X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FMemoryDependenceAnalysis.cpp;h=ce7674003fe013bb2f4b02fe7bd212eb2be53db0;hb=ea9d57bc9a8eb6aaa07b250cc59774d1a1981221;hp=2413bbc02516c10f0fea607a0a06e0a8242305d7;hpb=3f7eb5b795f933aecf0e0f1fea646c516f6dc1c5;p=oota-llvm.git diff --git a/lib/Analysis/MemoryDependenceAnalysis.cpp b/lib/Analysis/MemoryDependenceAnalysis.cpp index 2413bbc0251..ce7674003fe 100644 --- a/lib/Analysis/MemoryDependenceAnalysis.cpp +++ b/lib/Analysis/MemoryDependenceAnalysis.cpp @@ -16,14 +16,15 @@ #define DEBUG_TYPE "memdep" #include "llvm/Analysis/MemoryDependenceAnalysis.h" -#include "llvm/Constants.h" #include "llvm/Instructions.h" +#include "llvm/IntrinsicInst.h" #include "llvm/Function.h" #include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Analysis/MallocHelper.h" #include "llvm/ADT/Statistic.h" -#include "llvm/Support/CFG.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/Support/PredIteratorCache.h" #include "llvm/Support/Debug.h" -#include "llvm/Target/TargetData.h" using namespace llvm; STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); @@ -36,6 +37,8 @@ STATISTIC(NumCacheDirtyNonLocalPtr, "Number of cached, but dirty, non-local ptr responses"); STATISTIC(NumUncacheNonLocalPtr, "Number of uncached non-local ptr responses"); +STATISTIC(NumCacheCompleteNonLocalPtr, + "Number of block queries that were completely cached"); char MemoryDependenceAnalysis::ID = 0; @@ -43,17 +46,36 @@ char MemoryDependenceAnalysis::ID = 0; static RegisterPass X("memdep", "Memory Dependence Analysis", false, true); +MemoryDependenceAnalysis::MemoryDependenceAnalysis() +: FunctionPass(&ID), PredCache(0) { +} +MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { +} + +/// Clean up memory in between runs +void MemoryDependenceAnalysis::releaseMemory() { + LocalDeps.clear(); + NonLocalDeps.clear(); + NonLocalPointerDeps.clear(); + ReverseLocalDeps.clear(); + ReverseNonLocalDeps.clear(); + ReverseNonLocalPtrDeps.clear(); + PredCache->clear(); +} + + + /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. /// void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesAll(); AU.addRequiredTransitive(); - AU.addRequiredTransitive(); } bool MemoryDependenceAnalysis::runOnFunction(Function &) { AA = &getAnalysis(); - TD = &getAnalysis(); + if (PredCache == 0) + PredCache.reset(new PredIteratorCache()); return false; } @@ -61,9 +83,9 @@ bool MemoryDependenceAnalysis::runOnFunction(Function &) { /// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry. template static void RemoveFromReverseMap(DenseMap > &ReverseMap, - Instruction *Inst, KeyTy *Val) { - typename DenseMap >::iterator + SmallPtrSet > &ReverseMap, + Instruction *Inst, KeyTy Val) { + typename DenseMap >::iterator InstIt = ReverseMap.find(Inst); assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); bool Found = InstIt->second.erase(Val); @@ -76,8 +98,8 @@ static void RemoveFromReverseMap(DenseMapbegin()) { Instruction *Inst = --ScanIt; @@ -87,32 +109,49 @@ getCallSiteDependencyFrom(CallSite CS, BasicBlock::iterator ScanIt, uint64_t PointerSize = 0; if (StoreInst *S = dyn_cast(Inst)) { Pointer = S->getPointerOperand(); - PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); + PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); } else if (VAArgInst *V = dyn_cast(Inst)) { Pointer = V->getOperand(0); - PointerSize = TD->getTypeStoreSize(V->getType()); + PointerSize = AA->getTypeStoreSize(V->getType()); } else if (FreeInst *F = dyn_cast(Inst)) { Pointer = F->getPointerOperand(); // FreeInsts erase the entire structure PointerSize = ~0ULL; + } else if (isFreeCall(Inst)) { + Pointer = Inst->getOperand(0); + // calls to free() erase the entire structure + PointerSize = ~0ULL; } else if (isa(Inst) || isa(Inst)) { + // Debug intrinsics don't cause dependences. + if (isa(Inst)) continue; CallSite InstCS = CallSite::get(Inst); // If these two calls do not interfere, look past it. - if (AA->getModRefInfo(CS, InstCS) == AliasAnalysis::NoModRef) + switch (AA->getModRefInfo(CS, InstCS)) { + case AliasAnalysis::NoModRef: + // If the two calls don't interact (e.g. InstCS is readnone) keep + // scanning. continue; - - // FIXME: If this is a ref/ref result, we should ignore it! - // X = strlen(P); - // Y = strlen(Q); - // Z = strlen(P); // Z = X - - // If