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 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(AssumptionTracker)
63 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
64 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
65 "Memory Dependence Analysis", false, true)
67 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
68 : FunctionPass(ID), PredCache() {
69 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
71 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
74 /// Clean up memory in between runs
75 void MemoryDependenceAnalysis::releaseMemory() {
78 NonLocalPointerDeps.clear();
79 ReverseLocalDeps.clear();
80 ReverseNonLocalDeps.clear();
81 ReverseNonLocalPtrDeps.clear();
87 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
89 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
91 AU.addRequired<AssumptionTracker>();
92 AU.addRequiredTransitive<AliasAnalysis>();
95 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
96 AA = &getAnalysis<AliasAnalysis>();
97 AT = &getAnalysis<AssumptionTracker>();
98 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
99 DL = DLP ? &DLP->getDataLayout() : nullptr;
100 DominatorTreeWrapperPass *DTWP =
101 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
102 DT = DTWP ? &DTWP->getDomTree() : nullptr;
104 PredCache.reset(new PredIteratorCache());
108 /// RemoveFromReverseMap - This is a helper function that removes Val from
109 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
110 template <typename KeyTy>
111 static void RemoveFromReverseMap(DenseMap<Instruction*,
112 SmallPtrSet<KeyTy, 4> > &ReverseMap,
113 Instruction *Inst, KeyTy Val) {
114 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
115 InstIt = ReverseMap.find(Inst);
116 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
117 bool Found = InstIt->second.erase(Val);
118 assert(Found && "Invalid reverse map!"); (void)Found;
119 if (InstIt->second.empty())
120 ReverseMap.erase(InstIt);
123 /// GetLocation - If the given instruction references a specific memory
124 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
125 /// Return a ModRefInfo value describing the general behavior of the
128 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
129 AliasAnalysis::Location &Loc,
131 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
132 if (LI->isUnordered()) {
133 Loc = AA->getLocation(LI);
134 return AliasAnalysis::Ref;
136 if (LI->getOrdering() == Monotonic) {
137 Loc = AA->getLocation(LI);
138 return AliasAnalysis::ModRef;
140 Loc = AliasAnalysis::Location();
141 return AliasAnalysis::ModRef;
144 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
145 if (SI->isUnordered()) {
146 Loc = AA->getLocation(SI);
147 return AliasAnalysis::Mod;
149 if (SI->getOrdering() == Monotonic) {
150 Loc = AA->getLocation(SI);
151 return AliasAnalysis::ModRef;
153 Loc = AliasAnalysis::Location();
154 return AliasAnalysis::ModRef;
157 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
158 Loc = AA->getLocation(V);
159 return AliasAnalysis::ModRef;
162 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
163 // calls to free() deallocate the entire structure
164 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
165 return AliasAnalysis::Mod;
168 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
171 switch (II->getIntrinsicID()) {
172 case Intrinsic::lifetime_start:
173 case Intrinsic::lifetime_end:
174 case Intrinsic::invariant_start:
175 II->getAAMetadata(AAInfo);
176 Loc = AliasAnalysis::Location(II->getArgOperand(1),
177 cast<ConstantInt>(II->getArgOperand(0))
178 ->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;
182 case Intrinsic::invariant_end:
183 II->getAAMetadata(AAInfo);
184 Loc = AliasAnalysis::Location(II->getArgOperand(2),
185 cast<ConstantInt>(II->getArgOperand(1))
186 ->getZExtValue(), AAInfo);
187 // These intrinsics don't really modify the memory, but returning Mod
188 // will allow them to be handled conservatively.
189 return AliasAnalysis::Mod;
195 // Otherwise, just do the coarse-grained thing that always works.
196 if (Inst->mayWriteToMemory())
197 return AliasAnalysis::ModRef;
198 if (Inst->mayReadFromMemory())
199 return AliasAnalysis::Ref;
200 return AliasAnalysis::NoModRef;
203 /// getCallSiteDependencyFrom - Private helper for finding the local
204 /// dependencies of a call site.
205 MemDepResult MemoryDependenceAnalysis::
206 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
207 BasicBlock::iterator ScanIt, BasicBlock *BB) {
208 unsigned Limit = BlockScanLimit;
210 // Walk backwards through the block, looking for dependencies
211 while (ScanIt != BB->begin()) {
212 // Limit the amount of scanning we do so we don't end up with quadratic
213 // running time on extreme testcases.
216 return MemDepResult::getUnknown();
218 Instruction *Inst = --ScanIt;
220 // If this inst is a memory op, get the pointer it accessed
221 AliasAnalysis::Location Loc;
222 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
224 // A simple instruction.
225 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
226 return MemDepResult::getClobber(Inst);
230 if (CallSite InstCS = cast<Value>(Inst)) {
231 // Debug intrinsics don't cause dependences.
232 if (isa<DbgInfoIntrinsic>(Inst)) continue;
233 // If these two calls do not interfere, look past it.
234 switch (AA->getModRefInfo(CS, InstCS)) {
235 case AliasAnalysis::NoModRef:
236 // If the two calls are the same, return InstCS as a Def, so that
237 // CS can be found redundant and eliminated.
238 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
239 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
240 return MemDepResult::getDef(Inst);
242 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
246 return MemDepResult::getClobber(Inst);
250 // If we could not obtain a pointer for the instruction and the instruction
251 // touches memory then assume that this is a dependency.
252 if (MR != AliasAnalysis::NoModRef)
253 return MemDepResult::getClobber(Inst);
256 // No dependence found. If this is the entry block of the function, it is
257 // unknown, otherwise it is non-local.
258 if (BB != &BB->getParent()->getEntryBlock())
259 return MemDepResult::getNonLocal();
260 return MemDepResult::getNonFuncLocal();
263 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
264 /// would fully overlap MemLoc if done as a wider legal integer load.
266 /// MemLocBase, MemLocOffset are lazily computed here the first time the
267 /// base/offs of memloc is needed.
269 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
270 const Value *&MemLocBase,
273 const DataLayout *DL) {
274 // If we have no target data, we can't do this.
275 if (!DL) return false;
277 // If we haven't already computed the base/offset of MemLoc, do so now.
