1 //===- MemoryDependenceAnalysis.cpp - Mem Deps Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements an analysis that determines, for a given memory
11 // operation, what preceding memory operations it depends on. It builds on
12 // alias analysis information, and tries to provide a lazy, caching interface to
13 // a common kind of alias information query.
15 //===----------------------------------------------------------------------===//
17 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/AssumptionCache.h"
22 #include "llvm/Analysis/InstructionSimplify.h"
23 #include "llvm/Analysis/MemoryBuiltins.h"
24 #include "llvm/Analysis/PHITransAddr.h"
25 #include "llvm/Analysis/ValueTracking.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/Dominators.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/Instructions.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/PredIteratorCache.h"
33 #include "llvm/Support/Debug.h"
36 #define DEBUG_TYPE "memdep"
38 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
39 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
40 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
42 STATISTIC(NumCacheNonLocalPtr,
43 "Number of fully cached non-local ptr responses");
44 STATISTIC(NumCacheDirtyNonLocalPtr,
45 "Number of cached, but dirty, non-local ptr responses");
46 STATISTIC(NumUncacheNonLocalPtr,
47 "Number of uncached non-local ptr responses");
48 STATISTIC(NumCacheCompleteNonLocalPtr,
49 "Number of block queries that were completely cached");
51 // Limit for the number of instructions to scan in a block.
52 static const unsigned int BlockScanLimit = 100;
54 // Limit on the number of memdep results to process.
55 static const unsigned int NumResultsLimit = 100;
57 char MemoryDependenceAnalysis::ID = 0;
59 // Register this pass...
60 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
61 "Memory Dependence Analysis", false, true)
62 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
63 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
64 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
65 "Memory Dependence Analysis", false, true)
67 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
68 : FunctionPass(ID), PredCache() {
69 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
71 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
74 /// Clean up memory in between runs
75 void MemoryDependenceAnalysis::releaseMemory() {
78 NonLocalPointerDeps.clear();
79 ReverseLocalDeps.clear();
80 ReverseNonLocalDeps.clear();
81 ReverseNonLocalPtrDeps.clear();
85 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
87 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
89 AU.addRequired<AssumptionCacheTracker>();
90 AU.addRequiredTransitive<AliasAnalysis>();
93 bool MemoryDependenceAnalysis::runOnFunction(Function &F) {
94 AA = &getAnalysis<AliasAnalysis>();
95 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
96 DominatorTreeWrapperPass *DTWP =
97 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
98 DT = DTWP ? &DTWP->getDomTree() : nullptr;
100 PredCache.reset(new PredIteratorCache());
104 /// RemoveFromReverseMap - This is a helper function that removes Val from
105 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
106 template <typename KeyTy>
107 static void RemoveFromReverseMap(DenseMap<Instruction*,
108 SmallPtrSet<KeyTy, 4> > &ReverseMap,
109 Instruction *Inst, KeyTy Val) {
110 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
111 InstIt = ReverseMap.find(Inst);
112 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
113 bool Found = InstIt->second.erase(Val);
114 assert(Found && "Invalid reverse map!"); (void)Found;
115 if (InstIt->second.empty())
116 ReverseMap.erase(InstIt);
119 /// GetLocation - If the given instruction references a specific memory
120 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
121 /// Return a ModRefInfo value describing the general behavior of the
124 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
125 AliasAnalysis::Location &Loc,
127 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
128 if (LI->isUnordered()) {
129 Loc = AA->getLocation(LI);
130 return AliasAnalysis::Ref;
132 if (LI->getOrdering() == Monotonic) {
133 Loc = AA->getLocation(LI);
134 return AliasAnalysis::ModRef;
136 Loc = AliasAnalysis::Location();
137 return AliasAnalysis::ModRef;
140 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
141 if (SI->isUnordered()) {
142 Loc = AA->getLocation(SI);
143 return AliasAnalysis::Mod;
145 if (SI->getOrdering() == Monotonic) {
146 Loc = AA->getLocation(SI);
147 return AliasAnalysis::ModRef;
149 Loc = AliasAnalysis::Location();
150 return AliasAnalysis::ModRef;
153 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
154 Loc = AA->getLocation(V);
155 return AliasAnalysis::ModRef;
158 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
159 // calls to free() deallocate the entire structure
160 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
161 return AliasAnalysis::Mod;
164 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
167 switch (II->getIntrinsicID()) {
168 case Intrinsic::lifetime_start:
169 case Intrinsic::lifetime_end:
170 case Intrinsic::invariant_start:
171 II->getAAMetadata(AAInfo);
172 Loc = AliasAnalysis::Location(II->getArgOperand(1),
173 cast<ConstantInt>(II->getArgOperand(0))
174 ->getZExtValue(), AAInfo);
175 // These intrinsics don't really modify the memory, but returning Mod
176 // will allow them to be handled conservatively.
177 return AliasAnalysis::Mod;
178 case Intrinsic::invariant_end:
179 II->getAAMetadata(AAInfo);
180 Loc = AliasAnalysis::Location(II->getArgOperand(2),
181 cast<ConstantInt>(II->getArgOperand(1))
182 ->getZExtValue(), AAInfo);
183 // These intrinsics don't really modify the memory, but returning Mod
184 // will allow them to be handled conservatively.
185 return AliasAnalysis::Mod;
191 // Otherwise, just do the coarse-grained thing that always works.
192 if (Inst->mayWriteToMemory())
193 return AliasAnalysis::ModRef;
194 if (Inst->mayReadFromMemory())
195 return AliasAnalysis::Ref;
196 return AliasAnalysis::NoModRef;
199 /// getCallSiteDependencyFrom - Private helper for finding the local
200 /// dependencies of a call site.
201 MemDepResult MemoryDependenceAnalysis::
202 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
203 BasicBlock::iterator ScanIt, BasicBlock *BB) {
204 unsigned Limit = BlockScanLimit;
206 // Walk backwards through the block, looking for dependencies
207 while (ScanIt != BB->begin()) {
208 // Limit the amount of scanning we do so we don't end up with quadratic
209 // running time on extreme testcases.
212 return MemDepResult::getUnknown();
214 Instruction *Inst = --ScanIt;
216 // If this inst is a memory op, get the pointer it accessed
217 AliasAnalysis::Location Loc;
218 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
220 // A simple instruction.
221 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
222 return MemDepResult::getClobber(Inst);
226 if (auto InstCS = CallSite(Inst)) {
227 // Debug intrinsics don't cause dependences.