they interfere, we generally return clobber. However, if they are - // calls to the same read-only functions we return Def. - if (!AA->onlyReadsMemory(CS) || CS.getCalledFunction() == 0 || - CS.getCalledFunction() != InstCS.getCalledFunction()) + case AliasAnalysis::Ref: + // If the two calls read the same memory locations and CS is a readonly + // function, then we have two cases: 1) the calls may not interfere with + // each other at all. 2) the calls may produce the same value. In case + // #1 we want to ignore the values, in case #2, we want to return Inst + // as a Def dependence. This allows us to CSE in cases like: + // X = strlen(P); + // memchr(...); + // Y = strlen(P); // Y = X + if (isReadOnlyCall) { + if (CS.getCalledFunction() != 0 && + CS.getCalledFunction() == InstCS.getCalledFunction()) + return MemDepResult::getDef(Inst); + // Ignore unrelated read/read call dependences. + continue; + } + // FALL THROUGH + default: return MemDepResult::getClobber(Inst); - return MemDepResult::getDef(Inst); + } } else { // Non-memory instruction. continue; @@ -140,11 +179,14 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, while (ScanIt != BB->begin()) { Instruction *Inst = --ScanIt; + // Debug intrinsics don't cause dependences. + if (isa(Inst)) continue; + // Values depend on loads if the pointers are must aliased. This means that // a load depends on another must aliased load from the same value. if (LoadInst *LI = dyn_cast(Inst)) { Value *Pointer = LI->getPointerOperand(); - uint64_t PointerSize = TD->getTypeStoreSize(LI->getType()); + uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); // If we found a pointer, check if it could be the same as our pointer. AliasAnalysis::AliasResult R = @@ -161,9 +203,17 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, } if (StoreInst *SI = dyn_cast(Inst)) { - Value *Pointer = SI->getPointerOperand(); - uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); + // If alias analysis can tell that this store is guaranteed to not modify + // the query pointer, ignore it. Use getModRefInfo to handle cases where + // the query pointer points to constant memory etc. + if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) + continue; + // Ok, this store might clobber the query pointer. Check to see if it is + // a must alias: in this case, we want to return this as a def. + Value *Pointer = SI->getPointerOperand(); + uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); + // If we found a pointer, check if it could be the same as our pointer. AliasAnalysis::AliasResult R = AA->alias(Pointer, PointerSize, MemPtr, MemSize); @@ -179,22 +229,34 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, // the allocation, return Def. This means that there is no dependence and // the access can be optimized based on that. For example, a load could // turn into undef. - if (AllocationInst *AI = dyn_cast(Inst)) { + // Note: Only determine this to be a malloc if Inst is the malloc call, not + // a subsequent bitcast of the malloc call result. There can be stores to + // the malloced memory between the malloc call and its bitcast uses, and we + // need to continue scanning until the malloc call. + if (isa(Inst) || extractMallocCall(Inst)) { Value *AccessPtr = MemPtr->getUnderlyingObject(); - if (AccessPtr == AI || - AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) - return MemDepResult::getDef(AI); + if (AccessPtr == Inst || + AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) + return MemDepResult::getDef(Inst); continue; } - + // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. - // FIXME: If this is a load, we should ignore readonly calls! - if (AA->getModRefInfo(Inst, MemPtr, MemSize) == AliasAnalysis::NoModRef) + switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { + case AliasAnalysis::NoModRef: + // If the call has no effect on the queried pointer, just ignore it. continue; - - // Otherwise, there is a dependence. - return MemDepResult::getClobber(Inst); + case AliasAnalysis::Ref: + // If the call is known to never store to the pointer, and if this is a + // load query, we can safely ignore it (scan past it). + if (isLoad) + continue; + // FALL THROUGH. + default: + // Otherwise, there is a potential dependence. Return a clobber. + return