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()
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;
102 /// RemoveFromReverseMap - This is a helper function that removes Val from
103 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
104 template <typename KeyTy>
105 static void RemoveFromReverseMap(DenseMap<Instruction*,
106 SmallPtrSet<KeyTy, 4> > &ReverseMap,
107 Instruction *Inst, KeyTy Val) {
108 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
109 InstIt = ReverseMap.find(Inst);
110 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
111 bool Found = InstIt->second.erase(Val);
112 assert(Found && "Invalid reverse map!"); (void)Found;
113 if (InstIt->second.empty())
114 ReverseMap.erase(InstIt);
117 /// GetLocation - If the given instruction references a specific memory
118 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
119 /// Return a ModRefInfo value describing the general behavior of the
122 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
123 AliasAnalysis::Location &Loc,
125 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
126 if (LI->isUnordered()) {
127 Loc = MemoryLocation::get(LI);
128 return AliasAnalysis::Ref;
130 if (LI->getOrdering() == Monotonic) {
131 Loc = MemoryLocation::get(LI);
132 return AliasAnalysis::ModRef;
134 Loc = AliasAnalysis::Location();
135 return AliasAnalysis::ModRef;
138 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
139 if (SI->isUnordered()) {
140 Loc = MemoryLocation::get(SI);
141 return AliasAnalysis::Mod;
143 if (SI->getOrdering() == Monotonic) {
144 Loc = MemoryLocation::get(SI);
145 return AliasAnalysis::ModRef;
147 Loc = AliasAnalysis::Location();
148 return AliasAnalysis::ModRef;
151 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
152 Loc = MemoryLocation::get(V);
153 return AliasAnalysis::ModRef;
156 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
157 // calls to free() deallocate the entire structure
158 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
159 return AliasAnalysis::Mod;
162 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
165 switch (II->getIntrinsicID()) {
166 case Intrinsic::lifetime_start:
167 case Intrinsic::lifetime_end:
168 case Intrinsic::invariant_start:
169 II->getAAMetadata(AAInfo);
170 Loc = AliasAnalysis::Location(II->getArgOperand(1),
171 cast<ConstantInt>(II->getArgOperand(0))
172 ->getZExtValue(), AAInfo);
173 // These intrinsics don't really modify the memory, but returning Mod
174 // will allow them to be handled conservatively.
175 return AliasAnalysis::Mod;
176 case Intrinsic::invariant_end:
177 II->getAAMetadata(AAInfo);
178 Loc = AliasAnalysis::Location(II->getArgOperand(2),
179 cast<ConstantInt>(II->getArgOperand(1))
180 ->getZExtValue(), AAInfo);
181 // These intrinsics don't really modify the memory, but returning Mod
182 // will allow them to be handled conservatively.
183 return AliasAnalysis::Mod;
189 // Otherwise, just do the coarse-grained thing that always works.
190 if (Inst->mayWriteToMemory())
191 return AliasAnalysis::ModRef;
192 if (Inst->mayReadFromMemory())
193 return AliasAnalysis::Ref;
194 return AliasAnalysis::NoModRef;
197 /// getCallSiteDependencyFrom - Private helper for finding the local
198 /// dependencies of a call site.
199 MemDepResult MemoryDependenceAnalysis::
200 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
201 BasicBlock::iterator ScanIt, BasicBlock *BB) {
202 unsigned Limit = BlockScanLimit;
204 // Walk backwards through the block, looking for dependencies
205 while (ScanIt != BB->begin()) {
206 // Limit the amount of scanning we do so we don't end up with quadratic
207 // running time on extreme testcases.
210 return MemDepResult::getUnknown();
212 Instruction *Inst = --ScanIt;
214 // If this inst is a memory op, get the pointer it accessed
215 AliasAnalysis::Location Loc;
216 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
218 // A simple instruction.
219 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
220 return MemDepResult::getClobber(Inst);
224 if (auto InstCS = CallSite(Inst)) {
225 // Debug intrinsics don't cause dependences.
226 if (isa<DbgInfoIntrinsic>(Inst)) continue;
227 // If these two calls do not interfere, look past it.
228 switch (AA->getModRefInfo(CS, InstCS)) {
229 case AliasAnalysis::NoModRef:
230 // If the two calls are the same, return InstCS as a Def, so that
231 // CS can be found redundant and eliminated.
232 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
233 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
234 return MemDepResult::getDef(Inst);
236 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
240 return MemDepResult::getClobber(Inst);
244 // If we could not obtain a pointer for the instruction and the instruction
245 // touches memory then assume that this is a dependency.
246 if (MR != AliasAnalysis::NoModRef)
247 return MemDepResult::getClobber(Inst);
250 // No dependence found. If this is the entry block of the function, it is
251 // unknown, otherwise it is non-local.
252 if (BB != &BB->getParent()->getEntryBlock())
253 return MemDepResult::getNonLocal();
254 return MemDepResult::getNonFuncLocal();
257 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
258 /// would fully overlap MemLoc if done as a wider legal integer load.
260 /// MemLocBase, MemLocOffset are lazily computed here the first time the
261 /// base/offs of memloc is needed.
262 static bool isLoadLoadClobberIfExtendedToFullWidth(
263 const AliasAnalysis::Location &MemLoc, const Value *&MemLocBase,
264 int64_t &MemLocOffs, const LoadInst *LI) {
265 const DataLayout &DL = LI->getModule()->getDataLayout();
267 // If we haven't already computed the base/offset of MemLoc, do so now.
269 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
271 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
272 MemLocBase, MemLocOffs, MemLoc.Size, LI);
276 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
277 /// looks at a memory location for a load (specified by MemLocBase, Offs,
278 /// and Size) and compares it against a load. If the specified load could
279 /// be safely widened to a larger integer load that is 1) still efficient,
280 /// 2) safe for the target, and 3) would provide the specified memory
281 /// location value, then this function returns the size in bytes of the
282 /// load width to use. If not, this returns zero.
283 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
284 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
285 const LoadInst *LI) {
286 // We can only extend simple integer loads.
287 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
289 // Load widening is hostile to ThreadSanitizer: it may cause false positives
290 // or make the reports more cryptic (access sizes are wrong).
