1 //===- DeadStoreElimination.cpp - Fast Dead Store Elimination -------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements a trivial dead store elimination that only considers
11 // basic-block local redundant stores.
12 //
13 // FIXME: This should eventually be extended to be a post-dominator tree
14 // traversal. Doing so would be pretty trivial.
15 //
16 //===----------------------------------------------------------------------===//
17
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/STLExtras.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/CaptureTracking.h"
24 #include "llvm/Analysis/MemoryBuiltins.h"
25 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/Constants.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Target/TargetLibraryInfo.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 using namespace llvm;
39
40 #define DEBUG_TYPE "dse"
41
42 STATISTIC(NumFastStores, "Number of stores deleted");
43 STATISTIC(NumFastOther , "Number of other instrs removed");
44
45 namespace {
46 struct DSE : public FunctionPass {
47 AliasAnalysis *AA;
48 MemoryDependenceAnalysis *MD;
49 DominatorTree *DT;
50 const TargetLibraryInfo *TLI;
51
52 static char ID; // Pass identification, replacement for typeid
DSE__anon31d19a320111::DSE53 DSE() : FunctionPass(ID), AA(nullptr), MD(nullptr), DT(nullptr) {
54 initializeDSEPass(*PassRegistry::getPassRegistry());
55 }
56
runOnFunction__anon31d19a320111::DSE57 bool runOnFunction(Function &F) override {
58 if (skipOptnoneFunction(F))
59 return false;
60
61 AA = &getAnalysis<AliasAnalysis>();
62 MD = &getAnalysis<MemoryDependenceAnalysis>();
63 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
64 TLI = AA->getTargetLibraryInfo();
65
66 bool Changed = false;
67 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I)
68 // Only check non-dead blocks. Dead blocks may have strange pointer
69 // cycles that will confuse alias analysis.
70 if (DT->isReachableFromEntry(I))
71 Changed |= runOnBasicBlock(*I);
72
73 AA = nullptr; MD = nullptr; DT = nullptr;
74 return Changed;
75 }
76
77 bool runOnBasicBlock(BasicBlock &BB);
78 bool HandleFree(CallInst *F);
79 bool handleEndBlock(BasicBlock &BB);
80 void RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
81 SmallSetVector<Value*, 16> &DeadStackObjects);
82
getAnalysisUsage__anon31d19a320111::DSE83 void getAnalysisUsage(AnalysisUsage &AU) const override {
84 AU.setPreservesCFG();
85 AU.addRequired<DominatorTreeWrapperPass>();
86 AU.addRequired<AliasAnalysis>();
87 AU.addRequired<MemoryDependenceAnalysis>();
88 AU.addPreserved<AliasAnalysis>();
89 AU.addPreserved<DominatorTreeWrapperPass>();
90 AU.addPreserved<MemoryDependenceAnalysis>();
91 }
92 };
93 }
94
95 char DSE::ID = 0;
96 INITIALIZE_PASS_BEGIN(DSE, "dse", "Dead Store Elimination", false, false)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)97 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
98 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
99 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
100 INITIALIZE_PASS_END(DSE, "dse", "Dead Store Elimination", false, false)
101
102 FunctionPass *llvm::createDeadStoreEliminationPass() { return new DSE(); }
103
104 //===----------------------------------------------------------------------===//
105 // Helper functions
106 //===----------------------------------------------------------------------===//
107
108 /// DeleteDeadInstruction - Delete this instruction. Before we do, go through
109 /// and zero out all the operands of this instruction. If any of them become
110 /// dead, delete them and the computation tree that feeds them.
111 ///
112 /// If ValueSet is non-null, remove any deleted instructions from it as well.
113 ///
DeleteDeadInstruction(Instruction * I,MemoryDependenceAnalysis & MD,const TargetLibraryInfo * TLI,SmallSetVector<Value *,16> * ValueSet=nullptr)114 static void DeleteDeadInstruction(Instruction *I,
115 MemoryDependenceAnalysis &MD,
116 const TargetLibraryInfo *TLI,
117 SmallSetVector<Value*, 16> *ValueSet = nullptr) {
118 SmallVector<Instruction*, 32> NowDeadInsts;
119
120 NowDeadInsts.push_back(I);
121 --NumFastOther;
122
123 // Before we touch this instruction, remove it from memdep!
