpyStore. This is checked here.
915 // 4. If cpyDest is potentially captured, there may be accesses to it from
916 // another thread. In this case, we need to check that cpyStore is
917 // guaranteed to be executed if C is. As it is a non-atomic access, it
918 // renders accesses from other threads undefined.
919 // TODO: This is currently not checked.
920 if (mayBeVisibleThroughUnwinding(cpyDest, C, cpyStore))
921 return false;
922
923 // Check that dest points to memory that is at least as aligned as src.
924 Align srcAlign = srcAlloca->getAlign();
925 bool isDestSufficientlyAligned = srcAlign <= cpyAlign;
926 // If dest is not aligned enough and we can't increase its alignment then
927 // bail out.
928 if (!isDestSufficientlyAligned && !isa<AllocaInst>(cpyDest))
929 return false;
930
931 // Check that src is not accessed except via the call and the memcpy. This
932 // guarantees that it holds only undefined values when passed in (so the final
933 // memcpy can be dropped), that it is not read or written between the call and
934 // the memcpy, and that writing beyond the end of it is undefined.
935 SmallVector<User *, 8> srcUseList(srcAlloca->users());
936 while (!srcUseList.empty()) {
937 User *U = srcUseList.pop_back_val();
938
939 if (isa<BitCastInst>(U) || isa<AddrSpaceCastInst>(U)) {
940 append_range(srcUseList, U->users());
941 continue;
942 }
943 if (GetElementPtrInst *G = dyn_cast<GetElementPtrInst>(U)) {
944 if (!G->hasAllZeroIndices())
945 return false;
946
947 append_range(srcUseList, U->users());
948 continue;
949 }
950 if (const IntrinsicInst *IT = dyn_cast<IntrinsicInst>(U))
951 if (IT->isLifetimeStartOrEnd())
952 continue;
953
954 if (U != C && U != cpyLoad)
955 return false;
956 }
957
958 // Check that src isn't captured by the called function since the
959 // transformation can cause aliasing issues in that case.
960 for (unsigned ArgI = 0, E = C->arg_size(); ArgI != E; ++ArgI)
961 if (C->getArgOperand(ArgI) == cpySrc && !C->doesNotCapture(ArgI))
962 return false;
963
964 // Since we're changing the parameter to the callsite, we need to make sure
965 // that what would be the new parameter dominates the callsite.
966 if (!DT->dominates(cpyDest, C)) {
967 // Support moving a constant index GEP before the call.
968 auto *GEP = dyn_cast<GetElementPtrInst>(cpyDest);
969 if (GEP && GEP->hasAllConstantIndices() &&
970 DT->dominates(GEP->getPointerOperand(), C))
971 GEP->moveBefore(C);
972 else
973 return false;
974 }
975
976 // In addition to knowing that the call does not access src in some
977 // unexpected manner, for example via a global, which we deduce from
978 // the use analysis, we also need to know that it does not sneakily
979 // access dest. We rely on AA to figure this out for us.
980 ModRefInfo MR = AA->getModRefInfo(C, cpyDest, LocationSize::precise(srcSize));
981 // If necessary, perform additional analysis.
982 if (isModOrRefSet(MR))
983 MR = AA->callCapturesBefore(C, cpyDest, LocationSize::precise(srcSize), DT);
984 if (isModOrRefSet(MR))
985 return false;
986
987 // We can't create address space casts here because we don't know if they're
988 // safe for the target.
989 if (cpySrc->getType()->getPointerAddressSpace() !=
990 cpyDest->getType()->getPointerAddressSpace())
991 return false;
992 for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
993 if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc &&
994 cpySrc->getType()->getPointerAddressSpace() !=
995 C->getArgOperand(ArgI)->getType()->getPointerAddressSpace())
996 return false;
997
998 // All the checks have passed, so do the transformation.
999 bool changedArgument = false;
1000 for (unsigned ArgI = 0; ArgI < C->arg_size(); ++ArgI)
1001 if (C->getArgOperand(ArgI)->stripPointerCasts() == cpySrc) {
1002 Value *Dest = cpySrc->getType() == cpyDest->getType() ? cpyDest
1003 : CastInst::CreatePointerCast(cpyDest, cpySrc->getType(),
1004 cpyDest->getName(), C);
1005 changedArgument = true;
1006 if (C->getArgOperand(ArgI)->getType() == Dest->getType())
1007 C->setArgOperand(ArgI, Dest);
1008 else
1009 C->setArgOperand(ArgI, CastInst::CreatePointerCast(
1010 Dest, C->getArgOperand(ArgI)->getType(),
1011 Dest->getName(), C));
1012 }
1013
1014 if (!changedArgument)
1015 return false;
1016
1017 // If the destination wasn't sufficiently aligned then increase its alignment.
1018 if (!isDestSufficientlyAligned) {
1019 assert(isa<AllocaInst>(cpyDest) && "Can only increase alloca alignment!")(static_cast <bool> (isa<AllocaInst>(cpyDest) &&
"Can only increase alloca alignment!") ? void (0) : __assert_fail
("isa<AllocaInst>(cpyDest) && \"Can only increase alloca alignment!\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 1019, __extension__ __PRETTY_FUNCTION__))
;
1020 cast<AllocaInst>(cpyDest)->setAlignment(srcAlign);
1021 }
1022
1023 // Drop any cached information about the call, because we may have changed
1024 // its dependence information by changing its parameter.
