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
121 static AliasAnalysis::ModRefResult
122 GetLocation(const Instruction *Inst, MemoryLocation &Loc, AliasAnalysis *AA) {
123 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
124 if (LI->isUnordered()) {
125 Loc = MemoryLocation::get(LI);
126 return AliasAnalysis::Ref;
128 if (LI->getOrdering() == Monotonic) {
129 Loc = MemoryLocation::get(LI);
130 return AliasAnalysis::ModRef;
132 Loc = MemoryLocation();
133 return AliasAnalysis::ModRef;
136 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
137 if (SI->isUnordered()) {
138 Loc = MemoryLocation::get(SI);
139 return AliasAnalysis::Mod;
141 if (SI->getOrdering() == Monotonic) {
142 Loc = MemoryLocation::get(SI);
143 return AliasAnalysis::ModRef;
145 Loc = MemoryLocation();
146 return AliasAnalysis::ModRef;
149 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
150 Loc = MemoryLocation::get(V);
151 return AliasAnalysis::ModRef;
154 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
155 // calls to free() deallocate the entire structure
156 Loc = MemoryLocation(CI->getArgOperand(0));
157 return AliasAnalysis::Mod;
160 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
163 switch (II->getIntrinsicID()) {
164 case Intrinsic::lifetime_start:
165 case Intrinsic::lifetime_end:
166 case Intrinsic::invariant_start:
167 II->getAAMetadata(AAInfo);
168 Loc = MemoryLocation(
169 II->getArgOperand(1),
170 cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(), AAInfo);
171 // These intrinsics don't really modify the memory, but returning Mod
172 // will allow them to be handled conservatively.
173 return AliasAnalysis::Mod;
174 case Intrinsic::invariant_end:
175 II->getAAMetadata(AAInfo);
176 Loc = MemoryLocation(
177 II->getArgOperand(2),
178 cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(), AAInfo);
179 // These intrinsics don't really modify the memory, but returning Mod
180 // will allow them to be handled conservatively.
181 return AliasAnalysis::Mod;
187 // Otherwise, just do the coarse-grained thing that always works.
188 if (Inst->mayWriteToMemory())
189 return AliasAnalysis::ModRef;
190 if (Inst->mayReadFromMemory())
191 return AliasAnalysis::Ref;
192 return AliasAnalysis::NoModRef;
195 /// getCallSiteDependencyFrom - Private helper for finding the local
196 /// dependencies of a call site.
197 MemDepResult MemoryDependenceAnalysis::
198 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
199 BasicBlock::iterator ScanIt, BasicBlock *BB) {
200 unsigned Limit = BlockScanLimit;
202 // Walk backwards through the block, looking for dependencies
203 while (ScanIt != BB->begin()) {
204 // Limit the amount of scanning we do so we don't end up with quadratic
205 // running time on extreme testcases.
208 return MemDepResult::getUnknown();
210 Instruction *Inst = --ScanIt;
212 // If this inst is a memory op, get the pointer it accessed
214 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
216 // A simple instruction.
217 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
218 return MemDepResult::getClobber(Inst);
222 if (auto InstCS = CallSite(Inst)) {
223 // Debug intrinsics don't cause dependences.
224 if (isa<DbgInfoIntrinsic>(Inst)) continue;
225 // If these two calls do not interfere, look past it.
226 switch (AA->getModRefInfo(CS, InstCS)) {
227 case AliasAnalysis::NoModRef:
228 // If the two calls are the same, return InstCS as a Def, so that
229 // CS can be found redundant and eliminated.
230 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
231 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
232 return MemDepResult::getDef(Inst);
234 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
238 return MemDepResult::getClobber(Inst);
242 // If we could not obtain a pointer for the instruction and the instruction
243 // touches memory then assume that this is a dependency.
244 if (MR != AliasAnalysis::NoModRef)
245 return MemDepResult::getClobber(Inst);
248 // No dependence found. If this is the entry block of the function, it is
249 // unknown, otherwise it is non-local.
250 if (BB != &BB->getParent()->getEntryBlock())
251 return MemDepResult::getNonLocal();
252 return MemDepResult::getNonFuncLocal();
255 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
256 /// would fully overlap MemLoc if done as a wider legal integer load.
258 /// MemLocBase, MemLocOffset are lazily computed here the first time the
259 /// base/offs of memloc is needed.
260 static bool isLoadLoadClobberIfExtendedToFullWidth(const MemoryLocation &MemLoc,
261 const Value *&MemLocBase,
263 const LoadInst *LI) {
264 const DataLayout &DL = LI->getModule()->getDataLayout();
266 // If we haven't already computed the base/offset of MemLoc, do so now.
268 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
270 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
271 MemLocBase, MemLocOffs, MemLoc.Size, LI);
275 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
276 /// looks at a memory location for a load (specified by MemLocBase, Offs,
277 /// and Size) and compares it against a load. If the specified load could
278 /// be safely widened to a larger integer load that is 1) still efficient,
279 /// 2) safe for the target, and 3) would provide the specified memory
280 /// location value, then this function returns the size in bytes of the
281 /// load width to use. If not, this returns zero.
282 unsigned MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
283 const Value *MemLocBase, int64_t MemLocOffs, unsigned MemLocSize,
284 const LoadInst *LI) {
285 // We can only extend simple integer loads.
286 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
288 // Load widening is hostile to ThreadSanitizer: it may cause false positives
289 // or make the reports more cryptic (access sizes are wrong).
290 if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeThread))
293 const DataLayout &DL = LI->getModule()->getDataLayout();
295 // Get the base of this load.
297 const Value *LIBase =
298 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, DL);
300 // If the two pointers are not based on the same pointer, we can't tell that
302 if (LIBase != MemLocBase) return 0;
304 // Okay, the two values are based on the same pointer, but returned as
305 // no-alias. This happens when we have things like two byte loads at "P+1"
306 // and "P+3". Check to see if increasing the size of the "LI" load up to its
307 // alignment (or the largest native integer type) will allow us to load all
308 // the bits required by MemLoc.
310 // If MemLoc is before LI, then no widening of LI will help us out.