MemDepResult::getClobber(Inst); + } } // No dependence found. If this is the entry block of the function, it is a @@ -245,7 +307,7 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); else { MemPtr = SI->getPointerOperand(); - MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); + MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); } } else if (LoadInst *LI = dyn_cast(QueryInst)) { // If this is a volatile load, don't mess around with it. Just return the @@ -254,10 +316,16 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); else { MemPtr = LI->getPointerOperand(); - MemSize = TD->getTypeStoreSize(LI->getType()); + MemSize = AA->getTypeStoreSize(LI->getType()); } + } else if (isFreeCall(QueryInst)) { + MemPtr = QueryInst->getOperand(0); + // calls to free() erase the entire structure, not just a field. + MemSize = ~0UL; } else if (isa(QueryInst) || isa(QueryInst)) { - LocalCache = getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanPos, + CallSite QueryCS = CallSite::get(QueryInst); + bool isReadOnly = AA->onlyReadsMemory(QueryCS); + LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, QueryParent); } else if (FreeInst *FI = dyn_cast(QueryInst)) { MemPtr = FI->getPointerOperand(); @@ -281,22 +349,36 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { return LocalCache; } -/// getNonLocalDependency - Perform a full dependency query for the -/// specified instruction, returning the set of blocks that the value is +#ifndef NDEBUG +/// AssertSorted - This method is used when -debug is specified to verify that +/// cache arrays are properly kept sorted. +static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, + int Count = -1) { + if (Count == -1) Count = Cache.size(); + if (Count == 0) return; + + for (unsigned i = 1; i != unsigned(Count); ++i) + assert(Cache[i-1] <= Cache[i] && "Cache isn't sorted!"); +} +#endif + +/// getNonLocalCallDependency - Perform a full dependency query for the +/// specified call, returning the set of blocks that the value is /// potentially live across. The returned set of results will include a /// "NonLocal" result for all blocks where the value is live across. /// -/// This method assumes the instruction returns a "nonlocal" dependency +/// This method assumes the instruction returns a "NonLocal" dependency /// within its own block. /// +/// This returns a reference to an internal data structure that may be +/// invalidated on the next non-local query or when an instruction is +/// removed. Clients must copy this data if they want it around longer than +/// that. const MemoryDependenceAnalysis::NonLocalDepInfo & -MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { - // FIXME: Make this only be for callsites in the future. - assert(isa(QueryInst) || isa(QueryInst) || - isa(QueryInst) || isa(QueryInst)); - assert(getDependency(QueryInst).isNonLocal() && - "getNonLocalDependency should only be used on insts with non-local deps!"); - PerInstNLInfo &CacheP = NonLocalDeps[QueryInst]; +MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { + assert(getDependency(QueryCS.getInstruction()).isNonLocal() && + "getNonLocalCallDependency should only be used on calls with non-local deps!"); + PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()]; NonLocalDepInfo &Cache = CacheP.first; /// DirtyBlocks - This is the set of blocks that need to be recomputed. In @@ -328,15 +410,19 @@ MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { // << Cache.size() << " cached: " << *QueryInst; } else { // Seed DirtyBlocks with each of the preds of QueryInst's block. - BasicBlock *QueryBB = QueryInst->getParent(); - DirtyBlocks.append(pred_begin(QueryBB), pred_end(QueryBB)); + BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); + for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) + DirtyBlocks.push_back(*PI); NumUncacheNonLocal++; } - // Visited checked first, vector in sorted order. + // isReadonlyCall - If this is a read-only call, we can be more aggressive. + bool isReadonlyCall = AA->onlyReadsMemory(QueryCS); + SmallPtrSet Visited; unsigned NumSortedEntries = Cache.size(); + DEBUG(AssertSorted(Cache)); // Iterate while we still have blocks to update. while (!DirtyBlocks.empty()) { @@ -349,10 +435,11 @@ MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { // Do a binary search to see if we