279 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
281 unsigned Size = MemoryDependenceAnalysis::
282 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
287 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
288 /// looks at a memory location for a load (specified by MemLocBase, Offs,
289 /// and Size) and compares it against a load. If the specified load could
290 /// be safely widened to a larger integer load that is 1) still efficient,
291 /// 2) safe for the target, and 3) would provide the specified memory
292 /// location value, then this function returns the size in bytes of the
293 /// load width to use. If not, this returns zero.
294 unsigned MemoryDependenceAnalysis::
295 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
296 unsigned MemLocSize, const LoadInst *LI,
297 const DataLayout &DL) {
298 // We can only extend simple integer loads.
299 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
301 // Load widening is hostile to ThreadSanitizer: it may cause false positives
302 // or make the reports more cryptic (access sizes are wrong).
303 if (LI->getParent()->getParent()->getAttributes().
304 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
307 // Get the base of this load.
309 const Value *LIBase =
310 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &DL);
312 // If the two pointers are not based on the same pointer, we can't tell that
314 if (LIBase != MemLocBase) return 0;
316 // Okay, the two values are based on the same pointer, but returned as
317 // no-alias. This happens when we have things like two byte loads at "P+1"
318 // and "P+3". Check to see if increasing the size of the "LI" load up to its
319 // alignment (or the largest native integer type) will allow us to load all
320 // the bits required by MemLoc.
322 // If MemLoc is before LI, then no widening of LI will help us out.
323 if (MemLocOffs < LIOffs) return 0;
325 // Get the alignment of the load in bytes. We assume that it is safe to load
326 // any legal integer up to this size without a problem. For example, if we're
327 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
328 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
330 unsigned LoadAlign = LI->getAlignment();
332 int64_t MemLocEnd = MemLocOffs+MemLocSize;
334 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
335 if (LIOffs+LoadAlign < MemLocEnd) return 0;
337 // This is the size of the load to try. Start with the next larger power of
339 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
340 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
343 // If this load size is bigger than our known alignment or would not fit
344 // into a native integer register, then we fail.
345 if (NewLoadByteSize > LoadAlign ||
346 !DL.fitsInLegalInteger(NewLoadByteSize*8))
349 if (LIOffs+NewLoadByteSize > MemLocEnd &&
350 LI->getParent()->getParent()->getAttributes().
351 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
352 // We will be reading past the location accessed by the original program.
353 // While this is safe in a regular build, Address Safety analysis tools
354 // may start reporting false warnings. So, don't do widening.
357 // If a load of this width would include all of MemLoc, then we succeed.
358 if (LIOffs+NewLoadByteSize >= MemLocEnd)
359 return NewLoadByteSize;
361 NewLoadByteSize <<= 1;
365 /// getPointerDependencyFrom - Return the instruction on which a memory
366 /// location depends. If isLoad is true, this routine ignores may-aliases with
367 /// read-only operations. If isLoad is false, this routine ignores may-aliases
368 /// with reads from read-only locations. If possible, pass the query
369 /// instruction as well; this function may take advantage of the metadata
370 /// annotated to the query instruction to refine the result.
371 MemDepResult MemoryDependenceAnalysis::
372 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
373 BasicBlock::iterator ScanIt, BasicBlock *BB,
374 Instruction *QueryInst) {
376 const Value *MemLocBase = nullptr;
377 int64_t MemLocOffset = 0;
378 unsigned Limit = BlockScanLimit;
379 bool isInvariantLoad = false;
381 // We must be careful with atomic accesses, as they may allow another thread
382 // to touch this location, cloberring it. We are conservative: if the
383 // QueryInst is not a simple (non-atomic) memory access, we automatically
384 // return getClobber.
385 // If it is simple, we know based on the results of
386 // "Compiler testing via a theory of sound optimisations in the C11/C++11
387 // memory model" in PLDI 2013, that a non-atomic location can only be
388 // clobbered between a pair of a release and an acquire action, with no
389 // access to the location in between.
390 // Here is an example for giving the general intuition behind this rule.
391 // In the following code:
393 // release action; [1]
394 // acquire action; [4]
396 // It is unsafe to replace %val by 0 because another thread may be running:
397 // acquire action; [2]
399 // release action; [3]
400 // with synchronization from 1 to 2 and from 3 to 4, resulting in %val
401 // being 42. A key property of this program however is that if either
402 // 1 or 4 were missing, there would be a race between the store of 42
403 // either the store of 0 or the load (making the whole progam racy).
404 // The paper mentionned above shows that the same property is respected
405 // by every program that can detect any optimisation of that kind: either
406 // it is racy (undefined) or there is a release followed by an acquire
407 // between the pair of accesses under consideration.
408 bool HasSeenAcquire = false;
410 if (isLoad && QueryInst) {
411 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
412 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
413 isInvariantLoad = true;
416 // Walk backwards through the basic block, looking for dependencies.
417 while (ScanIt != BB->begin()) {
418 Instruction *Inst = --ScanIt;
420 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
421 // Debug intrinsics don't (and can't) cause dependencies.
422 if (isa<DbgInfoIntrinsic>(II)) continue;
424 // Limit the amount of scanning we do so we don't end up with quadratic
425 // running time on extreme testcases.
428 return MemDepResult::getUnknown();
430 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
431 // If we reach a lifetime begin or end marker, then the query ends here
432 // because the value is undefined.
433 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
434 // FIXME: This only considers queries directly on the invariant-tagged
435 // pointer, not on query pointers that are indexed off of them. It'd
436 // be nice to handle that at some point (the right approach is to use
437 // GetPointerBaseWithConstantOffset).
438 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
440 return MemDepResult::getDef(II);
445 // Values depend on loads if the pointers are must aliased. This means that
446 // a load depends on another must aliased load from the same value.
447 // One exception is atomic loads: a value can depend on an atomic load that it
448 // does not alias with when this atomic load indicates that another thread may
449 // be accessing the location.
450 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
451 // Atomic loads have complications involved.
452 // A Monotonic (or higher) load is OK if the query inst is itself not atomic.