311 if (MemLocOffs < LIOffs) return 0;
313 // Get the alignment of the load in bytes. We assume that it is safe to load
314 // any legal integer up to this size without a problem. For example, if we're
315 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
316 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
318 unsigned LoadAlign = LI->getAlignment();
320 int64_t MemLocEnd = MemLocOffs+MemLocSize;
322 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
323 if (LIOffs+LoadAlign < MemLocEnd) return 0;
325 // This is the size of the load to try. Start with the next larger power of
327 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
328 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
331 // If this load size is bigger than our known alignment or would not fit
332 // into a native integer register, then we fail.
333 if (NewLoadByteSize > LoadAlign ||
334 !DL.fitsInLegalInteger(NewLoadByteSize*8))
337 if (LIOffs + NewLoadByteSize > MemLocEnd &&
338 LI->getParent()->getParent()->hasFnAttribute(
339 Attribute::SanitizeAddress))
340 // We will be reading past the location accessed by the original program.
341 // While this is safe in a regular build, Address Safety analysis tools
342 // may start reporting false warnings. So, don't do widening.
345 // If a load of this width would include all of MemLoc, then we succeed.
346 if (LIOffs+NewLoadByteSize >= MemLocEnd)
347 return NewLoadByteSize;
349 NewLoadByteSize <<= 1;
353 static bool isVolatile(Instruction *Inst) {
354 if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
355 return LI->isVolatile();
356 else if (StoreInst *SI = dyn_cast<StoreInst>(Inst))
357 return SI->isVolatile();
358 else if (AtomicCmpXchgInst *AI = dyn_cast<AtomicCmpXchgInst>(Inst))
359 return AI->isVolatile();
364 /// getPointerDependencyFrom - Return the instruction on which a memory
365 /// location depends. If isLoad is true, this routine ignores may-aliases with
366 /// read-only operations. If isLoad is false, this routine ignores may-aliases
367 /// with reads from read-only locations. If possible, pass the query
368 /// instruction as well; this function may take advantage of the metadata
369 /// annotated to the query instruction to refine the result.
370 MemDepResult MemoryDependenceAnalysis::getPointerDependencyFrom(
371 const MemoryLocation &MemLoc, bool isLoad, BasicBlock::iterator ScanIt,
372 BasicBlock *BB, Instruction *QueryInst) {
374 const Value *MemLocBase = nullptr;
375 int64_t MemLocOffset = 0;
376 unsigned Limit = BlockScanLimit;
377 bool isInvariantLoad = false;
379 // We must be careful with atomic accesses, as they may allow another thread
380 // to touch this location, cloberring it. We are conservative: if the
381 // QueryInst is not a simple (non-atomic) memory access, we automatically
382 // return getClobber.
383 // If it is simple, we know based on the results of
384 // "Compiler testing via a theory of sound optimisations in the C11/C++11
385 // memory model" in PLDI 2013, that a non-atomic location can only be
386 // clobbered between a pair of a release and an acquire action, with no
387 // access to the location in between.
388 // Here is an example for giving the general intuition behind this rule.
389 // In the following code:
391 // release action; [1]
392 // acquire action; [4]
394 // It is unsafe to replace %val by 0 because another thread may be running:
395 // acquire action; [2]
397 // release action; [3]
398 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
399 // being 42. A key property of this program however is that if either
400 // 1 or 4 were missing, there would be a race between the store of 42
401 // either the store of 0 or the load (making the whole progam racy).
402 // The paper mentionned above shows that the same property is respected
403 // by every program that can detect any optimisation of that kind: either
404 // it is racy (undefined) or there is a release followed by an acquire
405 // between the pair of accesses under consideration.
407 // If the load is invariant, we "know" that it doesn't alias *any* write. We
408 // do want to respect mustalias results since defs are useful for value
409 // forwarding, but any mayalias write can be assumed to be noalias.
410 // Arguably, this logic should be pushed inside AliasAnalysis itself.
411 if (isLoad && QueryInst) {
412 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
413 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
414 isInvariantLoad = true;
417 const DataLayout &DL = BB->getModule()->getDataLayout();
419 // Walk backwards through the basic block, looking for dependencies.
420 while (ScanIt != BB->begin()) {
421 Instruction *Inst = --ScanIt;
423 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
424 // Debug intrinsics don't (and can't) cause dependencies.
425 if (isa<DbgInfoIntrinsic>(II)) continue;
427 // Limit the amount of scanning we do so we don't end up with quadratic
428 // running time on extreme testcases.
431 return MemDepResult::getUnknown();
433 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
434 // If we reach a lifetime begin or end marker, then the query ends here
435 // because the value is undefined.
436 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
437 // FIXME: This only considers queries directly on the invariant-tagged
438 // pointer, not on query pointers that are indexed off of them. It'd
439 // be nice to handle that at some point (the right approach is to use
440 // GetPointerBaseWithConstantOffset).
441 if (AA->isMustAlias(MemoryLocation(II->getArgOperand(1)), MemLoc))
442 return MemDepResult::getDef(II);
447 // Values depend on loads if the pointers are must aliased. This means that
448 // a load depends on another must aliased load from the same value.
449 // One exception is atomic loads: a value can depend on an atomic load that it
450 // does not alias with when this atomic load indicates that another thread may
451 // be accessing the location.
452 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
454 // While volatile access cannot be eliminated, they do not have to clobber
455 // non-aliasing locations, as normal accesses, for example, can be safely
456 // reordered with volatile accesses.
457 if (LI->isVolatile()) {
459 // Original QueryInst *may* be volatile
460 return MemDepResult::getClobber(LI);
461 if (isVolatile(QueryInst))
462 // Ordering required if QueryInst is itself volatile
463 return MemDepResult::getClobber(LI);
464 // Otherwise, volatile doesn't imply any special ordering
467 // Atomic loads have complications involved.
468 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
469 // FIXME: This is overly conservative.
