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)) {
164 switch (II->getIntrinsicID()) {
165 case Intrinsic::lifetime_start:
166 case Intrinsic::lifetime_end:
167 case Intrinsic::invariant_start:
168 II->getAAMetadata(AAInfo);
169 Loc = AliasAnalysis::Location(II->getArgOperand(1),
170 cast<ConstantInt>(II->getArgOperand(0))
171 ->getZExtValue(), AAInfo);
172 // These intrinsics don't really modify the memory, but returning Mod
173 // will allow them to be handled conservatively.
174 return AliasAnalysis::Mod;
175 case Intrinsic::invariant_end:
176 II->getAAMetadata(AAInfo);
177 Loc = AliasAnalysis::Location(II->getArgOperand(2),
178 cast<ConstantInt>(II->getArgOperand(1))
179 ->getZExtValue(), AAInfo);
180 // These intrinsics don't really modify the memory, but returning Mod
181 // will allow them to be handled conservatively.
182 return AliasAnalysis::Mod;
188 // Otherwise, just do the coarse-grained thing that always works.
189 if (Inst->mayWriteToMemory())
190 return AliasAnalysis::ModRef;
191 if (Inst->mayReadFromMemory())
192 return AliasAnalysis::Ref;
193 return AliasAnalysis::NoModRef;
196 /// getCallSiteDependencyFrom - Private helper for finding the local
197 /// dependencies of a call site.
198 MemDepResult MemoryDependenceAnalysis::
199 getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
200 BasicBlock::iterator ScanIt, BasicBlock *BB) {
201 unsigned Limit = BlockScanLimit;
203 // Walk backwards through the block, looking for dependencies
204 while (ScanIt != BB->begin()) {
205 // Limit the amount of scanning we do so we don't end up with quadratic
206 // running time on extreme testcases.
209 return MemDepResult::getUnknown();
211 Instruction *Inst = --ScanIt;
213 // If this inst is a memory op, get the pointer it accessed
214 AliasAnalysis::Location Loc;
215 AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA);
217 // A simple instruction.
218 if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
219 return MemDepResult::getClobber(Inst);
223 if (CallSite InstCS = cast<Value>(Inst)) {
224 // Debug intrinsics don't cause dependences.
225 if (isa<DbgInfoIntrinsic>(Inst)) continue;
226 // If these two calls do not interfere, look past it.
227 switch (AA->getModRefInfo(CS, InstCS)) {
228 case AliasAnalysis::NoModRef:
229 // If the two calls are the same, return InstCS as a Def, so that
230 // CS can be found redundant and eliminated.
231 if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) &&
232 CS.getInstruction()->isIdenticalToWhenDefined(Inst))
233 return MemDepResult::getDef(Inst);
235 // Otherwise if the two calls don't interact (e.g. InstCS is readnone)
239 return MemDepResult::getClobber(Inst);
243 // If we could not obtain a pointer for the instruction and the instruction
244 // touches memory then assume that this is a dependency.
245 if (MR != AliasAnalysis::NoModRef)
246 return MemDepResult::getClobber(Inst);
249 // No dependence found. If this is the entry block of the function, it is
250 // unknown, otherwise it is non-local.
251 if (BB != &BB->getParent()->getEntryBlock())
252 return MemDepResult::getNonLocal();
253 return MemDepResult::getNonFuncLocal();
256 /// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that
257 /// would fully overlap MemLoc if done as a wider legal integer load.
259 /// MemLocBase, MemLocOffset are lazily computed here the first time the
260 /// base/offs of memloc is needed.
262 isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc,
263 const Value *&MemLocBase,
266 const DataLayout *DL) {
267 // If we have no target data, we can't do this.
268 if (!DL) return false;
270 // If we haven't already computed the base/offset of MemLoc, do so now.
272 MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, DL);
274 unsigned Size = MemoryDependenceAnalysis::
275 getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size,
280 /// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that
281 /// looks at a memory location for a load (specified by MemLocBase, Offs,
282 /// and Size) and compares it against a load. If the specified load could
283 /// be safely widened to a larger integer load that is 1) still efficient,
284 /// 2) safe for the target, and 3) would provide the specified memory
285 /// location value, then this function returns the size in bytes of the
286 /// load width to use. If not, this returns zero.
287 unsigned MemoryDependenceAnalysis::
288 getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs,
289 unsigned MemLocSize, const LoadInst *LI,
290 const DataLayout &DL) {
291 // We can only extend simple integer loads.
292 if (!isa<IntegerType>(LI->getType()) || !LI->isSimple()) return 0;
294 // Load widening is hostile to ThreadSanitizer: it may cause false positives
295 // or make the reports more cryptic (access sizes are wrong).
296 if (LI->getParent()->getParent()->getAttributes().
297 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeThread))
300 // Get the base of this load.
302 const Value *LIBase =
303 GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, &DL);
305 // If the two pointers are not based on the same pointer, we can't tell that
307 if (LIBase != MemLocBase) return 0;
309 // Okay, the two values are based on the same pointer, but returned as
310 // no-alias. This happens when we have things like two byte loads at "P+1"
311 // and "P+3". Check to see if increasing the size of the "LI" load up to its
312 // alignment (or the largest native integer type) will allow us to load all
313 // the bits required by MemLoc.
315 // If MemLoc is before LI, then no widening of LI will help us out.
316 if (MemLocOffs < LIOffs) return 0;
318 // Get the alignment of the load in bytes. We assume that it is safe to load
319 // any legal integer up to this size without a problem. For example, if we're
320 // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can
321 // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it
323 unsigned LoadAlign = LI->getAlignment();
325 int64_t MemLocEnd = MemLocOffs+MemLocSize;
327 // If no amount of rounding up will let MemLoc fit into LI, then bail out.