453 // An Acquire (or higher) load sets the HasSeenAcquire flag, so that any
454 // release store will know to return getClobber.
455 // FIXME: This is overly conservative.
456 if (!LI->isUnordered()) {
458 return MemDepResult::getClobber(LI);
459 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
460 if (!QueryLI->isSimple())
461 return MemDepResult::getClobber(LI);
462 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
463 if (!QuerySI->isSimple())
464 return MemDepResult::getClobber(LI);
465 } else if (QueryInst->mayReadOrWriteMemory()) {
466 return MemDepResult::getClobber(LI);
469 if (isAtLeastAcquire(LI->getOrdering()))
470 HasSeenAcquire = true;
473 // FIXME: this is overly conservative.
474 // While volatile access cannot be eliminated, they do not have to clobber
475 // non-aliasing locations, as normal accesses can for example be reordered
476 // with volatile accesses.
477 if (LI->isVolatile())
478 return MemDepResult::getClobber(LI);
480 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
482 // If we found a pointer, check if it could be the same as our pointer.
483 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
486 if (R == AliasAnalysis::NoAlias) {
487 // If this is an over-aligned integer load (for example,
488 // "load i8* %P, align 4") see if it would obviously overlap with the
489 // queried location if widened to a larger load (e.g. if the queried
490 // location is 1 byte at P+1). If so, return it as a load/load
491 // clobber result, allowing the client to decide to widen the load if
493 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
494 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
495 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
496 MemLocOffset, LI, DL))
497 return MemDepResult::getClobber(Inst);
502 // Must aliased loads are defs of each other.
503 if (R == AliasAnalysis::MustAlias)
504 return MemDepResult::getDef(Inst);
506 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
507 // in terms of clobbering loads, but since it does this by looking
508 // at the clobbering load directly, it doesn't know about any
509 // phi translation that may have happened along the way.
511 // If we have a partial alias, then return this as a clobber for the
513 if (R == AliasAnalysis::PartialAlias)
514 return MemDepResult::getClobber(Inst);
517 // Random may-alias loads don't depend on each other without a
522 // Stores don't depend on other no-aliased accesses.
523 if (R == AliasAnalysis::NoAlias)
526 // Stores don't alias loads from read-only memory.
527 if (AA->pointsToConstantMemory(LoadLoc))
530 // Stores depend on may/must aliased loads.
531 return MemDepResult::getDef(Inst);
534 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
535 // Atomic stores have complications involved.
536 // A Monotonic store is OK if the query inst is itself not atomic.
537 // A Release (or higher) store further requires that no acquire load
539 // FIXME: This is overly conservative.
540 if (!SI->isUnordered()) {
542 return MemDepResult::getClobber(SI);
543 if (auto *QueryLI = dyn_cast<LoadInst>(QueryInst)) {
544 if (!QueryLI->isSimple())
545 return MemDepResult::getClobber(SI);
546 } else if (auto *QuerySI = dyn_cast<StoreInst>(QueryInst)) {
547 if (!QuerySI->isSimple())
548 return MemDepResult::getClobber(SI);
549 } else if (QueryInst->mayReadOrWriteMemory()) {
550 return MemDepResult::getClobber(SI);
553 if (HasSeenAcquire && isAtLeastRelease(SI->getOrdering()))
554 return MemDepResult::getClobber(SI);
557 // FIXME: this is overly conservative.
558 // While volatile access cannot be eliminated, they do not have to clobber
559 // non-aliasing locations, as normal accesses can for example be reordered
560 // with volatile accesses.
561 if (SI->isVolatile())
562 return MemDepResult::getClobber(SI);
564 // If alias analysis can tell that this store is guaranteed to not modify
565 // the query pointer, ignore it. Use getModRefInfo to handle cases where
566 // the query pointer points to constant memory etc.
567 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
570 // Ok, this store might clobber the query pointer. Check to see if it is
571 // a must alias: in this case, we want to return this as a def.
572 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
574 // If we found a pointer, check if it could be the same as our pointer.
575 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
577 if (R == AliasAnalysis::NoAlias)
579 if (R == AliasAnalysis::MustAlias)
580 return MemDepResult::getDef(Inst);
583 return MemDepResult::getClobber(Inst);
586 // If this is an allocation, and if we know that the accessed pointer is to
587 // the allocation, return Def. This means that there is no dependence and
588 // the access can be optimized based on that. For example, a load could
590 // Note: Only determine this to be a malloc if Inst is the malloc call, not
591 // a subsequent bitcast of the malloc call result. There can be stores to
592 // the malloced memory between the malloc call and its bitcast uses, and we
593 // need to continue scanning until the malloc call.
594 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
595 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
596 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
598 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
599 return MemDepResult::getDef(Inst);
600 // Be conservative if the accessed pointer may alias the allocation.
601 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
602 return MemDepResult::getClobber(Inst);
603 // If the allocation is not aliased and does not read memory (like
604 // strdup), it is safe to ignore.
605 if (isa<AllocaInst>(Inst) ||
606 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
610 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
611 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
612 // If necessary, perform additional analysis.
613 if (MR == AliasAnalysis::ModRef)
614 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
616 case AliasAnalysis::NoModRef:
617 // If the call has no effect on the queried pointer, just ignore it.
619 case AliasAnalysis::Mod:
620 return MemDepResult::getClobber(Inst);
621 case AliasAnalysis::Ref:
622 // If the call is known to never store to the pointer, and if this is a
623 // load query, we can safely ignore it (scan past it).
627 // Otherwise, there is a potential dependence. Return a clobber.
628 return MemDepResult::getClobber(Inst);
632 // No dependence found. If this is the entry block of the function, it is
633 // unknown, otherwise it is non-local.
634 if (BB != &BB->getParent()->getEntryBlock())
635 return MemDepResult::getNonLocal();
636 return MemDepResult::getNonFuncLocal();
639 /// getDependency - Return the instruction on which a memory operation
641 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
642 Instruction *ScanPos = QueryInst;
644 // Check for a cached result
645 MemDepResult &LocalCache = LocalDeps[QueryInst];
647 // If the cached entry is non-dirty, just return it. Note that this depends
648 // on MemDepResult's default constructing to 'dirty'.