124 do {
125 Instruction *DeadInst = NowDeadInsts.pop_back_val();
126 ++NumFastOther;
127
128 // This instruction is dead, zap it, in stages. Start by removing it from
129 // MemDep, which needs to know the operands and needs it to be in the
130 // function.
131 MD.removeInstruction(DeadInst);
132
133 for (unsigned op = 0, e = DeadInst->getNumOperands(); op != e; ++op) {
134 Value *Op = DeadInst->getOperand(op);
135 DeadInst->setOperand(op, nullptr);
136
137 // If this operand just became dead, add it to the NowDeadInsts list.
138 if (!Op->use_empty()) continue;
139
140 if (Instruction *OpI = dyn_cast<Instruction>(Op))
141 if (isInstructionTriviallyDead(OpI, TLI))
142 NowDeadInsts.push_back(OpI);
143 }
144
145 DeadInst->eraseFromParent();
146
147 if (ValueSet) ValueSet->remove(DeadInst);
148 } while (!NowDeadInsts.empty());
149 }
150
151
152 /// hasMemoryWrite - Does this instruction write some memory? This only returns
153 /// true for things that we can analyze with other helpers below.
hasMemoryWrite(Instruction * I,const TargetLibraryInfo * TLI)154 static bool hasMemoryWrite(Instruction *I, const TargetLibraryInfo *TLI) {
155 if (isa<StoreInst>(I))
156 return true;
157 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
158 switch (II->getIntrinsicID()) {
159 default:
160 return false;
161 case Intrinsic::memset:
162 case Intrinsic::memmove:
163 case Intrinsic::memcpy:
164 case Intrinsic::init_trampoline:
165 case Intrinsic::lifetime_end:
166 return true;
167 }
168 }
169 if (CallSite CS = I) {
170 if (Function *F = CS.getCalledFunction()) {
171 if (TLI && TLI->has(LibFunc::strcpy) &&
172 F->getName() == TLI->getName(LibFunc::strcpy)) {
173 return true;
174 }
175 if (TLI && TLI->has(LibFunc::strncpy) &&
176 F->getName() == TLI->getName(LibFunc::strncpy)) {
177 return true;
178 }
179 if (TLI && TLI->has(LibFunc::strcat) &&
180 F->getName() == TLI->getName(LibFunc::strcat)) {
181 return true;
182 }
183 if (TLI && TLI->has(LibFunc::strncat) &&
184 F->getName() == TLI->getName(LibFunc::strncat)) {
185 return true;
186 }
187 }
188 }
189 return false;
190 }
191
192 /// getLocForWrite - Return a Location stored to by the specified instruction.
193 /// If isRemovable returns true, this function and getLocForRead completely
194 /// describe the memory operations for this instruction.
195 static AliasAnalysis::Location
getLocForWrite(Instruction * Inst,AliasAnalysis & AA)196 getLocForWrite(Instruction *Inst, AliasAnalysis &AA) {
197 const DataLayout *DL = AA.getDataLayout();
198 if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
199 return AA.getLocation(SI);
200
201 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(Inst)) {
202 // memcpy/memmove/memset.
203 AliasAnalysis::Location Loc = AA.getLocationForDest(MI);
204 // If we don't have target data around, an unknown size in Location means
205 // that we should use the size of the pointee type. This isn't valid for
206 // memset/memcpy, which writes more than an i8.
207 if (Loc.Size == AliasAnalysis::UnknownSize && DL == nullptr)
208 return AliasAnalysis::Location();
209 return Loc;
210 }
211
212 IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
213 if (!II) return AliasAnalysis::Location();
214
215 switch (II->getIntrinsicID()) {
216 default: return AliasAnalysis::Location(); // Unhandled intrinsic.
217 case Intrinsic::init_trampoline:
218 // If we don't have target data around, an unknown size in Location means
219 // that we should use the size of the pointee type. This isn't valid for
220 // init.trampoline, which writes more than an i8.
221 if (!DL) return AliasAnalysis::Location();
222
223 // FIXME: We don't know the size of the trampoline, so we can't really
224 // handle it here.
225 return AliasAnalysis::Location(II->getArgOperand(0));
226 case Intrinsic::lifetime_end: {
227 uint64_t Len = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
228 return AliasAnalysis::Location(II->getArgOperand(1), Len);
229 }
230 }
231 }
232
233 /// getLocForRead - Return the location read by the specified "hasMemoryWrite"
234 /// instruction if any.