328 if (LIOffs+LoadAlign < MemLocEnd) return 0;
330 // This is the size of the load to try. Start with the next larger power of
332 unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U;
333 NewLoadByteSize = NextPowerOf2(NewLoadByteSize);
336 // If this load size is bigger than our known alignment or would not fit
337 // into a native integer register, then we fail.
338 if (NewLoadByteSize > LoadAlign ||
339 !DL.fitsInLegalInteger(NewLoadByteSize*8))
342 if (LIOffs+NewLoadByteSize > MemLocEnd &&
343 LI->getParent()->getParent()->getAttributes().
344 hasAttribute(AttributeSet::FunctionIndex, Attribute::SanitizeAddress))
345 // We will be reading past the location accessed by the original program.
346 // While this is safe in a regular build, Address Safety analysis tools
347 // may start reporting false warnings. So, don't do widening.
350 // If a load of this width would include all of MemLoc, then we succeed.
351 if (LIOffs+NewLoadByteSize >= MemLocEnd)
352 return NewLoadByteSize;
354 NewLoadByteSize <<= 1;
358 /// getPointerDependencyFrom - Return the instruction on which a memory
359 /// location depends. If isLoad is true, this routine ignores may-aliases with
360 /// read-only operations. If isLoad is false, this routine ignores may-aliases
361 /// with reads from read-only locations. If possible, pass the query
362 /// instruction as well; this function may take advantage of the metadata
363 /// annotated to the query instruction to refine the result.
364 MemDepResult MemoryDependenceAnalysis::
365 getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
366 BasicBlock::iterator ScanIt, BasicBlock *BB,
367 Instruction *QueryInst) {
369 const Value *MemLocBase = nullptr;
370 int64_t MemLocOffset = 0;
371 unsigned Limit = BlockScanLimit;
372 bool isInvariantLoad = false;
373 if (isLoad && QueryInst) {
374 LoadInst *LI = dyn_cast<LoadInst>(QueryInst);
375 if (LI && LI->getMetadata(LLVMContext::MD_invariant_load) != nullptr)
376 isInvariantLoad = true;
379 // Walk backwards through the basic block, looking for dependencies.
380 while (ScanIt != BB->begin()) {
381 Instruction *Inst = --ScanIt;
383 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst))
384 // Debug intrinsics don't (and can't) cause dependencies.
385 if (isa<DbgInfoIntrinsic>(II)) continue;
387 // Limit the amount of scanning we do so we don't end up with quadratic
388 // running time on extreme testcases.
391 return MemDepResult::getUnknown();
393 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
394 // If we reach a lifetime begin or end marker, then the query ends here
395 // because the value is undefined.
396 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
397 // FIXME: This only considers queries directly on the invariant-tagged
398 // pointer, not on query pointers that are indexed off of them. It'd
399 // be nice to handle that at some point (the right approach is to use
400 // GetPointerBaseWithConstantOffset).
401 if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
403 return MemDepResult::getDef(II);
408 // Values depend on loads if the pointers are must aliased. This means that
409 // a load depends on another must aliased load from the same value.
410 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
411 // Atomic loads have complications involved.
412 // FIXME: This is overly conservative.
413 if (!LI->isUnordered())
414 return MemDepResult::getClobber(LI);
416 AliasAnalysis::Location LoadLoc = AA->getLocation(LI);
418 // If we found a pointer, check if it could be the same as our pointer.
419 AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc);
422 if (R == AliasAnalysis::NoAlias) {
423 // If this is an over-aligned integer load (for example,
424 // "load i8* %P, align 4") see if it would obviously overlap with the
425 // queried location if widened to a larger load (e.g. if the queried
426 // location is 1 byte at P+1). If so, return it as a load/load
427 // clobber result, allowing the client to decide to widen the load if
429 if (IntegerType *ITy = dyn_cast<IntegerType>(LI->getType()))
430 if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() &&
431 isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase,
432 MemLocOffset, LI, DL))
433 return MemDepResult::getClobber(Inst);
438 // Must aliased loads are defs of each other.
439 if (R == AliasAnalysis::MustAlias)
440 return MemDepResult::getDef(Inst);
442 #if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads
443 // in terms of clobbering loads, but since it does this by looking
444 // at the clobbering load directly, it doesn't know about any
445 // phi translation that may have happened along the way.
447 // If we have a partial alias, then return this as a clobber for the
449 if (R == AliasAnalysis::PartialAlias)
450 return MemDepResult::getClobber(Inst);
453 // Random may-alias loads don't depend on each other without a
458 // Stores don't depend on other no-aliased accesses.
459 if (R == AliasAnalysis::NoAlias)
462 // Stores don't alias loads from read-only memory.
463 if (AA->pointsToConstantMemory(LoadLoc))
466 // Stores depend on may/must aliased loads.
467 return MemDepResult::getDef(Inst);
470 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
471 // Atomic stores have complications involved.
472 // FIXME: This is overly conservative.
473 if (!SI->isUnordered())
474 return MemDepResult::getClobber(SI);
476 // If alias analysis can tell that this store is guaranteed to not modify
477 // the query pointer, ignore it. Use getModRefInfo to handle cases where
478 // the query pointer points to constant memory etc.
479 if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
482 // Ok, this store might clobber the query pointer. Check to see if it is
483 // a must alias: in this case, we want to return this as a def.
484 AliasAnalysis::Location StoreLoc = AA->getLocation(SI);
486 // If we found a pointer, check if it could be the same as our pointer.