649 if (!LocalCache.isDirty())
652 // Otherwise, if we have a dirty entry, we know we can start the scan at that
653 // instruction, which may save us some work.
654 if (Instruction *Inst = LocalCache.getInst()) {
657 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
660 BasicBlock *QueryParent = QueryInst->getParent();
663 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
664 // No dependence found. If this is the entry block of the function, it is
665 // unknown, otherwise it is non-local.
666 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
667 LocalCache = MemDepResult::getNonLocal();
669 LocalCache = MemDepResult::getNonFuncLocal();
671 AliasAnalysis::Location MemLoc;
672 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
674 // If we can do a pointer scan, make it happen.
675 bool isLoad = !(MR & AliasAnalysis::Mod);
676 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
677 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
679 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
680 QueryParent, QueryInst);
681 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
682 CallSite QueryCS(QueryInst);
683 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
684 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
687 // Non-memory instruction.
688 LocalCache = MemDepResult::getUnknown();
691 // Remember the result!
692 if (Instruction *I = LocalCache.getInst())
693 ReverseLocalDeps[I].insert(QueryInst);
699 /// AssertSorted - This method is used when -debug is specified to verify that
700 /// cache arrays are properly kept sorted.
701 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
703 if (Count == -1) Count = Cache.size();
704 if (Count == 0) return;
706 for (unsigned i = 1; i != unsigned(Count); ++i)
707 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
711 /// getNonLocalCallDependency - Perform a full dependency query for the
712 /// specified call, returning the set of blocks that the value is
713 /// potentially live across. The returned set of results will include a
714 /// "NonLocal" result for all blocks where the value is live across.
716 /// This method assumes the instruction returns a "NonLocal" dependency
717 /// within its own block.
719 /// This returns a reference to an internal data structure that may be
720 /// invalidated on the next non-local query or when an instruction is
721 /// removed. Clients must copy this data if they want it around longer than
723 const MemoryDependenceAnalysis::NonLocalDepInfo &
724 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
725 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
726 "getNonLocalCallDependency should only be used on calls with non-local deps!");
727 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
728 NonLocalDepInfo &Cache = CacheP.first;
730 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
731 /// the cached case, this can happen due to instructions being deleted etc. In
732 /// the uncached case, this starts out as the set of predecessors we care
734 SmallVector<BasicBlock*, 32> DirtyBlocks;
736 if (!Cache.empty()) {
737 // Okay, we have a cache entry. If we know it is not dirty, just return it
738 // with no computation.
739 if (!CacheP.second) {
744 // If we already have a partially computed set of results, scan them to
745 // determine what is dirty, seeding our initial DirtyBlocks worklist.
746 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
748 if (I->getResult().isDirty())
749 DirtyBlocks.push_back(I->getBB());
751 // Sort the cache so that we can do fast binary search lookups below.
752 std::sort(Cache.begin(), Cache.end());
754 ++NumCacheDirtyNonLocal;
755 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
756 // << Cache.size() << " cached: " << *QueryInst;
758 // Seed DirtyBlocks with each of the preds of QueryInst's block.
759 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
760 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
761 DirtyBlocks.push_back(*PI);
762 ++NumUncacheNonLocal;
765 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
766 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
768 SmallPtrSet<BasicBlock*, 64> Visited;
770 unsigned NumSortedEntries = Cache.size();
771 DEBUG(AssertSorted(Cache));
773 // Iterate while we still have blocks to update.
774 while (!DirtyBlocks.empty()) {
775 BasicBlock *DirtyBB = DirtyBlocks.back();
776 DirtyBlocks.pop_back();
778 // Already processed this block?
779 if (!Visited.insert(DirtyBB))
782 // Do a binary search to see if we already have an entry for this block in
783 // the cache set. If so, find it.
784 DEBUG(AssertSorted(Cache, NumSortedEntries));
785 NonLocalDepInfo::iterator Entry =
786 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
787 NonLocalDepEntry(DirtyBB));
788 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
791 NonLocalDepEntry *ExistingResult = nullptr;
792 if (Entry != Cache.begin()+NumSortedEntries &&
793 Entry->getBB() == DirtyBB) {
794 // If we already have an entry, and if it isn't already dirty, the block
796 if (!Entry->getResult().isDirty())
799 // Otherwise, remember this slot so we can update the value.
800 ExistingResult = &*Entry;
803 // If the dirty entry has a pointer, start scanning from it so we don't have
804 // to rescan the entire block.
805 BasicBlock::iterator ScanPos = DirtyBB->end();
806 if (ExistingResult) {
807 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
809 // We're removing QueryInst's use of Inst.
810 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
811 QueryCS.getInstruction());
815 // Find out if this block has a local dependency for QueryInst.
818 if (ScanPos != DirtyBB->begin()) {
819 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
820 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
821 // No dependence found. If this is the entry block of the function, it is
822 // a clobber, otherwise it is unknown.
823 Dep = MemDepResult::getNonLocal();
825 Dep = MemDepResult::getNonFuncLocal();
828 // If we had a dirty entry for the block, update it. Otherwise, just add
831 ExistingResult->setResult(Dep);
833 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
835 // If the block has a dependency (i.e. it isn't completely transparent to
836 // the value), remember the association!
837 if (!Dep.isNonLocal()) {
838 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
839 // update this when we remove instructions.
840 if (Instruction *Inst = Dep.getInst())
841 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
844 // If the block *is* completely transparent to the load, we need to check
845 // the predecessors of this block. Add them to our worklist.
846 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
847 DirtyBlocks.push_back(*PI);
854 /// getNonLocalPointerDependency - Perform a full dependency query for an
855 /// access to the specified (non-volatile) memory location, returning the
856 /// set of instructions that either define or clobber the value.
