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/AssumptionTracker.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 int BlockScanLimit = 100;
54 char MemoryDependenceAnalysis::ID = 0;
56 // Register this pass...
57 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
58 "Memory Dependence Analysis", false, true)
59 INITIALIZE_PASS_DEPENDENCY(AssumptionTracker)
60 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
61 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
62 "Memory Dependence Analysis", false, true)
64 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
65 : FunctionPass(ID), PredCache() {
66 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
68 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
71 /// Clean up memory in between runs
72 void MemoryDependenceAnalysis::releaseMemory() {
75 NonLocalPointerDeps.clear();
76 ReverseLocalDeps.clear();
77 ReverseNonLocalDeps.clear();
78 ReverseNonLocalPtrDeps.clear();
84 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
86 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
88 AU.addRequired<AssumptionTracker>();
89 AU.addRequiredTransitive<AliasAnalysis>();
92 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
93 AA = &getAnalysis<AliasAnalysis>();
94 AT = &getAnalysis<AssumptionTracker>();
95 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
96 DL = DLP ? &DLP->getDataLayout() : nullptr;
97 DominatorTreeWrapperPass *DTWP =
98 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
99 DT = DTWP ? &DTWP->getDomTree() : nullptr;
101 PredCache.reset(new PredIteratorCache());
105 /// RemoveFromReverseMap - This is a helper function that removes Val from
106 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
107 template <typename KeyTy>
108 static void RemoveFromReverseMap(DenseMap<Instruction*,
109 SmallPtrSet<KeyTy, 4> > &ReverseMap,
110 Instruction *Inst, KeyTy Val) {
111 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
112 InstIt = ReverseMap.find(Inst);
113 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
114 bool Found = InstIt->second.erase(Val);
115 assert(Found && "Invalid reverse map!"); (void)Found;
116 if (InstIt->second.empty())
117 ReverseMap.erase(InstIt);
120 /// GetLocation - If the given instruction references a specific memory
121 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
122 /// Return a ModRefInfo value describing the general behavior of the
125 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
126 AliasAnalysis::Location &Loc,
128 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
129 if (LI->isUnordered()) {
130 Loc = AA->getLocation(LI);
131 return AliasAnalysis::Ref;
133 if (LI->getOrdering() == Monotonic) {
134 Loc = AA->getLocation(LI);
135 return AliasAnalysis::ModRef;
137 Loc = AliasAnalysis::Location();
138 return AliasAnalysis::ModRef;
141 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
142 if (SI->isUnordered()) {
143 Loc = AA->getLocation(SI);
144 return AliasAnalysis::Mod;
146 if (SI->getOrdering() == Monotonic) {
147 Loc = AA->getLocation(SI);
148 return AliasAnalysis::ModRef;
150 Loc = AliasAnalysis::Location();
151 return AliasAnalysis::ModRef;
154 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
155 Loc = AA->getLocation(V);
156 return AliasAnalysis::ModRef;
159 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
160 // calls to free() deallocate the entire structure
161 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
162 return AliasAnalysis::Mod;
165 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
168 switch (II->getIntrinsicID()) {
169 case Intrinsic::lifetime_start:
170 case Intrinsic::lifetime_end:
171 case Intrinsic::invariant_start:
172 II->getAAMetadata(AAInfo);
173 Loc = AliasAnalysis::Location(II->getArgOperand(1),
174 cast<ConstantInt>(II->getArgOperand(0))
175 ->getZExtValue(), AAInfo);
176 // These intrinsics don't really modify the memory, but returning Mod
177 // will allow them to be handled conservatively.
178 return AliasAnalysis::Mod;
179 case Intrinsic::invariant_end:
180 II->getAAMetadata(AAInfo);
181 Loc = AliasAnalysis::Location(II->getArgOperand(2),
182 cast<ConstantInt>(II->getArgOperand(1))
183 ->getZExtValue(), AAInfo);
184 // These intrinsics don't really modify the memory, but returning Mod
185 // will allow them to be handled conservatively.
186 return AliasAnalysis::Mod;
192 // Otherwise, just do the coarse-grained thing that always works.
193 if (Inst->mayWriteToMemory())
194 return AliasAnalysis::ModRef;
195 if (Inst->mayReadFromMemory())
196 return AliasAnalysis::Ref;
197 return AliasAnalysis::NoModRef;
200 /// getCallSiteDependencyFrom - Private helper for finding the local
201 /// dependencies of a call site.
202 MemDepResult MemoryDependenceAnalysis::
203 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
204 BasicBlock::iterator ScanIt, BasicBlock *BB) {
205 unsigned Limit = BlockScanLimit;
207 // Walk backwards through the block, looking for dependencies
208 while (ScanIt != BB->begin()) {
209 // Limit the amount of scanning we do so we don't end up with quadratic
210 // running time on extreme testcases.
213 return MemDepResult::getUnknown();
215 Instruction *Inst = --ScanIt;
217 // If this inst is a memory op, get the pointer it accessed
218 AliasAnalysis::Location Loc;
219 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
221 // A simple instruction.
222 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
223 return MemDepResult::getClobber(Inst);
227 if (CallSite InstCS = cast<Value>(Inst)) {
228 // Debug intrinsics don't cause dependences.
229 if (isa<DbgInfoIntrinsic>(Inst)) continue;
230 // If these two calls do not interfere, look past it.
231 switch (AA->getModRefInfo(CS, InstCS)) {
232 case AliasAnalysis::NoModRef:
233 // If the two calls are the same, return InstCS as a Def, so that
234 // CS can be found redundant and eliminated.
235 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
236 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
237 return MemDepResult::getDef(Inst);
239 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
243 return MemDepResult::getClobber(Inst);
247 // If we could not obtain a pointer for the instruction and the instruction
248 // touches memory then assume that this is a dependency.
249 if (MR != AliasAnalysis::NoModRef)
250 return MemDepResult::getClobber(Inst);
253 // No dependence found. If this is the entry block of the function, it is
254 // unknown, otherwise it is non-local.