470 if (LI->isAtomic() && LI->getOrdering() > Unordered) {
472 return MemDepResult::getClobber(LI);
473 if (LI->getOrdering() != Monotonic)
474 return MemDepResult::getClobber(LI);
475 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
476 if (!QueryLI->isSimple())
477 return MemDepResult::getClobber(LI);
478 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
479 if (!QuerySI->isSimple())
480 return MemDepResult::getClobber(LI);
481 } else if (QueryInst->mayReadOrWriteMemory()) {
482 return MemDepResult::getClobber(LI);
486 MemoryLocation LoadLoc = MemoryLocation::get(LI);
488 // If we found a pointer, check if it could be the same as our pointer.
489 AliasResult R = AA->alias(LoadLoc, MemLoc);
493 // If this is an over-aligned integer load (for example,
494 // "load i8* %P, align 4") see if it would obviously overlap with the
495 // queried location if widened to a larger load (e.g. if the queried
496 // location is 1 byte at P+1). If so, return it as a load/load
497 // clobber result, allowing the client to decide to widen the load if
499 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType())) {
500 if (LI->getAlignment() * 8 > ITy->getPrimitiveSizeInBits() &&
501 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
503 return MemDepResult::getClobber(Inst);
508 // Must aliased loads are defs of each other.
510 return MemDepResult::getDef(Inst);
512 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
513 // in terms of clobbering loads, but since it does this by looking
514 // at the clobbering load directly, it doesn't know about any
515 // phi translation that may have happened along the way.
517 // If we have a partial alias, then return this as a clobber for the
519 if (R == PartialAlias)
520 return MemDepResult::getClobber(Inst);
523 // Random may-alias loads don't depend on each other without a
528 // Stores don't depend on other no-aliased accesses.
532 // Stores don't alias loads from read-only memory.
533 if (AA->pointsToConstantMemory(LoadLoc))
536 // Stores depend on may/must aliased loads.
537 return MemDepResult::getDef(Inst);
540 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
541 // Atomic stores have complications involved.
542 // A Monotonic store is OK if the query inst is itself not atomic.
543 // FIXME: This is overly conservative.
544 if (!SI->isUnordered()) {
546 return MemDepResult::getClobber(SI);
547 if (SI->getOrdering() != Monotonic)
548 return MemDepResult::getClobber(SI);
549 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
550 if (!QueryLI->isSimple())
551 return MemDepResult::getClobber(SI);
552 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
553 if (!QuerySI->isSimple())
554 return MemDepResult::getClobber(SI);
555 } else if (QueryInst->mayReadOrWriteMemory()) {
556 return MemDepResult::getClobber(SI);
560 // FIXME: this is overly conservative.
561 // While volatile access cannot be eliminated, they do not have to clobber
562 // non-aliasing locations, as normal accesses can for example be reordered
563 // with volatile accesses.
564 if (SI->isVolatile())
565 return MemDepResult::getClobber(SI);
567 // If alias analysis can tell that this store is guaranteed to not modify
568 // the query pointer, ignore it. Use getModRefInfo to handle cases where
569 // the query pointer points to constant memory etc.
570 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
573 // Ok, this store might clobber the query pointer. Check to see if it is
574 // a must alias: in this case, we want to return this as a def.
575 MemoryLocation StoreLoc = MemoryLocation::get(SI);
577 // If we found a pointer, check if it could be the same as our pointer.
578 AliasResult R = AA->alias(StoreLoc, MemLoc);
583 return MemDepResult::getDef(Inst);
586 return MemDepResult::getClobber(Inst);
589 // If this is an allocation, and if we know that the accessed pointer is to
590 // the allocation, return Def. This means that there is no dependence and
591 // the access can be optimized based on that. For example, a load could
593 // Note: Only determine this to be a malloc if Inst is the malloc call, not
594 // a subsequent bitcast of the malloc call result. There can be stores to
595 // the malloced memory between the malloc call and its bitcast uses, and we
596 // need to continue scanning until the malloc call.
597 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
598 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
599 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
601 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
602 return MemDepResult::getDef(Inst);
605 // Be conservative if the accessed pointer may alias the allocation.
606 if (AA->alias(Inst, AccessPtr) != NoAlias)
607 return MemDepResult::getClobber(Inst);
608 // If the allocation is not aliased and does not read memory (like
609 // strdup), it is safe to ignore.
610 if (isa<AllocaInst>(Inst) ||
611 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
618 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
619 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
620 // If necessary, perform additional analysis.
621 if (MR == AliasAnalysis::ModRef)
622 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
624 case AliasAnalysis::NoModRef:
625 // If the call has no effect on the queried pointer, just ignore it.
627 case AliasAnalysis::Mod:
628 return MemDepResult::getClobber(Inst);
629 case AliasAnalysis::Ref:
630 // If the call is known to never store to the pointer, and if this is a
631 // load query, we can safely ignore it (scan past it).
635 // Otherwise, there is a potential dependence. Return a clobber.
636 return MemDepResult::getClobber(Inst);
640 // No dependence found. If this is the entry block of the function, it is
641 // unknown, otherwise it is non-local.
642 if (BB != &BB->getParent()->getEntryBlock())
643 return MemDepResult::getNonLocal();
644 return MemDepResult::getNonFuncLocal();
647 /// getDependency - Return the instruction on which a memory operation
649 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
650 Instruction *ScanPos = QueryInst;
652 // Check for a cached result
653 MemDepResult &LocalCache = LocalDeps[QueryInst];
655 // If the cached entry is non-dirty, just return it. Note that this depends
656 // on MemDepResult's default constructing to 'dirty'.
657 if (!LocalCache.isDirty())
660 // Otherwise, if we have a dirty entry, we know we can start the scan at that
661 // instruction, which may save us some work.
662 if (Instruction *Inst = LocalCache.getInst()) {
665 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
668 BasicBlock *QueryParent = QueryInst->getParent();
671 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
672 // No dependence found. If this is the entry block of the function, it is
673 // unknown, otherwise it is non-local.
674 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
675 LocalCache = MemDepResult::getNonLocal();
677 LocalCache = MemDepResult::getNonFuncLocal();
679 MemoryLocation MemLoc;
680 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
682 // If we can do a pointer scan, make it happen.
683 bool isLoad = !(MR & AliasAnalysis::Mod);
684 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
685 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
687 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
688 QueryParent, QueryInst);
689 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
690 CallSite QueryCS(QueryInst);
691 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
692 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
695 // Non-memory instruction.