487 AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc);
489 if (R == AliasAnalysis::NoAlias)
491 if (R == AliasAnalysis::MustAlias)
492 return MemDepResult::getDef(Inst);
495 return MemDepResult::getClobber(Inst);
498 // If this is an allocation, and if we know that the accessed pointer is to
499 // the allocation, return Def. This means that there is no dependence and
500 // the access can be optimized based on that. For example, a load could
502 // Note: Only determine this to be a malloc if Inst is the malloc call, not
503 // a subsequent bitcast of the malloc call result. There can be stores to
504 // the malloced memory between the malloc call and its bitcast uses, and we
505 // need to continue scanning until the malloc call.
506 const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
507 if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
508 const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, DL);
510 if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
511 return MemDepResult::getDef(Inst);
512 // Be conservative if the accessed pointer may alias the allocation.
513 if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias)
514 return MemDepResult::getClobber(Inst);
515 // If the allocation is not aliased and does not read memory (like
516 // strdup), it is safe to ignore.
517 if (isa<AllocaInst>(Inst) ||
518 isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
522 // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
523 AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc);
524 // If necessary, perform additional analysis.
525 if (MR == AliasAnalysis::ModRef)
526 MR = AA->callCapturesBefore(Inst, MemLoc, DT);
528 case AliasAnalysis::NoModRef:
529 // If the call has no effect on the queried pointer, just ignore it.
531 case AliasAnalysis::Mod:
532 return MemDepResult::getClobber(Inst);
533 case AliasAnalysis::Ref:
534 // If the call is known to never store to the pointer, and if this is a
535 // load query, we can safely ignore it (scan past it).
539 // Otherwise, there is a potential dependence. Return a clobber.
540 return MemDepResult::getClobber(Inst);
544 // No dependence found. If this is the entry block of the function, it is
545 // unknown, otherwise it is non-local.
546 if (BB != &BB->getParent()->getEntryBlock())
547 return MemDepResult::getNonLocal();
548 return MemDepResult::getNonFuncLocal();
551 /// getDependency - Return the instruction on which a memory operation
553 MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
554 Instruction *ScanPos = QueryInst;
556 // Check for a cached result
557 MemDepResult &LocalCache = LocalDeps[QueryInst];
559 // If the cached entry is non-dirty, just return it. Note that this depends
560 // on MemDepResult's default constructing to 'dirty'.
561 if (!LocalCache.isDirty())
564 // Otherwise, if we have a dirty entry, we know we can start the scan at that
565 // instruction, which may save us some work.
566 if (Instruction *Inst = LocalCache.getInst()) {
569 RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst);
572 BasicBlock *QueryParent = QueryInst->getParent();
575 if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
576 // No dependence found. If this is the entry block of the function, it is
577 // unknown, otherwise it is non-local.
578 if (QueryParent != &QueryParent->getParent()->getEntryBlock())
579 LocalCache = MemDepResult::getNonLocal();
581 LocalCache = MemDepResult::getNonFuncLocal();
583 AliasAnalysis::Location MemLoc;
584 AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA);
586 // If we can do a pointer scan, make it happen.
587 bool isLoad = !(MR & AliasAnalysis::Mod);
588 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst))
589 isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
591 LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
592 QueryParent, QueryInst);
593 } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
594 CallSite QueryCS(QueryInst);
595 bool isReadOnly = AA->onlyReadsMemory(QueryCS);
596 LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
599 // Non-memory instruction.
600 LocalCache = MemDepResult::getUnknown();
603 // Remember the result!
604 if (Instruction *I = LocalCache.getInst())
605 ReverseLocalDeps[I].insert(QueryInst);
611 /// AssertSorted - This method is used when -debug is specified to verify that
612 /// cache arrays are properly kept sorted.
613 static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
615 if (Count == -1) Count = Cache.size();
616 if (Count == 0) return;
618 for (unsigned i = 1; i != unsigned(Count); ++i)
619 assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
623 /// getNonLocalCallDependency - Perform a full dependency query for the
624 /// specified call, returning the set of blocks that the value is
625 /// potentially live across. The returned set of results will include a
626 /// "NonLocal" result for all blocks where the value is live across.
628 /// This method assumes the instruction returns a "NonLocal" dependency
629 /// within its own block.
631 /// This returns a reference to an internal data structure that may be
632 /// invalidated on the next non-local query or when an instruction is
633 /// removed. Clients must copy this data if they want it around longer than
635 const MemoryDependenceAnalysis::NonLocalDepInfo &
636 MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
637 assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
638 "getNonLocalCallDependency should only be used on calls with non-local deps!");
639 PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
640 NonLocalDepInfo &Cache = CacheP.first;
642 /// DirtyBlocks - This is the set of blocks that need to be recomputed. In
643 /// the cached case, this can happen due to instructions being deleted etc. In
644 /// the uncached case, this starts out as the set of predecessors we care
646 SmallVector<BasicBlock*, 32> DirtyBlocks;
648 if (!Cache.empty()) {
649 // Okay, we have a cache entry. If we know it is not dirty, just return it
650 // with no computation.
651 if (!CacheP.second) {
656 // If we already have a partially computed set of results, scan them to
657 // determine what is dirty, seeding our initial DirtyBlocks worklist.
658 for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
660 if (I->getResult().isDirty())
661 DirtyBlocks.push_back(I->getBB());
663 // Sort the cache so that we can do fast binary search lookups below.
664 std::sort(Cache.begin(), Cache.end());
666 ++NumCacheDirtyNonLocal;
667 //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
668 // << Cache.size() << " cached: " << *QueryInst;
670 // Seed DirtyBlocks with each of the preds of QueryInst's block.
671 BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
672 for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
673 DirtyBlocks.push_back(*PI);
674 ++NumUncacheNonLocal;
677 // isReadonlyCall - If this is a read-only call, we can be more aggressive.
678 bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
680 SmallPtrSet<BasicBlock*, 64> Visited;
682 unsigned NumSortedEntries = Cache.size();
683 DEBUG(AssertSorted(Cache));
685 // Iterate while we still have blocks to update.