228 if (isa<DbgInfoIntrinsic>(Inst)) continue;
229 // If these two calls do not interfere, look past it.
230 switch (AA->getModRefInfo(CS, InstCS)) {
231 case AliasAnalysis::NoModRef:
232 // If the two calls are the same, return InstCS as a Def, so that
233 // CS can be found redundant and eliminated.
234 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
235 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
236 return MemDepResult::getDef(Inst);
238 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
242 return MemDepResult::getClobber(Inst);
246 // If we could not obtain a pointer for the instruction and the instruction
247 // touches memory then assume that this is a dependency.
248 if (MR != AliasAnalysis::NoModRef)
249 return MemDepResult::getClobber(Inst);
252 // No dependence found. If this is the entry block of the function, it is
253 // unknown, otherwise it is non-local.
254 if (BB != &BB->getParent()->getEntryBlock())
255 return MemDepResult::getNonLocal();
256 return MemDepResult::getNonFuncLocal();
259 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
260 /// would fully overlap MemLoc if done as a wider legal integer load.
262 /// MemLocBase, MemLocOffset are lazily computed here the first time the
263 /// base/offs of memloc is needed.
264 static bool isLoadLoadClobberIfExtendedToFullWidth(
265 const AliasAnalysis::Location &MemLoc, const Value *&MemLocBase,
266 int64_t &MemLocOffs, const LoadInst *LI) {
267 const DataLayout &DL = LI->getModule()->getDataLayout();
269 // If we haven't already computed the base/offset of MemLoc, do so now.
271 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
273 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
274 MemLocBase, MemLocOffs, MemLoc.Size, LI);
278 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
279 /// looks at a memory location for a load (specified by MemLocBase, Offs,
280 /// and Size) and compares it against a load. If the specified load could
281 /// be safely widened to a larger integer load that is 1) still efficient,
282 /// 2) safe for the target, and 3) would provide the specified memory
283 /// location value, then this function returns the size in bytes of the
284 /// load width to use. If not, this returns zero.
285 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
286 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
287 const LoadInst *LI) {
288 // We can only extend simple integer loads.
289 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
291 // Load widening is hostile to ThreadSanitizer: it may cause false positives
292 // or make the reports more cryptic (access sizes are wrong).
293 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
296 const DataLayout &DL = LI->getModule()->getDataLayout();
298 // Get the base of this load.
300 const Value *LIBase =
301 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
303 // If the two pointers are not based on the same pointer, we can't tell that
305 if (LIBase != MemLocBase) return 0;
307 // Okay, the two values are based on the same pointer, but returned as
308 // no-alias. This happens when we have things like two byte loads at "P+1"
309 // and "P+3". Check to see if increasing the size of the "LI" load up to its
310 // alignment (or the largest native integer type) will allow us to load all
311 // the bits required by MemLoc.
313 // If MemLoc is before LI, then no widening of LI will help us out.
314 if (MemLocOffs < LIOffs) return 0;
316 // Get the alignment of the load in bytes. We assume that it is safe to load
317 // any legal integer up to this size without a problem. For example, if we're
318 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
319 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
321 unsigned LoadAlign = LI->getAlignment();
323 int64_t MemLocEnd = MemLocOffs+MemLocSize;
325 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
326 if (LIOffs+LoadAlign < MemLocEnd) return 0;
328 // This is the size of the load to try. Start with the next larger power of
330 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
331 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
334 // If this load size is bigger than our known alignment or would not fit
335 // into a native integer register, then we fail.
336 if (NewLoadByteSize > LoadAlign ||
337 !DL.fitsInLegalInteger(NewLoadByteSize*8))
340 if (LIOffs + NewLoadByteSize > MemLocEnd &&
341 LI->getParent()->getParent()->hasFnAttribute(
342 Attribute::SanitizeAddress))
343 // We will be reading past the location accessed by the original program.
344 // While this is safe in a regular build, Address Safety analysis tools
345 // may start reporting false warnings. So, don't do widening.
348 // If a load of this width would include all of MemLoc, then we succeed.
349 if (LIOffs+NewLoadByteSize >= MemLocEnd)
350 return NewLoadByteSize;
352 NewLoadByteSize <<= 1;
356 static bool isVolatile(Instruction *Inst) {
357 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
358 return LI->isVolatile();
359 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
360 return SI->isVolatile();
361 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
362 return AI->isVolatile();
367 /// getPointerDependencyFrom - Return the instruction on which a memory
368 /// location depends. If isLoad is true, this routine ignores may-aliases with
369 /// read-only operations. If isLoad is false, this routine ignores may-aliases
370 /// with reads from read-only locations. If possible, pass the query
371 /// instruction as well; this function may take advantage of the metadata
372 /// annotated to the query instruction to refine the result.
373 MemDepResult MemoryDependenceAnalysis::
374 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
375 BasicBlock::iterator ScanIt, BasicBlock *BB,
376 Instruction *QueryInst) {
378 const Value *MemLocBase = nullptr;
379 int64_t MemLocOffset = 0;
380 unsigned Limit = BlockScanLimit;
381 bool isInvariantLoad = false;
383 // We must be careful with atomic accesses, as they may allow another thread
384 // to touch this location, cloberring it. We are conservative: if the
385 // QueryInst is not a simple (non-atomic) memory access, we automatically
386 // return getClobber.
387 // If it is simple, we know based on the results of
388 // "Compiler testing via a theory of sound optimisations in the C11/C++11
389 // memory model" in PLDI 2013, that a non-atomic location can only be
390 // clobbered between a pair of a release and an acquire action, with no
391 // access to the location in between.
392 // Here is an example for giving the general intuition behind this rule.
393 // In the following code:
395 // release action; [1]
396 // acquire action; [4]
398 // It is unsafe to replace %val by 0 because another thread may be running:
399 // acquire action; [2]
401 // release action; [3]
402 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
403 // being 42. A key property of this program however is that if either
404 // 1 or 4 were missing, there would be a race between the store of 42
405 // either the store of 0 or the load (making the whole progam racy).
406 // The paper mentionned above shows that the same property is respected
407 // by every program that can detect any optimisation of that kind: either
408 // it is racy (undefined) or there is a release followed by an acquire
409 // between the pair of accesses under consideration.
411 // If the load is invariant, we "know" that it doesn't alias *any* write. We
412 // do want to respect mustalias results since defs are useful for value
413 // forwarding, but any mayalias write can be assumed to be noalias.
414 // Arguably, this logic should be pushed inside AliasAnalysis itself.