235 static AliasAnalysis::Location
getLocForRead(Instruction * Inst,AliasAnalysis & AA)236 getLocForRead(Instruction *Inst, AliasAnalysis &AA) {
237 assert(hasMemoryWrite(Inst, AA.getTargetLibraryInfo()) &&
238 "Unknown instruction case");
239
240 // The only instructions that both read and write are the mem transfer
241 // instructions (memcpy/memmove).
242 if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(Inst))
243 return AA.getLocationForSource(MTI);
244 return AliasAnalysis::Location();
245 }
246
247
248 /// isRemovable - If the value of this instruction and the memory it writes to
249 /// is unused, may we delete this instruction?
isRemovable(Instruction * I)250 static bool isRemovable(Instruction *I) {
251 // Don't remove volatile/atomic stores.
252 if (StoreInst *SI = dyn_cast<StoreInst>(I))
253 return SI->isUnordered();
254
255 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
256 switch (II->getIntrinsicID()) {
257 default: llvm_unreachable("doesn't pass 'hasMemoryWrite' predicate");
258 case Intrinsic::lifetime_end:
259 // Never remove dead lifetime_end's, e.g. because it is followed by a
260 // free.
261 return false;
262 case Intrinsic::init_trampoline:
263 // Always safe to remove init_trampoline.
264 return true;
265
266 case Intrinsic::memset:
267 case Intrinsic::memmove:
268 case Intrinsic::memcpy:
269 // Don't remove volatile memory intrinsics.
270 return !cast<MemIntrinsic>(II)->isVolatile();
271 }
272 }
273
274 if (CallSite CS = I)
275 return CS.getInstruction()->use_empty();
276
277 return false;
278 }
279
280
281 /// isShortenable - Returns true if this instruction can be safely shortened in
282 /// length.
isShortenable(Instruction * I)283 static bool isShortenable(Instruction *I) {
284 // Don't shorten stores for now
285 if (isa<StoreInst>(I))
286 return false;
287
288 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
289 switch (II->getIntrinsicID()) {
290 default: return false;
291 case Intrinsic::memset:
292 case Intrinsic::memcpy:
293 // Do shorten memory intrinsics.
294 return true;
295 }
296 }
297
298 // Don't shorten libcalls calls for now.
299
300 return false;
301 }
302
303 /// getStoredPointerOperand - Return the pointer that is being written to.
getStoredPointerOperand(Instruction * I)304 static Value *getStoredPointerOperand(Instruction *I) {
305 if (StoreInst *SI = dyn_cast<StoreInst>(I))
306 return SI->getPointerOperand();
307 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
308 return MI->getDest();
309
310 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
311 switch (II->getIntrinsicID()) {
312 default: llvm_unreachable("Unexpected intrinsic!");
313 case Intrinsic::init_trampoline:
314 return II->getArgOperand(0);
315 }
316 }
317
318 CallSite CS = I;
319 // All the supported functions so far happen to have dest as their first
320 // argument.
321 return CS.getArgument(0);
322 }
323
getPointerSize(const Value * V,AliasAnalysis & AA)324 static uint64_t getPointerSize(const Value *V, AliasAnalysis &AA) {
325 uint64_t Size;
326 if (getObjectSize(V, Size, AA.getDataLayout(), AA.getTargetLibraryInfo()))
327 return Size;
328 return AliasAnalysis::UnknownSize;
329 }
330
331 namespace {
332 enum OverwriteResult
333 {
334 OverwriteComplete,
335 OverwriteEnd,
336 OverwriteUnknown
337 };
338 }
339
340 /// isOverwrite - Return 'OverwriteComplete' if a store to the 'Later' location
341 /// completely overwrites a store to the 'Earlier' location.
342 /// 'OverwriteEnd' if the end of the 'Earlier' location is completely
343 /// overwritten by 'Later', or 'OverwriteUnknown' if nothing can be determined
isOverwrite(const AliasAnalysis::Location & Later,const AliasAnalysis::Location & Earlier,AliasAnalysis & AA,int64_t & EarlierOff,int64_t & LaterOff)344 static OverwriteResult isOverwrite(const AliasAnalysis::Location &Later,
345 const AliasAnalysis::Location &Earlier,
346 AliasAnalysis &AA,
347 int64_t &EarlierOff,
348 int64_t &LaterOff) {
349 const DataLayout *DL = AA.getDataLayout();
350 const Value *P1 = Earlier.Ptr->stripPointerCasts();
351 const Value *P2 = Later.Ptr->stripPointerCasts();
352
353 // If the start pointers are the same, we just have to compare sizes to see if
354 // the later store was larger than the earlier store.