696 LocalCache = MemDepResult::getUnknown();
699 // Remember the result!
700 if (Instruction *I = LocalCache.getInst())
701 ReverseLocalDeps[I].insert(QueryInst);
707 /// AssertSorted - This method is used when -debug is specified to verify that
708 /// cache arrays are properly kept sorted.
709 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
711 if (Count == -1) Count = Cache.size();
712 if (Count == 0) return;
714 for (unsigned i = 1; i != unsigned(Count); ++i)
715 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
719 /// getNonLocalCallDependency - Perform a full dependency query for the
720 /// specified call, returning the set of blocks that the value is
721 /// potentially live across. The returned set of results will include a
722 /// "NonLocal" result for all blocks where the value is live across.
724 /// This method assumes the instruction returns a "NonLocal" dependency
725 /// within its own block.
727 /// This returns a reference to an internal data structure that may be
728 /// invalidated on the next non-local query or when an instruction is
729 /// removed. Clients must copy this data if they want it around longer than
731 const MemoryDependenceAnalysis::NonLocalDepInfo &
732 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
733 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
734 "getNonLocalCallDependency should only be used on calls with non-local deps!");
735 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
736 NonLocalDepInfo &Cache = CacheP.first;
738 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
739 /// the cached case, this can happen due to instructions being deleted etc. In
740 /// the uncached case, this starts out as the set of predecessors we care
742 SmallVector<BasicBlock*, 32> DirtyBlocks;
744 if (!Cache.empty()) {
745 // Okay, we have a cache entry. If we know it is not dirty, just return it
746 // with no computation.
747 if (!CacheP.second) {
752 // If we already have a partially computed set of results, scan them to
753 // determine what is dirty, seeding our initial DirtyBlocks worklist.
754 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
756 if (I->getResult().isDirty())
757 DirtyBlocks.push_back(I->getBB());
759 // Sort the cache so that we can do fast binary search lookups below.
760 std::sort(Cache.begin(), Cache.end());
762 ++NumCacheDirtyNonLocal;
763 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
764 // << Cache.size() << " cached: " << *QueryInst;
766 // Seed DirtyBlocks with each of the preds of QueryInst's block.
767 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
768 for (BasicBlock *Pred : PredCache.get(QueryBB))
769 DirtyBlocks.push_back(Pred);
770 ++NumUncacheNonLocal;
773 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
774 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
776 SmallPtrSet<BasicBlock*, 64> Visited;
778 unsigned NumSortedEntries = Cache.size();
779 DEBUG(AssertSorted(Cache));
781 // Iterate while we still have blocks to update.
782 while (!DirtyBlocks.empty()) {
783 BasicBlock *DirtyBB = DirtyBlocks.back();
784 DirtyBlocks.pop_back();
786 // Already processed this block?
787 if (!Visited.insert(DirtyBB).second)
790 // Do a binary search to see if we already have an entry for this block in
791 // the cache set. If so, find it.
792 DEBUG(AssertSorted(Cache, NumSortedEntries));
793 NonLocalDepInfo::iterator Entry =
794 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
795 NonLocalDepEntry(DirtyBB));
796 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
799 NonLocalDepEntry *ExistingResult = nullptr;
800 if (Entry != Cache.begin()+NumSortedEntries &&
801 Entry->getBB() == DirtyBB) {
802 // If we already have an entry, and if it isn't already dirty, the block
804 if (!Entry->getResult().isDirty())
807 // Otherwise, remember this slot so we can update the value.
808 ExistingResult = &*Entry;
811 // If the dirty entry has a pointer, start scanning from it so we don't have
812 // to rescan the entire block.
813 BasicBlock::iterator ScanPos = DirtyBB->end();
814 if (ExistingResult) {
815 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
817 // We're removing QueryInst's use of Inst.
818 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
819 QueryCS.getInstruction());
823 // Find out if this block has a local dependency for QueryInst.
826 if (ScanPos != DirtyBB->begin()) {
827 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
828 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
829 // No dependence found. If this is the entry block of the function, it is
830 // a clobber, otherwise it is unknown.
831 Dep = MemDepResult::getNonLocal();
833 Dep = MemDepResult::getNonFuncLocal();
836 // If we had a dirty entry for the block, update it. Otherwise, just add
839 ExistingResult->setResult(Dep);
841 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
843 // If the block has a dependency (i.e. it isn't completely transparent to
844 // the value), remember the association!
845 if (!Dep.isNonLocal()) {
846 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
847 // update this when we remove instructions.
848 if (Instruction *Inst = Dep.getInst())
849 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
852 // If the block *is* completely transparent to the load, we need to check
853 // the predecessors of this block. Add them to our worklist.
854 for (BasicBlock *Pred : PredCache.get(DirtyBB))
855 DirtyBlocks.push_back(Pred);
862 /// getNonLocalPointerDependency - Perform a full dependency query for an
863 /// access to the specified (non-volatile) memory location, returning the
864 /// set of instructions that either define or clobber the value.
866 /// This method assumes the pointer has a "NonLocal" dependency within its
869 void MemoryDependenceAnalysis::
870 getNonLocalPointerDependency(Instruction *QueryInst,
871 SmallVectorImpl<NonLocalDepResult> &Result) {
872 const MemoryLocation Loc = MemoryLocation::get(QueryInst);
873 bool isLoad = isa<LoadInst>(QueryInst);
874 BasicBlock *FromBB = QueryInst->getParent();
877 assert(Loc.Ptr->getType()->isPointerTy() &&
878 "Can't get pointer deps of a non-pointer!");
881 // This routine does not expect to deal with volatile instructions.
882 // Doing so would require piping through the QueryInst all the way through.
883 // TODO: volatiles can't be elided, but they can be reordered with other
884 // non-volatile accesses.
886 // We currently give up on any instruction which is ordered, but we do handle
887 // atomic instructions which are unordered.