415 if (isLoad && QueryInst) {
416 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
417 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
418 isInvariantLoad = true;
421 const DataLayout &DL = BB->getModule()->getDataLayout();
423 // Walk backwards through the basic block, looking for dependencies.
424 while (ScanIt != BB->begin()) {
425 Instruction *Inst = --ScanIt;
427 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
428 // Debug intrinsics don't (and can't) cause dependencies.
429 if (isa<DbgInfoIntrinsic>(II)) continue;
431 // Limit the amount of scanning we do so we don't end up with quadratic
432 // running time on extreme testcases.
435 return MemDepResult::getUnknown();
437 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
438 // If we reach a lifetime begin or end marker, then the query ends here
439 // because the value is undefined.
440 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
441 // FIXME: This only considers queries directly on the invariant-tagged
442 // pointer, not on query pointers that are indexed off of them. It'd
443 // be nice to handle that at some point (the right approach is to use
444 // GetPointerBaseWithConstantOffset).
445 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
447 return MemDepResult::getDef(II);
452 // Values depend on loads if the pointers are must aliased. This means that
453 // a load depends on another must aliased load from the same value.
454 // One exception is atomic loads: a value can depend on an atomic load that it
455 // does not alias with when this atomic load indicates that another thread may
456 // be accessing the location.
457 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
459 // While volatile access cannot be eliminated, they do not have to clobber
460 // non-aliasing locations, as normal accesses, for example, can be safely
461 // reordered with volatile accesses.
462 if (LI->isVolatile()) {
464 // Original QueryInst *may* be volatile
465 return MemDepResult::getClobber(LI);
466 if (isVolatile(QueryInst))
467 // Ordering required if QueryInst is itself volatile
468 return MemDepResult::getClobber(LI);
469 // Otherwise, volatile doesn't imply any special ordering
472 // Atomic loads have complications involved.
473 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
474 // FIXME: This is overly conservative.
475 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
477 return MemDepResult::getClobber(LI);
478 if (LI->getOrdering() != Monotonic)
479 return MemDepResult::getClobber(LI);
480 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
481 if (!QueryLI->isSimple())
482 return MemDepResult::getClobber(LI);
483 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
484 if (!QuerySI->isSimple())
485 return MemDepResult::getClobber(LI);
486 } else if (QueryInst->mayReadOrWriteMemory()) {
487 return MemDepResult::getClobber(LI);
491 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
493 // If we found a pointer, check if it could be the same as our pointer.
494 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
497 if (R == AliasAnalysis::NoAlias) {
498 // If this is an over-aligned integer load (for example,
499 // "load i8* %P, align 4") see if it would obviously overlap with the
500 // queried location if widened to a larger load (e.g. if the queried
501 // location is 1 byte at P+1). If so, return it as a load/load
502 // clobber result, allowing the client to decide to widen the load if
504 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
505 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() &&
506 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
508 return MemDepResult::getClobber(Inst);
513 // Must aliased loads are defs of each other.
514 if (R == AliasAnalysis::MustAlias)
515 return MemDepResult::getDef(Inst);
517 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
518 // in terms of clobbering loads, but since it does this by looking
519 // at the clobbering load directly, it doesn't know about any
520 // phi translation that may have happened along the way.
522 // If we have a partial alias, then return this as a clobber for the
524 if (R == AliasAnalysis::PartialAlias)
525 return MemDepResult::getClobber(Inst);
528 // Random may-alias loads don't depend on each other without a
533 // Stores don't depend on other no-aliased accesses.
534 if (R == AliasAnalysis::NoAlias)
537 // Stores don't alias loads from read-only memory.
538 if (AA->pointsToConstantMemory(LoadLoc))
541 // Stores depend on may/must aliased loads.
542 return MemDepResult::getDef(Inst);
545 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
546 // Atomic stores have complications involved.
547 // A Monotonic store is OK if the query inst is itself not atomic.
548 // FIXME: This is overly conservative.
549 if (!SI->isUnordered()) {
551 return MemDepResult::getClobber(SI);
552 if (SI->getOrdering() != Monotonic)
553 return MemDepResult::getClobber(SI);
554 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
555 if (!QueryLI->isSimple())
556 return MemDepResult::getClobber(SI);
557 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
558 if (!QuerySI->isSimple())
559 return MemDepResult::getClobber(SI);
560 } else if (QueryInst->mayReadOrWriteMemory()) {
561 return MemDepResult::getClobber(SI);
565 // FIXME: this is overly conservative.
566 // While volatile access cannot be eliminated, they do not have to clobber
567 // non-aliasing locations, as normal accesses can for example be reordered
568 // with volatile accesses.
569 if (SI->isVolatile())
570 return MemDepResult::getClobber(SI);
572 // If alias analysis can tell that this store is guaranteed to not modify
573 // the query pointer, ignore it. Use getModRefInfo to handle cases where
574 // the query pointer points to constant memory etc.
575 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
578 // Ok, this store might clobber the query pointer. Check to see if it is
579 // a must alias: in this case, we want to return this as a def.
580 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
582 // If we found a pointer, check if it could be the same as our pointer.
583 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
585 if (R == AliasAnalysis::NoAlias)
587 if (R == AliasAnalysis::MustAlias)
588 return MemDepResult::getDef(Inst);
591 return MemDepResult::getClobber(Inst);
594 // If this is an allocation, and if we know that the accessed pointer is to
595 // the allocation, return Def. This means that there is no dependence and
596 // the access can be optimized based on that. For example, a load could
598 // Note: Only determine this to be a malloc if Inst is the malloc call, not
599 // a subsequent bitcast of the malloc call result. There can be stores to
600 // the malloced memory between the malloc call and its bitcast uses, and we
601 // need to continue scanning until the malloc call.
602 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
603 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
604 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
606 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
607 return MemDepResult::getDef(Inst);
610 // Be conservative if the accessed pointer may alias the allocation.
611 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
612 return MemDepResult::getClobber(Inst);
613 // If the allocation is not aliased and does not read memory (like
614 // strdup), it is safe to ignore.
615 if (isa<AllocaInst>(Inst) ||
616 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
623 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
624 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
625 // If necessary, perform additional analysis.
626 if (MR == AliasAnalysis::ModRef)
627 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
629 case AliasAnalysis::NoModRef:
630 // If the call has no effect on the queried pointer, just ignore it.
632 case AliasAnalysis::Mod:
633 return MemDepResult::getClobber(Inst);
634 case AliasAnalysis::Ref:
635 // If the call is known to never store to the pointer, and if this is a
636 // load query, we can safely ignore it (scan past it).