686 while (!DirtyBlocks.empty()) {
687 BasicBlock *DirtyBB = DirtyBlocks.back();
688 DirtyBlocks.pop_back();
690 // Already processed this block?
691 if (!Visited.insert(DirtyBB))
694 // Do a binary search to see if we already have an entry for this block in
695 // the cache set. If so, find it.
696 DEBUG(AssertSorted(Cache, NumSortedEntries));
697 NonLocalDepInfo::iterator Entry =
698 std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
699 NonLocalDepEntry(DirtyBB));
700 if (Entry != Cache.begin() && std::prev(Entry)->getBB() == DirtyBB)
703 NonLocalDepEntry *ExistingResult = nullptr;
704 if (Entry != Cache.begin()+NumSortedEntries &&
705 Entry->getBB() == DirtyBB) {
706 // If we already have an entry, and if it isn't already dirty, the block
708 if (!Entry->getResult().isDirty())
711 // Otherwise, remember this slot so we can update the value.
712 ExistingResult = &*Entry;
715 // If the dirty entry has a pointer, start scanning from it so we don't have
716 // to rescan the entire block.
717 BasicBlock::iterator ScanPos = DirtyBB->end();
718 if (ExistingResult) {
719 if (Instruction *Inst = ExistingResult->getResult().getInst()) {
721 // We're removing QueryInst's use of Inst.
722 RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
723 QueryCS.getInstruction());
727 // Find out if this block has a local dependency for QueryInst.
730 if (ScanPos != DirtyBB->begin()) {
731 Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
732 } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
733 // No dependence found. If this is the entry block of the function, it is
734 // a clobber, otherwise it is unknown.
735 Dep = MemDepResult::getNonLocal();
737 Dep = MemDepResult::getNonFuncLocal();
740 // If we had a dirty entry for the block, update it. Otherwise, just add
743 ExistingResult->setResult(Dep);
745 Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
747 // If the block has a dependency (i.e. it isn't completely transparent to
748 // the value), remember the association!
749 if (!Dep.isNonLocal()) {
750 // Keep the ReverseNonLocalDeps map up to date so we can efficiently
751 // update this when we remove instructions.
752 if (Instruction *Inst = Dep.getInst())
753 ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
756 // If the block *is* completely transparent to the load, we need to check
757 // the predecessors of this block. Add them to our worklist.
758 for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
759 DirtyBlocks.push_back(*PI);
766 /// getNonLocalPointerDependency - Perform a full dependency query for an
767 /// access to the specified (non-volatile) memory location, returning the
768 /// set of instructions that either define or clobber the value.
770 /// This method assumes the pointer has a "NonLocal" dependency within its
773 void MemoryDependenceAnalysis::
774 getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
776 SmallVectorImpl<NonLocalDepResult> &Result) {
777 assert(Loc.Ptr->getType()->isPointerTy() &&
778 "Can't get pointer deps of a non-pointer!");
781 PHITransAddr Address(const_cast<Value *>(Loc.Ptr), DL);
783 // This is the set of blocks we've inspected, and the pointer we consider in
784 // each block. Because of critical edges, we currently bail out if querying
785 // a block with multiple different pointers. This can happen during PHI
787 DenseMap<BasicBlock*, Value*> Visited;
788 if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
789 Result, Visited, true))
792 Result.push_back(NonLocalDepResult(FromBB,
793 MemDepResult::getUnknown(),
794 const_cast<Value *>(Loc.Ptr)));
797 /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
798 /// Pointer/PointeeSize using either cached information in Cache or by doing a
799 /// lookup (which may use dirty cache info if available). If we do a lookup,
800 /// add the result to the cache.
801 MemDepResult MemoryDependenceAnalysis::
802 GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
803 bool isLoad, BasicBlock *BB,
804 NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
806 // Do a binary search to see if we already have an entry for this block in
807 // the cache set. If so, find it.
808 NonLocalDepInfo::iterator Entry =
809 std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
810 NonLocalDepEntry(BB));
811 if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
814 NonLocalDepEntry *ExistingResult = nullptr;
815 if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
816 ExistingResult = &*Entry;
818 // If we have a cached entry, and it is non-dirty, use it as the value for
820 if (ExistingResult && !ExistingResult->getResult().isDirty()) {
821 ++NumCacheNonLocalPtr;
822 return ExistingResult->getResult();
825 // Otherwise, we have to scan for the value. If we have a dirty cache
826 // entry, start scanning from its position, otherwise we scan from the end
828 BasicBlock::iterator ScanPos = BB->end();
829 if (ExistingResult && ExistingResult->getResult().getInst()) {
830 assert(ExistingResult->getResult().getInst()->getParent() == BB &&
831 "Instruction invalidated?");
832 ++NumCacheDirtyNonLocalPtr;
833 ScanPos = ExistingResult->getResult().getInst();
835 // Eliminating the dirty entry from 'Cache', so update the reverse info.
836 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
837 RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
839 ++NumUncacheNonLocalPtr;
842 // Scan the block for the dependency.
843 MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
845 // If we had a dirty entry for the block, update it. Otherwise, just add
848 ExistingResult->setResult(Dep);
850 Cache->push_back(NonLocalDepEntry(BB, Dep));
852 // If the block has a dependency (i.e. it isn't completely transparent to
853 // the value), remember the reverse association because we just added it
855 if (!Dep.isDef() && !Dep.isClobber())
858 // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
859 // update MemDep when we remove instructions.
860 Instruction *Inst = Dep.getInst();
861 assert(Inst && "Didn't depend on anything?");
862 ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
863 ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
867 /// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
868 /// number of elements in the array that are already properly ordered. This is
869 /// optimized for the case when only a few entries are added.