888 // TODO: Handle ordered instructions
889 auto isOrdered = [](Instruction *Inst) {
890 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
891 return !LI->isUnordered();
892 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
893 return !SI->isUnordered();
897 if (isVolatile(QueryInst) || isOrdered(QueryInst)) {
898 Result.push_back(NonLocalDepResult(FromBB,
899 MemDepResult::getUnknown(),
900 const_cast<Value *>(Loc.Ptr)));
903 const DataLayout &DL = FromBB->getModule()->getDataLayout();
904 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AC);
906 // This is the set of blocks we've inspected, and the pointer we consider in
907 // each block. Because of critical edges, we currently bail out if querying
908 // a block with multiple different pointers. This can happen during PHI
910 DenseMap<BasicBlock*, Value*> Visited;
911 if (!getNonLocalPointerDepFromBB(QueryInst, Address, Loc, isLoad, FromBB,
912 Result, Visited, true))
915 Result.push_back(NonLocalDepResult(FromBB,
916 MemDepResult::getUnknown(),
917 const_cast<Value *>(Loc.Ptr)));
920 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
921 /// Pointer/PointeeSize using either cached information in Cache or by doing a
922 /// lookup (which may use dirty cache info if available). If we do a lookup,
923 /// add the result to the cache.
924 MemDepResult MemoryDependenceAnalysis::GetNonLocalInfoForBlock(
925 Instruction *QueryInst, const MemoryLocation &Loc, bool isLoad,
926 BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
928 // Do a binary search to see if we already have an entry for this block in
929 // the cache set. If so, find it.
930 NonLocalDepInfo::iterator Entry =
931 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
932 NonLocalDepEntry(BB));
933 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
936 NonLocalDepEntry *ExistingResult = nullptr;
937 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
938 ExistingResult = &*Entry;
940 // If we have a cached entry, and it is non-dirty, use it as the value for
942 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
943 ++NumCacheNonLocalPtr;
944 return ExistingResult->getResult();
947 // Otherwise, we have to scan for the value. If we have a dirty cache
948 // entry, start scanning from its position, otherwise we scan from the end
950 BasicBlock::iterator ScanPos = BB->end();
951 if (ExistingResult && ExistingResult->getResult().getInst()) {
952 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
953 "Instruction invalidated?");
954 ++NumCacheDirtyNonLocalPtr;
955 ScanPos = ExistingResult->getResult().getInst();
957 // Eliminating the dirty entry from 'Cache', so update the reverse info.
958 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
959 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
961 ++NumUncacheNonLocalPtr;
964 // Scan the block for the dependency.
965 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB,
968 // If we had a dirty entry for the block, update it. Otherwise, just add
971 ExistingResult->setResult(Dep);
973 Cache->push_back(NonLocalDepEntry(BB, Dep));
975 // If the block has a dependency (i.e. it isn't completely transparent to
976 // the value), remember the reverse association because we just added it
978 if (!Dep.isDef() && !Dep.isClobber())
981 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
982 // update MemDep when we remove instructions.
983 Instruction *Inst = Dep.getInst();
984 assert(Inst && "Didn't depend on anything?");
985 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
986 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
990 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
991 /// number of elements in the array that are already properly ordered. This is
992 /// optimized for the case when only a few entries are added.
994 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
995 unsigned NumSortedEntries) {
996 switch (Cache.size() - NumSortedEntries) {
998 // done, no new entries.
1001 // Two new entries, insert the last one into place.
1002 NonLocalDepEntry Val = Cache.back();
1004 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1005 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
1006 Cache.insert(Entry, Val);
1010 // One new entry, Just insert the new value at the appropriate position.
1011 if (Cache.size() != 1) {
1012 NonLocalDepEntry Val = Cache.back();
1014 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
1015 std::upper_bound(Cache.begin(), Cache.end(), Val);
1016 Cache.insert(Entry, Val);
1020 // Added many values, do a full scale sort.
1021 std::sort(Cache.begin(), Cache.end());
1026 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
1027 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
1028 /// results to the results vector and keep track of which blocks are visited in
1031 /// This has special behavior for the first block queries (when SkipFirstBlock
1032 /// is true). In this special case, it ignores the contents of the specified
1033 /// block and starts returning dependence info for its predecessors.
1035 /// This function returns false on success, or true to indicate that it could
1036 /// not compute dependence information for some reason. This should be treated
1037 /// as a clobber dependence on the first instruction in the predecessor block.
1038 bool MemoryDependenceAnalysis::getNonLocalPointerDepFromBB(
1039 Instruction *QueryInst, const PHITransAddr &Pointer,
1040 const MemoryLocation &Loc, bool isLoad, BasicBlock *StartBB,
1041 SmallVectorImpl<NonLocalDepResult> &Result,
1042 DenseMap<BasicBlock *, Value *> &Visited, bool SkipFirstBlock) {
1043 // Look up the cached info for Pointer.
1044 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1046 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1047 // CacheKey, this value will be inserted as the associated value. Otherwise,
1048 // it'll be ignored, and we'll have to check to see if the cached size and
1049 // aa tags are consistent with the current query.
1050 NonLocalPointerInfo InitialNLPI;
1051 InitialNLPI.Size = Loc.Size;
1052 InitialNLPI.AATags = Loc.AATags;
1054 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1055 // already have one.
1056 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1057 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1058 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1060 // If we already have a cache entry for this CacheKey, we may need to do some
1061 // work to reconcile the cache entry and the current query.
1063 if (CacheInfo->Size < Loc.Size) {
1064 // The query's Size is greater than the cached one. Throw out the
1065 // cached data and proceed with the query at the greater size.
1066 CacheInfo->Pair = BBSkipFirstBlockPair();
1067 CacheInfo->Size = Loc.Size;
1068 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1069 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1070 if (Instruction *Inst = DI->getResult().getInst())
1071 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1072 CacheInfo->NonLocalDeps.clear();
1073 } else if (CacheInfo->Size > Loc.Size) {
1074 // This query's Size is less than the cached one. Conservatively restart
1075 // the query using the greater size.
1076 return getNonLocalPointerDepFromBB(QueryInst, Pointer,
1077 Loc.getWithNewSize(CacheInfo->Size),
1078 isLoad, StartBB, Result, Visited,
1082 // If the query's AATags are inconsistent with the cached one,
1083 // conservatively throw out the cached data and restart the query with
1084 // no tag if needed.
