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/InstructionSimplify.h"
22 #include "llvm/Analysis/MemoryBuiltins.h"
23 #include "llvm/Analysis/PHITransAddr.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Instructions.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/IR/PredIteratorCache.h"
32 #include "llvm/Support/Debug.h"
35 #define DEBUG_TYPE "memdep"
37 STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
38 STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
39 STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
41 STATISTIC(NumCacheNonLocalPtr,
42 "Number of fully cached non-local ptr responses");
43 STATISTIC(NumCacheDirtyNonLocalPtr,
44 "Number of cached, but dirty, non-local ptr responses");
45 STATISTIC(NumUncacheNonLocalPtr,
46 "Number of uncached non-local ptr responses");
47 STATISTIC(NumCacheCompleteNonLocalPtr,
48 "Number of block queries that were completely cached");
50 // Limit for the number of instructions to scan in a block.
51 static const int BlockScanLimit = 100;
53 char MemoryDependenceAnalysis::ID = 0;
55 // Register this pass...
56 INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep",
57 "Memory Dependence Analysis", false, true)
58 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
59 INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep",
60 "Memory Dependence Analysis", false, true)
62 MemoryDependenceAnalysis::MemoryDependenceAnalysis()
63 : FunctionPass(ID), PredCache() {
64 initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry());
66 MemoryDependenceAnalysis::~MemoryDependenceAnalysis() {
69 /// Clean up memory in between runs
70 void MemoryDependenceAnalysis::releaseMemory() {
73 NonLocalPointerDeps.clear();
74 ReverseLocalDeps.clear();
75 ReverseNonLocalDeps.clear();
76 ReverseNonLocalPtrDeps.clear();
82 /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis.
84 void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
86 AU.addRequiredTransitive<AliasAnalysis>();
89 bool MemoryDependenceAnalysis::runOnFunction(Function &) {
90 AA = &getAnalysis<AliasAnalysis>();
91 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
92 DL = DLP ? &DLP->getDataLayout() : nullptr;
93 DominatorTreeWrapperPass *DTWP =
94 getAnalysisIfAvailable<DominatorTreeWrapperPass>();
95 DT = DTWP ? &DTWP->getDomTree() : nullptr;
97 PredCache.reset(new PredIteratorCache());
101 /// RemoveFromReverseMap - This is a helper function that removes Val from
102 /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
103 template <typename KeyTy>
104 static void RemoveFromReverseMap(DenseMap<Instruction*,
105 SmallPtrSet<KeyTy, 4> > &ReverseMap,
106 Instruction *Inst, KeyTy Val) {
107 typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::iterator
108 InstIt = ReverseMap.find(Inst);
109 assert(InstIt != ReverseMap.end() && "Reverse map out of sync?");
110 bool Found = InstIt->second.erase(Val);
111 assert(Found && "Invalid reverse map!"); (void)Found;
112 if (InstIt->second.empty())
113 ReverseMap.erase(InstIt);
116 /// GetLocation - If the given instruction references a specific memory
117 /// location, fill in Loc with the details, otherwise set Loc.Ptr to null.
118 /// Return a ModRefInfo value describing the general behavior of the
121 AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst,
122 AliasAnalysis::Location &Loc,
124 if (const LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
125 if (LI->isUnordered()) {
126 Loc = AA->getLocation(LI);
127 return AliasAnalysis::Ref;
129 if (LI->getOrdering() == Monotonic) {
130 Loc = AA->getLocation(LI);
131 return AliasAnalysis::ModRef;
133 Loc = AliasAnalysis::Location();
134 return AliasAnalysis::ModRef;
137 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
138 if (SI->isUnordered()) {
139 Loc = AA->getLocation(SI);
140 return AliasAnalysis::Mod;
142 if (SI->getOrdering() == Monotonic) {
143 Loc = AA->getLocation(SI);
144 return AliasAnalysis::ModRef;
146 Loc = AliasAnalysis::Location();
147 return AliasAnalysis::ModRef;
150 if (const VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
151 Loc = AA->getLocation(V);
152 return AliasAnalysis::ModRef;
155 if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) {
156 // calls to free() deallocate the entire structure
157 Loc = AliasAnalysis::Location(CI->getArgOperand(0));
158 return AliasAnalysis::Mod;
161 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
162 switch (II->getIntrinsicID()) {
163 case Intrinsic::lifetime_start:
164 case Intrinsic::lifetime_end:
165 case Intrinsic::invariant_start:
166 Loc = AliasAnalysis::Location(II->getArgOperand(1),
167 cast<ConstantInt>(II->getArgOperand(0))
169 II->getMetadata(LLVMContext::MD_tbaa));
170 // These intrinsics don't really modify the memory, but returning Mod
171 // will allow them to be handled conservatively.
172 return AliasAnalysis::Mod;
173 case Intrinsic::invariant_end:
174 Loc = AliasAnalysis::Location(II->getArgOperand(2),
175 cast<ConstantInt>(II->getArgOperand(1))
177 II->getMetadata(LLVMContext::MD_tbaa));
178 // These intrinsics don't really modify the memory, but returning Mod
179 // will allow them to be handled conservatively.
180 return AliasAnalysis::Mod;
185 // Otherwise, just do the coarse-grained thing that always works.
186 if (Inst->mayWriteToMemory())
187 return AliasAnalysis::ModRef;
188 if (Inst->mayReadFromMemory())
189 return AliasAnalysis::Ref;
190 return AliasAnalysis::NoModRef;
193 /// getCallSiteDependencyFrom - Private helper for finding the local
194 /// dependencies of a call site.