640 // Otherwise, there is a potential dependence. Return a clobber.
641 return MemDepResult::getClobber(Inst);
645 // No dependence found. If this is the entry block of the function, it is
646 // unknown, otherwise it is non-local.
647 if (BB != &BB->getParent()->getEntryBlock())
648 return MemDepResult::getNonLocal();
649 return MemDepResult::getNonFuncLocal();
652 /// getDependency - Return the instruction on which a memory operation
654 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
655 Instruction *ScanPos = QueryInst;
657 // Check for a cached result
658 MemDepResult &LocalCache = LocalDeps[QueryInst];
660 // If the cached entry is non-dirty, just return it. Note that this depends
661 // on MemDepResult's default constructing to 'dirty'.
662 if (!LocalCache.isDirty())
665 // Otherwise, if we have a dirty entry, we know we can start the scan at that
666 // instruction, which may save us some work.
667 if (Instruction *Inst = LocalCache.getInst()) {
670 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
673 BasicBlock *QueryParent = QueryInst->getParent();
676 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
677 // No dependence found. If this is the entry block of the function, it is
678 // unknown, otherwise it is non-local.
679 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
680 LocalCache = MemDepResult::getNonLocal();
682 LocalCache = MemDepResult::getNonFuncLocal();
684 AliasAnalysis::Location MemLoc;
685 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
687 // If we can do a pointer scan, make it happen.
688 bool isLoad = !(MR & AliasAnalysis::Mod);
689 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
690 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
692 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
693 QueryParent, QueryInst);
694 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
695 CallSite QueryCS(QueryInst);
696 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
697 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
700 // Non-memory instruction.
701 LocalCache = MemDepResult::getUnknown();
704 // Remember the result!
705 if (Instruction *I = LocalCache.getInst())
706 ReverseLocalDeps[I].insert(QueryInst);
712 /// AssertSorted - This method is used when -debug is specified to verify that
713 /// cache arrays are properly kept sorted.
714 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
716 if (Count == -1) Count = Cache.size();
717 if (Count == 0) return;
719 for (unsigned i = 1; i != unsigned(Count); ++i)
720 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
724 /// getNonLocalCallDependency - Perform a full dependency query for the
725 /// specified call, returning the set of blocks that the value is
726 /// potentially live across. The returned set of results will include a
727 /// "NonLocal" result for all blocks where the value is live across.
729 /// This method assumes the instruction returns a "NonLocal" dependency
730 /// within its own block.
732 /// This returns a reference to an internal data structure that may be
733 /// invalidated on the next non-local query or when an instruction is
734 /// removed. Clients must copy this data if they want it around longer than
736 const MemoryDependenceAnalysis::NonLocalDepInfo &
737 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
738 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
739 "getNonLocalCallDependency should only be used on calls with non-local deps!");
740 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
741 NonLocalDepInfo &Cache = CacheP.first;
743 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
744 /// the cached case, this can happen due to instructions being deleted etc. In
745 /// the uncached case, this starts out as the set of predecessors we care
747 SmallVector<BasicBlock*, 32> DirtyBlocks;
749 if (!Cache.empty()) {
750 // Okay, we have a cache entry. If we know it is not dirty, just return it
751 // with no computation.
752 if (!CacheP.second) {
757 // If we already have a partially computed set of results, scan them to
758 // determine what is dirty, seeding our initial DirtyBlocks worklist.
759 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
761 if (I->getResult().isDirty())
762 DirtyBlocks.push_back(I->getBB());
764 // Sort the cache so that we can do fast binary search lookups below.
765 std::sort(Cache.begin(), Cache.end());
767 ++NumCacheDirtyNonLocal;
768 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
769 // << Cache.size() << " cached: " << *QueryInst;
771 // Seed DirtyBlocks with each of the preds of QueryInst's block.
772 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
773 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
774 DirtyBlocks.push_back(*PI);
775 ++NumUncacheNonLocal;
778 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
779 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
781 SmallPtrSet<BasicBlock*, 64> Visited;
783 unsigned NumSortedEntries = Cache.size();
784 DEBUG(AssertSorted(Cache));
786 // Iterate while we still have blocks to update.
787 while (!DirtyBlocks.empty()) {
788 BasicBlock *DirtyBB = DirtyBlocks.back();
789 DirtyBlocks.pop_back();
791 // Already processed this block?
792 if (!Visited.insert(DirtyBB).second)
795 // Do a binary search to see if we already have an entry for this block in
796 // the cache set. If so, find it.
797 DEBUG(AssertSorted(Cache, NumSortedEntries));
798 NonLocalDepInfo::iterator Entry =
799 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
800 NonLocalDepEntry(DirtyBB));
801 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
804 NonLocalDepEntry *ExistingResult = nullptr;
805 if (Entry != Cache.begin()+NumSortedEntries &&
806 Entry->getBB() == DirtyBB) {
807 // If we already have an entry, and if it isn't already dirty, the block
809 if (!Entry->getResult().isDirty())
812 // Otherwise, remember this slot so we can update the value.
813 ExistingResult = &*Entry;
816 // If the dirty entry has a pointer, start scanning from it so we don't have
817 // to rescan the entire block.
818 BasicBlock::iterator ScanPos = DirtyBB->end();
819 if (ExistingResult) {
820 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
822 // We're removing QueryInst's use of Inst.
823 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
824 QueryCS.getInstruction());
828 // Find out if this block has a local dependency for QueryInst.
831 if (ScanPos != DirtyBB->begin()) {
832 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
833 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
834 // No dependence found. If this is the entry block of the function, it is
835 // a clobber, otherwise it is unknown.
836 Dep = MemDepResult::getNonLocal();
838 Dep = MemDepResult::getNonFuncLocal();
841 // If we had a dirty entry for the block, update it. Otherwise, just add
844 ExistingResult->setResult(Dep);
846 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
848 // If the block has a dependency (i.e. it isn't completely transparent to
849 // the value), remember the association!
850 if (!Dep.isNonLocal()) {
851 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
852 // update this when we remove instructions.
853 if (Instruction *Inst = Dep.getInst())
854 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
857 // If the block *is* completely transparent to the load, we need to check
858 // the predecessors of this block. Add them to our worklist.
859 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
860 DirtyBlocks.push_back(*PI);
867 /// getNonLocalPointerDependency - Perform a full dependency query for an
868 /// access to the specified (non-volatile) memory location, returning the
869 /// set of instructions that either define or clobber the value.