291 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
294 const DataLayout &DL = LI->getModule()->getDataLayout();
296 // Get the base of this load.
298 const Value *LIBase =
299 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
301 // If the two pointers are not based on the same pointer, we can't tell that
303 if (LIBase != MemLocBase) return 0;
305 // Okay, the two values are based on the same pointer, but returned as
306 // no-alias. This happens when we have things like two byte loads at "P+1"
307 // and "P+3". Check to see if increasing the size of the "LI" load up to its
308 // alignment (or the largest native integer type) will allow us to load all
309 // the bits required by MemLoc.
311 // If MemLoc is before LI, then no widening of LI will help us out.
312 if (MemLocOffs < LIOffs) return 0;
314 // Get the alignment of the load in bytes. We assume that it is safe to load
315 // any legal integer up to this size without a problem. For example, if we're
316 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
317 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
319 unsigned LoadAlign = LI->getAlignment();
321 int64_t MemLocEnd = MemLocOffs+MemLocSize;
323 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
324 if (LIOffs+LoadAlign < MemLocEnd) return 0;
326 // This is the size of the load to try. Start with the next larger power of
328 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
329 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
332 // If this load size is bigger than our known alignment or would not fit
333 // into a native integer register, then we fail.
334 if (NewLoadByteSize > LoadAlign ||
335 !DL.fitsInLegalInteger(NewLoadByteSize*8))
338 if (LIOffs + NewLoadByteSize > MemLocEnd &&
339 LI->getParent()->getParent()->hasFnAttribute(
340 Attribute::SanitizeAddress))
341 // We will be reading past the location accessed by the original program.
342 // While this is safe in a regular build, Address Safety analysis tools
343 // may start reporting false warnings. So, don't do widening.
346 // If a load of this width would include all of MemLoc, then we succeed.
347 if (LIOffs+NewLoadByteSize >= MemLocEnd)
348 return NewLoadByteSize;
350 NewLoadByteSize <<= 1;
354 static bool isVolatile(Instruction *Inst) {
355 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
356 return LI->isVolatile();
357 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
358 return SI->isVolatile();
359 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
360 return AI->isVolatile();
365 /// getPointerDependencyFrom - Return the instruction on which a memory
366 /// location depends. If isLoad is true, this routine ignores may-aliases with
367 /// read-only operations. If isLoad is false, this routine ignores may-aliases
368 /// with reads from read-only locations. If possible, pass the query
369 /// instruction as well; this function may take advantage of the metadata
370 /// annotated to the query instruction to refine the result.
371 MemDepResult MemoryDependenceAnalysis::
372 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
373 BasicBlock::iterator ScanIt, BasicBlock *BB,
374 Instruction *QueryInst) {
376 const Value *MemLocBase = nullptr;
377 int64_t MemLocOffset = 0;
378 unsigned Limit = BlockScanLimit;
379 bool isInvariantLoad = false;
381 // We must be careful with atomic accesses, as they may allow another thread
382 // to touch this location, cloberring it. We are conservative: if the
383 // QueryInst is not a simple (non-atomic) memory access, we automatically
384 // return getClobber.
385 // If it is simple, we know based on the results of
386 // "Compiler testing via a theory of sound optimisations in the C11/C++11
387 // memory model" in PLDI 2013, that a non-atomic location can only be
388 // clobbered between a pair of a release and an acquire action, with no
389 // access to the location in between.
390 // Here is an example for giving the general intuition behind this rule.
391 // In the following code:
393 // release action; [1]
394 // acquire action; [4]
396 // It is unsafe to replace %val by 0 because another thread may be running:
397 // acquire action; [2]
399 // release action; [3]
400 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
401 // being 42. A key property of this program however is that if either
402 // 1 or 4 were missing, there would be a race between the store of 42
403 // either the store of 0 or the load (making the whole progam racy).
404 // The paper mentionned above shows that the same property is respected
405 // by every program that can detect any optimisation of that kind: either
406 // it is racy (undefined) or there is a release followed by an acquire
407 // between the pair of accesses under consideration.
409 // If the load is invariant, we "know" that it doesn't alias *any* write. We
410 // do want to respect mustalias results since defs are useful for value
411 // forwarding, but any mayalias write can be assumed to be noalias.
412 // Arguably, this logic should be pushed inside AliasAnalysis itself.
413 if (isLoad && QueryInst) {
414 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
415 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
416 isInvariantLoad = true;
419 const DataLayout &DL = BB->getModule()->getDataLayout();
421 // Walk backwards through the basic block, looking for dependencies.
422 while (ScanIt != BB->begin()) {
423 Instruction *Inst = --ScanIt;
425 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
426 // Debug intrinsics don't (and can't) cause dependencies.
427 if (isa<DbgInfoIntrinsic>(II)) continue;
429 // Limit the amount of scanning we do so we don't end up with quadratic
430 // running time on extreme testcases.
433 return MemDepResult::getUnknown();
435 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
436 // If we reach a lifetime begin or end marker, then the query ends here
437 // because the value is undefined.
438 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
439 // FIXME: This only considers queries directly on the invariant-tagged
440 // pointer, not on query pointers that are indexed off of them. It'd
441 // be nice to handle that at some point (the right approach is to use
442 // GetPointerBaseWithConstantOffset).
443 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
445 return MemDepResult::getDef(II);
450 // Values depend on loads if the pointers are must aliased. This means that
451 // a load depends on another must aliased load from the same value.
452 // One exception is atomic loads: a value can depend on an atomic load that it
453 // does not alias with when this atomic load indicates that another thread may
454 // be accessing the location.
455 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
457 // While volatile access cannot be eliminated, they do not have to clobber
458 // non-aliasing locations, as normal accesses, for example, can be safely
459 // reordered with volatile accesses.
460 if (LI->isVolatile()) {
462 // Original QueryInst *may* be volatile
463 return MemDepResult::getClobber(LI);
464 if (isVolatile(QueryInst))
465 // Ordering required if QueryInst is itself volatile
466 return MemDepResult::getClobber(LI);
467 // Otherwise, volatile doesn't imply any special ordering
470 // Atomic loads have complications involved.