195 MemDepResult MemoryDependenceAnalysis::
196 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
197 BasicBlock::iterator ScanIt, BasicBlock *BB) {
198 unsigned Limit = BlockScanLimit;
200 // Walk backwards through the block, looking for dependencies
201 while (ScanIt != BB->begin()) {
202 // Limit the amount of scanning we do so we don't end up with quadratic
203 // running time on extreme testcases.
206 return MemDepResult::getUnknown();
208 Instruction *Inst = --ScanIt;
210 // If this inst is a memory op, get the pointer it accessed
211 AliasAnalysis::Location Loc;
212 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
214 // A simple instruction.
215 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
216 return MemDepResult::getClobber(Inst);
220 if (CallSite InstCS = cast<Value>(Inst)) {
221 // Debug intrinsics don't cause dependences.
222 if (isa<DbgInfoIntrinsic>(Inst)) continue;
223 // If these two calls do not interfere, look past it.
224 switch (AA->getModRefInfo(CS, InstCS)) {
225 case AliasAnalysis::NoModRef:
226 // If the two calls are the same, return InstCS as a Def, so that
227 // CS can be found redundant and eliminated.
228 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
229 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
230 return MemDepResult::getDef(Inst);
232 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
236 return MemDepResult::getClobber(Inst);
240 // If we could not obtain a pointer for the instruction and the instruction
241 // touches memory then assume that this is a dependency.
242 if (MR != AliasAnalysis::NoModRef)
243 return MemDepResult::getClobber(Inst);
246 // No dependence found. If this is the entry block of the function, it is
247 // unknown, otherwise it is non-local.
248 if (BB != &BB->getParent()->getEntryBlock())
249 return MemDepResult::getNonLocal();
250 return MemDepResult::getNonFuncLocal();
253 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
254 /// would fully overlap MemLoc if done as a wider legal integer load.
256 /// MemLocBase, MemLocOffset are lazily computed here the first time the
257 /// base/offs of memloc is needed.
259 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
260 const Value *&MemLocBase,
263 const DataLayout *DL) {
264 // If we have no target data, we can't do this.
265 if (!DL) return false;
267 // If we haven't already computed the base/offset of MemLoc, do so now.
269 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
271 unsigned Size = MemoryDependenceAnalysis::
272 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
277 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
278 /// looks at a memory location for a load (specified by MemLocBase, Offs,
279 /// and Size) and compares it against a load. If the specified load could
280 /// be safely widened to a larger integer load that is 1) still efficient,
281 /// 2) safe for the target, and 3) would provide the specified memory
282 /// location value, then this function returns the size in bytes of the
283 /// load width to use. If not, this returns zero.
284 unsigned MemoryDependenceAnalysis::
285 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
286 unsigned MemLocSize, const LoadInst *LI,
287 const DataLayout &DL) {
288 // We can only extend simple integer loads.
289 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
291 // Load widening is hostile to ThreadSanitizer: it may cause false positives
292 // or make the reports more cryptic (access sizes are wrong).
293 if (LI->getParent()->getParent()->getAttributes().
294 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
297 // Get the base of this load.
299 const Value *LIBase =
300 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &DL);
302 // If the two pointers are not based on the same pointer, we can't tell that
304 if (LIBase != MemLocBase) return 0;
306 // Okay, the two values are based on the same pointer, but returned as
307 // no-alias. This happens when we have things like two byte loads at "P+1"
308 // and "P+3". Check to see if increasing the size of the "LI" load up to its
309 // alignment (or the largest native integer type) will allow us to load all
310 // the bits required by MemLoc.
312 // If MemLoc is before LI, then no widening of LI will help us out.
313 if (MemLocOffs < LIOffs) return 0;
315 // Get the alignment of the load in bytes. We assume that it is safe to load
316 // any legal integer up to this size without a problem. For example, if we're
317 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
318 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
320 unsigned LoadAlign = LI->getAlignment();
322 int64_t MemLocEnd = MemLocOffs+MemLocSize;
324 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
325 if (LIOffs+LoadAlign < MemLocEnd) return 0;
327 // This is the size of the load to try. Start with the next larger power of
329 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
330 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
333 // If this load size is bigger than our known alignment or would not fit
334 // into a native integer register, then we fail.
335 if (NewLoadByteSize > LoadAlign ||
336 !DL.fitsInLegalInteger(NewLoadByteSize*8))
339 if (LIOffs+NewLoadByteSize > MemLocEnd &&
340 LI->getParent()->getParent()->getAttributes().
341 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
342 // We will be reading past the location accessed by the original program.
343 // While this is safe in a regular build, Address Safety analysis tools
344 // may start reporting false warnings. So, don't do widening.
347 // If a load of this width would include all of MemLoc, then we succeed.
348 if (LIOffs+NewLoadByteSize >= MemLocEnd)
349 return NewLoadByteSize;
351 NewLoadByteSize <<= 1;
355 /// getPointerDependencyFrom - Return the instruction on which a memory
356 /// location depends. If isLoad is true, this routine ignores may-aliases with
357 /// read-only operations. If isLoad is false, this routine ignores may-aliases
358 /// with reads from read-only locations. If possible, pass the query
359 /// instruction as well; this function may take advantage of the metadata
360 /// annotated to the query instruction to refine the result.
361 MemDepResult MemoryDependenceAnalysis::
362 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
363 BasicBlock::iterator ScanIt, BasicBlock *BB,
364 Instruction *QueryInst) {
366 const Value *MemLocBase = nullptr;
367 int64_t MemLocOffset = 0;
368 unsigned Limit = BlockScanLimit;
369 bool isInvariantLoad = false;
370 if (isLoad && QueryInst) {
371 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
372 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
373 isInvariantLoad = true;
376 // Walk backwards through the basic block, looking for dependencies.
377 while (ScanIt != BB->begin()) {
378 Instruction *Inst = --ScanIt;
380 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
381 // Debug intrinsics don't (and can't) cause dependencies.