255 if (BB != &BB->getParent()->getEntryBlock())
256 return MemDepResult::getNonLocal();
257 return MemDepResult::getNonFuncLocal();
260 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
261 /// would fully overlap MemLoc if done as a wider legal integer load.
263 /// MemLocBase, MemLocOffset are lazily computed here the first time the
264 /// base/offs of memloc is needed.
266 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
267 const Value *&MemLocBase,
270 const DataLayout *DL) {
271 // If we have no target data, we can't do this.
272 if (!DL) return false;
274 // If we haven't already computed the base/offset of MemLoc, do so now.
276 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
278 unsigned Size = MemoryDependenceAnalysis::
279 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
284 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
285 /// looks at a memory location for a load (specified by MemLocBase, Offs,
286 /// and Size) and compares it against a load. If the specified load could
287 /// be safely widened to a larger integer load that is 1) still efficient,
288 /// 2) safe for the target, and 3) would provide the specified memory
289 /// location value, then this function returns the size in bytes of the
290 /// load width to use. If not, this returns zero.
291 unsigned MemoryDependenceAnalysis::
292 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
293 unsigned MemLocSize, const LoadInst *LI,
294 const DataLayout &DL) {
295 // We can only extend simple integer loads.
296 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
298 // Load widening is hostile to ThreadSanitizer: it may cause false positives
299 // or make the reports more cryptic (access sizes are wrong).
300 if (LI->getParent()->getParent()->getAttributes().
301 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
304 // Get the base of this load.
306 const Value *LIBase =
307 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &DL);
309 // If the two pointers are not based on the same pointer, we can't tell that
311 if (LIBase != MemLocBase) return 0;
313 // Okay, the two values are based on the same pointer, but returned as
314 // no-alias. This happens when we have things like two byte loads at "P+1"
315 // and "P+3". Check to see if increasing the size of the "LI" load up to its
316 // alignment (or the largest native integer type) will allow us to load all
317 // the bits required by MemLoc.
319 // If MemLoc is before LI, then no widening of LI will help us out.
320 if (MemLocOffs < LIOffs) return 0;
322 // Get the alignment of the load in bytes. We assume that it is safe to load
323 // any legal integer up to this size without a problem. For example, if we're
324 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
325 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
327 unsigned LoadAlign = LI->getAlignment();
329 int64_t MemLocEnd = MemLocOffs+MemLocSize;
331 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
332 if (LIOffs+LoadAlign < MemLocEnd) return 0;
334 // This is the size of the load to try. Start with the next larger power of
336 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
337 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
340 // If this load size is bigger than our known alignment or would not fit
341 // into a native integer register, then we fail.
342 if (NewLoadByteSize > LoadAlign ||
343 !DL.fitsInLegalInteger(NewLoadByteSize*8))
346 if (LIOffs+NewLoadByteSize > MemLocEnd &&
347 LI->getParent()->getParent()->getAttributes().
348 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
349 // We will be reading past the location accessed by the original program.
350 // While this is safe in a regular build, Address Safety analysis tools
351 // may start reporting false warnings. So, don't do widening.
354 // If a load of this width would include all of MemLoc, then we succeed.
355 if (LIOffs+NewLoadByteSize >= MemLocEnd)
356 return NewLoadByteSize;
358 NewLoadByteSize <<= 1;
362 /// getPointerDependencyFrom - Return the instruction on which a memory
363 /// location depends. If isLoad is true, this routine ignores may-aliases with
364 /// read-only operations. If isLoad is false, this routine ignores may-aliases
365 /// with reads from read-only locations. If possible, pass the query
366 /// instruction as well; this function may take advantage of the metadata
367 /// annotated to the query instruction to refine the result.
368 MemDepResult MemoryDependenceAnalysis::
369 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
370 BasicBlock::iterator ScanIt, BasicBlock *BB,
371 Instruction *QueryInst) {
373 const Value *MemLocBase = nullptr;
374 int64_t MemLocOffset = 0;
375 unsigned Limit = BlockScanLimit;
376 bool isInvariantLoad = false;
378 // We must be careful with atomic accesses, as they may allow another thread
379 // to touch this location, cloberring it. We are conservative: if the
380 // QueryInst is not a simple (non-atomic) memory access, we automatically
381 // return getClobber.
382 // If it is simple, we know based on the results of
383 // "Compiler testing via a theory of sound optimisations in the C11/C++11
384 // memory model" in PLDI 2013, that a non-atomic location can only be
385 // clobbered between a pair of a release and an acquire action, with no
386 // access to the location in between.
387 // Here is an example for giving the general intuition behind this rule.
388 // In the following code:
390 // release action; [1]
391 // acquire action; [4]
393 // It is unsafe to replace %val by 0 because another thread may be running:
394 // acquire action; [2]
396 // release action; [3]
397 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
398 // being 42. A key property of this program however is that if either
399 // 1 or 4 were missing, there would be a race between the store of 42
400 // either the store of 0 or the load (making the whole progam racy).
401 // The paper mentionned above shows that the same property is respected
402 // by every program that can detect any optimisation of that kind: either
403 // it is racy (undefined) or there is a release followed by an acquire
404 // between the pair of accesses under consideration.
405 bool HasSeenAcquire = false;
407 if (isLoad && QueryInst) {
408 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
409 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
410 isInvariantLoad = true;
413 // Walk backwards through the basic block, looking for dependencies.
414 while (ScanIt != BB->begin()) {
415 Instruction *Inst = --ScanIt;
417 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
418 // Debug intrinsics don't (and can't) cause dependencies.
419 if (isa<DbgInfoIntrinsic>(II)) continue;
421 // Limit the amount of scanning we do so we don't end up with quadratic
422 // running time on extreme testcases.
425 return MemDepResult::getUnknown();
427 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
428 // If we reach a lifetime begin or end marker, then the query ends here
429 // because the value is undefined.