471 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
472 // FIXME: This is overly conservative.
473 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
475 return MemDepResult::getClobber(LI);
476 if (LI->getOrdering() != Monotonic)
477 return MemDepResult::getClobber(LI);
478 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
479 if (!QueryLI->isSimple())
480 return MemDepResult::getClobber(LI);
481 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
482 if (!QuerySI->isSimple())
483 return MemDepResult::getClobber(LI);
484 } else if (QueryInst->mayReadOrWriteMemory()) {
485 return MemDepResult::getClobber(LI);
489 AliasAnalysis::Location LoadLoc = MemoryLocation::get(LI);
491 // If we found a pointer, check if it could be the same as our pointer.
492 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
495 if (R == AliasAnalysis::NoAlias) {
496 // If this is an over-aligned integer load (for example,
497 // "load i8* %P, align 4") see if it would obviously overlap with the
498 // queried location if widened to a larger load (e.g. if the queried
499 // location is 1 byte at P+1). If so, return it as a load/load
500 // clobber result, allowing the client to decide to widen the load if
502 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
503 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() &&
504 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
506 return MemDepResult::getClobber(Inst);
511 // Must aliased loads are defs of each other.
512 if (R == AliasAnalysis::MustAlias)
513 return MemDepResult::getDef(Inst);
515 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
516 // in terms of clobbering loads, but since it does this by looking
517 // at the clobbering load directly, it doesn't know about any
518 // phi translation that may have happened along the way.
520 // If we have a partial alias, then return this as a clobber for the
522 if (R == AliasAnalysis::PartialAlias)
523 return MemDepResult::getClobber(Inst);
526 // Random may-alias loads don't depend on each other without a
531 // Stores don't depend on other no-aliased accesses.
532 if (R == AliasAnalysis::NoAlias)
535 // Stores don't alias loads from read-only memory.
536 if (AA->pointsToConstantMemory(LoadLoc))
539 // Stores depend on may/must aliased loads.
540 return MemDepResult::getDef(Inst);
543 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
544 // Atomic stores have complications involved.
545 // A Monotonic store is OK if the query inst is itself not atomic.
546 // FIXME: This is overly conservative.
547 if (!SI->isUnordered()) {
549 return MemDepResult::getClobber(SI);
550 if (SI->getOrdering() != Monotonic)
551 return MemDepResult::getClobber(SI);
552 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
553 if (!QueryLI->isSimple())
554 return MemDepResult::getClobber(SI);
555 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
556 if (!QuerySI->isSimple())
557 return MemDepResult::getClobber(SI);
558 } else if (QueryInst->mayReadOrWriteMemory()) {
559 return MemDepResult::getClobber(SI);
563 // FIXME: this is overly conservative.
564 // While volatile access cannot be eliminated, they do not have to clobber
565 // non-aliasing locations, as normal accesses can for example be reordered
566 // with volatile accesses.
567 if (SI->isVolatile())
568 return MemDepResult::getClobber(SI);
570 // If alias analysis can tell that this store is guaranteed to not modify
571 // the query pointer, ignore it. Use getModRefInfo to handle cases where
572 // the query pointer points to constant memory etc.
573 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
576 // Ok, this store might clobber the query pointer. Check to see if it is
577 // a must alias: in this case, we want to return this as a def.
578 AliasAnalysis::Location StoreLoc = MemoryLocation::get(SI);
580 // If we found a pointer, check if it could be the same as our pointer.
581 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
583 if (R == AliasAnalysis::NoAlias)
585 if (R == AliasAnalysis::MustAlias)
586 return MemDepResult::getDef(Inst);
589 return MemDepResult::getClobber(Inst);
592 // If this is an allocation, and if we know that the accessed pointer is to
593 // the allocation, return Def. This means that there is no dependence and
594 // the access can be optimized based on that. For example, a load could
596 // Note: Only determine this to be a malloc if Inst is the malloc call, not
597 // a subsequent bitcast of the malloc call result. There can be stores to
598 // the malloced memory between the malloc call and its bitcast uses, and we
599 // need to continue scanning until the malloc call.
600 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
601 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
602 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
604 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
605 return MemDepResult::getDef(Inst);
608 // Be conservative if the accessed pointer may alias the allocation.
609 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
610 return MemDepResult::getClobber(Inst);
611 // If the allocation is not aliased and does not read memory (like
612 // strdup), it is safe to ignore.
613 if (isa<AllocaInst>(Inst) ||
614 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
621 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
622 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
623 // If necessary, perform additional analysis.
624 if (MR == AliasAnalysis::ModRef)
625 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
627 case AliasAnalysis::NoModRef:
628 // If the call has no effect on the queried pointer, just ignore it.
630 case AliasAnalysis::Mod:
631 return MemDepResult::getClobber(Inst);
632 case AliasAnalysis::Ref:
633 // If the call is known to never store to the pointer, and if this is a
634 // load query, we can safely ignore it (scan past it).
638 // Otherwise, there is a potential dependence. Return a clobber.
639 return MemDepResult::getClobber(Inst);
643 // No dependence found. If this is the entry block of the function, it is
644 // unknown, otherwise it is non-local.
645 if (BB != &BB->getParent()->getEntryBlock())
646 return MemDepResult::getNonLocal();
647 return MemDepResult::getNonFuncLocal();
650 /// getDependency - Return the instruction on which a memory operation
652 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
653 Instruction *ScanPos = QueryInst;
655 // Check for a cached result
656 MemDepResult &LocalCache = LocalDeps[QueryInst];
658 // If the cached entry is non-dirty, just return it. Note that this depends
659 // on MemDepResult's default constructing to 'dirty'.
660 if (!LocalCache.isDirty())
663 // Otherwise, if we have a dirty entry, we know we can start the scan at that
664 // instruction, which may save us some work.
665 if (Instruction *Inst = LocalCache.getInst()) {
668 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
671 BasicBlock *QueryParent = QueryInst->getParent();
674 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
675 // No dependence found. If this is the entry block of the function, it is
676 // unknown, otherwise it is non-local.