1025 if (MD)
1026 MD->removeInstruction(C);
1027
1028 // Update AA metadata
1029 // FIXME: MD_tbaa_struct and MD_mem_parallel_loop_access should also be
1030 // handled here, but combineMetadata doesn't support them yet
1031 unsigned KnownIDs[] = {LLVMContext::MD_tbaa, LLVMContext::MD_alias_scope,
1032 LLVMContext::MD_noalias,
1033 LLVMContext::MD_invariant_group,
1034 LLVMContext::MD_access_group};
1035 combineMetadata(C, cpyLoad, KnownIDs, true);
1036
1037 ++NumCallSlot;
1038 return true;
1039}
1040
1041/// We've found that the (upward scanning) memory dependence of memcpy 'M' is
1042/// the memcpy 'MDep'. Try to simplify M to copy from MDep's input if we can.
1043bool MemCpyOptPass::processMemCpyMemCpyDependence(MemCpyInst *M,
1044 MemCpyInst *MDep) {
1045 // We can only transforms memcpy's where the dest of one is the source of the
1046 // other.
1047 if (M->getSource() != MDep->getDest() || MDep->isVolatile())
1048 return false;
1049
1050 // If dep instruction is reading from our current input, then it is a noop
1051 // transfer and substituting the input won't change this instruction. Just
1052 // ignore the input and let someone else zap MDep. This handles cases like:
1053 // memcpy(a <- a)
1054 // memcpy(b <- a)
1055 if (M->getSource() == MDep->getSource())
1056 return false;
1057
1058 // Second, the length of the memcpy's must be the same, or the preceding one
1059 // must be larger than the following one.
1060 if (MDep->getLength() != M->getLength()) {
1061 ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
1062 ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
1063 if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
1064 return false;
1065 }
1066
1067 // Verify that the copied-from memory doesn't change in between the two
1068 // transfers. For example, in:
1069 // memcpy(a <- b)
1070 // *b = 42;
1071 // memcpy(c <- a)
1072 // It would be invalid to transform the second memcpy into memcpy(c <- b).
1073 //
1074 // TODO: If the code between M and MDep is transparent to the destination "c",
1075 // then we could still perform the xform by moving M up to the first memcpy.
1076 if (EnableMemorySSA) {
1077 // TODO: It would be sufficient to check the MDep source up to the memcpy
1078 // size of M, rather than MDep.
1079 if (writtenBetween(MSSA, MemoryLocation::getForSource(MDep),
1080 MSSA->getMemoryAccess(MDep), MSSA->getMemoryAccess(M)))
1081 return false;
1082 } else {
1083 // NOTE: This is conservative, it will stop on any read from the source loc,
1084 // not just the defining memcpy.
1085 MemDepResult SourceDep =
1086 MD->getPointerDependencyFrom(MemoryLocation::getForSource(MDep), false,
1087 M->getIterator(), M->getParent());
1088 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
1089 return false;
1090 }
1091
1092 // If the dest of the second might alias the source of the first, then the
1093 // source and dest might overlap. We still want to eliminate the intermediate
1094 // value, but we have to generate a memmove instead of memcpy.
1095 bool UseMemMove = false;
1096 if (!AA->isNoAlias(MemoryLocation::getForDest(M),
1097 MemoryLocation::getForSource(MDep)))
1098 UseMemMove = true;
1099
1100 // If all checks passed, then we can transform M.
1101 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
<< *MDep << '\n' << *M << '\n'; } } while
(false)
1102 << *MDep << '\n' << *M << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Forwarding memcpy->memcpy src:\n"
<< *MDep << '\n' << *M << '\n'; } } while
(false)
;
1103
1104 // TODO: Is this worth it if we're creating a less aligned memcpy? For
1105 // example we could be moving from movaps -> movq on x86.
1106 IRBuilder<> Builder(M);
1107 Instruction *NewM;
1108 if (UseMemMove)
1109 NewM = Builder.CreateMemMove(M->getRawDest(), M->getDestAlign(),
1110 MDep->getRawSource(), MDep->getSourceAlign(),
1111 M->getLength(), M->isVolatile());
1112 else if (isa<MemCpyInlineInst>(M)) {
1113 // llvm.memcpy may be promoted to llvm.memcpy.inline, but the converse is
1114 // never allowed since that would allow the latter to be lowered as a call
1115 // to an external function.
1116 NewM = Builder.CreateMemCpyInline(
1117 M->getRawDest(), M->getDestAlign(), MDep->getRawSource(),
1118 MDep->getSourceAlign(), M->getLength(), M->isVolatile());
1119 } else
1120 NewM = Builder.CreateMemCpy(M->getRawDest(), M->getDestAlign(),
1121 MDep->getRawSource(), MDep->getSourceAlign(),
1122 M->getLength(), M->isVolatile());
1123
1124 if (MSSAU) {
1125 assert(isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M)))(static_cast <bool> (isa<MemoryDef>(MSSAU->getMemorySSA
()->getMemoryAccess(M))) ? void (0) : __assert_fail ("isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M))"
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 1125, __extension__ __PRETTY_FUNCTION__))
;
1126 auto *LastDef = cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M));
1127 auto *NewAccess = MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1128 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1129 }
1130
1131 // Remove the instruction we're replacing.
1132 eraseInstruction(M);
1133 ++NumMemCpyInstr;
1134 return true;
1135}
1136
1137/// We've found that the (upward scanning) memory dependence of \p MemCpy is
1138/// \p MemSet. Try to simplify \p MemSet to only set the trailing bytes that
1139/// weren't copied over by \p MemCpy.