858 /// This method assumes the pointer has a "NonLocal" dependency within its
861 void MemoryDependenceAnalysis::
862 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
864 SmallVectorImpl<NonLocalDepResult> &Result) {
865 assert(Loc.Ptr->getType()->isPointerTy() &&
866 "Can't get pointer deps of a non-pointer!");
869 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL, AT);
871 // This is the set of blocks we've inspected, and the pointer we consider in
872 // each block. Because of critical edges, we currently bail out if querying
873 // a block with multiple different pointers. This can happen during PHI
875 DenseMap<BasicBlock*, Value*> Visited;
876 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
877 Result, Visited, true))
880 Result.push_back(NonLocalDepResult(FromBB,
881 MemDepResult::getUnknown(),
882 const_cast<Value *>(Loc.Ptr)));
885 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
886 /// Pointer/PointeeSize using either cached information in Cache or by doing a
887 /// lookup (which may use dirty cache info if available). If we do a lookup,
888 /// add the result to the cache.
889 MemDepResult MemoryDependenceAnalysis::
890 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
891 bool isLoad, BasicBlock *BB,
892 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
894 // Do a binary search to see if we already have an entry for this block in
895 // the cache set. If so, find it.
896 NonLocalDepInfo::iterator Entry =
897 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
898 NonLocalDepEntry(BB));
899 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
902 NonLocalDepEntry *ExistingResult = nullptr;
903 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
904 ExistingResult = &*Entry;
906 // If we have a cached entry, and it is non-dirty, use it as the value for
908 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
909 ++NumCacheNonLocalPtr;
910 return ExistingResult->getResult();
913 // Otherwise, we have to scan for the value. If we have a dirty cache
914 // entry, start scanning from its position, otherwise we scan from the end
916 BasicBlock::iterator ScanPos = BB->end();
917 if (ExistingResult && ExistingResult->getResult().getInst()) {
918 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
919 "Instruction invalidated?");
920 ++NumCacheDirtyNonLocalPtr;
921 ScanPos = ExistingResult->getResult().getInst();
923 // Eliminating the dirty entry from 'Cache', so update the reverse info.
924 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
925 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
927 ++NumUncacheNonLocalPtr;
930 // Scan the block for the dependency.
931 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
933 // If we had a dirty entry for the block, update it. Otherwise, just add
936 ExistingResult->setResult(Dep);
938 Cache->push_back(NonLocalDepEntry(BB, Dep));
940 // If the block has a dependency (i.e. it isn't completely transparent to
941 // the value), remember the reverse association because we just added it
943 if (!Dep.isDef() && !Dep.isClobber())
946 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
947 // update MemDep when we remove instructions.
948 Instruction *Inst = Dep.getInst();
949 assert(Inst && "Didn't depend on anything?");
950 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
951 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
955 /// SortNonLocalDepInfoCache - Sort the NonLocalDepInfo cache, given a certain
956 /// number of elements in the array that are already properly ordered. This is
957 /// optimized for the case when only a few entries are added.
959 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
960 unsigned NumSortedEntries) {
961 switch (Cache.size() - NumSortedEntries) {
963 // done, no new entries.
966 // Two new entries, insert the last one into place.
967 NonLocalDepEntry Val = Cache.back();
969 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
970 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
971 Cache.insert(Entry, Val);
975 // One new entry, Just insert the new value at the appropriate position.
976 if (Cache.size() != 1) {
977 NonLocalDepEntry Val = Cache.back();
979 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
980 std::upper_bound(Cache.begin(), Cache.end(), Val);
981 Cache.insert(Entry, Val);
985 // Added many values, do a full scale sort.
986 std::sort(Cache.begin(), Cache.end());
991 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
992 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
993 /// results to the results vector and keep track of which blocks are visited in
996 /// This has special behavior for the first block queries (when SkipFirstBlock
997 /// is true). In this special case, it ignores the contents of the specified
998 /// block and starts returning dependence info for its predecessors.
1000 /// This function returns false on success, or true to indicate that it could
1001 /// not compute dependence information for some reason. This should be treated
1002 /// as a clobber dependence on the first instruction in the predecessor block.
1003 bool MemoryDependenceAnalysis::
1004 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
1005 const AliasAnalysis::Location &Loc,
1006 bool isLoad, BasicBlock *StartBB,
1007 SmallVectorImpl<NonLocalDepResult> &Result,
1008 DenseMap<BasicBlock*, Value*> &Visited,
1009 bool SkipFirstBlock) {
1010 // Look up the cached info for Pointer.
1011 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
1013 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
1014 // CacheKey, this value will be inserted as the associated value. Otherwise,
1015 // it'll be ignored, and we'll have to check to see if the cached size and
1016 // aa tags are consistent with the current query.
1017 NonLocalPointerInfo InitialNLPI;
1018 InitialNLPI.Size = Loc.Size;
1019 InitialNLPI.AATags = Loc.AATags;
1021 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
1022 // already have one.
1023 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
1024 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
1025 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
1027 // If we already have a cache entry for this CacheKey, we may need to do some
1028 // work to reconcile the cache entry and the current query.
1030 if (CacheInfo->Size < Loc.Size) {
1031 // The query's Size is greater than the cached one. Throw out the
1032 // cached data and proceed with the query at the greater size.
1033 CacheInfo->Pair = BBSkipFirstBlockPair();
1034 CacheInfo->Size = Loc.Size;
1035 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1036 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1037 if (Instruction *Inst = DI->getResult().getInst())
1038 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1039 CacheInfo->NonLocalDeps.clear();
1040 } else if (CacheInfo->Size > Loc.Size) {
1041 // This query's Size is less than the cached one. Conservatively restart
1042 // the query using the greater size.
1043 return getNonLocalPointerDepFromBB(Pointer,
1044 Loc.getWithNewSize(CacheInfo->Size),
1045 isLoad, StartBB, Result, Visited,
1049 // If the query's AATags are inconsistent with the cached one,
1050 // conservatively throw out the cached data and restart the query with
1051 // no tag if needed.
1052 if (CacheInfo->AATags != Loc.AATags) {
1053 if (CacheInfo->AATags) {
1054 CacheInfo->Pair = BBSkipFirstBlockPair();
1055 CacheInfo->AATags = AAMDNodes();
1056 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
1057 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
1058 if (Instruction *Inst = DI->getResult().getInst())
1059 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
1060 CacheInfo->NonLocalDeps.clear();
1063 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutAATags(),
1064 isLoad, StartBB, Result, Visited,
1069 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
1071 // If we have valid cached information for exactly the block we are
1072 // investigating, just return it with no recomputation.