430 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
431 // FIXME: This only considers queries directly on the invariant-tagged
432 // pointer, not on query pointers that are indexed off of them. It'd
433 // be nice to handle that at some point (the right approach is to use
434 // GetPointerBaseWithConstantOffset).
435 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
437 return MemDepResult::getDef(II);
442 // Values depend on loads if the pointers are must aliased. This means that
443 // a load depends on another must aliased load from the same value.
444 // One exception is atomic loads: a value can depend on an atomic load that it
445 // does not alias with when this atomic load indicates that another thread may
446 // be accessing the location.
447 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
448 // Atomic loads have complications involved.
449 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
450 // An Acquire (or higher) load sets the HasSeenAcquire flag, so that any
451 // release store will know to return getClobber.
452 // FIXME: This is overly conservative.
453 if (!LI->isUnordered()) {
455 return MemDepResult::getClobber(LI);
456 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
457 if (!QueryLI->isSimple())
458 return MemDepResult::getClobber(LI);
459 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
460 if (!QuerySI->isSimple())
461 return MemDepResult::getClobber(LI);
462 } else if (QueryInst->mayReadOrWriteMemory()) {
463 return MemDepResult::getClobber(LI);
466 if (isAtLeastAcquire(LI->getOrdering()))
467 HasSeenAcquire = true;
470 // FIXME: this is overly conservative.
471 // While volatile access cannot be eliminated, they do not have to clobber
472 // non-aliasing locations, as normal accesses can for example be reordered
473 // with volatile accesses.
474 if (LI->isVolatile())
475 return MemDepResult::getClobber(LI);
477 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
479 // If we found a pointer, check if it could be the same as our pointer.
480 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
483 if (R == AliasAnalysis::NoAlias) {
484 // If this is an over-aligned integer load (for example,
485 // "load i8* %P, align 4") see if it would obviously overlap with the
486 // queried location if widened to a larger load (e.g. if the queried
487 // location is 1 byte at P+1). If so, return it as a load/load
488 // clobber result, allowing the client to decide to widen the load if
490 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
491 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
492 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
493 MemLocOffset, LI, DL))
494 return MemDepResult::getClobber(Inst);
499 // Must aliased loads are defs of each other.
500 if (R == AliasAnalysis::MustAlias)
501 return MemDepResult::getDef(Inst);
503 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
504 // in terms of clobbering loads, but since it does this by looking
505 // at the clobbering load directly, it doesn't know about any
506 // phi translation that may have happened along the way.
508 // If we have a partial alias, then return this as a clobber for the
510 if (R == AliasAnalysis::PartialAlias)
511 return MemDepResult::getClobber(Inst);
514 // Random may-alias loads don't depend on each other without a
519 // Stores don't depend on other no-aliased accesses.
520 if (R == AliasAnalysis::NoAlias)
523 // Stores don't alias loads from read-only memory.
524 if (AA->pointsToConstantMemory(LoadLoc))
527 // Stores depend on may/must aliased loads.
528 return MemDepResult::getDef(Inst);
531 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
532 // Atomic stores have complications involved.
533 // A Monotonic store is OK if the query inst is itself not atomic.
534 // A Release (or higher) store further requires that no acquire load
536 // FIXME: This is overly conservative.
537 if (!SI->isUnordered()) {
539 return MemDepResult::getClobber(SI);
540 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
541 if (!QueryLI->isSimple())
542 return MemDepResult::getClobber(SI);
543 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
544 if (!QuerySI->isSimple())
545 return MemDepResult::getClobber(SI);
546 } else if (QueryInst->mayReadOrWriteMemory()) {
547 return MemDepResult::getClobber(SI);
550 if (HasSeenAcquire && isAtLeastRelease(SI->getOrdering()))
551 return MemDepResult::getClobber(SI);
554 // FIXME: this is overly conservative.
555 // While volatile access cannot be eliminated, they do not have to clobber
556 // non-aliasing locations, as normal accesses can for example be reordered
557 // with volatile accesses.
558 if (SI->isVolatile())
559 return MemDepResult::getClobber(SI);
561 // If alias analysis can tell that this store is guaranteed to not modify
562 // the query pointer, ignore it. Use getModRefInfo to handle cases where
563 // the query pointer points to constant memory etc.
564 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
567 // Ok, this store might clobber the query pointer. Check to see if it is
568 // a must alias: in this case, we want to return this as a def.
569 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
571 // If we found a pointer, check if it could be the same as our pointer.
572 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
574 if (R == AliasAnalysis::NoAlias)
576 if (R == AliasAnalysis::MustAlias)
577 return MemDepResult::getDef(Inst);
580 return MemDepResult::getClobber(Inst);
583 // If this is an allocation, and if we know that the accessed pointer is to
584 // the allocation, return Def. This means that there is no dependence and
585 // the access can be optimized based on that. For example, a load could
587 // Note: Only determine this to be a malloc if Inst is the malloc call, not
588 // a subsequent bitcast of the malloc call result. There can be stores to
589 // the malloced memory between the malloc call and its bitcast uses, and we
590 // need to continue scanning until the malloc call.
591 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
592 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
593 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
595 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
596 return MemDepResult::getDef(Inst);
597 // Be conservative if the accessed pointer may alias the allocation.
598 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
599 return MemDepResult::getClobber(Inst);
600 // If the allocation is not aliased and does not read memory (like
601 // strdup), it is safe to ignore.
602 if (isa<AllocaInst>(Inst) ||
603 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
607 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
608 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
609 // If necessary, perform additional analysis.
610 if (MR == AliasAnalysis::ModRef)
611 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
613 case AliasAnalysis::NoModRef:
614 // If the call has no effect on the queried pointer, just ignore it.
616 case AliasAnalysis::Mod:
617 return MemDepResult::getClobber(Inst);
618 case AliasAnalysis::Ref:
619 // If the call is known to never store to the pointer, and if this is a
620 // load query, we can safely ignore it (scan past it).
624 // Otherwise, there is a potential dependence. Return a clobber.