1140///
1141/// In other words, transform:
1142/// \code
1143/// memset(dst, c, dst_size);
1144/// memcpy(dst, src, src_size);
1145/// \endcode
1146/// into:
1147/// \code
1148/// memcpy(dst, src, src_size);
1149/// memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
1150/// \endcode
1151bool MemCpyOptPass::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
1152 MemSetInst *MemSet) {
1153 // We can only transform memset/memcpy with the same destination.
1154 if (!AA->isMustAlias(MemSet->getDest(), MemCpy->getDest()))
1155 return false;
1156
1157 // Check that src and dst of the memcpy aren't the same. While memcpy
1158 // operands cannot partially overlap, exact equality is allowed.
1159 if (!AA->isNoAlias(MemoryLocation(MemCpy->getSource(),
1160 LocationSize::precise(1)),
1161 MemoryLocation(MemCpy->getDest(),
1162 LocationSize::precise(1))))
1163 return false;
1164
1165 if (EnableMemorySSA) {
1166 // We know that dst up to src_size is not written. We now need to make sure
1167 // that dst up to dst_size is not accessed. (If we did not move the memset,
1168 // checking for reads would be sufficient.)
1169 if (accessedBetween(*AA, MemoryLocation::getForDest(MemSet),
1170 MSSA->getMemoryAccess(MemSet),
1171 MSSA->getMemoryAccess(MemCpy))) {
1172 return false;
1173 }
1174 } else {
1175 // We have already checked that dst up to src_size is not accessed. We
1176 // need to make sure that there are no accesses up to dst_size either.
1177 MemDepResult DstDepInfo = MD->getPointerDependencyFrom(
1178 MemoryLocation::getForDest(MemSet), false, MemCpy->getIterator(),
1179 MemCpy->getParent());
1180 if (DstDepInfo.getInst() != MemSet)
1181 return false;
1182 }
1183
1184 // Use the same i8* dest as the memcpy, killing the memset dest if different.
1185 Value *Dest = MemCpy->getRawDest();
1186 Value *DestSize = MemSet->getLength();
1187 Value *SrcSize = MemCpy->getLength();
1188
1189 if (mayBeVisibleThroughUnwinding(Dest, MemSet, MemCpy))
1190 return false;
1191
1192 // If the sizes are the same, simply drop the memset instead of generating
1193 // a replacement with zero size.
1194 if (DestSize == SrcSize) {
1195 eraseInstruction(MemSet);
1196 return true;
1197 }
1198
1199 // By default, create an unaligned memset.
1200 unsigned Align = 1;
1201 // If Dest is aligned, and SrcSize is constant, use the minimum alignment
1202 // of the sum.
1203 const unsigned DestAlign =
1204 std::max(MemSet->getDestAlignment(), MemCpy->getDestAlignment());
1205 if (DestAlign > 1)
1206 if (ConstantInt *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
1207 Align = MinAlign(SrcSizeC->getZExtValue(), DestAlign);
1208
1209 IRBuilder<> Builder(MemCpy);
1210
1211 // If the sizes have different types, zext the smaller one.
1212 if (DestSize->getType() != SrcSize->getType()) {
1213 if (DestSize->getType()->getIntegerBitWidth() >
1214 SrcSize->getType()->getIntegerBitWidth())
1215 SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
1216 else
1217 DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
1218 }
1219
1220 Value *Ule = Builder.CreateICmpULE(DestSize, SrcSize);
1221 Value *SizeDiff = Builder.CreateSub(DestSize, SrcSize);
1222 Value *MemsetLen = Builder.CreateSelect(
1223 Ule, ConstantInt::getNullValue(DestSize->getType()), SizeDiff);
1224 unsigned DestAS = Dest->getType()->getPointerAddressSpace();
1225 Instruction *NewMemSet = Builder.CreateMemSet(
1226 Builder.CreateGEP(Builder.getInt8Ty(),
1227 Builder.CreatePointerCast(Dest,
1228 Builder.getInt8PtrTy(DestAS)),
1229 SrcSize),
1230 MemSet->getOperand(1), MemsetLen, MaybeAlign(Align));
1231
1232 if (MSSAU) {
1233 assert(isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy)) &&(static_cast <bool> (isa<MemoryDef>(MSSAU->getMemorySSA
()->getMemoryAccess(MemCpy)) && "MemCpy must be a MemoryDef"
) ? void (0) : __assert_fail ("isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy)) && \"MemCpy must be a MemoryDef\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 1234, __extension__ __PRETTY_FUNCTION__))
1234 "MemCpy must be a MemoryDef")(static_cast <bool> (isa<MemoryDef>(MSSAU->getMemorySSA
()->getMemoryAccess(MemCpy)) && "MemCpy must be a MemoryDef"
) ? void (0) : __assert_fail ("isa<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy)) && \"MemCpy must be a MemoryDef\""
, "/build/llvm-toolchain-snapshot-13~++20210726100616+dead50d4427c/llvm/lib/Transforms/Scalar/MemCpyOptimizer.cpp"
, 1234, __extension__ __PRETTY_FUNCTION__))
;
1235 // The new memset is inserted after the memcpy, but it is known that its
1236 // defining access is the memset about to be removed which immediately
1237 // precedes the memcpy.
1238 auto *LastDef =
1239 cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy));
1240 auto *NewAccess = MSSAU->createMemoryAccessBefore(
1241 NewMemSet, LastDef->getDefiningAccess(), LastDef);
1242 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1243 }
1244
1245 eraseInstruction(MemSet);
1246 return true;
1247}
1248
1249/// Determine whether the instruction has undefined content for the given Size,
1250/// either because it was freshly alloca'd or started its lifetime.