871 /// This method assumes the pointer has a "NonLocal" dependency within its
874 void MemoryDependenceAnalysis::
875 getNonLocalPointerDependency(Instruction *QueryInst,
876 SmallVectorImpl<NonLocalDepResult> &Result) {
878 auto getLocation = [](AliasAnalysis *AA, Instruction *Inst) {
879 if (auto *I = dyn_cast<LoadInst>(Inst))
880 return AA->getLocation(I);
881 else if (auto *I = dyn_cast<StoreInst>(Inst))
882 return AA->getLocation(I);
883 else if (auto *I = dyn_cast<VAArgInst>(Inst))
884 return AA->getLocation(I);
885 else if (auto *I = dyn_cast<AtomicCmpXchgInst>(Inst))
886 return AA->getLocation(I);
887 else if (auto *I = dyn_cast<AtomicRMWInst>(Inst))
888 return AA->getLocation(I);
890 llvm_unreachable("unsupported memory instruction");
893 const AliasAnalysis::Location Loc = getLocation(AA, QueryInst);
894 bool isLoad = isa<LoadInst>(QueryInst);
895 BasicBlock *FromBB = QueryInst->getParent();
898 assert(Loc.Ptr->getType()->isPointerTy() &&
899 "Can't get pointer deps of a non-pointer!");
902 // This routine does not expect to deal with volatile instructions.
903 // Doing so would require piping through the QueryInst all the way through.
904 // TODO: volatiles can't be elided, but they can be reordered with other
905 // non-volatile accesses.
907 // We currently give up on any instruction which is ordered, but we do handle
908 // atomic instructions which are unordered.
909 // TODO: Handle ordered instructions
910 auto isOrdered = [](Instruction *Inst) {
911 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
912 return !LI->isUnordered();
913 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
914 return !SI->isUnordered();
918 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
919 Result.push_back(NonLocalDepResult(FromBB,
920 MemDepResult::getUnknown(),
921 const_cast<Value *>(Loc.Ptr)));
924 const DataLayout &DL = FromBB->getModule()->getDataLayout();
925 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
927 // This is the set of blocks we've inspected, and the pointer we consider in
928 // each block. Because of critical edges, we currently bail out if querying
929 // a block with multiple different pointers. This can happen during PHI
931 DenseMap<BasicBlock*, Value*> Visited;
932 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
933 Result, Visited, true))
936 Result.push_back(NonLocalDepResult(FromBB,
937 MemDepResult::getUnknown(),
938 const_cast<Value *>(Loc.Ptr)));
941 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
942 /// Pointer/PointeeSize using either cached information in Cache or by doing a
943 /// lookup (which may use dirty cache info if available). If we do a lookup,
944 /// add the result to the cache.
945 MemDepResult MemoryDependenceAnalysis::
946 GetNonLocalInfoForBlock(Instruction *QueryInst,
947 const AliasAnalysis::Location &Loc,
948 bool isLoad, BasicBlock *BB,
949 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
951 // Do a binary search to see if we already have an entry for this block in
952 // the cache set. If so, find it.
953 NonLocalDepInfo::iterator Entry =
954 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
955 NonLocalDepEntry(BB));
956 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
959 NonLocalDepEntry *ExistingResult = nullptr;
960 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
961 ExistingResult = &*Entry;
963 // If we have a cached entry, and it is non-dirty, use it as the value for
965 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
966 ++NumCacheNonLocalPtr;
967 return ExistingResult->getResult();
970 // Otherwise, we have to scan for the value. If we have a dirty cache
971 // entry, start scanning from its position, otherwise we scan from the end
973 BasicBlock::iterator ScanPos = BB->end();
974 if (ExistingResult && ExistingResult->getResult().getInst()) {
975 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
976 "Instruction invalidated?");
977 ++NumCacheDirtyNonLocalPtr;
978 ScanPos = ExistingResult->getResult().getInst();
980 // Eliminating the dirty entry from 'Cache', so update the reverse info.
981 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
982 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
984 ++NumUncacheNonLocalPtr;
987 // Scan the block for the dependency.
988 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
991 // If we had a dirty entry for the block, update it. Otherwise, just add
994 ExistingResult->setResult(Dep);
996 Cache->push_back(NonLocalDepEntry(BB, Dep));
998 // If the block has a dependency (i.e. it isn't completely transparent to
999 // the value), remember the reverse association because we just added it
1001 if (!Dep.isDef() && !Dep.isClobber())
1004 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
1005 // update MemDep when we remove instructions.
1006 Instruction *Inst = Dep.getInst();
1007 assert(Inst && "Didn't depend on anything?");
1008 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
1009 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
1013 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
1014 /// number of elements in the array that are already properly ordered. This is
1015 /// optimized for the case when only a few entries are added.
1017 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
1018 unsigned NumSortedEntries) {
1019 switch (Cache.size() - NumSortedEntries) {
1021 // done, no new entries.
1024 // Two new entries, insert the last one into place.
1025 NonLocalDepEntry Val = Cache.back();
1027 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1028 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1029 Cache.insert(Entry, Val);
1033 // One new entry, Just insert the new value at the appropriate position.
1034 if (Cache.size() != 1) {
1035 NonLocalDepEntry Val = Cache.back();
1037 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1038 std::upper_bound(Cache.begin(), Cache.end(), Val);
1039 Cache.insert(Entry, Val);
1043 // Added many values, do a full scale sort.
1044 std::sort(Cache.begin(), Cache.end());
1049 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1050 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1051 /// results to the results vector and keep track of which blocks are visited in
1054 /// This has special behavior for the first block queries (when SkipFirstBlock
1055 /// is true). In this special case, it ignores the contents of the specified
1056 /// block and starts returning dependence info for its predecessors.
1058 /// This function returns false on success, or true to indicate that it could
1059 /// not compute dependence information for some reason. This should be treated
1060 /// as a clobber dependence on the first instruction in the predecessor block.
1061 bool MemoryDependenceAnalysis::
1062 getNonLocalPointerDepFromBB(Instruction *QueryInst,
1063 const PHITransAddr &Pointer,
1064 const AliasAnalysis::Location &Loc,
1065 bool isLoad, BasicBlock *StartBB,
1066 SmallVectorImpl<NonLocalDepResult> &Result,
1067 DenseMap<BasicBlock*, Value*> &Visited,
1068 bool SkipFirstBlock) {
1069 // Look up the cached info for Pointer.
1070 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1072 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1073 // CacheKey, this value will be inserted as the associated value. Otherwise,
1074 // it'll be ignored, and we'll have to check to see if the cached size and
1075 // aa tags are consistent with the current query.