625 return MemDepResult::getClobber(Inst);
629 // No dependence found. If this is the entry block of the function, it is
630 // unknown, otherwise it is non-local.
631 if (BB != &BB->getParent()->getEntryBlock())
632 return MemDepResult::getNonLocal();
633 return MemDepResult::getNonFuncLocal();
636 /// getDependency - Return the instruction on which a memory operation
638 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
639 Instruction *ScanPos = QueryInst;
641 // Check for a cached result
642 MemDepResult &LocalCache = LocalDeps[QueryInst];
644 // If the cached entry is non-dirty, just return it. Note that this depends
645 // on MemDepResult's default constructing to 'dirty'.
646 if (!LocalCache.isDirty())
649 // Otherwise, if we have a dirty entry, we know we can start the scan at that
650 // instruction, which may save us some work.
651 if (Instruction *Inst = LocalCache.getInst()) {
654 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
657 BasicBlock *QueryParent = QueryInst->getParent();
660 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
661 // No dependence found. If this is the entry block of the function, it is
662 // unknown, otherwise it is non-local.
663 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
664 LocalCache = MemDepResult::getNonLocal();
666 LocalCache = MemDepResult::getNonFuncLocal();
668 AliasAnalysis::Location MemLoc;
669 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
671 // If we can do a pointer scan, make it happen.
672 bool isLoad = !(MR & AliasAnalysis::Mod);
673 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
674 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
676 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
677 QueryParent, QueryInst);
678 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
679 CallSite QueryCS(QueryInst);
680 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
681 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
684 // Non-memory instruction.
685 LocalCache = MemDepResult::getUnknown();
688 // Remember the result!
689 if (Instruction *I = LocalCache.getInst())
690 ReverseLocalDeps[I].insert(QueryInst);
696 /// AssertSorted - This method is used when -debug is specified to verify that
697 /// cache arrays are properly kept sorted.
698 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
700 if (Count == -1) Count = Cache.size();
701 if (Count == 0) return;
703 for (unsigned i = 1; i != unsigned(Count); ++i)
704 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
708 /// getNonLocalCallDependency - Perform a full dependency query for the
709 /// specified call, returning the set of blocks that the value is
710 /// potentially live across. The returned set of results will include a
711 /// "NonLocal" result for all blocks where the value is live across.
713 /// This method assumes the instruction returns a "NonLocal" dependency
714 /// within its own block.
716 /// This returns a reference to an internal data structure that may be
717 /// invalidated on the next non-local query or when an instruction is
718 /// removed. Clients must copy this data if they want it around longer than
720 const MemoryDependenceAnalysis::NonLocalDepInfo &
721 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
722 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
723 "getNonLocalCallDependency should only be used on calls with non-local deps!");
724 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
725 NonLocalDepInfo &Cache = CacheP.first;
727 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
728 /// the cached case, this can happen due to instructions being deleted etc. In
729 /// the uncached case, this starts out as the set of predecessors we care
731 SmallVector<BasicBlock*, 32> DirtyBlocks;
733 if (!Cache.empty()) {
734 // Okay, we have a cache entry. If we know it is not dirty, just return it
735 // with no computation.
736 if (!CacheP.second) {
741 // If we already have a partially computed set of results, scan them to
742 // determine what is dirty, seeding our initial DirtyBlocks worklist.
743 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
745 if (I->getResult().isDirty())
746 DirtyBlocks.push_back(I->getBB());
748 // Sort the cache so that we can do fast binary search lookups below.
749 std::sort(Cache.begin(), Cache.end());
751 ++NumCacheDirtyNonLocal;
752 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
753 // << Cache.size() << " cached: " << *QueryInst;
755 // Seed DirtyBlocks with each of the preds of QueryInst's block.
756 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
757 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
758 DirtyBlocks.push_back(*PI);
759 ++NumUncacheNonLocal;
762 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
763 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
765 SmallPtrSet<BasicBlock*, 64> Visited;
767 unsigned NumSortedEntries = Cache.size();
768 DEBUG(AssertSorted(Cache));
770 // Iterate while we still have blocks to update.
771 while (!DirtyBlocks.empty()) {
772 BasicBlock *DirtyBB = DirtyBlocks.back();
773 DirtyBlocks.pop_back();
775 // Already processed this block?
776 if (!Visited.insert(DirtyBB))
779 // Do a binary search to see if we already have an entry for this block in
780 // the cache set. If so, find it.
781 DEBUG(AssertSorted(Cache, NumSortedEntries));
782 NonLocalDepInfo::iterator Entry =
783 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
784 NonLocalDepEntry(DirtyBB));
785 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
788 NonLocalDepEntry *ExistingResult = nullptr;
789 if (Entry != Cache.begin()+NumSortedEntries &&
790 Entry->getBB() == DirtyBB) {
791 // If we already have an entry, and if it isn't already dirty, the block
793 if (!Entry->getResult().isDirty())
796 // Otherwise, remember this slot so we can update the value.
797 ExistingResult = &*Entry;
800 // If the dirty entry has a pointer, start scanning from it so we don't have
801 // to rescan the entire block.
802 BasicBlock::iterator ScanPos = DirtyBB->end();
803 if (ExistingResult) {
804 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
806 // We're removing QueryInst's use of Inst.
807 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
808 QueryCS.getInstruction());
812 // Find out if this block has a local dependency for QueryInst.
815 if (ScanPos != DirtyBB->begin()) {
816 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
817 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
818 // No dependence found. If this is the entry block of the function, it is
819 // a clobber, otherwise it is unknown.
820 Dep = MemDepResult::getNonLocal();
822 Dep = MemDepResult::getNonFuncLocal();
825 // If we had a dirty entry for the block, update it. Otherwise, just add
828 ExistingResult->setResult(Dep);
830 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
832 // If the block has a dependency (i.e. it isn't completely transparent to
833 // the value), remember the association!
834 if (!Dep.isNonLocal()) {
835 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
836 // update this when we remove instructions.