1251static bool hasUndefContents(Instruction *I, Value *Size) {
1252 if (isa<AllocaInst>(I))
1253 return true;
1254
1255 if (ConstantInt *CSize = dyn_cast<ConstantInt>(Size)) {
1256 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1257 if (II->getIntrinsicID() == Intrinsic::lifetime_start)
1258 if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
1259 if (LTSize->getZExtValue() >= CSize->getZExtValue())
1260 return true;
1261 }
1262
1263 return false;
1264}
1265
1266static bool hasUndefContentsMSSA(MemorySSA *MSSA, AliasAnalysis *AA, Value *V,
1267 MemoryDef *Def, Value *Size) {
1268 if (MSSA->isLiveOnEntryDef(Def))
1269 return isa<AllocaInst>(getUnderlyingObject(V));
1270
1271 if (IntrinsicInst *II =
1272 dyn_cast_or_null<IntrinsicInst>(Def->getMemoryInst())) {
1273 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1274 ConstantInt *LTSize = cast<ConstantInt>(II->getArgOperand(0));
1275
1276 if (ConstantInt *CSize = dyn_cast<ConstantInt>(Size)) {
1277 if (AA->isMustAlias(V, II->getArgOperand(1)) &&
1278 LTSize->getZExtValue() >= CSize->getZExtValue())
1279 return true;
1280 }
1281
1282 // If the lifetime.start covers a whole alloca (as it almost always
1283 // does) and we're querying a pointer based on that alloca, then we know
1284 // the memory is definitely undef, regardless of how exactly we alias.
1285 // The size also doesn't matter, as an out-of-bounds access would be UB.
1286 AllocaInst *Alloca = dyn_cast<AllocaInst>(getUnderlyingObject(V));
1287 if (getUnderlyingObject(II->getArgOperand(1)) == Alloca) {
1288 const DataLayout &DL = Alloca->getModule()->getDataLayout();
1289 if (Optional<TypeSize> AllocaSize = Alloca->getAllocationSizeInBits(DL))
1290 if (*AllocaSize == LTSize->getValue() * 8)
1291 return true;
1292 }
1293 }
1294 }
1295
1296 return false;
1297}
1298
1299/// Transform memcpy to memset when its source was just memset.
1300/// In other words, turn:
1301/// \code
1302/// memset(dst1, c, dst1_size);
1303/// memcpy(dst2, dst1, dst2_size);
1304/// \endcode
1305/// into:
1306/// \code
1307/// memset(dst1, c, dst1_size);
1308/// memset(dst2, c, dst2_size);
1309/// \endcode
1310/// When dst2_size <= dst1_size.
1311bool MemCpyOptPass::performMemCpyToMemSetOptzn(MemCpyInst *MemCpy,
1312 MemSetInst *MemSet) {
1313 // Make sure that memcpy(..., memset(...), ...), that is we are memsetting and
1314 // memcpying from the same address. Otherwise it is hard to reason about.
1315 if (!AA->isMustAlias(MemSet->getRawDest(), MemCpy->getRawSource()))
1316 return false;
1317
1318 Value *MemSetSize = MemSet->getLength();
1319 Value *CopySize = MemCpy->getLength();
1320
1321 if (MemSetSize != CopySize) {
1322 // Make sure the memcpy doesn't read any more than what the memset wrote.
1323 // Don't worry about sizes larger than i64.
1324
1325 // A known memset size is required.
1326 ConstantInt *CMemSetSize = dyn_cast<ConstantInt>(MemSetSize);
1327 if (!CMemSetSize)
1328 return false;
1329
1330 // A known memcpy size is also required.
1331 ConstantInt *CCopySize = dyn_cast<ConstantInt>(CopySize);
1332 if (!CCopySize)
1333 return false;
1334 if (CCopySize->getZExtValue() > CMemSetSize->getZExtValue()) {
1335 // If the memcpy is larger than the memset, but the memory was undef prior
1336 // to the memset, we can just ignore the tail. Technically we're only
1337 // interested in the bytes from MemSetSize..CopySize here, but as we can't
1338 // easily represent this location, we use the full 0..CopySize range.
1339 MemoryLocation MemCpyLoc = MemoryLocation::getForSource(MemCpy);
1340 bool CanReduceSize = false;
1341 if (EnableMemorySSA) {
1342 MemoryUseOrDef *MemSetAccess = MSSA->getMemoryAccess(MemSet);
1343 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
1344 MemSetAccess->getDefiningAccess(), MemCpyLoc);
1345 if (auto *MD = dyn_cast<MemoryDef>(Clobber))
1346 if (hasUndefContentsMSSA(MSSA, AA, MemCpy->getSource(), MD, CopySize))
1347 CanReduceSize = true;
1348 } else {
1349 MemDepResult DepInfo = MD->getPointerDependencyFrom(
1350 MemCpyLoc, true, MemSet->getIterator(), MemSet->getParent());
1351 if (DepInfo.isDef() && hasUndefContents(DepInfo.getInst(), CopySize))
1352 CanReduceSize = true;
1353 }
1354
1355 if (!CanReduceSize)
1356 return false;
1357 CopySize = MemSetSize;
1358 }
1359 }
1360
1361 IRBuilder<> Builder(MemCpy);
1362 Instruction *NewM =
1363 Builder.CreateMemSet(MemCpy->getRawDest(), MemSet->getOperand(1),
1364 CopySize, MaybeAlign(MemCpy->getDestAlignment()));
1365 if (MSSAU) {
1366 auto *LastDef =
1367 cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(MemCpy));
1368 auto *NewAccess = MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1369 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1370 }
1371
1372 return true;
1373}
1374
1375/// Perform simplification of memcpy's. If we have memcpy A
1376/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
1377/// B to be a memcpy from X to Z (or potentially a memmove, depending on
1378/// circumstances). This allows later passes to remove the first memcpy
1379/// altogether.