1073 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
1074 // We have a fully cached result for this query then we can just return the
1075 // cached results and populate the visited set. However, we have to verify
1076 // that we don't already have conflicting results for these blocks. Check
1077 // to ensure that if a block in the results set is in the visited set that
1078 // it was for the same pointer query.
1079 if (!Visited.empty()) {
1080 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1082 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
1083 if (VI == Visited.end() || VI->second == Pointer.getAddr())
1086 // We have a pointer mismatch in a block. Just return clobber, saying
1087 // that something was clobbered in this result. We could also do a
1088 // non-fully cached query, but there is little point in doing this.
1093 Value *Addr = Pointer.getAddr();
1094 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1096 Visited.insert(std::make_pair(I->getBB(), Addr));
1097 if (I->getResult().isNonLocal()) {
1102 Result.push_back(NonLocalDepResult(I->getBB(),
1103 MemDepResult::getUnknown(),
1105 } else if (DT->isReachableFromEntry(I->getBB())) {
1106 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1109 ++NumCacheCompleteNonLocalPtr;
1113 // Otherwise, either this is a new block, a block with an invalid cache
1114 // pointer or one that we're about to invalidate by putting more info into it
1115 // than its valid cache info. If empty, the result will be valid cache info,
1116 // otherwise it isn't.
1118 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1120 CacheInfo->Pair = BBSkipFirstBlockPair();
1122 SmallVector<BasicBlock*, 32> Worklist;
1123 Worklist.push_back(StartBB);
1125 // PredList used inside loop.
1126 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1128 // Keep track of the entries that we know are sorted. Previously cached
1129 // entries will all be sorted. The entries we add we only sort on demand (we
1130 // don't insert every element into its sorted position). We know that we
1131 // won't get any reuse from currently inserted values, because we don't
1132 // revisit blocks after we insert info for them.
1133 unsigned NumSortedEntries = Cache->size();
1134 DEBUG(AssertSorted(*Cache));
1136 while (!Worklist.empty()) {
1137 BasicBlock *BB = Worklist.pop_back_val();
1139 // If we do process a large number of blocks it becomes very expensive and
1140 // likely it isn't worth worrying about
1141 if (Result.size() > NumResultsLimit) {
1143 // Sort it now (if needed) so that recursive invocations of
1144 // getNonLocalPointerDepFromBB and other routines that could reuse the
1145 // cache value will only see properly sorted cache arrays.
1146 if (Cache && NumSortedEntries != Cache->size()) {
1147 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1148 NumSortedEntries = Cache->size();
1150 // Since we bail out, the "Cache" set won't contain all of the
1151 // results for the query. This is ok (we can still use it to accelerate
1152 // specific block queries) but we can't do the fastpath "return all
1153 // results from the set". Clear out the indicator for this.
1154 CacheInfo->Pair = BBSkipFirstBlockPair();
1158 // Skip the first block if we have it.
1159 if (!SkipFirstBlock) {
1160 // Analyze the dependency of *Pointer in FromBB. See if we already have
1162 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1164 // Get the dependency info for Pointer in BB. If we have cached
1165 // information, we will use it, otherwise we compute it.
1166 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1167 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1170 // If we got a Def or Clobber, add this to the list of results.
1171 if (!Dep.isNonLocal()) {
1173 Result.push_back(NonLocalDepResult(BB,
1174 MemDepResult::getUnknown(),
1175 Pointer.getAddr()));
1177 } else if (DT->isReachableFromEntry(BB)) {
1178 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1184 // If 'Pointer' is an instruction defined in this block, then we need to do
1185 // phi translation to change it into a value live in the predecessor block.
1186 // If not, we just add the predecessors to the worklist and scan them with
1187 // the same Pointer.
1188 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1189 SkipFirstBlock = false;
1190 SmallVector<BasicBlock*, 16> NewBlocks;
1191 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1192 // Verify that we haven't looked at this block yet.
1193 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1194 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1195 if (InsertRes.second) {
1196 // First time we've looked at *PI.
1197 NewBlocks.push_back(*PI);
1201 // If we have seen this block before, but it was with a different
1202 // pointer then we have a phi translation failure and we have to treat
1203 // this as a clobber.
1204 if (InsertRes.first->second != Pointer.getAddr()) {
1205 // Make sure to clean up the Visited map before continuing on to
1206 // PredTranslationFailure.
1207 for (unsigned i = 0; i < NewBlocks.size(); i++)
1208 Visited.erase(NewBlocks[i]);
1209 goto PredTranslationFailure;
1212 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1216 // We do need to do phi translation, if we know ahead of time we can't phi
1217 // translate this value, don't even try.
1218 if (!Pointer.IsPotentiallyPHITranslatable())
1219 goto PredTranslationFailure;
1221 // We may have added values to the cache list before this PHI translation.
1222 // If so, we haven't done anything to ensure that the cache remains sorted.
1223 // Sort it now (if needed) so that recursive invocations of
1224 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1225 // value will only see properly sorted cache arrays.
1226 if (Cache && NumSortedEntries != Cache->size()) {
1227 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1228 NumSortedEntries = Cache->size();
1233 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1234 BasicBlock *Pred = *PI;
1235 PredList.push_back(std::make_pair(Pred, Pointer));
1237 // Get the PHI translated pointer in this predecessor. This can fail if
1238 // not translatable, in which case the getAddr() returns null.
1239 PHITransAddr &PredPointer = PredList.back().second;
1240 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1242 Value *PredPtrVal = PredPointer.getAddr();
1244 // Check to see if we have already visited this pred block with another
1245 // pointer. If so, we can't do this lookup. This failure can occur
1246 // with PHI translation when a critical edge exists and the PHI node in
1247 // the successor translates to a pointer value different than the
1248 // pointer the block was first analyzed with.
1249 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1250 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1252 if (!InsertRes.second) {
1253 // We found the pred; take it off the list of preds to visit.