355 if (P1 == P2) {
356 // If we don't know the sizes of either access, then we can't do a
357 // comparison.
358 if (Later.Size == AliasAnalysis::UnknownSize ||
359 Earlier.Size == AliasAnalysis::UnknownSize)
360 return OverwriteUnknown;
361
362 // Make sure that the Later size is >= the Earlier size.
363 if (Later.Size >= Earlier.Size)
364 return OverwriteComplete;
365 }
366
367 // Otherwise, we have to have size information, and the later store has to be
368 // larger than the earlier one.
369 if (Later.Size == AliasAnalysis::UnknownSize ||
370 Earlier.Size == AliasAnalysis::UnknownSize || DL == nullptr)
371 return OverwriteUnknown;
372
373 // Check to see if the later store is to the entire object (either a global,
374 // an alloca, or a byval/inalloca argument). If so, then it clearly
375 // overwrites any other store to the same object.
376 const Value *UO1 = GetUnderlyingObject(P1, DL),
377 *UO2 = GetUnderlyingObject(P2, DL);
378
379 // If we can't resolve the same pointers to the same object, then we can't
380 // analyze them at all.
381 if (UO1 != UO2)
382 return OverwriteUnknown;
383
384 // If the "Later" store is to a recognizable object, get its size.
385 uint64_t ObjectSize = getPointerSize(UO2, AA);
386 if (ObjectSize != AliasAnalysis::UnknownSize)
387 if (ObjectSize == Later.Size && ObjectSize >= Earlier.Size)
388 return OverwriteComplete;
389
390 // Okay, we have stores to two completely different pointers. Try to
391 // decompose the pointer into a "base + constant_offset" form. If the base
392 // pointers are equal, then we can reason about the two stores.
393 EarlierOff = 0;
394 LaterOff = 0;
395 const Value *BP1 = GetPointerBaseWithConstantOffset(P1, EarlierOff, DL);
396 const Value *BP2 = GetPointerBaseWithConstantOffset(P2, LaterOff, DL);
397
398 // If the base pointers still differ, we have two completely different stores.
399 if (BP1 != BP2)
400 return OverwriteUnknown;
401
402 // The later store completely overlaps the earlier store if:
403 //
404 // 1. Both start at the same offset and the later one's size is greater than
405 // or equal to the earlier one's, or
406 //
407 // |--earlier--|
408 // |-- later --|
409 //
410 // 2. The earlier store has an offset greater than the later offset, but which
411 // still lies completely within the later store.
412 //
413 // |--earlier--|
414 // |----- later ------|
415 //
416 // We have to be careful here as *Off is signed while *.Size is unsigned.
417 if (EarlierOff >= LaterOff &&
418 Later.Size >= Earlier.Size &&
419 uint64_t(EarlierOff - LaterOff) + Earlier.Size <= Later.Size)
420 return OverwriteComplete;
421
422 // The other interesting case is if the later store overwrites the end of
423 // the earlier store
424 //
425 // |--earlier--|
426 // |-- later --|
427 //
428 // In this case we may want to trim the size of earlier to avoid generating
429 // writes to addresses which will definitely be overwritten later
430 if (LaterOff > EarlierOff &&
431 LaterOff < int64_t(EarlierOff + Earlier.Size) &&
432 int64_t(LaterOff + Later.Size) >= int64_t(EarlierOff + Earlier.Size))
433 return OverwriteEnd;
434
435 // Otherwise, they don't completely overlap.
436 return OverwriteUnknown;
437 }
438
439 /// isPossibleSelfRead - If 'Inst' might be a self read (i.e. a noop copy of a
440 /// memory region into an identical pointer) then it doesn't actually make its
441 /// input dead in the traditional sense. Consider this case:
442 ///
443 /// memcpy(A <- B)
444 /// memcpy(A <- A)
445 ///
446 /// In this case, the second store to A does not make the first store to A dead.
447 /// The usual situation isn't an explicit A<-A store like this (which can be
448 /// trivially removed) but a case where two pointers may alias.