1380bool MemCpyOptPass::processMemCpy(MemCpyInst *M, BasicBlock::iterator &BBI) {
1381 // We can only optimize non-volatile memcpy's.
1382 if (M->isVolatile()) return false;
1383
1384 // If the source and destination of the memcpy are the same, then zap it.
1385 if (M->getSource() == M->getDest()) {
1386 ++BBI;
1387 eraseInstruction(M);
1388 return true;
1389 }
1390
1391 // If copying from a constant, try to turn the memcpy into a memset.
1392 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
1393 if (GV->isConstant() && GV->hasDefinitiveInitializer())
1394 if (Value *ByteVal = isBytewiseValue(GV->getInitializer(),
1395 M->getModule()->getDataLayout())) {
1396 IRBuilder<> Builder(M);
1397 Instruction *NewM =
1398 Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
1399 MaybeAlign(M->getDestAlignment()), false);
1400 if (MSSAU) {
1401 auto *LastDef =
1402 cast<MemoryDef>(MSSAU->getMemorySSA()->getMemoryAccess(M));
1403 auto *NewAccess =
1404 MSSAU->createMemoryAccessAfter(NewM, LastDef, LastDef);
1405 MSSAU->insertDef(cast<MemoryDef>(NewAccess), /*RenameUses=*/true);
1406 }
1407
1408 eraseInstruction(M);
1409 ++NumCpyToSet;
1410 return true;
1411 }
1412
1413 if (EnableMemorySSA) {
1414 MemoryUseOrDef *MA = MSSA->getMemoryAccess(M);
1415 MemoryAccess *AnyClobber = MSSA->getWalker()->getClobberingMemoryAccess(MA);
1416 MemoryLocation DestLoc = MemoryLocation::getForDest(M);
1417 const MemoryAccess *DestClobber =
1418 MSSA->getWalker()->getClobberingMemoryAccess(AnyClobber, DestLoc);
1419
1420 // Try to turn a partially redundant memset + memcpy into
1421 // memcpy + smaller memset. We don't need the memcpy size for this.
1422 // The memcpy most post-dom the memset, so limit this to the same basic
1423 // block. A non-local generalization is likely not worthwhile.
1424 if (auto *MD = dyn_cast<MemoryDef>(DestClobber))
1425 if (auto *MDep = dyn_cast_or_null<MemSetInst>(MD->getMemoryInst()))
1426 if (DestClobber->getBlock() == M->getParent())
1427 if (processMemSetMemCpyDependence(M, MDep))
1428 return true;
1429
1430 MemoryAccess *SrcClobber = MSSA->getWalker()->getClobberingMemoryAccess(
1431 AnyClobber, MemoryLocation::getForSource(M));
1432
1433 // There are four possible optimizations we can do for memcpy:
1434 // a) memcpy-memcpy xform which exposes redundance for DSE.
1435 // b) call-memcpy xform for return slot optimization.
1436 // c) memcpy from freshly alloca'd space or space that has just started
1437 // its lifetime copies undefined data, and we can therefore eliminate
1438 // the memcpy in favor of the data that was already at the destination.
1439 // d) memcpy from a just-memset'd source can be turned into memset.
1440 if (auto *MD = dyn_cast<MemoryDef>(SrcClobber)) {
1441 if (Instruction *MI = MD->getMemoryInst()) {
1442 if (ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength())) {
1443 if (auto *C = dyn_cast<CallInst>(MI)) {
1444 // The memcpy must post-dom the call. Limit to the same block for
1445 // now. Additionally, we need to ensure that there are no accesses
1446 // to dest between the call and the memcpy. Accesses to src will be
1447 // checked by performCallSlotOptzn().
1448 // TODO: Support non-local call-slot optimization?
1449 if (C->getParent() == M->getParent() &&
1450 !accessedBetween(*AA, DestLoc, MD, MA)) {
1451 // FIXME: Can we pass in either of dest/src alignment here instead
1452 // of conservatively taking the minimum?
1453 Align Alignment = std::min(M->getDestAlign().valueOrOne(),
1454 M->getSourceAlign().valueOrOne());
1455 if (performCallSlotOptzn(M, M, M->getDest(), M->getSource(),
1456 CopySize->getZExtValue(), Alignment,
1457 C)) {
1458 LLVM_DEBUG(dbgs() << "Performed call slot optimization:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Performed call slot optimization:\n"
<< " call: " << *C << "\n" << " memcpy: "
<< *M << "\n"; } } while (false)
1459 << " call: " << *C << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Performed call slot optimization:\n"
<< " call: " << *C << "\n" << " memcpy: "
<< *M << "\n"; } } while (false)
1460 << " memcpy: " << *M << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Performed call slot optimization:\n"
<< " call: " << *C << "\n" << " memcpy: "
<< *M << "\n"; } } while (false)
;
1461 eraseInstruction(M);
1462 ++NumMemCpyInstr;
1463 return true;
1464 }
1465 }
1466 }
1467 }
1468 if (auto *MDep = dyn_cast<MemCpyInst>(MI))
1469 return processMemCpyMemCpyDependence(M, MDep);
1470 if (auto *MDep = dyn_cast<MemSetInst>(MI)) {
1471 if (performMemCpyToMemSetOptzn(M, MDep)) {
1472 LLVM_DEBUG(dbgs() << "Converted memcpy to memset\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Converted memcpy to memset\n"
; } } while (false)
;
1473 eraseInstruction(M);
1474 ++NumCpyToSet;
1475 return true;
1476 }
1477 }
1478 }
1479
1480 if (hasUndefContentsMSSA(MSSA, AA, M->getSource(), MD, M->getLength())) {
1481 LLVM_DEBUG(dbgs() << "Removed memcpy from undef\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "Removed memcpy from undef\n"
; } } while (false)
;
1482 eraseInstruction(M);
1483 ++NumMemCpyInstr;
1484 return true;
1485 }
1486 }
1487 } else {
1488 MemDepResult DepInfo = MD->getDependency(M);
1489
1490 // Try to turn a partially redundant memset + memcpy into
1491 // memcpy + smaller memset. We don't need the memcpy size for this.