382 if (isa<DbgInfoIntrinsic>(II)) continue;
384 // Limit the amount of scanning we do so we don't end up with quadratic
385 // running time on extreme testcases.
388 return MemDepResult::getUnknown();
390 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
391 // If we reach a lifetime begin or end marker, then the query ends here
392 // because the value is undefined.
393 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
394 // FIXME: This only considers queries directly on the invariant-tagged
395 // pointer, not on query pointers that are indexed off of them. It'd
396 // be nice to handle that at some point (the right approach is to use
397 // GetPointerBaseWithConstantOffset).
398 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
400 return MemDepResult::getDef(II);
405 // Values depend on loads if the pointers are must aliased. This means that
406 // a load depends on another must aliased load from the same value.
407 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
408 // Atomic loads have complications involved.
409 // FIXME: This is overly conservative.
410 if (!LI->isUnordered())
411 return MemDepResult::getClobber(LI);
413 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
415 // If we found a pointer, check if it could be the same as our pointer.
416 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
419 if (R == AliasAnalysis::NoAlias) {
420 // If this is an over-aligned integer load (for example,
421 // "load i8* %P, align 4") see if it would obviously overlap with the
422 // queried location if widened to a larger load (e.g. if the queried
423 // location is 1 byte at P+1). If so, return it as a load/load
424 // clobber result, allowing the client to decide to widen the load if
426 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
427 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
428 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
429 MemLocOffset, LI, DL))
430 return MemDepResult::getClobber(Inst);
435 // Must aliased loads are defs of each other.
436 if (R == AliasAnalysis::MustAlias)
437 return MemDepResult::getDef(Inst);
439 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
440 // in terms of clobbering loads, but since it does this by looking
441 // at the clobbering load directly, it doesn't know about any
442 // phi translation that may have happened along the way.
444 // If we have a partial alias, then return this as a clobber for the
446 if (R == AliasAnalysis::PartialAlias)
447 return MemDepResult::getClobber(Inst);
450 // Random may-alias loads don't depend on each other without a
455 // Stores don't depend on other no-aliased accesses.
456 if (R == AliasAnalysis::NoAlias)
459 // Stores don't alias loads from read-only memory.
460 if (AA->pointsToConstantMemory(LoadLoc))
463 // Stores depend on may/must aliased loads.
464 return MemDepResult::getDef(Inst);
467 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
468 // Atomic stores have complications involved.
469 // FIXME: This is overly conservative.
470 if (!SI->isUnordered())
471 return MemDepResult::getClobber(SI);
473 // If alias analysis can tell that this store is guaranteed to not modify
474 // the query pointer, ignore it. Use getModRefInfo to handle cases where
475 // the query pointer points to constant memory etc.
476 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
479 // Ok, this store might clobber the query pointer. Check to see if it is
480 // a must alias: in this case, we want to return this as a def.
481 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
483 // If we found a pointer, check if it could be the same as our pointer.
484 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
486 if (R == AliasAnalysis::NoAlias)
488 if (R == AliasAnalysis::MustAlias)
489 return MemDepResult::getDef(Inst);
492 return MemDepResult::getClobber(Inst);
495 // If this is an allocation, and if we know that the accessed pointer is to
496 // the allocation, return Def. This means that there is no dependence and
497 // the access can be optimized based on that. For example, a load could
499 // Note: Only determine this to be a malloc if Inst is the malloc call, not
500 // a subsequent bitcast of the malloc call result. There can be stores to
501 // the malloced memory between the malloc call and its bitcast uses, and we
502 // need to continue scanning until the malloc call.
503 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
504 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
505 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
507 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
508 return MemDepResult::getDef(Inst);
509 // Be conservative if the accessed pointer may alias the allocation.
510 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
511 return MemDepResult::getClobber(Inst);
512 // If the allocation is not aliased and does not read memory (like
513 // strdup), it is safe to ignore.
514 if (isa<AllocaInst>(Inst) ||
515 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
519 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
520 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
521 // If necessary, perform additional analysis.
522 if (MR == AliasAnalysis::ModRef)
523 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
525 case AliasAnalysis::NoModRef:
526 // If the call has no effect on the queried pointer, just ignore it.
528 case AliasAnalysis::Mod:
529 return MemDepResult::getClobber(Inst);
530 case AliasAnalysis::Ref:
531 // If the call is known to never store to the pointer, and if this is a
532 // load query, we can safely ignore it (scan past it).
536 // Otherwise, there is a potential dependence. Return a clobber.
537 return MemDepResult::getClobber(Inst);
541 // No dependence found. If this is the entry block of the function, it is
542 // unknown, otherwise it is non-local.
543 if (BB != &BB->getParent()->getEntryBlock())
544 return MemDepResult::getNonLocal();
545 return MemDepResult::getNonFuncLocal();
548 /// getDependency - Return the instruction on which a memory operation
550 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
551 Instruction *ScanPos = QueryInst;
553 // Check for a cached result
554 MemDepResult &LocalCache = LocalDeps[QueryInst];
556 // If the cached entry is non-dirty, just return it. Note that this depends
557 // on MemDepResult's default constructing to 'dirty'.
558 if (!LocalCache.isDirty())
561 // Otherwise, if we have a dirty entry, we know we can start the scan at that
562 // instruction, which may save us some work.
563 if (Instruction *Inst = LocalCache.getInst()) {
566 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
569 BasicBlock *QueryParent = QueryInst->getParent();
572 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
573 // No dependence found. If this is the entry block of the function, it is
574 // unknown, otherwise it is non-local.
575 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
576 LocalCache = MemDepResult::getNonLocal();
578 LocalCache = MemDepResult::getNonFuncLocal();
580 AliasAnalysis::Location MemLoc;
581 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
583 // If we can do a pointer scan, make it happen.