837 if (Instruction *Inst = Dep.getInst())
838 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
841 // If the block *is* completely transparent to the load, we need to check
842 // the predecessors of this block. Add them to our worklist.
843 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
844 DirtyBlocks.push_back(*PI);
851 /// getNonLocalPointerDependency - Perform a full dependency query for an
852 /// access to the specified (non-volatile) memory location, returning the
853 /// set of instructions that either define or clobber the value.
855 /// This method assumes the pointer has a "NonLocal" dependency within its
858 void MemoryDependenceAnalysis::
859 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
861 SmallVectorImpl<NonLocalDepResult> &Result) {
862 assert(Loc.Ptr->getType()->isPointerTy() &&
863 "Can't get pointer deps of a non-pointer!");
866 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AT);
868 // This is the set of blocks we've inspected, and the pointer we consider in
869 // each block. Because of critical edges, we currently bail out if querying
870 // a block with multiple different pointers. This can happen during PHI
872 DenseMap<BasicBlock*, Value*> Visited;
873 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
874 Result, Visited, true))
877 Result.push_back(NonLocalDepResult(FromBB,
878 MemDepResult::getUnknown(),
879 const_cast<Value *>(Loc.Ptr)));
882 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
883 /// Pointer/PointeeSize using either cached information in Cache or by doing a
884 /// lookup (which may use dirty cache info if available). If we do a lookup,
885 /// add the result to the cache.
886 MemDepResult MemoryDependenceAnalysis::
887 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
888 bool isLoad, BasicBlock *BB,
889 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
891 // Do a binary search to see if we already have an entry for this block in
892 // the cache set. If so, find it.
893 NonLocalDepInfo::iterator Entry =
894 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
895 NonLocalDepEntry(BB));
896 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
899 NonLocalDepEntry *ExistingResult = nullptr;
900 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
901 ExistingResult = &*Entry;
903 // If we have a cached entry, and it is non-dirty, use it as the value for
905 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
906 ++NumCacheNonLocalPtr;
907 return ExistingResult->getResult();
910 // Otherwise, we have to scan for the value. If we have a dirty cache
911 // entry, start scanning from its position, otherwise we scan from the end
913 BasicBlock::iterator ScanPos = BB->end();
914 if (ExistingResult && ExistingResult->getResult().getInst()) {
915 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
916 "Instruction invalidated?");
917 ++NumCacheDirtyNonLocalPtr;
918 ScanPos = ExistingResult->getResult().getInst();
920 // Eliminating the dirty entry from 'Cache', so update the reverse info.
921 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
922 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
924 ++NumUncacheNonLocalPtr;
927 // Scan the block for the dependency.
928 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
930 // If we had a dirty entry for the block, update it. Otherwise, just add
933 ExistingResult->setResult(Dep);
935 Cache->push_back(NonLocalDepEntry(BB, Dep));
937 // If the block has a dependency (i.e. it isn't completely transparent to
938 // the value), remember the reverse association because we just added it
940 if (!Dep.isDef() && !Dep.isClobber())
943 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
944 // update MemDep when we remove instructions.
945 Instruction *Inst = Dep.getInst();
946 assert(Inst && "Didn't depend on anything?");
947 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
948 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
952 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
953 /// number of elements in the array that are already properly ordered. This is
954 /// optimized for the case when only a few entries are added.
956 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
957 unsigned NumSortedEntries) {
958 switch (Cache.size() - NumSortedEntries) {
960 // done, no new entries.
963 // Two new entries, insert the last one into place.
964 NonLocalDepEntry Val = Cache.back();
966 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
967 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
968 Cache.insert(Entry, Val);
972 // One new entry, Just insert the new value at the appropriate position.
973 if (Cache.size() != 1) {
974 NonLocalDepEntry Val = Cache.back();
976 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
977 std::upper_bound(Cache.begin(), Cache.end(), Val);
978 Cache.insert(Entry, Val);
982 // Added many values, do a full scale sort.
983 std::sort(Cache.begin(), Cache.end());
988 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
989 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
990 /// results to the results vector and keep track of which blocks are visited in
993 /// This has special behavior for the first block queries (when SkipFirstBlock
994 /// is true). In this special case, it ignores the contents of the specified
995 /// block and starts returning dependence info for its predecessors.
997 /// This function returns false on success, or true to indicate that it could
998 /// not compute dependence information for some reason. This should be treated
999 /// as a clobber dependence on the first instruction in the predecessor block.
1000 bool MemoryDependenceAnalysis::
1001 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
1002 const AliasAnalysis::Location &Loc,
1003 bool isLoad, BasicBlock *StartBB,
1004 SmallVectorImpl<NonLocalDepResult> &Result,
1005 DenseMap<BasicBlock*, Value*> &Visited,
1006 bool SkipFirstBlock) {
1007 // Look up the cached info for Pointer.
1008 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1010 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1011 // CacheKey, this value will be inserted as the associated value. Otherwise,
1012 // it'll be ignored, and we'll have to check to see if the cached size and
1013 // aa tags are consistent with the current query.
1014 NonLocalPointerInfo InitialNLPI;
1015 InitialNLPI.Size = Loc.Size;
1016 InitialNLPI.AATags = Loc.AATags;
1018 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1019 // already have one.
1020 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1021 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1022 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1024 // If we already have a cache entry for this CacheKey, we may need to do some
1025 // work to reconcile the cache entry and the current query.
1027 if (CacheInfo->Size < Loc.Size) {
1028 // The query's Size is greater than the cached one. Throw out the
1029 // cached data and proceed with the query at the greater size.
1030 CacheInfo->Pair = BBSkipFirstBlockPair();
1031 CacheInfo->Size = Loc.Size;
1032 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1033 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1034 if (Instruction *Inst = DI->getResult().getInst())
1035 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1036 CacheInfo->NonLocalDeps.clear();
1037 } else if (CacheInfo->Size > Loc.Size) {
1038 // This query's Size is less than the cached one. Conservatively restart
1039 // the query using the greater size.