871 SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
872 unsigned NumSortedEntries) {
873 switch (Cache.size() - NumSortedEntries) {
875 // done, no new entries.
878 // Two new entries, insert the last one into place.
879 NonLocalDepEntry Val = Cache.back();
881 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
882 std::upper_bound(Cache.begin(), Cache.end()-1, Val);
883 Cache.insert(Entry, Val);
887 // One new entry, Just insert the new value at the appropriate position.
888 if (Cache.size() != 1) {
889 NonLocalDepEntry Val = Cache.back();
891 MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
892 std::upper_bound(Cache.begin(), Cache.end(), Val);
893 Cache.insert(Entry, Val);
897 // Added many values, do a full scale sort.
898 std::sort(Cache.begin(), Cache.end());
903 /// getNonLocalPointerDepFromBB - Perform a dependency query based on
904 /// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
905 /// results to the results vector and keep track of which blocks are visited in
908 /// This has special behavior for the first block queries (when SkipFirstBlock
909 /// is true). In this special case, it ignores the contents of the specified
910 /// block and starts returning dependence info for its predecessors.
912 /// This function returns false on success, or true to indicate that it could
913 /// not compute dependence information for some reason. This should be treated
914 /// as a clobber dependence on the first instruction in the predecessor block.
915 bool MemoryDependenceAnalysis::
916 getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
917 const AliasAnalysis::Location &Loc,
918 bool isLoad, BasicBlock *StartBB,
919 SmallVectorImpl<NonLocalDepResult> &Result,
920 DenseMap<BasicBlock*, Value*> &Visited,
921 bool SkipFirstBlock) {
922 // Look up the cached info for Pointer.
923 ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
925 // Set up a temporary NLPI value. If the map doesn't yet have an entry for
926 // CacheKey, this value will be inserted as the associated value. Otherwise,
927 // it'll be ignored, and we'll have to check to see if the cached size and
928 // aa tags are consistent with the current query.
929 NonLocalPointerInfo InitialNLPI;
930 InitialNLPI.Size = Loc.Size;
931 InitialNLPI.AATags = Loc.AATags;
933 // Get the NLPI for CacheKey, inserting one into the map if it doesn't
935 std::pair<CachedNonLocalPointerInfo::iterator, bool> Pair =
936 NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI));
937 NonLocalPointerInfo *CacheInfo = &Pair.first->second;
939 // If we already have a cache entry for this CacheKey, we may need to do some
940 // work to reconcile the cache entry and the current query.
942 if (CacheInfo->Size < Loc.Size) {
943 // The query's Size is greater than the cached one. Throw out the
944 // cached data and proceed with the query at the greater size.
945 CacheInfo->Pair = BBSkipFirstBlockPair();
946 CacheInfo->Size = Loc.Size;
947 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
948 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
949 if (Instruction *Inst = DI->getResult().getInst())
950 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
951 CacheInfo->NonLocalDeps.clear();
952 } else if (CacheInfo->Size > Loc.Size) {
953 // This query's Size is less than the cached one. Conservatively restart
954 // the query using the greater size.
955 return getNonLocalPointerDepFromBB(Pointer,
956 Loc.getWithNewSize(CacheInfo->Size),
957 isLoad, StartBB, Result, Visited,
961 // If the query's AATags are inconsistent with the cached one,
962 // conservatively throw out the cached data and restart the query with
964 if (CacheInfo->AATags != Loc.AATags) {
965 if (CacheInfo->AATags) {
966 CacheInfo->Pair = BBSkipFirstBlockPair();
967 CacheInfo->AATags = AAMDNodes();
968 for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(),
969 DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI)
970 if (Instruction *Inst = DI->getResult().getInst())
971 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey);
972 CacheInfo->NonLocalDeps.clear();
975 return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutAATags(),
976 isLoad, StartBB, Result, Visited,
981 NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps;
983 // If we have valid cached information for exactly the block we are
984 // investigating, just return it with no recomputation.
985 if (CacheInfo->Pair == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
986 // We have a fully cached result for this query then we can just return the
987 // cached results and populate the visited set. However, we have to verify
988 // that we don't already have conflicting results for these blocks. Check
989 // to ensure that if a block in the results set is in the visited set that
990 // it was for the same pointer query.
991 if (!Visited.empty()) {
992 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
994 DenseMap<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
995 if (VI == Visited.end() || VI->second == Pointer.getAddr())
998 // We have a pointer mismatch in a block. Just return clobber, saying
999 // that something was clobbered in this result. We could also do a
1000 // non-fully cached query, but there is little point in doing this.
1005 Value *Addr = Pointer.getAddr();
1006 for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
1008 Visited.insert(std::make_pair(I->getBB(), Addr));
1009 if (I->getResult().isNonLocal()) {
1014 Result.push_back(NonLocalDepResult(I->getBB(),
1015 MemDepResult::getUnknown(),
1017 } else if (DT->isReachableFromEntry(I->getBB())) {
1018 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
1021 ++NumCacheCompleteNonLocalPtr;
1025 // Otherwise, either this is a new block, a block with an invalid cache
1026 // pointer or one that we're about to invalidate by putting more info into it
1027 // than its valid cache info. If empty, the result will be valid cache info,
1028 // otherwise it isn't.
1030 CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
1032 CacheInfo->Pair = BBSkipFirstBlockPair();
1034 SmallVector<BasicBlock*, 32> Worklist;
1035 Worklist.push_back(StartBB);
1037 // PredList used inside loop.
1038 SmallVector<std::pair<BasicBlock*, PHITransAddr>, 16> PredList;
1040 // Keep track of the entries that we know are sorted. Previously cached
1041 // entries will all be sorted. The entries we add we only sort on demand (we
1042 // don't insert every element into its sorted position). We know that we
1043 // won't get any reuse from currently inserted values, because we don't
1044 // revisit blocks after we insert info for them.