584 bool isLoad = !(MR & AliasAnalysis::Mod);
585 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
586 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
588 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
589 QueryParent, QueryInst);
590 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
591 CallSite QueryCS(QueryInst);
592 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
593 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
596 // Non-memory instruction.
597 LocalCache = MemDepResult::getUnknown();
600 // Remember the result!
601 if (Instruction *I = LocalCache.getInst())
602 ReverseLocalDeps[I].insert(QueryInst);
608 /// AssertSorted - This method is used when -debug is specified to verify that
609 /// cache arrays are properly kept sorted.
610 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
612 if (Count == -1) Count = Cache.size();
613 if (Count == 0) return;
615 for (unsigned i = 1; i != unsigned(Count); ++i)
616 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
620 /// getNonLocalCallDependency - Perform a full dependency query for the
621 /// specified call, returning the set of blocks that the value is
622 /// potentially live across. The returned set of results will include a
623 /// "NonLocal" result for all blocks where the value is live across.
625 /// This method assumes the instruction returns a "NonLocal" dependency
626 /// within its own block.
628 /// This returns a reference to an internal data structure that may be
629 /// invalidated on the next non-local query or when an instruction is
630 /// removed. Clients must copy this data if they want it around longer than
632 const MemoryDependenceAnalysis::NonLocalDepInfo &
633 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
634 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
635 "getNonLocalCallDependency should only be used on calls with non-local deps!");
636 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
637 NonLocalDepInfo &Cache = CacheP.first;
639 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
640 /// the cached case, this can happen due to instructions being deleted etc. In
641 /// the uncached case, this starts out as the set of predecessors we care
643 SmallVector<BasicBlock*, 32> DirtyBlocks;
645 if (!Cache.empty()) {
646 // Okay, we have a cache entry. If we know it is not dirty, just return it
647 // with no computation.
648 if (!CacheP.second) {
653 // If we already have a partially computed set of results, scan them to
654 // determine what is dirty, seeding our initial DirtyBlocks worklist.
655 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
657 if (I->getResult().isDirty())
658 DirtyBlocks.push_back(I->getBB());
660 // Sort the cache so that we can do fast binary search lookups below.
661 std::sort(Cache.begin(), Cache.end());
663 ++NumCacheDirtyNonLocal;
664 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
665 // << Cache.size() << " cached: " << *QueryInst;
667 // Seed DirtyBlocks with each of the preds of QueryInst's block.
668 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
669 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
670 DirtyBlocks.push_back(*PI);
671 ++NumUncacheNonLocal;
674 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
675 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
677 SmallPtrSet<BasicBlock*, 64> Visited;
679 unsigned NumSortedEntries = Cache.size();
680 DEBUG(AssertSorted(Cache));
682 // Iterate while we still have blocks to update.
683 while (!DirtyBlocks.empty()) {
684 BasicBlock *DirtyBB = DirtyBlocks.back();
685 DirtyBlocks.pop_back();
687 // Already processed this block?
688 if (!Visited.insert(DirtyBB))
691 // Do a binary search to see if we already have an entry for this block in
692 // the cache set. If so, find it.
693 DEBUG(AssertSorted(Cache, NumSortedEntries));
694 NonLocalDepInfo::iterator Entry =
695 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
696 NonLocalDepEntry(DirtyBB));
697 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
700 NonLocalDepEntry *ExistingResult = nullptr;
701 if (Entry != Cache.begin()+NumSortedEntries &&
702 Entry->getBB() == DirtyBB) {
703 // If we already have an entry, and if it isn't already dirty, the block
705 if (!Entry->getResult().isDirty())
708 // Otherwise, remember this slot so we can update the value.
709 ExistingResult = &*Entry;
712 // If the dirty entry has a pointer, start scanning from it so we don't have
713 // to rescan the entire block.
714 BasicBlock::iterator ScanPos = DirtyBB->end();
715 if (ExistingResult) {
716 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
718 // We're removing QueryInst's use of Inst.
719 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
720 QueryCS.getInstruction());
724 // Find out if this block has a local dependency for QueryInst.
727 if (ScanPos != DirtyBB->begin()) {
728 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
729 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
730 // No dependence found. If this is the entry block of the function, it is
731 // a clobber, otherwise it is unknown.
732 Dep = MemDepResult::getNonLocal();
734 Dep = MemDepResult::getNonFuncLocal();
737 // If we had a dirty entry for the block, update it. Otherwise, just add
740 ExistingResult->setResult(Dep);
742 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
744 // If the block has a dependency (i.e. it isn't completely transparent to
745 // the value), remember the association!
746 if (!Dep.isNonLocal()) {
747 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
748 // update this when we remove instructions.
749 if (Instruction *Inst = Dep.getInst())
750 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
753 // If the block *is* completely transparent to the load, we need to check
754 // the predecessors of this block. Add them to our worklist.
755 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
756 DirtyBlocks.push_back(*PI);
763 /// getNonLocalPointerDependency - Perform a full dependency query for an
764 /// access to the specified (non-volatile) memory location, returning the
765 /// set of instructions that either define or clobber the value.
767 /// This method assumes the pointer has a "NonLocal" dependency within its
770 void MemoryDependenceAnalysis::
771 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
773 SmallVectorImpl<NonLocalDepResult> &Result) {
774 assert(Loc.Ptr->getType()->isPointerTy() &&
775 "Can't get pointer deps of a non-pointer!");
778 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL);
780 // This is the set of blocks we've inspected, and the pointer we consider in
781 // each block. Because of critical edges, we currently bail out if querying
782 // a block with multiple different pointers. This can happen during PHI
784 DenseMap<BasicBlock*, Value*> Visited;
785 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
786 Result, Visited, true))
789 Result.push_back(NonLocalDepResult(FromBB,
790 MemDepResult::getUnknown(),
791 const_cast<Value *>(Loc.Ptr)));
794 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
795 /// Pointer/PointeeSize using either cached information in Cache or by doing a
796 /// lookup (which may use dirty cache info if available). If we do a lookup,
797 /// add the result to the cache.