1254 PredList.pop_back();
1256 // If the predecessor was visited with PredPtr, then we already did
1257 // the analysis and can ignore it.
1258 if (InsertRes.first->second == PredPtrVal)
1261 // Otherwise, the block was previously analyzed with a different
1262 // pointer. We can't represent the result of this case, so we just
1263 // treat this as a phi translation failure.
1265 // Make sure to clean up the Visited map before continuing on to
1266 // PredTranslationFailure.
1267 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1268 Visited.erase(PredList[i].first);
1270 goto PredTranslationFailure;
1274 // Actually process results here; this need to be a separate loop to avoid
1275 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1276 // any results for. (getNonLocalPointerDepFromBB will modify our
1277 // datastructures in ways the code after the PredTranslationFailure label
1279 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1280 BasicBlock *Pred = PredList[i].first;
1281 PHITransAddr &PredPointer = PredList[i].second;
1282 Value *PredPtrVal = PredPointer.getAddr();
1284 bool CanTranslate = true;
1285 // If PHI translation was unable to find an available pointer in this
1286 // predecessor, then we have to assume that the pointer is clobbered in
1287 // that predecessor. We can still do PRE of the load, which would insert
1288 // a computation of the pointer in this predecessor.
1290 CanTranslate = false;
1292 // FIXME: it is entirely possible that PHI translating will end up with
1293 // the same value. Consider PHI translating something like:
1294 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1295 // to recurse here, pedantically speaking.
1297 // If getNonLocalPointerDepFromBB fails here, that means the cached
1298 // result conflicted with the Visited list; we have to conservatively
1299 // assume it is unknown, but this also does not block PRE of the load.
1300 if (!CanTranslate ||
1301 getNonLocalPointerDepFromBB(PredPointer,
1302 Loc.getWithNewPtr(PredPtrVal),
1305 // Add the entry to the Result list.
1306 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1307 Result.push_back(Entry);
1309 // Since we had a phi translation failure, the cache for CacheKey won't
1310 // include all of the entries that we need to immediately satisfy future
1311 // queries. Mark this in NonLocalPointerDeps by setting the
1312 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1313 // cached value to do more work but not miss the phi trans failure.
1314 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1315 NLPI.Pair = BBSkipFirstBlockPair();
1320 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1321 CacheInfo = &NonLocalPointerDeps[CacheKey];
1322 Cache = &CacheInfo->NonLocalDeps;
1323 NumSortedEntries = Cache->size();
1325 // Since we did phi translation, the "Cache" set won't contain all of the
1326 // results for the query. This is ok (we can still use it to accelerate
1327 // specific block queries) but we can't do the fastpath "return all
1328 // results from the set" Clear out the indicator for this.
1329 CacheInfo->Pair = BBSkipFirstBlockPair();
1330 SkipFirstBlock = false;
1333 PredTranslationFailure:
1334 // The following code is "failure"; we can't produce a sane translation
1335 // for the given block. It assumes that we haven't modified any of
1336 // our datastructures while processing the current block.
1339 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1340 CacheInfo = &NonLocalPointerDeps[CacheKey];
1341 Cache = &CacheInfo->NonLocalDeps;
1342 NumSortedEntries = Cache->size();
1345 // Since we failed phi translation, the "Cache" set won't contain all of the
1346 // results for the query. This is ok (we can still use it to accelerate
1347 // specific block queries) but we can't do the fastpath "return all
1348 // results from the set". Clear out the indicator for this.
1349 CacheInfo->Pair = BBSkipFirstBlockPair();
1351 // If *nothing* works, mark the pointer as unknown.
1353 // If this is the magic first block, return this as a clobber of the whole
1354 // incoming value. Since we can't phi translate to one of the predecessors,
1355 // we have to bail out.
1359 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1360 assert(I != Cache->rend() && "Didn't find current block??");
1361 if (I->getBB() != BB)
1364 assert(I->getResult().isNonLocal() &&
1365 "Should only be here with transparent block");
1366 I->setResult(MemDepResult::getUnknown());
1367 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1368 Pointer.getAddr()));
1373 // Okay, we're done now. If we added new values to the cache, re-sort it.
1374 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1375 DEBUG(AssertSorted(*Cache));
1379 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1380 /// CachedNonLocalPointerInfo, remove it.
1381 void MemoryDependenceAnalysis::
1382 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1383 CachedNonLocalPointerInfo::iterator It =
1384 NonLocalPointerDeps.find(P);
1385 if (It == NonLocalPointerDeps.end()) return;
1387 // Remove all of the entries in the BB->val map. This involves removing
1388 // instructions from the reverse map.
1389 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1391 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1392 Instruction *Target = PInfo[i].getResult().getInst();
1393 if (!Target) continue; // Ignore non-local dep results.
1394 assert(Target->getParent() == PInfo[i].getBB());
1396 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1397 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1400 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1401 NonLocalPointerDeps.erase(It);
1405 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1406 /// information about the specified pointer, because it may be too
1407 /// conservative in memdep. This is an optional call that can be used when
1408 /// the client detects an equivalence between the pointer and some other
1409 /// value and replaces the other value with ptr. This can make Ptr available
1410 /// in more places that cached info does not necessarily keep.
1411 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1412 // If Ptr isn't really a pointer, just ignore it.
1413 if (!Ptr->getType()->isPointerTy()) return;
1414 // Flush store info for the pointer.
1415 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1416 // Flush load info for the pointer.
1417 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1420 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1421 /// This needs to be done when the CFG changes, e.g., due to splitting
1423 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1427 /// removeInstruction - Remove an instruction from the dependence analysis,
1428 /// updating the dependence of instructions that previously depended on it.
1429 /// This method attempts to keep the cache coherent using the reverse map.
1430 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1431 // Walk through the Non-local dependencies, removing this one as the value
1432 // for any cached queries.
1433 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1434 if (NLDI != NonLocalDeps.end()) {
1435 NonLocalDepInfo &BlockMap = NLDI->second.first;
1436 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1438 if (Instruction *Inst = DI->getResult().getInst())
1439 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1440 NonLocalDeps.erase(NLDI);
1443 // If we have a cached local dependence query for this instruction, remove it.