1045 unsigned NumSortedEntries = Cache->size();
1046 DEBUG(AssertSorted(*Cache));
1048 while (!Worklist.empty()) {
1049 BasicBlock *BB = Worklist.pop_back_val();
1051 // Skip the first block if we have it.
1052 if (!SkipFirstBlock) {
1053 // Analyze the dependency of *Pointer in FromBB. See if we already have
1055 assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
1057 // Get the dependency info for Pointer in BB. If we have cached
1058 // information, we will use it, otherwise we compute it.
1059 DEBUG(AssertSorted(*Cache, NumSortedEntries));
1060 MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
1063 // If we got a Def or Clobber, add this to the list of results.
1064 if (!Dep.isNonLocal()) {
1066 Result.push_back(NonLocalDepResult(BB,
1067 MemDepResult::getUnknown(),
1068 Pointer.getAddr()));
1070 } else if (DT->isReachableFromEntry(BB)) {
1071 Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
1077 // If 'Pointer' is an instruction defined in this block, then we need to do
1078 // phi translation to change it into a value live in the predecessor block.
1079 // If not, we just add the predecessors to the worklist and scan them with
1080 // the same Pointer.
1081 if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
1082 SkipFirstBlock = false;
1083 SmallVector<BasicBlock*, 16> NewBlocks;
1084 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1085 // Verify that we haven't looked at this block yet.
1086 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1087 InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
1088 if (InsertRes.second) {
1089 // First time we've looked at *PI.
1090 NewBlocks.push_back(*PI);
1094 // If we have seen this block before, but it was with a different
1095 // pointer then we have a phi translation failure and we have to treat
1096 // this as a clobber.
1097 if (InsertRes.first->second != Pointer.getAddr()) {
1098 // Make sure to clean up the Visited map before continuing on to
1099 // PredTranslationFailure.
1100 for (unsigned i = 0; i < NewBlocks.size(); i++)
1101 Visited.erase(NewBlocks[i]);
1102 goto PredTranslationFailure;
1105 Worklist.append(NewBlocks.begin(), NewBlocks.end());
1109 // We do need to do phi translation, if we know ahead of time we can't phi
1110 // translate this value, don't even try.
1111 if (!Pointer.IsPotentiallyPHITranslatable())
1112 goto PredTranslationFailure;
1114 // We may have added values to the cache list before this PHI translation.
1115 // If so, we haven't done anything to ensure that the cache remains sorted.
1116 // Sort it now (if needed) so that recursive invocations of
1117 // getNonLocalPointerDepFromBB and other routines that could reuse the cache
1118 // value will only see properly sorted cache arrays.
1119 if (Cache && NumSortedEntries != Cache->size()) {
1120 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1121 NumSortedEntries = Cache->size();
1126 for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
1127 BasicBlock *Pred = *PI;
1128 PredList.push_back(std::make_pair(Pred, Pointer));
1130 // Get the PHI translated pointer in this predecessor. This can fail if
1131 // not translatable, in which case the getAddr() returns null.
1132 PHITransAddr &PredPointer = PredList.back().second;
1133 PredPointer.PHITranslateValue(BB, Pred, nullptr);
1135 Value *PredPtrVal = PredPointer.getAddr();
1137 // Check to see if we have already visited this pred block with another
1138 // pointer. If so, we can't do this lookup. This failure can occur
1139 // with PHI translation when a critical edge exists and the PHI node in
1140 // the successor translates to a pointer value different than the
1141 // pointer the block was first analyzed with.
1142 std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
1143 InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal));
1145 if (!InsertRes.second) {
1146 // We found the pred; take it off the list of preds to visit.
1147 PredList.pop_back();
1149 // If the predecessor was visited with PredPtr, then we already did
1150 // the analysis and can ignore it.
1151 if (InsertRes.first->second == PredPtrVal)
1154 // Otherwise, the block was previously analyzed with a different
1155 // pointer. We can't represent the result of this case, so we just
1156 // treat this as a phi translation failure.
1158 // Make sure to clean up the Visited map before continuing on to
1159 // PredTranslationFailure.
1160 for (unsigned i = 0, n = PredList.size(); i < n; ++i)
1161 Visited.erase(PredList[i].first);
1163 goto PredTranslationFailure;
1167 // Actually process results here; this need to be a separate loop to avoid
1168 // calling getNonLocalPointerDepFromBB for blocks we don't want to return
1169 // any results for. (getNonLocalPointerDepFromBB will modify our
1170 // datastructures in ways the code after the PredTranslationFailure label
1172 for (unsigned i = 0, n = PredList.size(); i < n; ++i) {
1173 BasicBlock *Pred = PredList[i].first;
1174 PHITransAddr &PredPointer = PredList[i].second;
1175 Value *PredPtrVal = PredPointer.getAddr();
1177 bool CanTranslate = true;
1178 // If PHI translation was unable to find an available pointer in this
1179 // predecessor, then we have to assume that the pointer is clobbered in
1180 // that predecessor. We can still do PRE of the load, which would insert
1181 // a computation of the pointer in this predecessor.
1183 CanTranslate = false;
1185 // FIXME: it is entirely possible that PHI translating will end up with
1186 // the same value. Consider PHI translating something like:
1187 // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
1188 // to recurse here, pedantically speaking.
1190 // If getNonLocalPointerDepFromBB fails here, that means the cached
1191 // result conflicted with the Visited list; we have to conservatively
1192 // assume it is unknown, but this also does not block PRE of the load.
1193 if (!CanTranslate ||
1194 getNonLocalPointerDepFromBB(PredPointer,
1195 Loc.getWithNewPtr(PredPtrVal),
1198 // Add the entry to the Result list.