1492 if (DepInfo.isClobber())
1493 if (MemSetInst *MDep = dyn_cast<MemSetInst>(DepInfo.getInst()))
1494 if (processMemSetMemCpyDependence(M, MDep))
1495 return true;
1496
1497 // There are four possible optimizations we can do for memcpy:
1498 // a) memcpy-memcpy xform which exposes redundance for DSE.
1499 // b) call-memcpy xform for return slot optimization.
1500 // c) memcpy from freshly alloca'd space or space that has just started
1501 // its lifetime copies undefined data, and we can therefore eliminate
1502 // the memcpy in favor of the data that was already at the destination.
1503 // d) memcpy from a just-memset'd source can be turned into memset.
1504 if (ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength())) {
1505 if (DepInfo.isClobber()) {
1506 if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
1507 // FIXME: Can we pass in either of dest/src alignment here instead
1508 // of conservatively taking the minimum?
1509 Align Alignment = std::min(M->getDestAlign().valueOrOne(),
1510 M->getSourceAlign().valueOrOne());
1511 if (performCallSlotOptzn(M, M, M->getDest(), M->getSource(),
1512 CopySize->getZExtValue(), Alignment, C)) {
1513 eraseInstruction(M);
1514 ++NumMemCpyInstr;
1515 return true;
1516 }
1517 }
1518 }
1519 }
1520
1521 MemoryLocation SrcLoc = MemoryLocation::getForSource(M);
1522 MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(
1523 SrcLoc, true, M->getIterator(), M->getParent());
1524
1525 if (SrcDepInfo.isClobber()) {
1526 if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
1527 return processMemCpyMemCpyDependence(M, MDep);
1528 } else if (SrcDepInfo.isDef()) {
1529 if (hasUndefContents(SrcDepInfo.getInst(), M->getLength())) {
1530 eraseInstruction(M);
1531 ++NumMemCpyInstr;
1532 return true;
1533 }
1534 }
1535
1536 if (SrcDepInfo.isClobber())
1537 if (MemSetInst *MDep = dyn_cast<MemSetInst>(SrcDepInfo.getInst()))
1538 if (performMemCpyToMemSetOptzn(M, MDep)) {
1539 eraseInstruction(M);
1540 ++NumCpyToSet;
1541 return true;
1542 }
1543 }
1544
1545 return false;
1546}
1547
1548/// Transforms memmove calls to memcpy calls when the src/dst are guaranteed
1549/// not to alias.
1550bool MemCpyOptPass::processMemMove(MemMoveInst *M) {
1551 if (!TLI->has(LibFunc_memmove))
1552 return false;
1553
1554 // See if the pointers alias.
1555 if (!AA->isNoAlias(MemoryLocation::getForDest(M),
1556 MemoryLocation::getForSource(M)))
1557 return false;
1558
1559 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: " << *Mdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: "
<< *M << "\n"; } } while (false)
1560 << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Optimizing memmove -> memcpy: "
<< *M << "\n"; } } while (false)
;
1561
1562 // If not, then we know we can transform this.
1563 Type *ArgTys[3] = { M->getRawDest()->getType(),
1564 M->getRawSource()->getType(),
1565 M->getLength()->getType() };
1566 M->setCalledFunction(Intrinsic::getDeclaration(M->getModule(),
1567 Intrinsic::memcpy, ArgTys));
1568
1569 // For MemorySSA nothing really changes (except that memcpy may imply stricter
1570 // aliasing guarantees).
1571
1572 // MemDep may have over conservative information about this instruction, just
1573 // conservatively flush it from the cache.
1574 if (MD)
1575 MD->removeInstruction(M);
1576
1577 ++NumMoveToCpy;
1578 return true;
1579}
1580
1581/// This is called on every byval argument in call sites.
1582bool MemCpyOptPass::processByValArgument(CallBase &CB, unsigned ArgNo) {
1583 const DataLayout &DL = CB.getCaller()->getParent()->getDataLayout();
1584 // Find out what feeds this byval argument.