449 ///
450 /// This function detects when it is unsafe to remove a dependent instruction
451 /// because the DSE inducing instruction may be a self-read.
isPossibleSelfRead(Instruction * Inst,const AliasAnalysis::Location & InstStoreLoc,Instruction * DepWrite,AliasAnalysis & AA)452 static bool isPossibleSelfRead(Instruction *Inst,
453 const AliasAnalysis::Location &InstStoreLoc,
454 Instruction *DepWrite, AliasAnalysis &AA) {
455 // Self reads can only happen for instructions that read memory. Get the
456 // location read.
457 AliasAnalysis::Location InstReadLoc = getLocForRead(Inst, AA);
458 if (!InstReadLoc.Ptr) return false; // Not a reading instruction.
459
460 // If the read and written loc obviously don't alias, it isn't a read.
461 if (AA.isNoAlias(InstReadLoc, InstStoreLoc)) return false;
462
463 // Okay, 'Inst' may copy over itself. However, we can still remove a the
464 // DepWrite instruction if we can prove that it reads from the same location
465 // as Inst. This handles useful cases like:
466 // memcpy(A <- B)
467 // memcpy(A <- B)
468 // Here we don't know if A/B may alias, but we do know that B/B are must
469 // aliases, so removing the first memcpy is safe (assuming it writes <= #
470 // bytes as the second one.
471 AliasAnalysis::Location DepReadLoc = getLocForRead(DepWrite, AA);
472
473 if (DepReadLoc.Ptr && AA.isMustAlias(InstReadLoc.Ptr, DepReadLoc.Ptr))
474 return false;
475
476 // If DepWrite doesn't read memory or if we can't prove it is a must alias,
477 // then it can't be considered dead.
478 return true;
479 }
480
481
482 //===----------------------------------------------------------------------===//
483 // DSE Pass
484 //===----------------------------------------------------------------------===//
485
runOnBasicBlock(BasicBlock & BB)486 bool DSE::runOnBasicBlock(BasicBlock &BB) {
487 bool MadeChange = false;
488
489 // Do a top-down walk on the BB.
490 for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
491 Instruction *Inst = BBI++;
492
493 // Handle 'free' calls specially.
494 if (CallInst *F = isFreeCall(Inst, TLI)) {
495 MadeChange |= HandleFree(F);
496 continue;
497 }
498
499 // If we find something that writes memory, get its memory dependence.
500 if (!hasMemoryWrite(Inst, TLI))
501 continue;
502
503 MemDepResult InstDep = MD->getDependency(Inst);
504
505 // Ignore any store where we can't find a local dependence.
506 // FIXME: cross-block DSE would be fun. :)
507 if (!InstDep.isDef() && !InstDep.isClobber())
508 continue;
509
510 // If we're storing the same value back to a pointer that we just
511 // loaded from, then the store can be removed.
512 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
513 if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
514 if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
515 SI->getOperand(0) == DepLoad && isRemovable(SI)) {
516 DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
517 << "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
518
519 // DeleteDeadInstruction can delete the current instruction. Save BBI
520 // in case we need it.
521 WeakVH NextInst(BBI);
522
523 DeleteDeadInstruction(SI, *MD, TLI);
524
525 if (!NextInst) // Next instruction deleted.
526 BBI = BB.begin();
527 else if (BBI != BB.begin()) // Revisit this instruction if possible.
528 --BBI;
529 ++NumFastStores;
530 MadeChange = true;
531 continue;
532 }
533 }
534 }
535
536 // Figure out what location is being stored to.
537 AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
538
539 // If we didn't get a useful location, fail.
540 if (!Loc.Ptr)
541 continue;
542
543 while (InstDep.isDef() || InstDep.isClobber()) {
544 // Get the memory clobbered by the instruction we depend on. MemDep will
545 // skip any instructions that 'Loc' clearly doesn't interact with. If we
546 // end up depending on a may- or must-aliased load, then we can't optimize
547 // away the store and we bail out. However, if we depend on on something
548 // that overwrites the memory location we *can* potentially optimize it.
549 //
550 // Find out what memory location the dependent instruction stores.
551 Instruction *DepWrite = InstDep.getInst();
552 AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
553 // If we didn't get a useful location, or if it isn't a size, bail out.
554 if (!DepLoc.Ptr)
555 break;
556
557 // If we find a write that is a) removable (i.e., non-volatile), b) is
558 // completely obliterated by the store to 'Loc', and c) which we know that
559 // 'Inst' doesn't load from, then we can remove it.