798 MemDepResult MemoryDependenceAnalysis::
799 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
800 bool isLoad, BasicBlock *BB,
801 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
803 // Do a binary search to see if we already have an entry for this block in
804 // the cache set. If so, find it.
805 NonLocalDepInfo::iterator Entry =
806 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
807 NonLocalDepEntry(BB));
808 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
811 NonLocalDepEntry *ExistingResult = nullptr;
812 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
813 ExistingResult = &*Entry;
815 // If we have a cached entry, and it is non-dirty, use it as the value for
817 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
818 ++NumCacheNonLocalPtr;
819 return ExistingResult->getResult();
822 // Otherwise, we have to scan for the value. If we have a dirty cache
823 // entry, start scanning from its position, otherwise we scan from the end
825 BasicBlock::iterator ScanPos = BB->end();
826 if (ExistingResult && ExistingResult->getResult().getInst()) {
827 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
828 "Instruction invalidated?");
829 ++NumCacheDirtyNonLocalPtr;
830 ScanPos = ExistingResult->getResult().getInst();
832 // Eliminating the dirty entry from 'Cache', so update the reverse info.
833 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
834 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
836 ++NumUncacheNonLocalPtr;
839 // Scan the block for the dependency.
840 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
842 // If we had a dirty entry for the block, update it. Otherwise, just add
845 ExistingResult->setResult(Dep);
847 Cache->push_back(NonLocalDepEntry(BB, Dep));
849 // If the block has a dependency (i.e. it isn't completely transparent to
850 // the value), remember the reverse association because we just added it
852 if (!Dep.isDef() && !Dep.isClobber())
855 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
856 // update MemDep when we remove instructions.
857 Instruction *Inst = Dep.getInst();
858 assert(Inst && "Didn't depend on anything?");
859 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
860 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
864 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
865 /// number of elements in the array that are already properly ordered. This is
866 /// optimized for the case when only a few entries are added.
868 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
869 unsigned NumSortedEntries) {
870 switch (Cache.size() - NumSortedEntries) {
872 // done, no new entries.
875 // Two new entries, insert the last one into place.
876 NonLocalDepEntry Val = Cache.back();
878 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
879 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
880 Cache.insert(Entry, Val);
884 // One new entry, Just insert the new value at the appropriate position.
885 if (Cache.size() != 1) {
886 NonLocalDepEntry Val = Cache.back();
888 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
889 std::upper_bound(Cache.begin(), Cache.end(), Val);
890 Cache.insert(Entry, Val);
894 // Added many values, do a full scale sort.
895 std::sort(Cache.begin(), Cache.end());
900 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
901 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
902 /// results to the results vector and keep track of which blocks are visited in
905 /// This has special behavior for the first block queries (when SkipFirstBlock
906 /// is true). In this special case, it ignores the contents of the specified
907 /// block and starts returning dependence info for its predecessors.
909 /// This function returns false on success, or true to indicate that it could
910 /// not compute dependence information for some reason. This should be treated
911 /// as a clobber dependence on the first instruction in the predecessor block.
912 bool MemoryDependenceAnalysis::
913 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
914 const AliasAnalysis::Location &Loc,
915 bool isLoad, BasicBlock *StartBB,
916 SmallVectorImpl<NonLocalDepResult> &Result,
917 DenseMap<BasicBlock*, Value*> &Visited,
918 bool SkipFirstBlock) {
919 // Look up the cached info for Pointer.
920 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
922 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
923 // CacheKey, this value will be inserted as the associated value. Otherwise,
924 // it'll be ignored, and we'll have to check to see if the cached size and
925 // tbaa tag are consistent with the current query.
926 NonLocalPointerInfo InitialNLPI;
927 InitialNLPI.Size = Loc.Size;
928 InitialNLPI.TBAATag = Loc.TBAATag;
930 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
932 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
933 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
934 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
936 // If we already have a cache entry for this CacheKey, we may need to do some
937 // work to reconcile the cache entry and the current query.
939 if (CacheInfo->Size < Loc.Size) {
940 // The query's Size is greater than the cached one. Throw out the
941 // cached data and proceed with the query at the greater size.
942 CacheInfo->Pair = BBSkipFirstBlockPair();
943 CacheInfo->Size = Loc.Size;
944 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
945 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
946 if (Instruction *Inst = DI->getResult().getInst())
947 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
948 CacheInfo->NonLocalDeps.clear();
949 } else if (CacheInfo->Size > Loc.Size) {
950 // This query's Size is less than the cached one. Conservatively restart
951 // the query using the greater size.
952 return getNonLocalPointerDepFromBB(Pointer,
953 Loc.getWithNewSize(CacheInfo->Size),
954 isLoad, StartBB, Result, Visited,
958 // If the query's TBAATag is inconsistent with the cached one,
959 // conservatively throw out the cached data and restart the query with
961 if (CacheInfo->TBAATag != Loc.TBAATag) {
962 if (CacheInfo->TBAATag) {
963 CacheInfo->Pair = BBSkipFirstBlockPair();
964 CacheInfo->TBAATag = nullptr;
965 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
966 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
967 if (Instruction *Inst = DI->getResult().getInst())
968 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
969 CacheInfo->NonLocalDeps.clear();
972 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(),
973 isLoad, StartBB, Result, Visited,
978 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
980 // If we have valid cached information for exactly the block we are
981 // investigating, just return it with no recomputation.