already have an entry for this block in // the cache set. If so, find it. + DEBUG(AssertSorted(Cache, NumSortedEntries)); NonLocalDepInfo::iterator Entry = std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries, std::make_pair(DirtyBB, MemDepResult())); - if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB) + if (Entry != Cache.begin() && prior(Entry)->first == DirtyBB) --Entry; MemDepResult *ExistingResult = 0; @@ -374,51 +461,24 @@ MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { if (Instruction *Inst = ExistingResult->getInst()) { ScanPos = Inst; // We're removing QueryInst's use of Inst. - RemoveFromReverseMap(ReverseNonLocalDeps, Inst, QueryInst); + RemoveFromReverseMap(ReverseNonLocalDeps, Inst, + QueryCS.getInstruction()); } } // Find out if this block has a local dependency for QueryInst. MemDepResult Dep; - Value *MemPtr = 0; - uint64_t MemSize = 0; - - if (ScanPos == DirtyBB->begin()) { - // No dependence found. If this is the entry block of the function, it is a - // clobber, otherwise it is non-local. - if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) - Dep = MemDepResult::getNonLocal(); - else - Dep = MemDepResult::getClobber(ScanPos); - } else if (StoreInst *SI = dyn_cast(QueryInst)) { - // If this is a volatile store, don't mess around with it. Just return the - // previous instruction as a clobber. - if (SI->isVolatile()) - Dep = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); - else { - MemPtr = SI->getPointerOperand(); - MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType()); - } - } else if (LoadInst *LI = dyn_cast(QueryInst)) { - // If this is a volatile load, don't mess around with it. Just return the - // previous instruction as a clobber. - if (LI->isVolatile()) - Dep = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); - else { - MemPtr = LI->getPointerOperand(); - MemSize = TD->getTypeStoreSize(LI->getType()); - } + if (ScanPos != DirtyBB->begin()) { + Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB); + } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) { + // No dependence found. If this is the entry block of the function, it is + // a clobber, otherwise it is non-local. + Dep = MemDepResult::getNonLocal(); } else { - assert(isa(QueryInst) || isa(QueryInst)); - Dep = getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanPos, - DirtyBB); + Dep = MemDepResult::getClobber(ScanPos); } - if (MemPtr) - Dep = getPointerDependencyFrom(MemPtr, MemSize, isa(QueryInst), - ScanPos, DirtyBB); - // If we had a dirty entry for the block, update it. Otherwise, just add // a new entry. if (ExistingResult) @@ -432,12 +492,13 @@ MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) { // Keep the ReverseNonLocalDeps map up to date so we can efficiently // update this when we remove instructions. if (Instruction *Inst = Dep.getInst()) - ReverseNonLocalDeps[Inst].insert(QueryInst); + ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction()); } else { // If the block *is* completely transparent to the load, we need to check // the predecessors of this block. Add them to our worklist. - DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB)); + for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) + DirtyBlocks.push_back(*PI); } } @@ -461,30 +522,197 @@ getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, // We know that the pointer value is live into FromBB find the def/clobbers // from presecessors. const Type *EltTy = cast(Pointer->getType())->getElementType(); - uint64_t PointeeSize = TD->getTypeStoreSize(EltTy); + uint64_t PointeeSize = AA->getTypeStoreSize(EltTy); - // While we have blocks to analyze, get their values. - SmallPtrSet Visited; + // This is the set of blocks we've inspected, and the pointer we consider in + // each block. Because of critical edges, we currently bail out if querying + // a block with multiple different pointers. This can happen during PHI + // translation. + DenseMap Visited; + if (!getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB, + Result, Visited, true)) + return; + Result.clear(); + Result.push_back(std::make_pair(FromBB, + MemDepResult::getClobber(FromBB->begin()))); +} + +/// GetNonLocalInfoForBlock - Compute the memdep value for BB with +/// Pointer/PointeeSize using either cached information in Cache or by