677 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
678 LocalCache = MemDepResult::getNonLocal();
680 LocalCache = MemDepResult::getNonFuncLocal();
682 AliasAnalysis::Location MemLoc;
683 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
685 // If we can do a pointer scan, make it happen.
686 bool isLoad = !(MR & AliasAnalysis::Mod);
687 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
688 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
690 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
691 QueryParent, QueryInst);
692 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
693 CallSite QueryCS(QueryInst);
694 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
695 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
698 // Non-memory instruction.
699 LocalCache = MemDepResult::getUnknown();
702 // Remember the result!
703 if (Instruction *I = LocalCache.getInst())
704 ReverseLocalDeps[I].insert(QueryInst);
710 /// AssertSorted - This method is used when -debug is specified to verify that
711 /// cache arrays are properly kept sorted.
712 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
714 if (Count == -1) Count = Cache.size();
715 if (Count == 0) return;
717 for (unsigned i = 1; i != unsigned(Count); ++i)
718 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
722 /// getNonLocalCallDependency - Perform a full dependency query for the
723 /// specified call, returning the set of blocks that the value is
724 /// potentially live across. The returned set of results will include a
725 /// "NonLocal" result for all blocks where the value is live across.
727 /// This method assumes the instruction returns a "NonLocal" dependency
728 /// within its own block.
730 /// This returns a reference to an internal data structure that may be
731 /// invalidated on the next non-local query or when an instruction is
732 /// removed. Clients must copy this data if they want it around longer than
734 const MemoryDependenceAnalysis::NonLocalDepInfo &
735 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
736 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
737 "getNonLocalCallDependency should only be used on calls with non-local deps!");
738 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
739 NonLocalDepInfo &Cache = CacheP.first;
741 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
742 /// the cached case, this can happen due to instructions being deleted etc. In
743 /// the uncached case, this starts out as the set of predecessors we care
745 SmallVector<BasicBlock*, 32> DirtyBlocks;
747 if (!Cache.empty()) {
748 // Okay, we have a cache entry. If we know it is not dirty, just return it
749 // with no computation.
750 if (!CacheP.second) {
755 // If we already have a partially computed set of results, scan them to
756 // determine what is dirty, seeding our initial DirtyBlocks worklist.
757 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
759 if (I->getResult().isDirty())
760 DirtyBlocks.push_back(I->getBB());
762 // Sort the cache so that we can do fast binary search lookups below.
763 std::sort(Cache.begin(), Cache.end());
765 ++NumCacheDirtyNonLocal;
766 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
767 // << Cache.size() << " cached: " << *QueryInst;
769 // Seed DirtyBlocks with each of the preds of QueryInst's block.
770 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
771 for (BasicBlock *Pred : PredCache.get(QueryBB))
772 DirtyBlocks.push_back(Pred);
773 ++NumUncacheNonLocal;
776 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
777 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
779 SmallPtrSet<BasicBlock*, 64> Visited;
781 unsigned NumSortedEntries = Cache.size();
782 DEBUG(AssertSorted(Cache));
784 // Iterate while we still have blocks to update.
785 while (!DirtyBlocks.empty()) {
786 BasicBlock *DirtyBB = DirtyBlocks.back();
787 DirtyBlocks.pop_back();
789 // Already processed this block?
790 if (!Visited.insert(DirtyBB).second)
793 // Do a binary search to see if we already have an entry for this block in
794 // the cache set. If so, find it.
795 DEBUG(AssertSorted(Cache, NumSortedEntries));
796 NonLocalDepInfo::iterator Entry =
797 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
798 NonLocalDepEntry(DirtyBB));
799 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
802 NonLocalDepEntry *ExistingResult = nullptr;
803 if (Entry != Cache.begin()+NumSortedEntries &&
804 Entry->getBB() == DirtyBB) {
805 // If we already have an entry, and if it isn't already dirty, the block
807 if (!Entry->getResult().isDirty())
810 // Otherwise, remember this slot so we can update the value.
811 ExistingResult = &*Entry;
814 // If the dirty entry has a pointer, start scanning from it so we don't have
815 // to rescan the entire block.
816 BasicBlock::iterator ScanPos = DirtyBB->end();
817 if (ExistingResult) {
818 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
820 // We're removing QueryInst's use of Inst.
821 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
822 QueryCS.getInstruction());
826 // Find out if this block has a local dependency for QueryInst.
829 if (ScanPos != DirtyBB->begin()) {
830 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
831 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
832 // No dependence found. If this is the entry block of the function, it is
833 // a clobber, otherwise it is unknown.
834 Dep = MemDepResult::getNonLocal();
836 Dep = MemDepResult::getNonFuncLocal();
839 // If we had a dirty entry for the block, update it. Otherwise, just add
842 ExistingResult->setResult(Dep);
844 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
846 // If the block has a dependency (i.e. it isn't completely transparent to
847 // the value), remember the association!
848 if (!Dep.isNonLocal()) {
849 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
850 // update this when we remove instructions.
851 if (Instruction *Inst = Dep.getInst())
852 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
855 // If the block *is* completely transparent to the load, we need to check
856 // the predecessors of this block. Add them to our worklist.
857 for (BasicBlock *Pred : PredCache.get(DirtyBB))
858 DirtyBlocks.push_back(Pred);
865 /// getNonLocalPointerDependency - Perform a full dependency query for an
866 /// access to the specified (non-volatile) memory location, returning the
867 /// set of instructions that either define or clobber the value.
869 /// This method assumes the pointer has a "NonLocal" dependency within its
872 void MemoryDependenceAnalysis::
873 getNonLocalPointerDependency(Instruction *QueryInst,
874 SmallVectorImpl<NonLocalDepResult> &Result) {
875 const AliasAnalysis::Location Loc = MemoryLocation::get(QueryInst);
876 bool isLoad = isa<LoadInst>(QueryInst);
877 BasicBlock *FromBB = QueryInst->getParent();
880 assert(Loc.Ptr->getType()->isPointerTy() &&
881 "Can't get pointer deps of a non-pointer!");
884 // This routine does not expect to deal with volatile instructions.
885 // Doing so would require piping through the QueryInst all the way through.