982 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
983 // We have a fully cached result for this query then we can just return the
984 // cached results and populate the visited set. However, we have to verify
985 // that we don't already have conflicting results for these blocks. Check
986 // to ensure that if a block in the results set is in the visited set that
987 // it was for the same pointer query.
988 if (!Visited.empty()) {
989 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
991 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
992 if (VI == Visited.end() || VI->second == Pointer.getAddr())
995 // We have a pointer mismatch in a block. Just return clobber, saying
996 // that something was clobbered in this result. We could also do a
997 // non-fully cached query, but there is little point in doing this.
1002 Value *Addr = Pointer.getAddr();
1003 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1005 Visited.insert(std::make_pair(I->getBB(), Addr));
1006 if (I->getResult().isNonLocal()) {
1011 Result.push_back(NonLocalDepResult(I->getBB(),
1012 MemDepResult::getUnknown(),
1014 } else if (DT->isReachableFromEntry(I->getBB())) {
1015 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1018 ++NumCacheCompleteNonLocalPtr;
1022 // Otherwise, either this is a new block, a block with an invalid cache
1023 // pointer or one that we're about to invalidate by putting more info into it
1024 // than its valid cache info. If empty, the result will be valid cache info,
1025 // otherwise it isn't.
1027 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1029 CacheInfo->Pair = BBSkipFirstBlockPair();
1031 SmallVector<BasicBlock*, 32> Worklist;
1032 Worklist.push_back(StartBB);
1034 // PredList used inside loop.
1035 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1037 // Keep track of the entries that we know are sorted. Previously cached
1038 // entries will all be sorted. The entries we add we only sort on demand (we
1039 // don't insert every element into its sorted position). We know that we
1040 // won't get any reuse from currently inserted values, because we don't
1041 // revisit blocks after we insert info for them.
1042 unsigned NumSortedEntries = Cache->size();
1043 DEBUG(AssertSorted(*Cache));
1045 while (!Worklist.empty()) {
1046 BasicBlock *BB = Worklist.pop_back_val();
1048 // Skip the first block if we have it.
1049 if (!SkipFirstBlock) {
1050 // Analyze the dependency of *Pointer in FromBB. See if we already have
1052 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1054 // Get the dependency info for Pointer in BB. If we have cached
1055 // information, we will use it, otherwise we compute it.
1056 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1057 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1060 // If we got a Def or Clobber, add this to the list of results.
1061 if (!Dep.isNonLocal()) {
1063 Result.push_back(NonLocalDepResult(BB,
1064 MemDepResult::getUnknown(),
1065 Pointer.getAddr()));
1067 } else if (DT->isReachableFromEntry(BB)) {
1068 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1074 // If 'Pointer' is an instruction defined in this block, then we need to do
1075 // phi translation to change it into a value live in the predecessor block.
1076 // If not, we just add the predecessors to the worklist and scan them with
1077 // the same Pointer.
1078 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1079 SkipFirstBlock = false;
1080 SmallVector<BasicBlock*, 16> NewBlocks;
1081 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1082 // Verify that we haven't looked at this block yet.
1083 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1084 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1085 if (InsertRes.second) {
1086 // First time we've looked at *PI.
1087 NewBlocks.push_back(*PI);
1091 // If we have seen this block before, but it was with a different
1092 // pointer then we have a phi translation failure and we have to treat
1093 // this as a clobber.
1094 if (InsertRes.first->second != Pointer.getAddr()) {
1095 // Make sure to clean up the Visited map before continuing on to
1096 // PredTranslationFailure.
1097 for (unsigned i = 0; i < NewBlocks.size(); i++)
1098 Visited.erase(NewBlocks[i]);
1099 goto PredTranslationFailure;
1102 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1106 // We do need to do phi translation, if we know ahead of time we can't phi
1107 // translate this value, don't even try.
1108 if (!Pointer.IsPotentiallyPHITranslatable())
1109 goto PredTranslationFailure;
1111 // We may have added values to the cache list before this PHI translation.
1112 // If so, we haven't done anything to ensure that the cache remains sorted.
1113 // Sort it now (if needed) so that recursive invocations of
1114 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1115 // value will only see properly sorted cache arrays.
1116 if (Cache && NumSortedEntries != Cache->size()) {
1117 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1118 NumSortedEntries = Cache->size();
1123 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1124 BasicBlock *Pred = *PI;
1125 PredList.push_back(std::make_pair(Pred, Pointer));
1127 // Get the PHI translated pointer in this predecessor. This can fail if
1128 // not translatable, in which case the getAddr() returns null.
1129 PHITransAddr &PredPointer = PredList.back().second;
1130 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1132 Value *PredPtrVal = PredPointer.getAddr();
1134 // Check to see if we have already visited this pred block with another
1135 // pointer. If so, we can't do this lookup. This failure can occur
1136 // with PHI translation when a critical edge exists and the PHI node in
1137 // the successor translates to a pointer value different than the
1138 // pointer the block was first analyzed with.
1139 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1140 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1142 if (!InsertRes.second) {
1143 // We found the pred; take it off the list of preds to visit.
1144 PredList.pop_back();
1146 // If the predecessor was visited with PredPtr, then we already did
1147 // the analysis and can ignore it.
1148 if (InsertRes.first->second == PredPtrVal)
1151 // Otherwise, the block was previously analyzed with a different
1152 // pointer. We can't represent the result of this case, so we just
1153 // treat this as a phi translation failure.
1155 // Make sure to clean up the Visited map before continuing on to
1156 // PredTranslationFailure.