1076 NonLocalPointerInfo InitialNLPI;
1077 InitialNLPI.Size = Loc.Size;
1078 InitialNLPI.AATags = Loc.AATags;
1080 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1081 // already have one.
1082 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1083 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1084 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1086 // If we already have a cache entry for this CacheKey, we may need to do some
1087 // work to reconcile the cache entry and the current query.
1089 if (CacheInfo->Size < Loc.Size) {
1090 // The query's Size is greater than the cached one. Throw out the
1091 // cached data and proceed with the query at the greater size.
1092 CacheInfo->Pair = BBSkipFirstBlockPair();
1093 CacheInfo->Size = Loc.Size;
1094 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1095 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1096 if (Instruction *Inst = DI->getResult().getInst())
1097 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1098 CacheInfo->NonLocalDeps.clear();
1099 } else if (CacheInfo->Size > Loc.Size) {
1100 // This query's Size is less than the cached one. Conservatively restart
1101 // the query using the greater size.
1102 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1103 Loc.getWithNewSize(CacheInfo->Size),
1104 isLoad, StartBB, Result, Visited,
1108 // If the query's AATags are inconsistent with the cached one,
1109 // conservatively throw out the cached data and restart the query with
1110 // no tag if needed.
1111 if (CacheInfo->AATags != Loc.AATags) {
1112 if (CacheInfo->AATags) {
1113 CacheInfo->Pair = BBSkipFirstBlockPair();
1114 CacheInfo->AATags = AAMDNodes();
1115 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1116 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1117 if (Instruction *Inst = DI->getResult().getInst())
1118 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1119 CacheInfo->NonLocalDeps.clear();
1122 return getNonLocalPointerDepFromBB(QueryInst,
1123 Pointer, Loc.getWithoutAATags(),
1124 isLoad, StartBB, Result, Visited,
1129 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1131 // If we have valid cached information for exactly the block we are
1132 // investigating, just return it with no recomputation.
1133 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1134 // We have a fully cached result for this query then we can just return the
1135 // cached results and populate the visited set. However, we have to verify
1136 // that we don't already have conflicting results for these blocks. Check
1137 // to ensure that if a block in the results set is in the visited set that
1138 // it was for the same pointer query.
1139 if (!Visited.empty()) {
1140 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1142 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1143 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1146 // We have a pointer mismatch in a block. Just return clobber, saying
1147 // that something was clobbered in this result. We could also do a
1148 // non-fully cached query, but there is little point in doing this.
1153 Value *Addr = Pointer.getAddr();
1154 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1156 Visited.insert(std::make_pair(I->getBB(), Addr));
1157 if (I->getResult().isNonLocal()) {
1162 Result.push_back(NonLocalDepResult(I->getBB(),
1163 MemDepResult::getUnknown(),
1165 } else if (DT->isReachableFromEntry(I->getBB())) {
1166 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1169 ++NumCacheCompleteNonLocalPtr;
1173 // Otherwise, either this is a new block, a block with an invalid cache
1174 // pointer or one that we're about to invalidate by putting more info into it
1175 // than its valid cache info. If empty, the result will be valid cache info,
1176 // otherwise it isn't.
1178 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1180 CacheInfo->Pair = BBSkipFirstBlockPair();
1182 SmallVector<BasicBlock*, 32> Worklist;
1183 Worklist.push_back(StartBB);
1185 // PredList used inside loop.
1186 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1188 // Keep track of the entries that we know are sorted. Previously cached
1189 // entries will all be sorted. The entries we add we only sort on demand (we
1190 // don't insert every element into its sorted position). We know that we
1191 // won't get any reuse from currently inserted values, because we don't
1192 // revisit blocks after we insert info for them.
1193 unsigned NumSortedEntries = Cache->size();
1194 DEBUG(AssertSorted(*Cache));
1196 while (!Worklist.empty()) {
1197 BasicBlock *BB = Worklist.pop_back_val();
1199 // If we do process a large number of blocks it becomes very expensive and
1200 // likely it isn't worth worrying about
1201 if (Result.size() > NumResultsLimit) {
1203 // Sort it now (if needed) so that recursive invocations of
1204 // getNonLocalPointerDepFromBB and other routines that could reuse the
1205 // cache value will only see properly sorted cache arrays.
1206 if (Cache && NumSortedEntries != Cache->size()) {
1207 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1209 // Since we bail out, the "Cache" set won't contain all of the
1210 // results for the query. This is ok (we can still use it to accelerate
1211 // specific block queries) but we can't do the fastpath "return all
1212 // results from the set". Clear out the indicator for this.
1213 CacheInfo->Pair = BBSkipFirstBlockPair();
1217 // Skip the first block if we have it.
1218 if (!SkipFirstBlock) {
1219 // Analyze the dependency of *Pointer in FromBB. See if we already have
1221 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1223 // Get the dependency info for Pointer in BB. If we have cached
1224 // information, we will use it, otherwise we compute it.
1225 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1226 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1227 Loc, isLoad, BB, Cache,
1230 // If we got a Def or Clobber, add this to the list of results.
1231 if (!Dep.isNonLocal()) {
1233 Result.push_back(NonLocalDepResult(BB,
1234 MemDepResult::getUnknown(),
1235 Pointer.getAddr()));
1237 } else if (DT->isReachableFromEntry(BB)) {
1238 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1244 // If 'Pointer' is an instruction defined in this block, then we need to do
1245 // phi translation to change it into a value live in the predecessor block.
1246 // If not, we just add the predecessors to the worklist and scan them with
1247 // the same Pointer.
1248 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1249 SkipFirstBlock = false;
1250 SmallVector<BasicBlock*, 16> NewBlocks;
1251 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1252 // Verify that we haven't looked at this block yet.
1253 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1254 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1255 if (InsertRes.second) {
1256 // First time we've looked at *PI.
1257 NewBlocks.push_back(*PI);
1261 // If we have seen this block before, but it was with a different
1262 // pointer then we have a phi translation failure and we have to treat
1263 // this as a clobber.
1264 if (InsertRes.first->second != Pointer.getAddr()) {
1265 // Make sure to clean up the Visited map before continuing on to
1266 // PredTranslationFailure.
1267 for (unsigned i = 0; i < NewBlocks.size(); i++)
1268 Visited.erase(NewBlocks[i]);
1269 goto PredTranslationFailure;
1272 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1276 // We do need to do phi translation, if we know ahead of time we can't phi
1277 // translate this value, don't even try.