1040 return getNonLocalPointerDepFromBB(Pointer,
1041 Loc.getWithNewSize(CacheInfo->Size),
1042 isLoad, StartBB, Result, Visited,
1046 // If the query's AATags are inconsistent with the cached one,
1047 // conservatively throw out the cached data and restart the query with
1048 // no tag if needed.
1049 if (CacheInfo->AATags != Loc.AATags) {
1050 if (CacheInfo->AATags) {
1051 CacheInfo->Pair = BBSkipFirstBlockPair();
1052 CacheInfo->AATags = AAMDNodes();
1053 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1054 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1055 if (Instruction *Inst = DI->getResult().getInst())
1056 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1057 CacheInfo->NonLocalDeps.clear();
1060 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutAATags(),
1061 isLoad, StartBB, Result, Visited,
1066 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1068 // If we have valid cached information for exactly the block we are
1069 // investigating, just return it with no recomputation.
1070 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1071 // We have a fully cached result for this query then we can just return the
1072 // cached results and populate the visited set. However, we have to verify
1073 // that we don't already have conflicting results for these blocks. Check
1074 // to ensure that if a block in the results set is in the visited set that
1075 // it was for the same pointer query.
1076 if (!Visited.empty()) {
1077 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1079 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1080 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1083 // We have a pointer mismatch in a block. Just return clobber, saying
1084 // that something was clobbered in this result. We could also do a
1085 // non-fully cached query, but there is little point in doing this.
1090 Value *Addr = Pointer.getAddr();
1091 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1093 Visited.insert(std::make_pair(I->getBB(), Addr));
1094 if (I->getResult().isNonLocal()) {
1099 Result.push_back(NonLocalDepResult(I->getBB(),
1100 MemDepResult::getUnknown(),
1102 } else if (DT->isReachableFromEntry(I->getBB())) {
1103 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1106 ++NumCacheCompleteNonLocalPtr;
1110 // Otherwise, either this is a new block, a block with an invalid cache
1111 // pointer or one that we're about to invalidate by putting more info into it
1112 // than its valid cache info. If empty, the result will be valid cache info,
1113 // otherwise it isn't.
1115 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1117 CacheInfo->Pair = BBSkipFirstBlockPair();
1119 SmallVector<BasicBlock*, 32> Worklist;
1120 Worklist.push_back(StartBB);
1122 // PredList used inside loop.
1123 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1125 // Keep track of the entries that we know are sorted. Previously cached
1126 // entries will all be sorted. The entries we add we only sort on demand (we
1127 // don't insert every element into its sorted position). We know that we
1128 // won't get any reuse from currently inserted values, because we don't
1129 // revisit blocks after we insert info for them.
1130 unsigned NumSortedEntries = Cache->size();
1131 DEBUG(AssertSorted(*Cache));
1133 while (!Worklist.empty()) {
1134 BasicBlock *BB = Worklist.pop_back_val();
1136 // Skip the first block if we have it.
1137 if (!SkipFirstBlock) {
1138 // Analyze the dependency of *Pointer in FromBB. See if we already have
1140 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1142 // Get the dependency info for Pointer in BB. If we have cached
1143 // information, we will use it, otherwise we compute it.
1144 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1145 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1148 // If we got a Def or Clobber, add this to the list of results.
1149 if (!Dep.isNonLocal()) {
1151 Result.push_back(NonLocalDepResult(BB,
1152 MemDepResult::getUnknown(),
1153 Pointer.getAddr()));
1155 } else if (DT->isReachableFromEntry(BB)) {
1156 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1162 // If 'Pointer' is an instruction defined in this block, then we need to do
1163 // phi translation to change it into a value live in the predecessor block.
1164 // If not, we just add the predecessors to the worklist and scan them with
1165 // the same Pointer.
1166 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1167 SkipFirstBlock = false;
1168 SmallVector<BasicBlock*, 16> NewBlocks;
1169 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1170 // Verify that we haven't looked at this block yet.
1171 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1172 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1173 if (InsertRes.second) {
1174 // First time we've looked at *PI.
1175 NewBlocks.push_back(*PI);
1179 // If we have seen this block before, but it was with a different
1180 // pointer then we have a phi translation failure and we have to treat
1181 // this as a clobber.
1182 if (InsertRes.first->second != Pointer.getAddr()) {
1183 // Make sure to clean up the Visited map before continuing on to
1184 // PredTranslationFailure.
1185 for (unsigned i = 0; i < NewBlocks.size(); i++)
1186 Visited.erase(NewBlocks[i]);
1187 goto PredTranslationFailure;
1190 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1194 // We do need to do phi translation, if we know ahead of time we can't phi
1195 // translate this value, don't even try.
1196 if (!Pointer.IsPotentiallyPHITranslatable())
1197 goto PredTranslationFailure;
1199 // We may have added values to the cache list before this PHI translation.
1200 // If so, we haven't done anything to ensure that the cache remains sorted.
1201 // Sort it now (if needed) so that recursive invocations of
1202 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1203 // value will only see properly sorted cache arrays.
1204 if (Cache && NumSortedEntries != Cache->size()) {
1205 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1206 NumSortedEntries = Cache->size();
1211 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1212 BasicBlock *Pred = *PI;
1213 PredList.push_back(std::make_pair(Pred, Pointer));
1215 // Get the PHI translated pointer in this predecessor. This can fail if
1216 // not translatable, in which case the getAddr() returns null.
1217 PHITransAddr &PredPointer = PredList.back().second;
1218 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1220 Value *PredPtrVal = PredPointer.getAddr();
1222 // Check to see if we have already visited this pred block with another
1223 // pointer. If so, we can't do this lookup. This failure can occur
1224 // with PHI translation when a critical edge exists and the PHI node in
1225 // the successor translates to a pointer value different than the
1226 // pointer the block was first analyzed with.