1157 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1158 Visited.erase(PredList[i].first);
1160 goto PredTranslationFailure;
1164 // Actually process results here; this need to be a separate loop to avoid
1165 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1166 // any results for. (getNonLocalPointerDepFromBB will modify our
1167 // datastructures in ways the code after the PredTranslationFailure label
1169 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1170 BasicBlock *Pred = PredList[i].first;
1171 PHITransAddr &PredPointer = PredList[i].second;
1172 Value *PredPtrVal = PredPointer.getAddr();
1174 bool CanTranslate = true;
1175 // If PHI translation was unable to find an available pointer in this
1176 // predecessor, then we have to assume that the pointer is clobbered in
1177 // that predecessor. We can still do PRE of the load, which would insert
1178 // a computation of the pointer in this predecessor.
1180 CanTranslate = false;
1182 // FIXME: it is entirely possible that PHI translating will end up with
1183 // the same value. Consider PHI translating something like:
1184 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1185 // to recurse here, pedantically speaking.
1187 // If getNonLocalPointerDepFromBB fails here, that means the cached
1188 // result conflicted with the Visited list; we have to conservatively
1189 // assume it is unknown, but this also does not block PRE of the load.
1190 if (!CanTranslate ||
1191 getNonLocalPointerDepFromBB(PredPointer,
1192 Loc.getWithNewPtr(PredPtrVal),
1195 // Add the entry to the Result list.
1196 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1197 Result.push_back(Entry);
1199 // Since we had a phi translation failure, the cache for CacheKey won't
1200 // include all of the entries that we need to immediately satisfy future
1201 // queries. Mark this in NonLocalPointerDeps by setting the
1202 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1203 // cached value to do more work but not miss the phi trans failure.
1204 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1205 NLPI.Pair = BBSkipFirstBlockPair();
1210 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1211 CacheInfo = &NonLocalPointerDeps[CacheKey];
1212 Cache = &CacheInfo->NonLocalDeps;
1213 NumSortedEntries = Cache->size();
1215 // Since we did phi translation, the "Cache" set won't contain all of the
1216 // results for the query. This is ok (we can still use it to accelerate
1217 // specific block queries) but we can't do the fastpath "return all
1218 // results from the set" Clear out the indicator for this.
1219 CacheInfo->Pair = BBSkipFirstBlockPair();
1220 SkipFirstBlock = false;
1223 PredTranslationFailure:
1224 // The following code is "failure"; we can't produce a sane translation
1225 // for the given block. It assumes that we haven't modified any of
1226 // our datastructures while processing the current block.
1229 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1230 CacheInfo = &NonLocalPointerDeps[CacheKey];
1231 Cache = &CacheInfo->NonLocalDeps;
1232 NumSortedEntries = Cache->size();
1235 // Since we failed phi translation, the "Cache" set won't contain all of the
1236 // results for the query. This is ok (we can still use it to accelerate
1237 // specific block queries) but we can't do the fastpath "return all
1238 // results from the set". Clear out the indicator for this.
1239 CacheInfo->Pair = BBSkipFirstBlockPair();
1241 // If *nothing* works, mark the pointer as unknown.
1243 // If this is the magic first block, return this as a clobber of the whole
1244 // incoming value. Since we can't phi translate to one of the predecessors,
1245 // we have to bail out.
1249 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1250 assert(I != Cache->rend() && "Didn't find current block??");
1251 if (I->getBB() != BB)
1254 assert(I->getResult().isNonLocal() &&
1255 "Should only be here with transparent block");
1256 I->setResult(MemDepResult::getUnknown());
1257 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1258 Pointer.getAddr()));
1263 // Okay, we're done now. If we added new values to the cache, re-sort it.
1264 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1265 DEBUG(AssertSorted(*Cache));
1269 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1270 /// CachedNonLocalPointerInfo, remove it.
1271 void MemoryDependenceAnalysis::
1272 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1273 CachedNonLocalPointerInfo::iterator It =
1274 NonLocalPointerDeps.find(P);
1275 if (It == NonLocalPointerDeps.end()) return;
1277 // Remove all of the entries in the BB->val map. This involves removing
1278 // instructions from the reverse map.
1279 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1281 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1282 Instruction *Target = PInfo[i].getResult().getInst();
1283 if (!Target) continue; // Ignore non-local dep results.
1284 assert(Target->getParent() == PInfo[i].getBB());
1286 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1287 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1290 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1291 NonLocalPointerDeps.erase(It);
1295 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1296 /// information about the specified pointer, because it may be too
1297 /// conservative in memdep. This is an optional call that can be used when
1298 /// the client detects an equivalence between the pointer and some other
1299 /// value and replaces the other value with ptr. This can make Ptr available
1300 /// in more places that cached info does not necessarily keep.
1301 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1302 // If Ptr isn't really a pointer, just ignore it.
1303 if (!Ptr->getType()->isPointerTy()) return;
1304 // Flush store info for the pointer.
1305 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1306 // Flush load info for the pointer.
1307 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1310 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1311 /// This needs to be done when the CFG changes, e.g., due to splitting
1313 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1317 /// removeInstruction - Remove an instruction from the dependence analysis,
1318 /// updating the dependence of instructions that previously depended on it.
1319 /// This method attempts to keep the cache coherent using the reverse map.
1320 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1321 // Walk through the Non-local dependencies, removing this one as the value
1322 // for any cached queries.
1323 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1324 if (NLDI != NonLocalDeps.end()) {
1325 NonLocalDepInfo &BlockMap = NLDI->second.first;
1326 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1328 if (Instruction *Inst = DI->getResult().getInst())
1329 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1330 NonLocalDeps.erase(NLDI);
1333 // If we have a cached local dependence query for this instruction, remove it.
1335 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1336 if (LocalDepEntry != LocalDeps.end()) {
1337 // Remove us from DepInst's reverse set now that the local dep info is gone.