1585 Value *ByValArg = CB.getArgOperand(ArgNo);
1586 Type *ByValTy = CB.getParamByValType(ArgNo);
1587 uint64_t ByValSize = DL.getTypeAllocSize(ByValTy);
1588 MemoryLocation Loc(ByValArg, LocationSize::precise(ByValSize));
1589 MemCpyInst *MDep = nullptr;
1590 if (EnableMemorySSA) {
35
Assuming the condition is true
36
Taking true branch
1591 MemoryUseOrDef *CallAccess = MSSA->getMemoryAccess(&CB);
37
Called C++ object pointer is null
1592 if (!CallAccess)
1593 return false;
1594 MemoryAccess *Clobber = MSSA->getWalker()->getClobberingMemoryAccess(
1595 CallAccess->getDefiningAccess(), Loc);
1596 if (auto *MD = dyn_cast<MemoryDef>(Clobber))
1597 MDep = dyn_cast_or_null<MemCpyInst>(MD->getMemoryInst());
1598 } else {
1599 MemDepResult DepInfo = MD->getPointerDependencyFrom(
1600 Loc, true, CB.getIterator(), CB.getParent());
1601 if (!DepInfo.isClobber())
1602 return false;
1603 MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
1604 }
1605
1606 // If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
1607 // a memcpy, see if we can byval from the source of the memcpy instead of the
1608 // result.
1609 if (!MDep || MDep->isVolatile() ||
1610 ByValArg->stripPointerCasts() != MDep->getDest())
1611 return false;
1612
1613 // The length of the memcpy must be larger or equal to the size of the byval.
1614 ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
1615 if (!C1 || C1->getValue().getZExtValue() < ByValSize)
1616 return false;
1617
1618 // Get the alignment of the byval. If the call doesn't specify the alignment,
1619 // then it is some target specific value that we can't know.
1620 MaybeAlign ByValAlign = CB.getParamAlign(ArgNo);
1621 if (!ByValAlign) return false;
1622
1623 // If it is greater than the memcpy, then we check to see if we can force the
1624 // source of the memcpy to the alignment we need. If we fail, we bail out.
1625 MaybeAlign MemDepAlign = MDep->getSourceAlign();
1626 if ((!MemDepAlign || *MemDepAlign < *ByValAlign) &&
1627 getOrEnforceKnownAlignment(MDep->getSource(), ByValAlign, DL, &CB, AC,
1628 DT) < *ByValAlign)
1629 return false;
1630
1631 // The address space of the memcpy source must match the byval argument
1632 if (MDep->getSource()->getType()->getPointerAddressSpace() !=
1633 ByValArg->getType()->getPointerAddressSpace())
1634 return false;
1635
1636 // Verify that the copied-from memory doesn't change in between the memcpy and
1637 // the byval call.
1638 // memcpy(a <- b)
1639 // *b = 42;
1640 // foo(*a)
1641 // It would be invalid to transform the second memcpy into foo(*b).
1642 if (EnableMemorySSA) {
1643 if (writtenBetween(MSSA, MemoryLocation::getForSource(MDep),
1644 MSSA->getMemoryAccess(MDep), MSSA->getMemoryAccess(&CB)))
1645 return false;
1646 } else {
1647 // NOTE: This is conservative, it will stop on any read from the source loc,
1648 // not just the defining memcpy.
1649 MemDepResult SourceDep = MD->getPointerDependencyFrom(
1650 MemoryLocation::getForSource(MDep), false,
1651 CB.getIterator(), MDep->getParent());
1652 if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
1653 return false;
1654 }
1655
1656 Value *TmpCast = MDep->getSource();
1657 if (MDep->getSource()->getType() != ByValArg->getType()) {
1658 BitCastInst *TmpBitCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
1659 "tmpcast", &CB);
1660 // Set the tmpcast's DebugLoc to MDep's
1661 TmpBitCast->setDebugLoc(MDep->getDebugLoc());
1662 TmpCast = TmpBitCast;
1663 }
1664
1665 LLVM_DEBUG(dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
<< " " << *MDep << "\n" << " " <<
CB << "\n"; } } while (false)
1666 << " " << *MDep << "\n"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
<< " " << *MDep << "\n" << " " <<
CB << "\n"; } } while (false)
1667 << " " << CB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType
("memcpyopt")) { dbgs() << "MemCpyOptPass: Forwarding memcpy to byval:\n"
<< " " << *MDep << "\n" << " " <<
CB << "\n"; } } while (false)
;
1668
1669 // Otherwise we're good! Update the byval argument.
1670 CB.setArgOperand(ArgNo, TmpCast);
1671 ++NumMemCpyInstr;
1672 return true;
1673}
1674
1675/// Executes one iteration of MemCpyOptPass.
1676bool MemCpyOptPass::iterateOnFunction(Function &F) {
1677 bool MadeChange = false;
1678
1679 // Walk all instruction in the function.
1680 for (BasicBlock &BB : F) {
1681 // Skip unreachable blocks. For example processStore assumes that an
1682 // instruction in a BB can't be dominated by a later instruction in the
1683 // same BB (which is a scenario that can happen for an unreachable BB that
1684 // has itself as a predecessor).
1685 if (!DT->isReachableFromEntry(&BB))
17
Assuming the condition is false
18
Taking false branch
1686 continue;
1687
1688 for (BasicBlock::iterator BI = BB.begin(), BE = BB.end(); BI != BE;) {
19
Loop condition is true. Entering loop body
1689 // Avoid invalidating the iterator.