1199 NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal);
1200 Result.push_back(Entry);
1202 // Since we had a phi translation failure, the cache for CacheKey won't
1203 // include all of the entries that we need to immediately satisfy future
1204 // queries. Mark this in NonLocalPointerDeps by setting the
1205 // BBSkipFirstBlockPair pointer to null. This requires reuse of the
1206 // cached value to do more work but not miss the phi trans failure.
1207 NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
1208 NLPI.Pair = BBSkipFirstBlockPair();
1213 // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
1214 CacheInfo = &NonLocalPointerDeps[CacheKey];
1215 Cache = &CacheInfo->NonLocalDeps;
1216 NumSortedEntries = Cache->size();
1218 // Since we did phi translation, the "Cache" set won't contain all of the
1219 // results for the query. This is ok (we can still use it to accelerate
1220 // specific block queries) but we can't do the fastpath "return all
1221 // results from the set" Clear out the indicator for this.
1222 CacheInfo->Pair = BBSkipFirstBlockPair();
1223 SkipFirstBlock = false;
1226 PredTranslationFailure:
1227 // The following code is "failure"; we can't produce a sane translation
1228 // for the given block. It assumes that we haven't modified any of
1229 // our datastructures while processing the current block.
1232 // Refresh the CacheInfo/Cache pointer if it got invalidated.
1233 CacheInfo = &NonLocalPointerDeps[CacheKey];
1234 Cache = &CacheInfo->NonLocalDeps;
1235 NumSortedEntries = Cache->size();
1238 // Since we failed phi translation, the "Cache" set won't contain all of the
1239 // results for the query. This is ok (we can still use it to accelerate
1240 // specific block queries) but we can't do the fastpath "return all
1241 // results from the set". Clear out the indicator for this.
1242 CacheInfo->Pair = BBSkipFirstBlockPair();
1244 // If *nothing* works, mark the pointer as unknown.
1246 // If this is the magic first block, return this as a clobber of the whole
1247 // incoming value. Since we can't phi translate to one of the predecessors,
1248 // we have to bail out.
1252 for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
1253 assert(I != Cache->rend() && "Didn't find current block??");
1254 if (I->getBB() != BB)
1257 assert(I->getResult().isNonLocal() &&
1258 "Should only be here with transparent block");
1259 I->setResult(MemDepResult::getUnknown());
1260 Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
1261 Pointer.getAddr()));
1266 // Okay, we're done now. If we added new values to the cache, re-sort it.
1267 SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
1268 DEBUG(AssertSorted(*Cache));
1272 /// RemoveCachedNonLocalPointerDependencies - If P exists in
1273 /// CachedNonLocalPointerInfo, remove it.
1274 void MemoryDependenceAnalysis::
1275 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
1276 CachedNonLocalPointerInfo::iterator It =
1277 NonLocalPointerDeps.find(P);
1278 if (It == NonLocalPointerDeps.end()) return;
1280 // Remove all of the entries in the BB->val map. This involves removing
1281 // instructions from the reverse map.
1282 NonLocalDepInfo &PInfo = It->second.NonLocalDeps;
1284 for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
1285 Instruction *Target = PInfo[i].getResult().getInst();
1286 if (!Target) continue; // Ignore non-local dep results.
1287 assert(Target->getParent() == PInfo[i].getBB());
1289 // Eliminating the dirty entry from 'Cache', so update the reverse info.
1290 RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
1293 // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
1294 NonLocalPointerDeps.erase(It);
1298 /// invalidateCachedPointerInfo - This method is used to invalidate cached
1299 /// information about the specified pointer, because it may be too
1300 /// conservative in memdep. This is an optional call that can be used when
1301 /// the client detects an equivalence between the pointer and some other
1302 /// value and replaces the other value with ptr. This can make Ptr available
1303 /// in more places that cached info does not necessarily keep.
1304 void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
1305 // If Ptr isn't really a pointer, just ignore it.
1306 if (!Ptr->getType()->isPointerTy()) return;
1307 // Flush store info for the pointer.
1308 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
1309 // Flush load info for the pointer.
1310 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
1313 /// invalidateCachedPredecessors - Clear the PredIteratorCache info.
1314 /// This needs to be done when the CFG changes, e.g., due to splitting
1316 void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
1320 /// removeInstruction - Remove an instruction from the dependence analysis,
1321 /// updating the dependence of instructions that previously depended on it.
1322 /// This method attempts to keep the cache coherent using the reverse map.
1323 void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
1324 // Walk through the Non-local dependencies, removing this one as the value
1325 // for any cached queries.
1326 NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
1327 if (NLDI != NonLocalDeps.end()) {
1328 NonLocalDepInfo &BlockMap = NLDI->second.first;
1329 for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
1331 if (Instruction *Inst = DI->getResult().getInst())
1332 RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
1333 NonLocalDeps.erase(NLDI);
1336 // If we have a cached local dependence query for this instruction, remove it.
1338 LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
1339 if (LocalDepEntry != LocalDeps.end()) {
1340 // Remove us from DepInst's reverse set now that the local dep info is gone.
1341 if (Instruction *Inst = LocalDepEntry->second.getInst())
1342 RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
1344 // Remove this local dependency info.
1345 LocalDeps.erase(LocalDepEntry);
1348 // If we have any cached pointer dependencies on this instruction, remove
1349 // them. If the instruction has non-pointer type, then it can't be a pointer
1352 // Remove it from both the load info and the store info. The instruction
1353 // can't be in either of these maps if it is non-pointer.
1354 if (RemInst->getType()->isPointerTy()) {
1355 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
1356 RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
1359 // Loop over all of the things that depend on the instruction we're removing.