1227 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1228 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1230 if (!InsertRes.second) {
1231 // We found the pred; take it off the list of preds to visit.
1232 PredList.pop_back();
1234 // If the predecessor was visited with PredPtr, then we already did
1235 // the analysis and can ignore it.
1236 if (InsertRes.first->second == PredPtrVal)
1239 // Otherwise, the block was previously analyzed with a different
1240 // pointer. We can't represent the result of this case, so we just
1241 // treat this as a phi translation failure.
1243 // Make sure to clean up the Visited map before continuing on to
1244 // PredTranslationFailure.
1245 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1246 Visited.erase(PredList[i].first);
1248 goto PredTranslationFailure;
1252 // Actually process results here; this need to be a separate loop to avoid
1253 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1254 // any results for. (getNonLocalPointerDepFromBB will modify our
1255 // datastructures in ways the code after the PredTranslationFailure label
1257 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1258 BasicBlock *Pred = PredList[i].first;
1259 PHITransAddr &PredPointer = PredList[i].second;
1260 Value *PredPtrVal = PredPointer.getAddr();
1262 bool CanTranslate = true;
1263 // If PHI translation was unable to find an available pointer in this
1264 // predecessor, then we have to assume that the pointer is clobbered in
1265 // that predecessor. We can still do PRE of the load, which would insert
1266 // a computation of the pointer in this predecessor.
1268 CanTranslate = false;
1270 // FIXME: it is entirely possible that PHI translating will end up with
1271 // the same value. Consider PHI translating something like:
1272 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1273 // to recurse here, pedantically speaking.
1275 // If getNonLocalPointerDepFromBB fails here, that means the cached
1276 // result conflicted with the Visited list; we have to conservatively
1277 // assume it is unknown, but this also does not block PRE of the load.
1278 if (!CanTranslate ||
1279 getNonLocalPointerDepFromBB(PredPointer,
1280 Loc.getWithNewPtr(PredPtrVal),
1283 // Add the entry to the Result list.
1284 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1285 Result.push_back(Entry);
1287 // Since we had a phi translation failure, the cache for CacheKey won't
1288 // include all of the entries that we need to immediately satisfy future
1289 // queries. Mark this in NonLocalPointerDeps by setting the
1290 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1291 // cached value to do more work but not miss the phi trans failure.
1292 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1293 NLPI.Pair = BBSkipFirstBlockPair();
1298 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1299 CacheInfo = &NonLocalPointerDeps[CacheKey];
1300 Cache = &CacheInfo->NonLocalDeps;
1301 NumSortedEntries = Cache->size();
1303 // Since we did phi translation, the "Cache" set won't contain all of the
1304 // results for the query. This is ok (we can still use it to accelerate
1305 // specific block queries) but we can't do the fastpath "return all
1306 // results from the set" Clear out the indicator for this.
1307 CacheInfo->Pair = BBSkipFirstBlockPair();
1308 SkipFirstBlock = false;
1311 PredTranslationFailure:
1312 // The following code is "failure"; we can't produce a sane translation
1313 // for the given block. It assumes that we haven't modified any of
1314 // our datastructures while processing the current block.
1317 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1318 CacheInfo = &NonLocalPointerDeps[CacheKey];
1319 Cache = &CacheInfo->NonLocalDeps;
1320 NumSortedEntries = Cache->size();
1323 // Since we failed phi translation, the "Cache" set won't contain all of the
1324 // results for the query. This is ok (we can still use it to accelerate
1325 // specific block queries) but we can't do the fastpath "return all
1326 // results from the set". Clear out the indicator for this.
1327 CacheInfo->Pair = BBSkipFirstBlockPair();
1329 // If *nothing* works, mark the pointer as unknown.
1331 // If this is the magic first block, return this as a clobber of the whole
1332 // incoming value. Since we can't phi translate to one of the predecessors,
1333 // we have to bail out.
1337 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1338 assert(I != Cache->rend() && "Didn't find current block??");
1339 if (I->getBB() != BB)
1342 assert(I->getResult().isNonLocal() &&
1343 "Should only be here with transparent block");
1344 I->setResult(MemDepResult::getUnknown());
1345 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1346 Pointer.getAddr()));
1351 // Okay, we're done now. If we added new values to the cache, re-sort it.
1352 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1353 DEBUG(AssertSorted(*Cache));
1357 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1358 /// CachedNonLocalPointerInfo, remove it.
1359 void MemoryDependenceAnalysis::
1360 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1361 CachedNonLocalPointerInfo::iterator It =
1362 NonLocalPointerDeps.find(P);
1363 if (It == NonLocalPointerDeps.end()) return;
1365 // Remove all of the entries in the BB->val map. This involves removing
1366 // instructions from the reverse map.
1367 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1369 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1370 Instruction *Target = PInfo[i].getResult().getInst();
1371 if (!Target) continue; // Ignore non-local dep results.
1372 assert(Target->getParent() == PInfo[i].getBB());
1374 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1375 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1378 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1379 NonLocalPointerDeps.erase(It);
1383 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1384 /// information about the specified pointer, because it may be too
1385 /// conservative in memdep. This is an optional call that can be used when
1386 /// the client detects an equivalence between the pointer and some other
1387 /// value and replaces the other value with ptr. This can make Ptr available
1388 /// in more places that cached info does not necessarily keep.
1389 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1390 // If Ptr isn't really a pointer, just ignore it.
1391 if (!Ptr->getType()->isPointerTy()) return;
1392 // Flush store info for the pointer.
1393 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1394 // Flush load info for the pointer.
1395 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1398 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1399 /// This needs to be done when the CFG changes, e.g., due to splitting
1401 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1405 /// removeInstruction - Remove an instruction from the dependence analysis,
1406 /// updating the dependence of instructions that previously depended on it.
1407 /// This method attempts to keep the cache coherent using the reverse map.
1408 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1409 // Walk through the Non-local dependencies, removing this one as the value
1410 // for any cached queries.