1338 if (Instruction *Inst = LocalDepEntry->second.getInst())
1339 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1341 // Remove this local dependency info.
1342 LocalDeps.erase(LocalDepEntry);
1345 // If we have any cached pointer dependencies on this instruction, remove
1346 // them. If the instruction has non-pointer type, then it can't be a pointer
1349 // Remove it from both the load info and the store info. The instruction
1350 // can't be in either of these maps if it is non-pointer.
1351 if (RemInst->getType()->isPointerTy()) {
1352 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1353 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1356 // Loop over all of the things that depend on the instruction we're removing.
1358 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1360 // If we find RemInst as a clobber or Def in any of the maps for other values,
1361 // we need to replace its entry with a dirty version of the instruction after
1362 // it. If RemInst is a terminator, we use a null dirty value.
1364 // Using a dirty version of the instruction after RemInst saves having to scan
1365 // the entire block to get to this point.
1366 MemDepResult NewDirtyVal;
1367 if (!RemInst->isTerminator())
1368 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1370 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1371 if (ReverseDepIt != ReverseLocalDeps.end()) {
1372 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1373 // RemInst can't be the terminator if it has local stuff depending on it.
1374 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1375 "Nothing can locally depend on a terminator");
1377 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1378 E = ReverseDeps.end(); I != E; ++I) {
1379 Instruction *InstDependingOnRemInst = *I;
1380 assert(InstDependingOnRemInst != RemInst &&
1381 "Already removed our local dep info");
1383 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1385 // Make sure to remember that new things depend on NewDepInst.
1386 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1387 "a local dep on this if it is a terminator!");
1388 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1389 InstDependingOnRemInst));
1392 ReverseLocalDeps.erase(ReverseDepIt);
1394 // Add new reverse deps after scanning the set, to avoid invalidating the
1395 // 'ReverseDeps' reference.
1396 while (!ReverseDepsToAdd.empty()) {
1397 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1398 .insert(ReverseDepsToAdd.back().second);
1399 ReverseDepsToAdd.pop_back();
1403 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1404 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1405 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1406 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1408 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1410 PerInstNLInfo &INLD = NonLocalDeps[*I];
1411 // The information is now dirty!
1414 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1415 DE = INLD.first.end(); DI != DE; ++DI) {
1416 if (DI->getResult().getInst() != RemInst) continue;
1418 // Convert to a dirty entry for the subsequent instruction.
1419 DI->setResult(NewDirtyVal);
1421 if (Instruction *NextI = NewDirtyVal.getInst())
1422 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1426 ReverseNonLocalDeps.erase(ReverseDepIt);
1428 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1429 while (!ReverseDepsToAdd.empty()) {
1430 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1431 .insert(ReverseDepsToAdd.back().second);
1432 ReverseDepsToAdd.pop_back();
1436 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1437 // value in the NonLocalPointerDeps info.
1438 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1439 ReverseNonLocalPtrDeps.find(RemInst);
1440 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1441 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1442 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1444 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1445 E = Set.end(); I != E; ++I) {
1446 ValueIsLoadPair P = *I;
1447 assert(P.getPointer() != RemInst &&
1448 "Already removed NonLocalPointerDeps info for RemInst");
1450 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1452 // The cache is not valid for any specific block anymore.
1453 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1455 // Update any entries for RemInst to use the instruction after it.
1456 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1458 if (DI->getResult().getInst() != RemInst) continue;
1460 // Convert to a dirty entry for the subsequent instruction.
1461 DI->setResult(NewDirtyVal);
1463 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1464 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1467 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1468 // subsequent value may invalidate the sortedness.
1469 std::sort(NLPDI.begin(), NLPDI.end());
1472 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1474 while (!ReversePtrDepsToAdd.empty()) {
1475 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1476 .insert(ReversePtrDepsToAdd.back().second);
1477 ReversePtrDepsToAdd.pop_back();
1482 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1483 AA->deleteValue(RemInst);
1484 DEBUG(verifyRemoved(RemInst));
1486 /// verifyRemoved - Verify that the specified instruction does not occur
1487 /// in our internal data structures.
1488 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1489 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1490 E = LocalDeps.end(); I != E; ++I) {
1491 assert(I->first != D && "Inst occurs in data structures");
1492 assert(I->second.getInst() != D &&
1493 "Inst occurs in data structures");
1496 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1497 E = NonLocalPointerDeps.end(); I != E; ++I) {
1498 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1499 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1500 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1502 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1505 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1506 E = NonLocalDeps.end(); I != E; ++I) {
1507 assert(I->first != D && "Inst occurs in data structures");
1508 const PerInstNLInfo &INLD = I->second;
1509 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1510 EE = INLD.first.end(); II != EE; ++II)
1511 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1514 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1515 E = ReverseLocalDeps.end(); I != E; ++I) {
1516 assert(I->first != D && "Inst occurs in data structures");
1517 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1518 EE = I->second.end(); II != EE; ++II)
1519 assert(*II != D && "Inst occurs in data structures");
1522 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1523 E = ReverseNonLocalDeps.end();
1525 assert(I->first != D && "Inst occurs in data structures");
1526 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1527 EE = I->second.end(); II != EE; ++II)
1528 assert(*II != D && "Inst occurs in data structures");
1531 for (ReverseNonLocalPtrDepTy::const_iterator
1532 I = ReverseNonLocalPtrDeps.begin(),
1533 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1534 assert(I->first != D && "Inst occurs in rev NLPD map");
1536 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1537 E = I->second.end(); II != E; ++II)
1538 assert(*II != ValueIsLoadPair(D, false) &&
1539 *II != ValueIsLoadPair(D, true) &&
1540 "Inst occurs in ReverseNonLocalPtrDeps map");