1690 Instruction *I = &*BI++;
1691
1692 bool RepeatInstruction = false;
1693
1694 if (StoreInst *SI
20.1
'SI' is null
= dyn_cast<StoreInst>(I))
20
Assuming 'I' is not a 'StoreInst'
21
Taking false branch
1695 MadeChange |= processStore(SI, BI);
1696 else if (MemSetInst *M
22.1
'M' is null
= dyn_cast<MemSetInst>(I))
22
Assuming 'I' is not a 'MemSetInst'
23
Taking false branch
1697 RepeatInstruction = processMemSet(M, BI);
1698 else if (MemCpyInst *M
24.1
'M' is null
= dyn_cast<MemCpyInst>(I))
24
Assuming 'I' is not a 'MemCpyInst'
25
Taking false branch
1699 RepeatInstruction = processMemCpy(M, BI);
1700 else if (MemMoveInst *M
26.1
'M' is null
= dyn_cast<MemMoveInst>(I))
26
Assuming 'I' is not a 'MemMoveInst'
27
Taking false branch
1701 RepeatInstruction = processMemMove(M);
1702 else if (auto *CB
28.1
'CB' is non-null
= dyn_cast<CallBase>(I)) {
28
Assuming 'I' is a 'CallBase'
29
Taking true branch
1703 for (unsigned i = 0, e = CB->arg_size(); i != e; ++i)
30
Assuming 'i' is not equal to 'e'
31
Loop condition is true. Entering loop body
1704 if (CB->isByValArgument(i))
32
Assuming the condition is true
33
Taking true branch
1705 MadeChange |= processByValArgument(*CB, i);
34
Calling 'MemCpyOptPass::processByValArgument'
1706 }
1707
1708 // Reprocess the instruction if desired.
1709 if (RepeatInstruction) {
1710 if (BI != BB.begin())
1711 --BI;
1712 MadeChange = true;
1713 }
1714 }
1715 }
1716
1717 return MadeChange;
1718}
1719
1720PreservedAnalyses MemCpyOptPass::run(Function &F, FunctionAnalysisManager &AM) {
1721 auto *MD = !EnableMemorySSA ? &AM.getResult<MemoryDependenceAnalysis>(F)
1722 : AM.getCachedResult<MemoryDependenceAnalysis>(F);
1723 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
1724 auto *AA = &AM.getResult<AAManager>(F);
1725 auto *AC = &AM.getResult<AssumptionAnalysis>(F);
1726 auto *DT = &AM.getResult<DominatorTreeAnalysis>(F);
1727 auto *MSSA = EnableMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F)
1728 : AM.getCachedResult<MemorySSAAnalysis>(F);
1729
1730 bool MadeChange =
1731 runImpl(F, MD, &TLI, AA, AC, DT, MSSA ? &MSSA->getMSSA() : nullptr);
1732 if (!MadeChange)
1733 return PreservedAnalyses::all();
1734
1735 PreservedAnalyses PA;
1736 PA.preserveSet<CFGAnalyses>();
1737 if (MD)
1738 PA.preserve<MemoryDependenceAnalysis>();
1739 if (MSSA)
1740 PA.preserve<MemorySSAAnalysis>();
1741 return PA;
1742}
1743
1744bool MemCpyOptPass::runImpl(Function &F, MemoryDependenceResults *MD_,
1745 TargetLibraryInfo *TLI_, AliasAnalysis *AA_,
1746 AssumptionCache *AC_, DominatorTree *DT_,
1747 MemorySSA *MSSA_) {
1748 bool MadeChange = false;
1749 MD = MD_;
1750 TLI = TLI_;
1751 AA = AA_;
1752 AC = AC_;
1753 DT = DT_;
1754 MSSA = MSSA_;
12
Null pointer value stored to field 'MSSA'
1755 MemorySSAUpdater MSSAU_(MSSA_);
1756 MSSAU = MSSA_
12.1
'MSSA_' is null
? &MSSAU_ : nullptr;
13
'?' condition is false
1757 // If we don't have at least memset and memcpy, there is little point of doing
1758 // anything here. These are required by a freestanding implementation, so if
1759 // even they are disabled, there is no point in trying hard.
1760 if (!TLI->has(LibFunc_memset) || !TLI->has(LibFunc_memcpy))
14
Taking false branch
1761 return false;
1762
1763 while (true) {
15
Loop condition is true. Entering loop body
1764 if (!iterateOnFunction(F))
16
Calling 'MemCpyOptPass::iterateOnFunction'
1765 break;
1766 MadeChange = true;
1767 }
1768
1769 if (MSSA_ && VerifyMemorySSA)
1770 MSSA_->verifyMemorySSA();
1771
1772 MD = nullptr;
1773 return MadeChange;
1774}
1775
1776/// This is the main transformation entry point for a function.
1777bool MemCpyOptLegacyPass::runOnFunction(Function &F) {
1778 if (skipFunction(F))
1
Assuming the condition is false
2
Taking false branch
1779 return false;
1780
1781 auto *MDWP = !EnableMemorySSA
3
Assuming the condition is false
4
'?' condition is false
1782 ? &getAnalysis<MemoryDependenceWrapperPass>()
1783 : getAnalysisIfAvailable<MemoryDependenceWrapperPass>();
1784 auto *TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
1785 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
1786 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
1787 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1788 auto *MSSAWP = EnableMemorySSA
5
Assuming the condition is false
6
'?' condition is false
1789 ? &getAnalysis<MemorySSAWrapperPass>()
1790 : getAnalysisIfAvailable<MemorySSAWrapperPass>();
1791
1792 return Impl.runImpl(F, MDWP ? & MDWP->getMemDep() : nullptr, TLI, AA, AC, DT,
7
Assuming 'MDWP' is null
8
'?' condition is false
11
Calling 'MemCpyOptPass::runImpl'
1793 MSSAWP
8.1
'MSSAWP' is null
? &MSSAWP->getMSSA() : nullptr)
;
9
'?' condition is false
10
Passing null pointer value via 7th parameter 'MSSA_'
1794}