1445 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1446 if (LocalDepEntry != LocalDeps.end()) {
1447 // Remove us from DepInst's reverse set now that the local dep info is gone.
1448 if (Instruction *Inst = LocalDepEntry->second.getInst())
1449 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1451 // Remove this local dependency info.
1452 LocalDeps.erase(LocalDepEntry);
1455 // If we have any cached pointer dependencies on this instruction, remove
1456 // them. If the instruction has non-pointer type, then it can't be a pointer
1459 // Remove it from both the load info and the store info. The instruction
1460 // can't be in either of these maps if it is non-pointer.
1461 if (RemInst->getType()->isPointerTy()) {
1462 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1463 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1466 // Loop over all of the things that depend on the instruction we're removing.
1468 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1470 // If we find RemInst as a clobber or Def in any of the maps for other values,
1471 // we need to replace its entry with a dirty version of the instruction after
1472 // it. If RemInst is a terminator, we use a null dirty value.
1474 // Using a dirty version of the instruction after RemInst saves having to scan
1475 // the entire block to get to this point.
1476 MemDepResult NewDirtyVal;
1477 if (!RemInst->isTerminator())
1478 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1480 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1481 if (ReverseDepIt != ReverseLocalDeps.end()) {
1482 // RemInst can't be the terminator if it has local stuff depending on it.
1483 assert(!ReverseDepIt->second.empty() && !isa<TerminatorInst>(RemInst) &&
1484 "Nothing can locally depend on a terminator");
1486 for (Instruction *InstDependingOnRemInst : ReverseDepIt->second) {
1487 assert(InstDependingOnRemInst != RemInst &&
1488 "Already removed our local dep info");
1490 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1492 // Make sure to remember that new things depend on NewDepInst.
1493 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1494 "a local dep on this if it is a terminator!");
1495 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1496 InstDependingOnRemInst));
1499 ReverseLocalDeps.erase(ReverseDepIt);
1501 // Add new reverse deps after scanning the set, to avoid invalidating the
1502 // 'ReverseDeps' reference.
1503 while (!ReverseDepsToAdd.empty()) {
1504 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1505 .insert(ReverseDepsToAdd.back().second);
1506 ReverseDepsToAdd.pop_back();
1510 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1511 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1512 for (Instruction *I : ReverseDepIt->second) {
1513 assert(I != RemInst && "Already removed NonLocalDep info for RemInst");
1515 PerInstNLInfo &INLD = NonLocalDeps[I];
1516 // The information is now dirty!
1519 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1520 DE = INLD.first.end(); DI != DE; ++DI) {
1521 if (DI->getResult().getInst() != RemInst) continue;
1523 // Convert to a dirty entry for the subsequent instruction.
1524 DI->setResult(NewDirtyVal);
1526 if (Instruction *NextI = NewDirtyVal.getInst())
1527 ReverseDepsToAdd.push_back(std::make_pair(NextI, I));
1531 ReverseNonLocalDeps.erase(ReverseDepIt);
1533 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1534 while (!ReverseDepsToAdd.empty()) {
1535 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1536 .insert(ReverseDepsToAdd.back().second);
1537 ReverseDepsToAdd.pop_back();
1541 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1542 // value in the NonLocalPointerDeps info.
1543 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1544 ReverseNonLocalPtrDeps.find(RemInst);
1545 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1546 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1548 for (ValueIsLoadPair P : ReversePtrDepIt->second) {
1549 assert(P.getPointer() != RemInst &&
1550 "Already removed NonLocalPointerDeps info for RemInst");
1552 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1554 // The cache is not valid for any specific block anymore.
1555 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1557 // Update any entries for RemInst to use the instruction after it.
1558 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1560 if (DI->getResult().getInst() != RemInst) continue;
1562 // Convert to a dirty entry for the subsequent instruction.
1563 DI->setResult(NewDirtyVal);
1565 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1566 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1569 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1570 // subsequent value may invalidate the sortedness.
1571 std::sort(NLPDI.begin(), NLPDI.end());
1574 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1576 while (!ReversePtrDepsToAdd.empty()) {
1577 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1578 .insert(ReversePtrDepsToAdd.back().second);
1579 ReversePtrDepsToAdd.pop_back();
1584 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1585 AA->deleteValue(RemInst);
1586 DEBUG(verifyRemoved(RemInst));
1588 /// verifyRemoved - Verify that the specified instruction does not occur
1589 /// in our internal data structures. This function verifies by asserting in
1591 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1593 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1594 E = LocalDeps.end(); I != E; ++I) {
1595 assert(I->first != D && "Inst occurs in data structures");
1596 assert(I->second.getInst() != D &&
1597 "Inst occurs in data structures");
1600 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1601 E = NonLocalPointerDeps.end(); I != E; ++I) {
1602 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1603 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1604 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1606 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1609 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1610 E = NonLocalDeps.end(); I != E; ++I) {
1611 assert(I->first != D && "Inst occurs in data structures");
1612 const PerInstNLInfo &INLD = I->second;
1613 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1614 EE = INLD.first.end(); II != EE; ++II)
1615 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1618 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1619 E = ReverseLocalDeps.end(); I != E; ++I) {
1620 assert(I->first != D && "Inst occurs in data structures");
1621 for (Instruction *Inst : I->second)
1622 assert(Inst != D && "Inst occurs in data structures");
1625 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1626 E = ReverseNonLocalDeps.end();
1628 assert(I->first != D && "Inst occurs in data structures");
1629 for (Instruction *Inst : I->second)
1630 assert(Inst != D && "Inst occurs in data structures");
1633 for (ReverseNonLocalPtrDepTy::const_iterator
1634 I = ReverseNonLocalPtrDeps.begin(),
1635 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1636 assert(I->first != D && "Inst occurs in rev NLPD map");
1638 for (ValueIsLoadPair P : I->second)
1639 assert(P != ValueIsLoadPair(D, false) &&
1640 P != ValueIsLoadPair(D, true) &&
1641 "Inst occurs in ReverseNonLocalPtrDeps map");