560 if (isRemovable(DepWrite) &&
561 !isPossibleSelfRead(Inst, Loc, DepWrite, *AA)) {
562 int64_t InstWriteOffset, DepWriteOffset;
563 OverwriteResult OR = isOverwrite(Loc, DepLoc, *AA,
564 DepWriteOffset, InstWriteOffset);
565 if (OR == OverwriteComplete) {
566 DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: "
567 << *DepWrite << "\n KILLER: " << *Inst << '\n');
568
569 // Delete the store and now-dead instructions that feed it.
570 DeleteDeadInstruction(DepWrite, *MD, TLI);
571 ++NumFastStores;
572 MadeChange = true;
573
574 // DeleteDeadInstruction can delete the current instruction in loop
575 // cases, reset BBI.
576 BBI = Inst;
577 if (BBI != BB.begin())
578 --BBI;
579 break;
580 } else if (OR == OverwriteEnd && isShortenable(DepWrite)) {
581 // TODO: base this on the target vector size so that if the earlier
582 // store was too small to get vector writes anyway then its likely
583 // a good idea to shorten it
584 // Power of 2 vector writes are probably always a bad idea to optimize
585 // as any store/memset/memcpy is likely using vector instructions so
586 // shortening it to not vector size is likely to be slower
587 MemIntrinsic* DepIntrinsic = cast<MemIntrinsic>(DepWrite);
588 unsigned DepWriteAlign = DepIntrinsic->getAlignment();
589 if (llvm::isPowerOf2_64(InstWriteOffset) ||
590 ((DepWriteAlign != 0) && InstWriteOffset % DepWriteAlign == 0)) {
591
592 DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW END: "
593 << *DepWrite << "\n KILLER (offset "
594 << InstWriteOffset << ", "
595 << DepLoc.Size << ")"
596 << *Inst << '\n');
597
598 Value* DepWriteLength = DepIntrinsic->getLength();
599 Value* TrimmedLength = ConstantInt::get(DepWriteLength->getType(),
600 InstWriteOffset -
601 DepWriteOffset);
602 DepIntrinsic->setLength(TrimmedLength);
603 MadeChange = true;
604 }
605 }
606 }
607
608 // If this is a may-aliased store that is clobbering the store value, we
609 // can keep searching past it for another must-aliased pointer that stores
610 // to the same location. For example, in:
611 // store -> P
612 // store -> Q
613 // store -> P
614 // we can remove the first store to P even though we don't know if P and Q
615 // alias.
616 if (DepWrite == &BB.front()) break;
617
618 // Can't look past this instruction if it might read 'Loc'.
619 if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
620 break;
621
622 InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
623 }
624 }
625
626 // If this block ends in a return, unwind, or unreachable, all allocas are
627 // dead at its end, which means stores to them are also dead.
628 if (BB.getTerminator()->getNumSuccessors() == 0)
629 MadeChange |= handleEndBlock(BB);
630
631 return MadeChange;
632 }
633
634 /// Find all blocks that will unconditionally lead to the block BB and append
635 /// them to F.
FindUnconditionalPreds(SmallVectorImpl<BasicBlock * > & Blocks,BasicBlock * BB,DominatorTree * DT)636 static void FindUnconditionalPreds(SmallVectorImpl<BasicBlock *> &Blocks,
637 BasicBlock *BB, DominatorTree *DT) {
638 for (pred_iterator I = pred_begin(BB), E = pred_end(BB); I != E; ++I) {
639 BasicBlock *Pred = *I;
640 if (Pred == BB) continue;
641 TerminatorInst *PredTI = Pred->getTerminator();
642 if (PredTI->getNumSuccessors() != 1)
643 continue;
644
645 if (DT->isReachableFromEntry(Pred))
646 Blocks.push_back(Pred);
647 }
648 }
649
650 /// HandleFree - Handle frees of entire structures whose dependency is a store
651 /// to a field of that structure.