1411 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1412 if (NLDI != NonLocalDeps.end()) {
1413 NonLocalDepInfo &BlockMap = NLDI->second.first;
1414 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1416 if (Instruction *Inst = DI->getResult().getInst())
1417 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1418 NonLocalDeps.erase(NLDI);
1421 // If we have a cached local dependence query for this instruction, remove it.
1423 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1424 if (LocalDepEntry != LocalDeps.end()) {
1425 // Remove us from DepInst's reverse set now that the local dep info is gone.
1426 if (Instruction *Inst = LocalDepEntry->second.getInst())
1427 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1429 // Remove this local dependency info.
1430 LocalDeps.erase(LocalDepEntry);
1433 // If we have any cached pointer dependencies on this instruction, remove
1434 // them. If the instruction has non-pointer type, then it can't be a pointer
1437 // Remove it from both the load info and the store info. The instruction
1438 // can't be in either of these maps if it is non-pointer.
1439 if (RemInst->getType()->isPointerTy()) {
1440 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1441 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1444 // Loop over all of the things that depend on the instruction we're removing.
1446 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1448 // If we find RemInst as a clobber or Def in any of the maps for other values,
1449 // we need to replace its entry with a dirty version of the instruction after
1450 // it. If RemInst is a terminator, we use a null dirty value.
1452 // Using a dirty version of the instruction after RemInst saves having to scan
1453 // the entire block to get to this point.
1454 MemDepResult NewDirtyVal;
1455 if (!RemInst->isTerminator())
1456 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1458 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1459 if (ReverseDepIt != ReverseLocalDeps.end()) {
1460 // RemInst can't be the terminator if it has local stuff depending on it.
1461 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1462 "Nothing can locally depend on a terminator");
1464 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1465 assert(InstDependingOnRemInst != RemInst &&
1466 "Already removed our local dep info");
1468 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1470 // Make sure to remember that new things depend on NewDepInst.
1471 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1472 "a local dep on this if it is a terminator!");
1473 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1474 InstDependingOnRemInst));
1477 ReverseLocalDeps.erase(ReverseDepIt);
1479 // Add new reverse deps after scanning the set, to avoid invalidating the
1480 // 'ReverseDeps' reference.
1481 while (!ReverseDepsToAdd.empty()) {
1482 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1483 .insert(ReverseDepsToAdd.back().second);
1484 ReverseDepsToAdd.pop_back();
1488 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1489 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1490 for (Instruction *I : ReverseDepIt->second) {
1491 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1493 PerInstNLInfo &INLD = NonLocalDeps[I];
1494 // The information is now dirty!
1497 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1498 DE = INLD.first.end(); DI != DE; ++DI) {
1499 if (DI->getResult().getInst() != RemInst) continue;
1501 // Convert to a dirty entry for the subsequent instruction.
1502 DI->setResult(NewDirtyVal);
1504 if (Instruction *NextI = NewDirtyVal.getInst())
1505 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1509 ReverseNonLocalDeps.erase(ReverseDepIt);
1511 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1512 while (!ReverseDepsToAdd.empty()) {
1513 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1514 .insert(ReverseDepsToAdd.back().second);
1515 ReverseDepsToAdd.pop_back();
1519 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1520 // value in the NonLocalPointerDeps info.
1521 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1522 ReverseNonLocalPtrDeps.find(RemInst);
1523 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1524 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1526 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1527 assert(P.getPointer() != RemInst &&
1528 "Already removed NonLocalPointerDeps info for RemInst");
1530 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1532 // The cache is not valid for any specific block anymore.
1533 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1535 // Update any entries for RemInst to use the instruction after it.
1536 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1538 if (DI->getResult().getInst() != RemInst) continue;
1540 // Convert to a dirty entry for the subsequent instruction.
1541 DI->setResult(NewDirtyVal);
1543 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1544 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1547 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1548 // subsequent value may invalidate the sortedness.
1549 std::sort(NLPDI.begin(), NLPDI.end());
1552 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1554 while (!ReversePtrDepsToAdd.empty()) {
1555 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1556 .insert(ReversePtrDepsToAdd.back().second);
1557 ReversePtrDepsToAdd.pop_back();
1562 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1563 AA->deleteValue(RemInst);
1564 DEBUG(verifyRemoved(RemInst));
1566 /// verifyRemoved - Verify that the specified instruction does not occur
1567 /// in our internal data structures. This function verifies by asserting in
1569 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1571 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1572 E = LocalDeps.end(); I != E; ++I) {
1573 assert(I->first != D && "Inst occurs in data structures");
1574 assert(I->second.getInst() != D &&
1575 "Inst occurs in data structures");
1578 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1579 E = NonLocalPointerDeps.end(); I != E; ++I) {
1580 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1581 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1582 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1584 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1587 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1588 E = NonLocalDeps.end(); I != E; ++I) {
1589 assert(I->first != D && "Inst occurs in data structures");
1590 const PerInstNLInfo &INLD = I->second;
1591 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1592 EE = INLD.first.end(); II != EE; ++II)
1593 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1596 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1597 E = ReverseLocalDeps.end(); I != E; ++I) {
1598 assert(I->first != D && "Inst occurs in data structures");
1599 for (Instruction *Inst : I->second)
1600 assert(Inst != D && "Inst occurs in data structures");
1603 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1604 E = ReverseNonLocalDeps.end();
1606 assert(I->first != D && "Inst occurs in data structures");
1607 for (Instruction *Inst : I->second)
1608 assert(Inst != D && "Inst occurs in data structures");
1611 for (ReverseNonLocalPtrDepTy::const_iterator
1612 I = ReverseNonLocalPtrDeps.begin(),
1613 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1614 assert(I->first != D && "Inst occurs in rev NLPD map");
1616 for (ValueIsLoadPair P : I->second)
1617 assert(P != ValueIsLoadPair(D, false) &&
1618 P != ValueIsLoadPair(D, true) &&
1619 "Inst occurs in ReverseNonLocalPtrDeps map");