1085 if (CacheInfo->AATags != Loc.AATags) {
1086 if (CacheInfo->AATags) {
1087 CacheInfo->Pair = BBSkipFirstBlockPair();
1088 CacheInfo->AATags = AAMDNodes();
1089 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1090 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1091 if (Instruction *Inst = DI->getResult().getInst())
1092 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1093 CacheInfo->NonLocalDeps.clear();
1096 return getNonLocalPointerDepFromBB(QueryInst,
1097 Pointer, Loc.getWithoutAATags(),
1098 isLoad, StartBB, Result, Visited,
1103 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1105 // If we have valid cached information for exactly the block we are
1106 // investigating, just return it with no recomputation.
1107 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1108 // We have a fully cached result for this query then we can just return the
1109 // cached results and populate the visited set. However, we have to verify
1110 // that we don't already have conflicting results for these blocks. Check
1111 // to ensure that if a block in the results set is in the visited set that
1112 // it was for the same pointer query.
1113 if (!Visited.empty()) {
1114 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1116 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1117 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1120 // We have a pointer mismatch in a block. Just return clobber, saying
1121 // that something was clobbered in this result. We could also do a
1122 // non-fully cached query, but there is little point in doing this.
1127 Value *Addr = Pointer.getAddr();
1128 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1130 Visited.insert(std::make_pair(I->getBB(), Addr));
1131 if (I->getResult().isNonLocal()) {
1136 Result.push_back(NonLocalDepResult(I->getBB(),
1137 MemDepResult::getUnknown(),
1139 } else if (DT->isReachableFromEntry(I->getBB())) {
1140 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1143 ++NumCacheCompleteNonLocalPtr;
1147 // Otherwise, either this is a new block, a block with an invalid cache
1148 // pointer or one that we're about to invalidate by putting more info into it
1149 // than its valid cache info. If empty, the result will be valid cache info,
1150 // otherwise it isn't.
1152 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1154 CacheInfo->Pair = BBSkipFirstBlockPair();
1156 SmallVector<BasicBlock*, 32> Worklist;
1157 Worklist.push_back(StartBB);
1159 // PredList used inside loop.
1160 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1162 // Keep track of the entries that we know are sorted. Previously cached
1163 // entries will all be sorted. The entries we add we only sort on demand (we
1164 // don't insert every element into its sorted position). We know that we
1165 // won't get any reuse from currently inserted values, because we don't
1166 // revisit blocks after we insert info for them.
1167 unsigned NumSortedEntries = Cache->size();
1168 DEBUG(AssertSorted(*Cache));
1170 while (!Worklist.empty()) {
1171 BasicBlock *BB = Worklist.pop_back_val();
1173 // If we do process a large number of blocks it becomes very expensive and
1174 // likely it isn't worth worrying about
1175 if (Result.size() > NumResultsLimit) {
1177 // Sort it now (if needed) so that recursive invocations of
1178 // getNonLocalPointerDepFromBB and other routines that could reuse the
1179 // cache value will only see properly sorted cache arrays.
1180 if (Cache && NumSortedEntries != Cache->size()) {
1181 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1183 // Since we bail out, the "Cache" set won't contain all of the
1184 // results for the query. This is ok (we can still use it to accelerate
1185 // specific block queries) but we can't do the fastpath "return all
1186 // results from the set". Clear out the indicator for this.
1187 CacheInfo->Pair = BBSkipFirstBlockPair();
1191 // Skip the first block if we have it.
1192 if (!SkipFirstBlock) {
1193 // Analyze the dependency of *Pointer in FromBB. See if we already have
1195 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1197 // Get the dependency info for Pointer in BB. If we have cached
1198 // information, we will use it, otherwise we compute it.
1199 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1200 MemDepResult Dep = GetNonLocalInfoForBlock(QueryInst,
1201 Loc, isLoad, BB, Cache,
1204 // If we got a Def or Clobber, add this to the list of results.
1205 if (!Dep.isNonLocal()) {
1207 Result.push_back(NonLocalDepResult(BB,
1208 MemDepResult::getUnknown(),
1209 Pointer.getAddr()));
1211 } else if (DT->isReachableFromEntry(BB)) {
1212 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1218 // If 'Pointer' is an instruction defined in this block, then we need to do
1219 // phi translation to change it into a value live in the predecessor block.
1220 // If not, we just add the predecessors to the worklist and scan them with
1221 // the same Pointer.
1222 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1223 SkipFirstBlock = false;
1224 SmallVector<BasicBlock*, 16> NewBlocks;
1225 for (BasicBlock *Pred : PredCache.get(BB)) {
1226 // Verify that we haven't looked at this block yet.
1227 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1228 InsertRes = Visited.insert(std::make_pair(Pred, Pointer.getAddr()));
1229 if (InsertRes.second) {
1230 // First time we've looked at *PI.
1231 NewBlocks.push_back(Pred);
1235 // If we have seen this block before, but it was with a different
1236 // pointer then we have a phi translation failure and we have to treat
1237 // this as a clobber.
1238 if (InsertRes.first->second != Pointer.getAddr()) {
1239 // Make sure to clean up the Visited map before continuing on to
1240 // PredTranslationFailure.
1241 for (unsigned i = 0; i < NewBlocks.size(); i++)
1242 Visited.erase(NewBlocks[i]);
1243 goto PredTranslationFailure;
1246 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1250 // We do need to do phi translation, if we know ahead of time we can't phi
1251 // translate this value, don't even try.
1252 if (!Pointer.IsPotentiallyPHITranslatable())
1253 goto PredTranslationFailure;
1255 // We may have added values to the cache list before this PHI translation.
1256 // If so, we haven't done anything to ensure that the cache remains sorted.
1257 // Sort it now (if needed) so that recursive invocations of
1258 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1259 // value will only see properly sorted cache arrays.
1260 if (Cache && NumSortedEntries != Cache->size()) {
1261 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1262 NumSortedEntries = Cache->size();
1267 for (BasicBlock *Pred : PredCache.get(BB)) {
1268 PredList.push_back(std::make_pair(Pred, Pointer));
1270 // Get the PHI translated pointer in this predecessor. This can fail if
1271 // not translatable, in which case the getAddr() returns null.