886 // TODO: volatiles can't be elided, but they can be reordered with other
887 // non-volatile accesses.
889 // We currently give up on any instruction which is ordered, but we do handle
890 // atomic instructions which are unordered.
891 // TODO: Handle ordered instructions
892 auto isOrdered = [](Instruction *Inst) {
893 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
894 return !LI->isUnordered();
895 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
896 return !SI->isUnordered();
900 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
901 Result.push_back(NonLocalDepResult(FromBB,
902 MemDepResult::getUnknown(),
903 const_cast<Value *>(Loc.Ptr)));
906 const DataLayout &DL = FromBB->getModule()->getDataLayout();
907 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
909 // This is the set of blocks we've inspected, and the pointer we consider in
910 // each block. Because of critical edges, we currently bail out if querying
911 // a block with multiple different pointers. This can happen during PHI
913 DenseMap<BasicBlock*, Value*> Visited;
914 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
915 Result, Visited, true))
918 Result.push_back(NonLocalDepResult(FromBB,
919 MemDepResult::getUnknown(),
920 const_cast<Value *>(Loc.Ptr)));
923 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
924 /// Pointer/PointeeSize using either cached information in Cache or by doing a
925 /// lookup (which may use dirty cache info if available). If we do a lookup,
926 /// add the result to the cache.
927 MemDepResult MemoryDependenceAnalysis::
928 GetNonLocalInfoForBlock(Instruction *QueryInst,
929 const AliasAnalysis::Location &Loc,
930 bool isLoad, BasicBlock *BB,
931 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
933 // Do a binary search to see if we already have an entry for this block in
934 // the cache set. If so, find it.
935 NonLocalDepInfo::iterator Entry =
936 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
937 NonLocalDepEntry(BB));
938 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
941 NonLocalDepEntry *ExistingResult = nullptr;
942 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
943 ExistingResult = &*Entry;
945 // If we have a cached entry, and it is non-dirty, use it as the value for
947 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
948 ++NumCacheNonLocalPtr;
949 return ExistingResult->getResult();
952 // Otherwise, we have to scan for the value. If we have a dirty cache
953 // entry, start scanning from its position, otherwise we scan from the end
955 BasicBlock::iterator ScanPos = BB->end();
956 if (ExistingResult && ExistingResult->getResult().getInst()) {
957 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
958 "Instruction invalidated?");
959 ++NumCacheDirtyNonLocalPtr;
960 ScanPos = ExistingResult->getResult().getInst();
962 // Eliminating the dirty entry from 'Cache', so update the reverse info.
963 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
964 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
966 ++NumUncacheNonLocalPtr;
969 // Scan the block for the dependency.
970 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
973 // If we had a dirty entry for the block, update it. Otherwise, just add
976 ExistingResult->setResult(Dep);
978 Cache->push_back(NonLocalDepEntry(BB, Dep));
980 // If the block has a dependency (i.e. it isn't completely transparent to
981 // the value), remember the reverse association because we just added it
983 if (!Dep.isDef() && !Dep.isClobber())
986 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
987 // update MemDep when we remove instructions.
988 Instruction *Inst = Dep.getInst();
989 assert(Inst && "Didn't depend on anything?");
990 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
991 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
995 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
996 /// number of elements in the array that are already properly ordered. This is
997 /// optimized for the case when only a few entries are added.
999 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
1000 unsigned NumSortedEntries) {
1001 switch (Cache.size() - NumSortedEntries) {
1003 // done, no new entries.
1006 // Two new entries, insert the last one into place.
1007 NonLocalDepEntry Val = Cache.back();
1009 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1010 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1011 Cache.insert(Entry, Val);
1015 // One new entry, Just insert the new value at the appropriate position.
1016 if (Cache.size() != 1) {
1017 NonLocalDepEntry Val = Cache.back();
1019 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1020 std::upper_bound(Cache.begin(), Cache.end(), Val);
1021 Cache.insert(Entry, Val);
1025 // Added many values, do a full scale sort.
1026 std::sort(Cache.begin(), Cache.end());
1031 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1032 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1033 /// results to the results vector and keep track of which blocks are visited in
1036 /// This has special behavior for the first block queries (when SkipFirstBlock
1037 /// is true). In this special case, it ignores the contents of the specified
1038 /// block and starts returning dependence info for its predecessors.
1040 /// This function returns false on success, or true to indicate that it could
1041 /// not compute dependence information for some reason. This should be treated
1042 /// as a clobber dependence on the first instruction in the predecessor block.
1043 bool MemoryDependenceAnalysis::
1044 getNonLocalPointerDepFromBB(Instruction *QueryInst,
1045 const PHITransAddr &Pointer,
1046 const AliasAnalysis::Location &Loc,
1047 bool isLoad, BasicBlock *StartBB,
1048 SmallVectorImpl<NonLocalDepResult> &Result,
1049 DenseMap<BasicBlock*, Value*> &Visited,
1050 bool SkipFirstBlock) {
1051 // Look up the cached info for Pointer.
1052 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1054 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1055 // CacheKey, this value will be inserted as the associated value. Otherwise,
1056 // it'll be ignored, and we'll have to check to see if the cached size and
1057 // aa tags are consistent with the current query.
1058 NonLocalPointerInfo InitialNLPI;
1059 InitialNLPI.Size = Loc.Size;
1060 InitialNLPI.AATags = Loc.AATags;
1062 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1063 // already have one.
1064 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1065 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1066 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1068 // If we already have a cache entry for this CacheKey, we may need to do some
1069 // work to reconcile the cache entry and the current query.
1071 if (CacheInfo->Size < Loc.Size) {
1072 // The query's Size is greater than the cached one. Throw out the
1073 // cached data and proceed with the query at the greater size.
1074 CacheInfo->Pair = BBSkipFirstBlockPair();
1075 CacheInfo->Size = Loc.Size;
1076 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1077 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1078 if (Instruction *Inst = DI->getResult().getInst())
1079 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1080 CacheInfo->NonLocalDeps.clear();
1081 } else if (CacheInfo->Size > Loc.Size) {
1082 // This query's Size is less than the cached one. Conservatively restart
1083 // the query using the greater size.