HandleFree(CallInst * F)652 bool DSE::HandleFree(CallInst *F) {
653 bool MadeChange = false;
654
655 AliasAnalysis::Location Loc = AliasAnalysis::Location(F->getOperand(0));
656 SmallVector<BasicBlock *, 16> Blocks;
657 Blocks.push_back(F->getParent());
658
659 while (!Blocks.empty()) {
660 BasicBlock *BB = Blocks.pop_back_val();
661 Instruction *InstPt = BB->getTerminator();
662 if (BB == F->getParent()) InstPt = F;
663
664 MemDepResult Dep = MD->getPointerDependencyFrom(Loc, false, InstPt, BB);
665 while (Dep.isDef() || Dep.isClobber()) {
666 Instruction *Dependency = Dep.getInst();
667 if (!hasMemoryWrite(Dependency, TLI) || !isRemovable(Dependency))
668 break;
669
670 Value *DepPointer =
671 GetUnderlyingObject(getStoredPointerOperand(Dependency));
672
673 // Check for aliasing.
674 if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
675 break;
676
677 Instruction *Next = std::next(BasicBlock::iterator(Dependency));
678
679 // DCE instructions only used to calculate that store
680 DeleteDeadInstruction(Dependency, *MD, TLI);
681 ++NumFastStores;
682 MadeChange = true;
683
684 // Inst's old Dependency is now deleted. Compute the next dependency,
685 // which may also be dead, as in
686 // s[0] = 0;
687 // s[1] = 0; // This has just been deleted.
688 // free(s);
689 Dep = MD->getPointerDependencyFrom(Loc, false, Next, BB);
690 }
691
692 if (Dep.isNonLocal())
693 FindUnconditionalPreds(Blocks, BB, DT);
694 }
695
696 return MadeChange;
697 }
698
699 /// handleEndBlock - Remove dead stores to stack-allocated locations in the
700 /// function end block. Ex:
701 /// %A = alloca i32
702 /// ...
703 /// store i32 1, i32* %A
704 /// ret void
handleEndBlock(BasicBlock & BB)705 bool DSE::handleEndBlock(BasicBlock &BB) {
706 bool MadeChange = false;
707
708 // Keep track of all of the stack objects that are dead at the end of the
709 // function.
710 SmallSetVector<Value*, 16> DeadStackObjects;
711
712 // Find all of the alloca'd pointers in the entry block.
713 BasicBlock *Entry = BB.getParent()->begin();
714 for (BasicBlock::iterator I = Entry->begin(), E = Entry->end(); I != E; ++I) {
715 if (isa<AllocaInst>(I))
716 DeadStackObjects.insert(I);
717
718 // Okay, so these are dead heap objects, but if the pointer never escapes
719 // then it's leaked by this function anyways.
720 else if (isAllocLikeFn(I, TLI) && !PointerMayBeCaptured(I, true, true))
721 DeadStackObjects.insert(I);
722 }
723
724 // Treat byval or inalloca arguments the same, stores to them are dead at the
725 // end of the function.
726 for (Function::arg_iterator AI = BB.getParent()->arg_begin(),
727 AE = BB.getParent()->arg_end(); AI != AE; ++AI)
728 if (AI->hasByValOrInAllocaAttr())
729 DeadStackObjects.insert(AI);
730
731 // Scan the basic block backwards
732 for (BasicBlock::iterator BBI = BB.end(); BBI != BB.begin(); ){
733 --BBI;
734
735 // If we find a store, check to see if it points into a dead stack value.
736 if (hasMemoryWrite(BBI, TLI) && isRemovable(BBI)) {
737 // See through pointer-to-pointer bitcasts
738 SmallVector<Value *, 4> Pointers;
739 GetUnderlyingObjects(getStoredPointerOperand(BBI), Pointers);
740
741 // Stores to stack values are valid candidates for removal.
742 bool AllDead = true;
743 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
744 E = Pointers.end(); I != E; ++I)
745 if (!DeadStackObjects.count(*I)) {
746 AllDead = false;
747 break;
748 }
749
750 if (AllDead) {
751 Instruction *Dead = BBI++;
752
753 DEBUG(dbgs() << "DSE: Dead Store at End of Block:\n DEAD: "
754 << *Dead << "\n Objects: ";
755 for (SmallVectorImpl<Value *>::iterator I = Pointers.begin(),
756 E = Pointers.end(); I != E; ++I) {
757 dbgs() << **I;
758 if (std::next(I) != E)
759 dbgs() << ", ";
760 }
761 dbgs() << '\n');
762
763 // DCE instructions only used to calculate that store.