1278 if (!Pointer.IsPotentiallyPHITranslatable())
1279 goto PredTranslationFailure;
1281 // We may have added values to the cache list before this PHI translation.
1282 // If so, we haven't done anything to ensure that the cache remains sorted.
1283 // Sort it now (if needed) so that recursive invocations of
1284 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1285 // value will only see properly sorted cache arrays.
1286 if (Cache && NumSortedEntries != Cache->size()) {
1287 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1288 NumSortedEntries = Cache->size();
1293 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1294 BasicBlock *Pred = *PI;
1295 PredList.push_back(std::make_pair(Pred, Pointer));
1297 // Get the PHI translated pointer in this predecessor. This can fail if
1298 // not translatable, in which case the getAddr() returns null.
1299 PHITransAddr &PredPointer = PredList.back().second;
1300 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1302 Value *PredPtrVal = PredPointer.getAddr();
1304 // Check to see if we have already visited this pred block with another
1305 // pointer. If so, we can't do this lookup. This failure can occur
1306 // with PHI translation when a critical edge exists and the PHI node in
1307 // the successor translates to a pointer value different than the
1308 // pointer the block was first analyzed with.
1309 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1310 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1312 if (!InsertRes.second) {
1313 // We found the pred; take it off the list of preds to visit.
1314 PredList.pop_back();
1316 // If the predecessor was visited with PredPtr, then we already did
1317 // the analysis and can ignore it.
1318 if (InsertRes.first->second == PredPtrVal)
1321 // Otherwise, the block was previously analyzed with a different
1322 // pointer. We can't represent the result of this case, so we just
1323 // treat this as a phi translation failure.
1325 // Make sure to clean up the Visited map before continuing on to
1326 // PredTranslationFailure.
1327 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1328 Visited.erase(PredList[i].first);
1330 goto PredTranslationFailure;
1334 // Actually process results here; this need to be a separate loop to avoid
1335 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1336 // any results for. (getNonLocalPointerDepFromBB will modify our
1337 // datastructures in ways the code after the PredTranslationFailure label
1339 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1340 BasicBlock *Pred = PredList[i].first;
1341 PHITransAddr &PredPointer = PredList[i].second;
1342 Value *PredPtrVal = PredPointer.getAddr();
1344 bool CanTranslate = true;
1345 // If PHI translation was unable to find an available pointer in this
1346 // predecessor, then we have to assume that the pointer is clobbered in
1347 // that predecessor. We can still do PRE of the load, which would insert
1348 // a computation of the pointer in this predecessor.
1350 CanTranslate = false;
1352 // FIXME: it is entirely possible that PHI translating will end up with
1353 // the same value. Consider PHI translating something like:
1354 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1355 // to recurse here, pedantically speaking.
1357 // If getNonLocalPointerDepFromBB fails here, that means the cached
1358 // result conflicted with the Visited list; we have to conservatively
1359 // assume it is unknown, but this also does not block PRE of the load.
1360 if (!CanTranslate ||
1361 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1362 Loc.getWithNewPtr(PredPtrVal),
1365 // Add the entry to the Result list.
1366 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1367 Result.push_back(Entry);
1369 // Since we had a phi translation failure, the cache for CacheKey won't
1370 // include all of the entries that we need to immediately satisfy future
1371 // queries. Mark this in NonLocalPointerDeps by setting the
1372 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1373 // cached value to do more work but not miss the phi trans failure.
1374 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1375 NLPI.Pair = BBSkipFirstBlockPair();
1380 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1381 CacheInfo = &NonLocalPointerDeps[CacheKey];
1382 Cache = &CacheInfo->NonLocalDeps;
1383 NumSortedEntries = Cache->size();
1385 // Since we did phi translation, the "Cache" set won't contain all of the
1386 // results for the query. This is ok (we can still use it to accelerate
1387 // specific block queries) but we can't do the fastpath "return all
1388 // results from the set" Clear out the indicator for this.
1389 CacheInfo->Pair = BBSkipFirstBlockPair();
1390 SkipFirstBlock = false;
1393 PredTranslationFailure:
1394 // The following code is "failure"; we can't produce a sane translation
1395 // for the given block. It assumes that we haven't modified any of
1396 // our datastructures while processing the current block.
1399 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1400 CacheInfo = &NonLocalPointerDeps[CacheKey];
1401 Cache = &CacheInfo->NonLocalDeps;
1402 NumSortedEntries = Cache->size();
1405 // Since we failed phi translation, the "Cache" set won't contain all of the
1406 // results for the query. This is ok (we can still use it to accelerate
1407 // specific block queries) but we can't do the fastpath "return all
1408 // results from the set". Clear out the indicator for this.
1409 CacheInfo->Pair = BBSkipFirstBlockPair();
1411 // If *nothing* works, mark the pointer as unknown.
1413 // If this is the magic first block, return this as a clobber of the whole
1414 // incoming value. Since we can't phi translate to one of the predecessors,
1415 // we have to bail out.
1419 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1420 assert(I != Cache->rend() && "Didn't find current block??");
1421 if (I->getBB() != BB)
1424 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1425 "Should only be here with transparent block");
1426 I->setResult(MemDepResult::getUnknown());
1427 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1428 Pointer.getAddr()));
1433 // Okay, we're done now. If we added new values to the cache, re-sort it.
1434 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1435 DEBUG(AssertSorted(*Cache));
1439 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1440 /// CachedNonLocalPointerInfo, remove it.
1441 void MemoryDependenceAnalysis::
1442 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1443 CachedNonLocalPointerInfo::iterator It =
1444 NonLocalPointerDeps.find(P);
1445 if (It == NonLocalPointerDeps.end()) return;
1447 // Remove all of the entries in the BB->val map. This involves removing
1448 // instructions from the reverse map.
1449 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1451 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1452 Instruction *Target = PInfo[i].getResult().getInst();
1453 if (!Target) continue; // Ignore non-local dep results.
1454 assert(Target->getParent() == PInfo[i].getBB());
1456 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1457 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1460 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1461 NonLocalPointerDeps.erase(It);
1465 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1466 /// information about the specified pointer, because it may be too
1467 /// conservative in memdep. This is an optional call that can be used when
1468 /// the client detects an equivalence between the pointer and some other
1469 /// value and replaces the other value with ptr. This can make Ptr available
1470 /// in more places that cached info does not necessarily keep.
1471 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1472 // If Ptr isn't really a pointer, just ignore it.