1084 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1085 Loc.getWithNewSize(CacheInfo->Size),
1086 isLoad, StartBB, Result, Visited,
1090 // If the query's AATags are inconsistent with the cached one,
1091 // conservatively throw out the cached data and restart the query with
1092 // no tag if needed.
1093 if (CacheInfo->AATags != Loc.AATags) {
1094 if (CacheInfo->AATags) {
1095 CacheInfo->Pair = BBSkipFirstBlockPair();
1096 CacheInfo->AATags = AAMDNodes();
1097 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1098 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1099 if (Instruction *Inst = DI->getResult().getInst())
1100 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1101 CacheInfo->NonLocalDeps.clear();
1104 return getNonLocalPointerDepFromBB(QueryInst,
1105 Pointer, Loc.getWithoutAATags(),
1106 isLoad, StartBB, Result, Visited,
1111 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1113 // If we have valid cached information for exactly the block we are
1114 // investigating, just return it with no recomputation.
1115 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1116 // We have a fully cached result for this query then we can just return the
1117 // cached results and populate the visited set. However, we have to verify
1118 // that we don't already have conflicting results for these blocks. Check
1119 // to ensure that if a block in the results set is in the visited set that
1120 // it was for the same pointer query.
1121 if (!Visited.empty()) {
1122 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1124 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1125 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1128 // We have a pointer mismatch in a block. Just return clobber, saying
1129 // that something was clobbered in this result. We could also do a
1130 // non-fully cached query, but there is little point in doing this.
1135 Value *Addr = Pointer.getAddr();
1136 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1138 Visited.insert(std::make_pair(I->getBB(), Addr));
1139 if (I->getResult().isNonLocal()) {
1144 Result.push_back(NonLocalDepResult(I->getBB(),
1145 MemDepResult::getUnknown(),
1147 } else if (DT->isReachableFromEntry(I->getBB())) {
1148 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1151 ++NumCacheCompleteNonLocalPtr;
1155 // Otherwise, either this is a new block, a block with an invalid cache
1156 // pointer or one that we're about to invalidate by putting more info into it
1157 // than its valid cache info. If empty, the result will be valid cache info,
1158 // otherwise it isn't.
1160 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1162 CacheInfo->Pair = BBSkipFirstBlockPair();
1164 SmallVector<BasicBlock*, 32> Worklist;
1165 Worklist.push_back(StartBB);
1167 // PredList used inside loop.
1168 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1170 // Keep track of the entries that we know are sorted. Previously cached
1171 // entries will all be sorted. The entries we add we only sort on demand (we
1172 // don't insert every element into its sorted position). We know that we
1173 // won't get any reuse from currently inserted values, because we don't
1174 // revisit blocks after we insert info for them.
1175 unsigned NumSortedEntries = Cache->size();
1176 DEBUG(AssertSorted(*Cache));
1178 while (!Worklist.empty()) {
1179 BasicBlock *BB = Worklist.pop_back_val();
1181 // If we do process a large number of blocks it becomes very expensive and
1182 // likely it isn't worth worrying about
1183 if (Result.size() > NumResultsLimit) {
1185 // Sort it now (if needed) so that recursive invocations of
1186 // getNonLocalPointerDepFromBB and other routines that could reuse the
1187 // cache value will only see properly sorted cache arrays.
1188 if (Cache && NumSortedEntries != Cache->size()) {
1189 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1191 // Since we bail out, the "Cache" set won't contain all of the
1192 // results for the query. This is ok (we can still use it to accelerate
1193 // specific block queries) but we can't do the fastpath "return all
1194 // results from the set". Clear out the indicator for this.
1195 CacheInfo->Pair = BBSkipFirstBlockPair();
1199 // Skip the first block if we have it.
1200 if (!SkipFirstBlock) {
1201 // Analyze the dependency of *Pointer in FromBB. See if we already have
1203 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1205 // Get the dependency info for Pointer in BB. If we have cached
1206 // information, we will use it, otherwise we compute it.
1207 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1208 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1209 Loc, isLoad, BB, Cache,
1212 // If we got a Def or Clobber, add this to the list of results.
1213 if (!Dep.isNonLocal()) {
1215 Result.push_back(NonLocalDepResult(BB,
1216 MemDepResult::getUnknown(),
1217 Pointer.getAddr()));
1219 } else if (DT->isReachableFromEntry(BB)) {
1220 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1226 // If 'Pointer' is an instruction defined in this block, then we need to do
1227 // phi translation to change it into a value live in the predecessor block.
1228 // If not, we just add the predecessors to the worklist and scan them with
1229 // the same Pointer.
1230 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1231 SkipFirstBlock = false;
1232 SmallVector<BasicBlock*, 16> NewBlocks;
1233 for (BasicBlock *Pred : PredCache.get(BB)) {
1234 // Verify that we haven't looked at this block yet.
1235 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1236 InsertRes = Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1237 if (InsertRes.second) {
1238 // First time we've looked at *PI.
1239 NewBlocks.push_back(Pred);
1243 // If we have seen this block before, but it was with a different
1244 // pointer then we have a phi translation failure and we have to treat
1245 // this as a clobber.
1246 if (InsertRes.first->second != Pointer.getAddr()) {
1247 // Make sure to clean up the Visited map before continuing on to
1248 // PredTranslationFailure.
1249 for (unsigned i = 0; i < NewBlocks.size(); i++)
1250 Visited.erase(NewBlocks[i]);
1251 goto PredTranslationFailure;
1254 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1258 // We do need to do phi translation, if we know ahead of time we can't phi
1259 // translate this value, don't even try.
1260 if (!Pointer.IsPotentiallyPHITranslatable())
1261 goto PredTranslationFailure;
1263 // We may have added values to the cache list before this PHI translation.
1264 // If so, we haven't done anything to ensure that the cache remains sorted.
1265 // Sort it now (if needed) so that recursive invocations of
1266 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1267 // value will only see properly sorted cache arrays.