1272 PHITransAddr &PredPointer = PredList.back().second;
1273 PredPointer.PHITranslateValue(BB, Pred, DT, /*MustDominate=*/false);
1274 Value *PredPtrVal = PredPointer.getAddr();
1276 // Check to see if we have already visited this pred block with another
1277 // pointer. If so, we can't do this lookup. This failure can occur
1278 // with PHI translation when a critical edge exists and the PHI node in
1279 // the successor translates to a pointer value different than the
1280 // pointer the block was first analyzed with.
1281 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1282 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1284 if (!InsertRes.second) {
1285 // We found the pred; take it off the list of preds to visit.
1286 PredList.pop_back();
1288 // If the predecessor was visited with PredPtr, then we already did
1289 // the analysis and can ignore it.
1290 if (InsertRes.first->second == PredPtrVal)
1293 // Otherwise, the block was previously analyzed with a different
1294 // pointer. We can't represent the result of this case, so we just
1295 // treat this as a phi translation failure.
1297 // Make sure to clean up the Visited map before continuing on to
1298 // PredTranslationFailure.
1299 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1300 Visited.erase(PredList[i].first);
1302 goto PredTranslationFailure;
1306 // Actually process results here; this need to be a separate loop to avoid
1307 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1308 // any results for. (getNonLocalPointerDepFromBB will modify our
1309 // datastructures in ways the code after the PredTranslationFailure label
1311 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1312 BasicBlock *Pred = PredList[i].first;
1313 PHITransAddr &PredPointer = PredList[i].second;
1314 Value *PredPtrVal = PredPointer.getAddr();
1316 bool CanTranslate = true;
1317 // If PHI translation was unable to find an available pointer in this
1318 // predecessor, then we have to assume that the pointer is clobbered in
1319 // that predecessor. We can still do PRE of the load, which would insert
1320 // a computation of the pointer in this predecessor.
1322 CanTranslate = false;
1324 // FIXME: it is entirely possible that PHI translating will end up with
1325 // the same value. Consider PHI translating something like:
1326 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1327 // to recurse here, pedantically speaking.
1329 // If getNonLocalPointerDepFromBB fails here, that means the cached
1330 // result conflicted with the Visited list; we have to conservatively
1331 // assume it is unknown, but this also does not block PRE of the load.
1332 if (!CanTranslate ||
1333 getNonLocalPointerDepFromBB(QueryInst, PredPointer,
1334 Loc.getWithNewPtr(PredPtrVal),
1337 // Add the entry to the Result list.
1338 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1339 Result.push_back(Entry);
1341 // Since we had a phi translation failure, the cache for CacheKey won't
1342 // include all of the entries that we need to immediately satisfy future
1343 // queries. Mark this in NonLocalPointerDeps by setting the
1344 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1345 // cached value to do more work but not miss the phi trans failure.
1346 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1347 NLPI.Pair = BBSkipFirstBlockPair();
1352 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1353 CacheInfo = &NonLocalPointerDeps[CacheKey];
1354 Cache = &CacheInfo->NonLocalDeps;
1355 NumSortedEntries = Cache->size();
1357 // Since we did phi translation, the "Cache" set won't contain all of the
1358 // results for the query. This is ok (we can still use it to accelerate
1359 // specific block queries) but we can't do the fastpath "return all
1360 // results from the set" Clear out the indicator for this.
1361 CacheInfo->Pair = BBSkipFirstBlockPair();
1362 SkipFirstBlock = false;
1365 PredTranslationFailure:
1366 // The following code is "failure"; we can't produce a sane translation
1367 // for the given block. It assumes that we haven't modified any of
1368 // our datastructures while processing the current block.
1371 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1372 CacheInfo = &NonLocalPointerDeps[CacheKey];
1373 Cache = &CacheInfo->NonLocalDeps;
1374 NumSortedEntries = Cache->size();
1377 // Since we failed phi translation, the "Cache" set won't contain all of the
1378 // results for the query. This is ok (we can still use it to accelerate
1379 // specific block queries) but we can't do the fastpath "return all
1380 // results from the set". Clear out the indicator for this.
1381 CacheInfo->Pair = BBSkipFirstBlockPair();
1383 // If *nothing* works, mark the pointer as unknown.
1385 // If this is the magic first block, return this as a clobber of the whole
1386 // incoming value. Since we can't phi translate to one of the predecessors,
1387 // we have to bail out.
1391 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1392 assert(I != Cache->rend() && "Didn't find current block??");
1393 if (I->getBB() != BB)
1396 assert((I->getResult().isNonLocal() || !DT->isReachableFromEntry(BB)) &&
1397 "Should only be here with transparent block");
1398 I->setResult(MemDepResult::getUnknown());
1399 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1400 Pointer.getAddr()));
1405 // Okay, we're done now. If we added new values to the cache, re-sort it.
1406 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1407 DEBUG(AssertSorted(*Cache));
1411 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1412 /// CachedNonLocalPointerInfo, remove it.
1413 void MemoryDependenceAnalysis::
1414 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1415 CachedNonLocalPointerInfo::iterator It =
1416 NonLocalPointerDeps.find(P);
1417 if (It == NonLocalPointerDeps.end()) return;
1419 // Remove all of the entries in the BB->val map. This involves removing
1420 // instructions from the reverse map.
1421 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1423 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1424 Instruction *Target = PInfo[i].getResult().getInst();
1425 if (!Target) continue; // Ignore non-local dep results.
1426 assert(Target->getParent() == PInfo[i].getBB());
1428 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1429 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1432 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1433 NonLocalPointerDeps.erase(It);
1437 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1438 /// information about the specified pointer, because it may be too
1439 /// conservative in memdep. This is an optional call that can be used when
1440 /// the client detects an equivalence between the pointer and some other
1441 /// value and replaces the other value with ptr. This can make Ptr available
1442 /// in more places that cached info does not necessarily keep.
1443 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1444 // If Ptr isn't really a pointer, just ignore it.
1445 if (!Ptr->getType()->isPointerTy()) return;
1446 // Flush store info for the pointer.
1447 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1448 // Flush load info for the pointer.