1473 if (!Ptr->getType()->isPointerTy()) return;
1474 // Flush store info for the pointer.
1475 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1476 // Flush load info for the pointer.
1477 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1480 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1481 /// This needs to be done when the CFG changes, e.g., due to splitting
1483 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1487 /// removeInstruction - Remove an instruction from the dependence analysis,
1488 /// updating the dependence of instructions that previously depended on it.
1489 /// This method attempts to keep the cache coherent using the reverse map.
1490 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1491 // Walk through the Non-local dependencies, removing this one as the value
1492 // for any cached queries.
1493 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1494 if (NLDI != NonLocalDeps.end()) {
1495 NonLocalDepInfo &BlockMap = NLDI->second.first;
1496 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1498 if (Instruction *Inst = DI->getResult().getInst())
1499 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1500 NonLocalDeps.erase(NLDI);
1503 // If we have a cached local dependence query for this instruction, remove it.
1505 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1506 if (LocalDepEntry != LocalDeps.end()) {
1507 // Remove us from DepInst's reverse set now that the local dep info is gone.
1508 if (Instruction *Inst = LocalDepEntry->second.getInst())
1509 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1511 // Remove this local dependency info.
1512 LocalDeps.erase(LocalDepEntry);
1515 // If we have any cached pointer dependencies on this instruction, remove
1516 // them. If the instruction has non-pointer type, then it can't be a pointer
1519 // Remove it from both the load info and the store info. The instruction
1520 // can't be in either of these maps if it is non-pointer.
1521 if (RemInst->getType()->isPointerTy()) {
1522 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1523 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1526 // Loop over all of the things that depend on the instruction we're removing.
1528 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1530 // If we find RemInst as a clobber or Def in any of the maps for other values,
1531 // we need to replace its entry with a dirty version of the instruction after
1532 // it. If RemInst is a terminator, we use a null dirty value.
1534 // Using a dirty version of the instruction after RemInst saves having to scan
1535 // the entire block to get to this point.
1536 MemDepResult NewDirtyVal;
1537 if (!RemInst->isTerminator())
1538 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1540 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1541 if (ReverseDepIt != ReverseLocalDeps.end()) {
1542 // RemInst can't be the terminator if it has local stuff depending on it.
1543 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1544 "Nothing can locally depend on a terminator");
1546 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1547 assert(InstDependingOnRemInst != RemInst &&
1548 "Already removed our local dep info");
1550 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1552 // Make sure to remember that new things depend on NewDepInst.
1553 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1554 "a local dep on this if it is a terminator!");
1555 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1556 InstDependingOnRemInst));
1559 ReverseLocalDeps.erase(ReverseDepIt);
1561 // Add new reverse deps after scanning the set, to avoid invalidating the
1562 // 'ReverseDeps' reference.
1563 while (!ReverseDepsToAdd.empty()) {
1564 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1565 .insert(ReverseDepsToAdd.back().second);
1566 ReverseDepsToAdd.pop_back();
1570 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1571 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1572 for (Instruction *I : ReverseDepIt->second) {
1573 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1575 PerInstNLInfo &INLD = NonLocalDeps[I];
1576 // The information is now dirty!
1579 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1580 DE = INLD.first.end(); DI != DE; ++DI) {
1581 if (DI->getResult().getInst() != RemInst) continue;
1583 // Convert to a dirty entry for the subsequent instruction.
1584 DI->setResult(NewDirtyVal);
1586 if (Instruction *NextI = NewDirtyVal.getInst())
1587 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1591 ReverseNonLocalDeps.erase(ReverseDepIt);
1593 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1594 while (!ReverseDepsToAdd.empty()) {
1595 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1596 .insert(ReverseDepsToAdd.back().second);
1597 ReverseDepsToAdd.pop_back();
1601 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1602 // value in the NonLocalPointerDeps info.
1603 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1604 ReverseNonLocalPtrDeps.find(RemInst);
1605 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1606 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1608 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1609 assert(P.getPointer() != RemInst &&
1610 "Already removed NonLocalPointerDeps info for RemInst");
1612 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1614 // The cache is not valid for any specific block anymore.
1615 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1617 // Update any entries for RemInst to use the instruction after it.
1618 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1620 if (DI->getResult().getInst() != RemInst) continue;
1622 // Convert to a dirty entry for the subsequent instruction.
1623 DI->setResult(NewDirtyVal);
1625 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1626 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1629 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1630 // subsequent value may invalidate the sortedness.
1631 std::sort(NLPDI.begin(), NLPDI.end());
1634 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1636 while (!ReversePtrDepsToAdd.empty()) {
1637 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1638 .insert(ReversePtrDepsToAdd.back().second);
1639 ReversePtrDepsToAdd.pop_back();
1644 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1645 AA->deleteValue(RemInst);
1646 DEBUG(verifyRemoved(RemInst));
1648 /// verifyRemoved - Verify that the specified instruction does not occur
1649 /// in our internal data structures. This function verifies by asserting in
1651 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1653 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1654 E = LocalDeps.end(); I != E; ++I) {
1655 assert(I->first != D && "Inst occurs in data structures");
1656 assert(I->second.getInst() != D &&
1657 "Inst occurs in data structures");
1660 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1661 E = NonLocalPointerDeps.end(); I != E; ++I) {
1662 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1663 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1664 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1666 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1669 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1670 E = NonLocalDeps.end(); I != E; ++I) {
1671 assert(I->first != D && "Inst occurs in data structures");
1672 const PerInstNLInfo &INLD = I->second;
1673 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1674 EE = INLD.first.end(); II != EE; ++II)
1675 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1678 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1679 E = ReverseLocalDeps.end(); I != E; ++I) {
1680 assert(I->first != D && "Inst occurs in data structures");
1681 for (Instruction *Inst : I->second)
1682 assert(Inst != D && "Inst occurs in data structures");
1685 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1686 E = ReverseNonLocalDeps.end();
1688 assert(I->first != D && "Inst occurs in data structures");
1689 for (Instruction *Inst : I->second)
1690 assert(Inst != D && "Inst occurs in data structures");
1693 for (ReverseNonLocalPtrDepTy::const_iterator
1694 I = ReverseNonLocalPtrDeps.begin(),
1695 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1696 assert(I->first != D && "Inst occurs in rev NLPD map");
1698 for (ValueIsLoadPair P : I->second)
1699 assert(P != ValueIsLoadPair(D, false) &&
1700 P != ValueIsLoadPair(D, true) &&
1701 "Inst occurs in ReverseNonLocalPtrDeps map");