1268 if (Cache && NumSortedEntries != Cache->size()) {
1269 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1270 NumSortedEntries = Cache->size();
1275 for (BasicBlock *Pred : PredCache.get(BB)) {
1276 PredList.push_back(std::make_pair(Pred, Pointer));
1278 // Get the PHI translated pointer in this predecessor. This can fail if
1279 // not translatable, in which case the getAddr() returns null.
1280 PHITransAddr &PredPointer = PredList.back().second;
1281 PredPointer.PHITranslateValue(BB, Pred, DT, /*MustDominate=*/false);
1282 Value *PredPtrVal = PredPointer.getAddr();
1284 // Check to see if we have already visited this pred block with another
1285 // pointer. If so, we can't do this lookup. This failure can occur
1286 // with PHI translation when a critical edge exists and the PHI node in
1287 // the successor translates to a pointer value different than the
1288 // pointer the block was first analyzed with.
1289 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1290 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1292 if (!InsertRes.second) {
1293 // We found the pred; take it off the list of preds to visit.
1294 PredList.pop_back();
1296 // If the predecessor was visited with PredPtr, then we already did
1297 // the analysis and can ignore it.
1298 if (InsertRes.first->second == PredPtrVal)
1301 // Otherwise, the block was previously analyzed with a different
1302 // pointer. We can't represent the result of this case, so we just
1303 // treat this as a phi translation failure.
1305 // Make sure to clean up the Visited map before continuing on to
1306 // PredTranslationFailure.
1307 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1308 Visited.erase(PredList[i].first);
1310 goto PredTranslationFailure;
1314 // Actually process results here; this need to be a separate loop to avoid
1315 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1316 // any results for. (getNonLocalPointerDepFromBB will modify our
1317 // datastructures in ways the code after the PredTranslationFailure label
1319 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1320 BasicBlock *Pred = PredList[i].first;
1321 PHITransAddr &PredPointer = PredList[i].second;
1322 Value *PredPtrVal = PredPointer.getAddr();
1324 bool CanTranslate = true;
1325 // If PHI translation was unable to find an available pointer in this
1326 // predecessor, then we have to assume that the pointer is clobbered in
1327 // that predecessor. We can still do PRE of the load, which would insert
1328 // a computation of the pointer in this predecessor.
1330 CanTranslate = false;
1332 // FIXME: it is entirely possible that PHI translating will end up with
1333 // the same value. Consider PHI translating something like:
1334 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1335 // to recurse here, pedantically speaking.
1337 // If getNonLocalPointerDepFromBB fails here, that means the cached
1338 // result conflicted with the Visited list; we have to conservatively
1339 // assume it is unknown, but this also does not block PRE of the load.
1340 if (!CanTranslate ||
1341 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1342 Loc.getWithNewPtr(PredPtrVal),
1345 // Add the entry to the Result list.
1346 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1347 Result.push_back(Entry);
1349 // Since we had a phi translation failure, the cache for CacheKey won't
1350 // include all of the entries that we need to immediately satisfy future
1351 // queries. Mark this in NonLocalPointerDeps by setting the
1352 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1353 // cached value to do more work but not miss the phi trans failure.
1354 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1355 NLPI.Pair = BBSkipFirstBlockPair();
1360 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1361 CacheInfo = &NonLocalPointerDeps[CacheKey];
1362 Cache = &CacheInfo->NonLocalDeps;
1363 NumSortedEntries = Cache->size();
1365 // Since we did phi translation, the "Cache" set won't contain all of the
1366 // results for the query. This is ok (we can still use it to accelerate
1367 // specific block queries) but we can't do the fastpath "return all
1368 // results from the set" Clear out the indicator for this.
1369 CacheInfo->Pair = BBSkipFirstBlockPair();
1370 SkipFirstBlock = false;
1373 PredTranslationFailure:
1374 // The following code is "failure"; we can't produce a sane translation
1375 // for the given block. It assumes that we haven't modified any of
1376 // our datastructures while processing the current block.
1379 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1380 CacheInfo = &NonLocalPointerDeps[CacheKey];
1381 Cache = &CacheInfo->NonLocalDeps;
1382 NumSortedEntries = Cache->size();
1385 // Since we failed 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();
1391 // If *nothing* works, mark the pointer as unknown.
1393 // If this is the magic first block, return this as a clobber of the whole
1394 // incoming value. Since we can't phi translate to one of the predecessors,
1395 // we have to bail out.
1399 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1400 assert(I != Cache->rend() && "Didn't find current block??");
1401 if (I->getBB() != BB)
1404 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1405 "Should only be here with transparent block");
1406 I->setResult(MemDepResult::getUnknown());
1407 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1408 Pointer.getAddr()));
1413 // Okay, we're done now. If we added new values to the cache, re-sort it.
1414 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1415 DEBUG(AssertSorted(*Cache));
1419 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1420 /// CachedNonLocalPointerInfo, remove it.
1421 void MemoryDependenceAnalysis::
1422 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1423 CachedNonLocalPointerInfo::iterator It =
1424 NonLocalPointerDeps.find(P);
1425 if (It == NonLocalPointerDeps.end()) return;
1427 // Remove all of the entries in the BB->val map. This involves removing
1428 // instructions from the reverse map.
1429 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1431 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1432 Instruction *Target = PInfo[i].getResult().getInst();
1433 if (!Target) continue; // Ignore non-local dep results.
1434 assert(Target->getParent() == PInfo[i].getBB());
1436 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1437 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1440 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1441 NonLocalPointerDeps.erase(It);
1445 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1446 /// information about the specified pointer, because it may be too
1447 /// conservative in memdep. This is an optional call that can be used when
1448 /// the client detects an equivalence between the pointer and some other
1449 /// value and replaces the other value with ptr. This can make Ptr available
1450 /// in more places that cached info does not necessarily keep.
1451 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1452 // If Ptr isn't really a pointer, just ignore it.
1453 if (!Ptr->getType()->isPointerTy()) return;
1454 // Flush store info for the pointer.
1455 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1456 // Flush load info for the pointer.
1457 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1460 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1461 /// This needs to be done when the CFG changes, e.g., due to splitting
1463 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1467 /// removeInstruction - Remove an instruction from the dependence analysis,
1468 /// updating the dependence of instructions that previously depended on it.