1361 SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
1363 // If we find RemInst as a clobber or Def in any of the maps for other values,
1364 // we need to replace its entry with a dirty version of the instruction after
1365 // it. If RemInst is a terminator, we use a null dirty value.
1367 // Using a dirty version of the instruction after RemInst saves having to scan
1368 // the entire block to get to this point.
1369 MemDepResult NewDirtyVal;
1370 if (!RemInst->isTerminator())
1371 NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
1373 ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
1374 if (ReverseDepIt != ReverseLocalDeps.end()) {
1375 SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
1376 // RemInst can't be the terminator if it has local stuff depending on it.
1377 assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
1378 "Nothing can locally depend on a terminator");
1380 for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
1381 E = ReverseDeps.end(); I != E; ++I) {
1382 Instruction *InstDependingOnRemInst = *I;
1383 assert(InstDependingOnRemInst != RemInst &&
1384 "Already removed our local dep info");
1386 LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
1388 // Make sure to remember that new things depend on NewDepInst.
1389 assert(NewDirtyVal.getInst() && "There is no way something else can have "
1390 "a local dep on this if it is a terminator!");
1391 ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
1392 InstDependingOnRemInst));
1395 ReverseLocalDeps.erase(ReverseDepIt);
1397 // Add new reverse deps after scanning the set, to avoid invalidating the
1398 // 'ReverseDeps' reference.
1399 while (!ReverseDepsToAdd.empty()) {
1400 ReverseLocalDeps[ReverseDepsToAdd.back().first]
1401 .insert(ReverseDepsToAdd.back().second);
1402 ReverseDepsToAdd.pop_back();
1406 ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
1407 if (ReverseDepIt != ReverseNonLocalDeps.end()) {
1408 SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
1409 for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
1411 assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
1413 PerInstNLInfo &INLD = NonLocalDeps[*I];
1414 // The information is now dirty!
1417 for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
1418 DE = INLD.first.end(); DI != DE; ++DI) {
1419 if (DI->getResult().getInst() != RemInst) continue;
1421 // Convert to a dirty entry for the subsequent instruction.
1422 DI->setResult(NewDirtyVal);
1424 if (Instruction *NextI = NewDirtyVal.getInst())
1425 ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
1429 ReverseNonLocalDeps.erase(ReverseDepIt);
1431 // Add new reverse deps after scanning the set, to avoid invalidating 'Set'
1432 while (!ReverseDepsToAdd.empty()) {
1433 ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
1434 .insert(ReverseDepsToAdd.back().second);
1435 ReverseDepsToAdd.pop_back();
1439 // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
1440 // value in the NonLocalPointerDeps info.
1441 ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
1442 ReverseNonLocalPtrDeps.find(RemInst);
1443 if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
1444 SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
1445 SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
1447 for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
1448 E = Set.end(); I != E; ++I) {
1449 ValueIsLoadPair P = *I;
1450 assert(P.getPointer() != RemInst &&
1451 "Already removed NonLocalPointerDeps info for RemInst");
1453 NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
1455 // The cache is not valid for any specific block anymore.
1456 NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
1458 // Update any entries for RemInst to use the instruction after it.
1459 for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
1461 if (DI->getResult().getInst() != RemInst) continue;
1463 // Convert to a dirty entry for the subsequent instruction.
1464 DI->setResult(NewDirtyVal);
1466 if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
1467 ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
1470 // Re-sort the NonLocalDepInfo. Changing the dirty entry to its
1471 // subsequent value may invalidate the sortedness.
1472 std::sort(NLPDI.begin(), NLPDI.end());
1475 ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
1477 while (!ReversePtrDepsToAdd.empty()) {
1478 ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
1479 .insert(ReversePtrDepsToAdd.back().second);
1480 ReversePtrDepsToAdd.pop_back();
1485 assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
1486 AA->deleteValue(RemInst);
1487 DEBUG(verifyRemoved(RemInst));
1489 /// verifyRemoved - Verify that the specified instruction does not occur
1490 /// in our internal data structures.
1491 void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
1492 for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
1493 E = LocalDeps.end(); I != E; ++I) {
1494 assert(I->first != D && "Inst occurs in data structures");
1495 assert(I->second.getInst() != D &&
1496 "Inst occurs in data structures");
1499 for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
1500 E = NonLocalPointerDeps.end(); I != E; ++I) {
1501 assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
1502 const NonLocalDepInfo &Val = I->second.NonLocalDeps;
1503 for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
1505 assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
1508 for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
1509 E = NonLocalDeps.end(); I != E; ++I) {
1510 assert(I->first != D && "Inst occurs in data structures");
1511 const PerInstNLInfo &INLD = I->second;
1512 for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
1513 EE = INLD.first.end(); II != EE; ++II)
1514 assert(II->getResult().getInst() != D && "Inst occurs in data structures");
1517 for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
1518 E = ReverseLocalDeps.end(); I != E; ++I) {
1519 assert(I->first != D && "Inst occurs in data structures");
1520 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1521 EE = I->second.end(); II != EE; ++II)
1522 assert(*II != D && "Inst occurs in data structures");
1525 for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
1526 E = ReverseNonLocalDeps.end();
1528 assert(I->first != D && "Inst occurs in data structures");
1529 for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
1530 EE = I->second.end(); II != EE; ++II)
1531 assert(*II != D && "Inst occurs in data structures");
1534 for (ReverseNonLocalPtrDepTy::const_iterator
1535 I = ReverseNonLocalPtrDeps.begin(),
1536 E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
1537 assert(I->first != D && "Inst occurs in rev NLPD map");
1539 for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
1540 E = I->second.end(); II != E; ++II)
1541 assert(*II != ValueIsLoadPair(D, false) &&
1542 *II != ValueIsLoadPair(D, true) &&
1543 "Inst occurs in ReverseNonLocalPtrDeps map");