764 DeleteDeadInstruction(Dead, *MD, TLI, &DeadStackObjects);
765 ++NumFastStores;
766 MadeChange = true;
767 continue;
768 }
769 }
770
771 // Remove any dead non-memory-mutating instructions.
772 if (isInstructionTriviallyDead(BBI, TLI)) {
773 Instruction *Inst = BBI++;
774 DeleteDeadInstruction(Inst, *MD, TLI, &DeadStackObjects);
775 ++NumFastOther;
776 MadeChange = true;
777 continue;
778 }
779
780 if (isa<AllocaInst>(BBI)) {
781 // Remove allocas from the list of dead stack objects; there can't be
782 // any references before the definition.
783 DeadStackObjects.remove(BBI);
784 continue;
785 }
786
787 if (CallSite CS = cast<Value>(BBI)) {
788 // Remove allocation function calls from the list of dead stack objects;
789 // there can't be any references before the definition.
790 if (isAllocLikeFn(BBI, TLI))
791 DeadStackObjects.remove(BBI);
792
793 // If this call does not access memory, it can't be loading any of our
794 // pointers.
795 if (AA->doesNotAccessMemory(CS))
796 continue;
797
798 // If the call might load from any of our allocas, then any store above
799 // the call is live.
800 DeadStackObjects.remove_if([&](Value *I) {
801 // See if the call site touches the value.
802 AliasAnalysis::ModRefResult A =
803 AA->getModRefInfo(CS, I, getPointerSize(I, *AA));
804
805 return A == AliasAnalysis::ModRef || A == AliasAnalysis::Ref;
806 });
807
808 // If all of the allocas were clobbered by the call then we're not going
809 // to find anything else to process.
810 if (DeadStackObjects.empty())
811 break;
812
813 continue;
814 }
815
816 AliasAnalysis::Location LoadedLoc;
817
818 // If we encounter a use of the pointer, it is no longer considered dead
819 if (LoadInst *L = dyn_cast<LoadInst>(BBI)) {
820 if (!L->isUnordered()) // Be conservative with atomic/volatile load
821 break;
822 LoadedLoc = AA->getLocation(L);
823 } else if (VAArgInst *V = dyn_cast<VAArgInst>(BBI)) {
824 LoadedLoc = AA->getLocation(V);
825 } else if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(BBI)) {
826 LoadedLoc = AA->getLocationForSource(MTI);
827 } else if (!BBI->mayReadFromMemory()) {
828 // Instruction doesn't read memory. Note that stores that weren't removed
829 // above will hit this case.
830 continue;
831 } else {
832 // Unknown inst; assume it clobbers everything.
833 break;
834 }
835
836 // Remove any allocas from the DeadPointer set that are loaded, as this
837 // makes any stores above the access live.
838 RemoveAccessedObjects(LoadedLoc, DeadStackObjects);
839
840 // If all of the allocas were clobbered by the access then we're not going
841 // to find anything else to process.
842 if (DeadStackObjects.empty())
843 break;
844 }
845
846 return MadeChange;
847 }
848
849 /// RemoveAccessedObjects - Check to see if the specified location may alias any
850 /// of the stack objects in the DeadStackObjects set. If so, they become live
851 /// because the location is being loaded.
RemoveAccessedObjects(const AliasAnalysis::Location & LoadedLoc,SmallSetVector<Value *,16> & DeadStackObjects)852 void DSE::RemoveAccessedObjects(const AliasAnalysis::Location &LoadedLoc,
853 SmallSetVector<Value*, 16> &DeadStackObjects) {
854 const Value *UnderlyingPointer = GetUnderlyingObject(LoadedLoc.Ptr);
855
856 // A constant can't be in the dead pointer set.
857 if (isa<Constant>(UnderlyingPointer))
858 return;
859
860 // If the kill pointer can be easily reduced to an alloca, don't bother doing
861 // extraneous AA queries.
862 if (isa<AllocaInst>(UnderlyingPointer) || isa<Argument>(UnderlyingPointer)) {
863 DeadStackObjects.remove(const_cast<Value*>(UnderlyingPointer));
864 return;
865 }
866
867 // Remove objects that could alias LoadedLoc.
868 DeadStackObjects.remove_if([&](Value *I) {
869 // See if the loaded location could alias the stack location.
870 AliasAnalysis::Location StackLoc(I, getPointerSize(I, *AA));
871 return !AA->isNoAlias(StackLoc, LoadedLoc);
872 });
873 }
874