1469 /// This method attempts to keep the cache coherent using the reverse map.
1470 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1471 // Walk through the Non-local dependencies, removing this one as the value
1472 // for any cached queries.
1473 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1474 if (NLDI != NonLocalDeps.end()) {
1475 NonLocalDepInfo &BlockMap = NLDI->second.first;
1476 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1478 if (Instruction *Inst = DI->getResult().getInst())
1479 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1480 NonLocalDeps.erase(NLDI);
1483 // If we have a cached local dependence query for this instruction, remove it.
1485 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1486 if (LocalDepEntry != LocalDeps.end()) {
1487 // Remove us from DepInst's reverse set now that the local dep info is gone.
1488 if (Instruction *Inst = LocalDepEntry->second.getInst())
1489 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1491 // Remove this local dependency info.
1492 LocalDeps.erase(LocalDepEntry);
1495 // If we have any cached pointer dependencies on this instruction, remove
1496 // them. If the instruction has non-pointer type, then it can't be a pointer
1499 // Remove it from both the load info and the store info. The instruction
1500 // can't be in either of these maps if it is non-pointer.
1501 if (RemInst->getType()->isPointerTy()) {
1502 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1503 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1506 // Loop over all of the things that depend on the instruction we're removing.
1508 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1510 // If we find RemInst as a clobber or Def in any of the maps for other values,
1511 // we need to replace its entry with a dirty version of the instruction after
1512 // it. If RemInst is a terminator, we use a null dirty value.
1514 // Using a dirty version of the instruction after RemInst saves having to scan
1515 // the entire block to get to this point.
1516 MemDepResult NewDirtyVal;
1517 if (!RemInst->isTerminator())
1518 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1520 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1521 if (ReverseDepIt != ReverseLocalDeps.end()) {
1522 // RemInst can't be the terminator if it has local stuff depending on it.
1523 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1524 "Nothing can locally depend on a terminator");
1526 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1527 assert(InstDependingOnRemInst != RemInst &&
1528 "Already removed our local dep info");
1530 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1532 // Make sure to remember that new things depend on NewDepInst.
1533 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1534 "a local dep on this if it is a terminator!");
1535 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1536 InstDependingOnRemInst));
1539 ReverseLocalDeps.erase(ReverseDepIt);
1541 // Add new reverse deps after scanning the set, to avoid invalidating the
1542 // 'ReverseDeps' reference.
1543 while (!ReverseDepsToAdd.empty()) {
1544 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1545 .insert(ReverseDepsToAdd.back().second);
1546 ReverseDepsToAdd.pop_back();
1550 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1551 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1552 for (Instruction *I : ReverseDepIt->second) {
1553 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1555 PerInstNLInfo &INLD = NonLocalDeps[I];
1556 // The information is now dirty!
1559 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1560 DE = INLD.first.end(); DI != DE; ++DI) {
1561 if (DI->getResult().getInst() != RemInst) continue;
1563 // Convert to a dirty entry for the subsequent instruction.
1564 DI->setResult(NewDirtyVal);
1566 if (Instruction *NextI = NewDirtyVal.getInst())
1567 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1571 ReverseNonLocalDeps.erase(ReverseDepIt);
1573 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1574 while (!ReverseDepsToAdd.empty()) {
1575 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1576 .insert(ReverseDepsToAdd.back().second);
1577 ReverseDepsToAdd.pop_back();
1581 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1582 // value in the NonLocalPointerDeps info.
1583 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1584 ReverseNonLocalPtrDeps.find(RemInst);
1585 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1586 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1588 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1589 assert(P.getPointer() != RemInst &&
1590 "Already removed NonLocalPointerDeps info for RemInst");
1592 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1594 // The cache is not valid for any specific block anymore.
1595 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1597 // Update any entries for RemInst to use the instruction after it.
1598 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1600 if (DI->getResult().getInst() != RemInst) continue;
1602 // Convert to a dirty entry for the subsequent instruction.
1603 DI->setResult(NewDirtyVal);
1605 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1606 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1609 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1610 // subsequent value may invalidate the sortedness.
1611 std::sort(NLPDI.begin(), NLPDI.end());
1614 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1616 while (!ReversePtrDepsToAdd.empty()) {
1617 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1618 .insert(ReversePtrDepsToAdd.back().second);
1619 ReversePtrDepsToAdd.pop_back();
1624 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1625 AA->deleteValue(RemInst);
1626 DEBUG(verifyRemoved(RemInst));
1628 /// verifyRemoved - Verify that the specified instruction does not occur
1629 /// in our internal data structures. This function verifies by asserting in
1631 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1633 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1634 E = LocalDeps.end(); I != E; ++I) {
1635 assert(I->first != D && "Inst occurs in data structures");
1636 assert(I->second.getInst() != D &&
1637 "Inst occurs in data structures");
1640 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1641 E = NonLocalPointerDeps.end(); I != E; ++I) {
1642 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1643 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1644 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1646 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1649 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1650 E = NonLocalDeps.end(); I != E; ++I) {
1651 assert(I->first != D && "Inst occurs in data structures");
1652 const PerInstNLInfo &INLD = I->second;
1653 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1654 EE = INLD.first.end(); II != EE; ++II)
1655 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1658 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1659 E = ReverseLocalDeps.end(); I != E; ++I) {
1660 assert(I->first != D && "Inst occurs in data structures");
1661 for (Instruction *Inst : I->second)
1662 assert(Inst != D && "Inst occurs in data structures");
1665 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1666 E = ReverseNonLocalDeps.end();
1668 assert(I->first != D && "Inst occurs in data structures");
1669 for (Instruction *Inst : I->second)
1670 assert(Inst != D && "Inst occurs in data structures");
1673 for (ReverseNonLocalPtrDepTy::const_iterator
1674 I = ReverseNonLocalPtrDeps.begin(),
1675 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1676 assert(I->first != D && "Inst occurs in rev NLPD map");
1678 for (ValueIsLoadPair P : I->second)
1679 assert(P != ValueIsLoadPair(D, false) &&
1680 P != ValueIsLoadPair(D, true) &&
1681 "Inst occurs in ReverseNonLocalPtrDeps map");