1449 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1452 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1453 /// This needs to be done when the CFG changes, e.g., due to splitting
1455 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1459 /// removeInstruction - Remove an instruction from the dependence analysis,
1460 /// updating the dependence of instructions that previously depended on it.
1461 /// This method attempts to keep the cache coherent using the reverse map.
1462 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1463 // Walk through the Non-local dependencies, removing this one as the value
1464 // for any cached queries.
1465 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1466 if (NLDI != NonLocalDeps.end()) {
1467 NonLocalDepInfo &BlockMap = NLDI->second.first;
1468 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1470 if (Instruction *Inst = DI->getResult().getInst())
1471 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1472 NonLocalDeps.erase(NLDI);
1475 // If we have a cached local dependence query for this instruction, remove it.
1477 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1478 if (LocalDepEntry != LocalDeps.end()) {
1479 // Remove us from DepInst's reverse set now that the local dep info is gone.
1480 if (Instruction *Inst = LocalDepEntry->second.getInst())
1481 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1483 // Remove this local dependency info.
1484 LocalDeps.erase(LocalDepEntry);
1487 // If we have any cached pointer dependencies on this instruction, remove
1488 // them. If the instruction has non-pointer type, then it can't be a pointer
1491 // Remove it from both the load info and the store info. The instruction
1492 // can't be in either of these maps if it is non-pointer.
1493 if (RemInst->getType()->isPointerTy()) {
1494 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1495 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1498 // Loop over all of the things that depend on the instruction we're removing.
1500 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1502 // If we find RemInst as a clobber or Def in any of the maps for other values,
1503 // we need to replace its entry with a dirty version of the instruction after
1504 // it. If RemInst is a terminator, we use a null dirty value.
1506 // Using a dirty version of the instruction after RemInst saves having to scan
1507 // the entire block to get to this point.
1508 MemDepResult NewDirtyVal;
1509 if (!RemInst->isTerminator())
1510 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1512 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1513 if (ReverseDepIt != ReverseLocalDeps.end()) {
1514 // RemInst can't be the terminator if it has local stuff depending on it.
1515 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1516 "Nothing can locally depend on a terminator");
1518 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1519 assert(InstDependingOnRemInst != RemInst &&
1520 "Already removed our local dep info");
1522 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1524 // Make sure to remember that new things depend on NewDepInst.
1525 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1526 "a local dep on this if it is a terminator!");
1527 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1528 InstDependingOnRemInst));
1531 ReverseLocalDeps.erase(ReverseDepIt);
1533 // Add new reverse deps after scanning the set, to avoid invalidating the
1534 // 'ReverseDeps' reference.
1535 while (!ReverseDepsToAdd.empty()) {
1536 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1537 .insert(ReverseDepsToAdd.back().second);
1538 ReverseDepsToAdd.pop_back();
1542 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1543 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1544 for (Instruction *I : ReverseDepIt->second) {
1545 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1547 PerInstNLInfo &INLD = NonLocalDeps[I];
1548 // The information is now dirty!
1551 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1552 DE = INLD.first.end(); DI != DE; ++DI) {
1553 if (DI->getResult().getInst() != RemInst) continue;
1555 // Convert to a dirty entry for the subsequent instruction.
1556 DI->setResult(NewDirtyVal);
1558 if (Instruction *NextI = NewDirtyVal.getInst())
1559 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1563 ReverseNonLocalDeps.erase(ReverseDepIt);
1565 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1566 while (!ReverseDepsToAdd.empty()) {
1567 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1568 .insert(ReverseDepsToAdd.back().second);
1569 ReverseDepsToAdd.pop_back();
1573 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1574 // value in the NonLocalPointerDeps info.
1575 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1576 ReverseNonLocalPtrDeps.find(RemInst);
1577 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1578 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1580 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1581 assert(P.getPointer() != RemInst &&
1582 "Already removed NonLocalPointerDeps info for RemInst");
1584 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1586 // The cache is not valid for any specific block anymore.
1587 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1589 // Update any entries for RemInst to use the instruction after it.
1590 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1592 if (DI->getResult().getInst() != RemInst) continue;
1594 // Convert to a dirty entry for the subsequent instruction.
1595 DI->setResult(NewDirtyVal);
1597 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1598 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1601 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1602 // subsequent value may invalidate the sortedness.
1603 std::sort(NLPDI.begin(), NLPDI.end());
1606 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1608 while (!ReversePtrDepsToAdd.empty()) {
1609 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1610 .insert(ReversePtrDepsToAdd.back().second);
1611 ReversePtrDepsToAdd.pop_back();
1616 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1617 AA->deleteValue(RemInst);
1618 DEBUG(verifyRemoved(RemInst));
1620 /// verifyRemoved - Verify that the specified instruction does not occur
1621 /// in our internal data structures. This function verifies by asserting in
1623 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1625 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1626 E = LocalDeps.end(); I != E; ++I) {
1627 assert(I->first != D && "Inst occurs in data structures");
1628 assert(I->second.getInst() != D &&
1629 "Inst occurs in data structures");
1632 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1633 E = NonLocalPointerDeps.end(); I != E; ++I) {
1634 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1635 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1636 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1638 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1641 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1642 E = NonLocalDeps.end(); I != E; ++I) {
1643 assert(I->first != D && "Inst occurs in data structures");
1644 const PerInstNLInfo &INLD = I->second;
1645 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1646 EE = INLD.first.end(); II != EE; ++II)
1647 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1650 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1651 E = ReverseLocalDeps.end(); I != E; ++I) {
1652 assert(I->first != D && "Inst occurs in data structures");
1653 for (Instruction *Inst : I->second)
1654 assert(Inst != D && "Inst occurs in data structures");
1657 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1658 E = ReverseNonLocalDeps.end();
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 (ReverseNonLocalPtrDepTy::const_iterator
1666 I = ReverseNonLocalPtrDeps.begin(),
1667 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1668 assert(I->first != D && "Inst occurs in rev NLPD map");
1670 for (ValueIsLoadPair P : I->second)
1671 assert(P != ValueIsLoadPair(D, false) &&
1672 P != ValueIsLoadPair(D, true) &&
1673 "Inst occurs in ReverseNonLocalPtrDeps map");