doing a +/// lookup (which may use dirty cache info if available). If we do a lookup, +/// add the result to the cache. +MemDepResult MemoryDependenceAnalysis:: +GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, + bool isLoad, BasicBlock *BB, + NonLocalDepInfo *Cache, unsigned NumSortedEntries) { + + // Do a binary search to see if we already have an entry for this block in + // the cache set. If so, find it. + NonLocalDepInfo::iterator Entry = + std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, + std::make_pair(BB, MemDepResult())); + if (Entry != Cache->begin() && prior(Entry)->first == BB) + --Entry; + + MemDepResult *ExistingResult = 0; + if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) + ExistingResult = &Entry->second; - for (pred_iterator PI = pred_begin(FromBB), E = pred_end(FromBB); PI != E; - ++PI) { - // TODO: PHI TRANSLATE. - getNonLocalPointerDepInternal(Pointer, PointeeSize, isLoad, *PI, - Result, Visited); + // If we have a cached entry, and it is non-dirty, use it as the value for + // this dependency. + if (ExistingResult && !ExistingResult->isDirty()) { + ++NumCacheNonLocalPtr; + return *ExistingResult; + } + + // Otherwise, we have to scan for the value. If we have a dirty cache + // entry, start scanning from its position, otherwise we scan from the end + // of the block. + BasicBlock::iterator ScanPos = BB->end(); + if (ExistingResult && ExistingResult->getInst()) { + assert(ExistingResult->getInst()->getParent() == BB && + "Instruction invalidated?"); + ++NumCacheDirtyNonLocalPtr; + ScanPos = ExistingResult->getInst(); + + // Eliminating the dirty entry from 'Cache', so update the reverse info. + ValueIsLoadPair CacheKey(Pointer, isLoad); + RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); + } else { + ++NumUncacheNonLocalPtr; } + + // Scan the block for the dependency. + MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, + ScanPos, BB); + + // If we had a dirty entry for the block, update it. Otherwise, just add + // a new entry. + if (ExistingResult) + *ExistingResult = Dep; + else + Cache->push_back(std::make_pair(BB, Dep)); + + // If the block has a dependency (i.e. it isn't completely transparent to + // the value), remember the reverse association because we just added it + // to Cache! + if (Dep.isNonLocal()) + return Dep; + + // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently + // update MemDep when we remove instructions. + Instruction *Inst = Dep.getInst(); + assert(Inst && "Didn't depend on anything?"); + ValueIsLoadPair CacheKey(Pointer, isLoad); + ReverseNonLocalPtrDeps[Inst].insert(CacheKey); + return Dep; } -void MemoryDependenceAnalysis:: -getNonLocalPointerDepInternal(Value *Pointer, uint64_t PointeeSize, - bool isLoad, BasicBlock *StartBB, - SmallVectorImpl &Result, - SmallPtrSet &Visited) { - SmallVector Worklist; - Worklist.push_back(StartBB); +/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain +/// number of elements in the array that are already properly ordered. This is +/// optimized for the case when only a few entries are added. +static void +SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, + unsigned NumSortedEntries) { + switch (Cache.size() - NumSortedEntries) { + case 0: + // done, no new entries. + break; + case 2: { + // Two new entries, insert the last one into place. + MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); + Cache.pop_back(); + MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = + std::upper_bound(Cache.begin(), Cache.end()-1, Val); + Cache.insert(Entry, Val); + // FALL THROUGH. + } + case 1: + // One new entry, Just insert the new value at the appropriate position. + if (Cache.size() != 1) { + MemoryDependenceAnalysis::NonLocalDepEntry Val = Cache.back(); + Cache.pop_back(); + MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry = + std::upper_bound(Cache.begin(), Cache.end(), Val); + Cache.insert(Entry, Val); + } + break; + default: + // Added many values, do a full scale sort. + std::sort(Cache.begin(), Cache.end()); + break; + } +} + + +/// getNonLocalPointerDepFromBB - Perform a dependency query based on +/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def +/// results to the results vector and keep track of which blocks are visited in +/// 'Visited'. +/// +/// This has special behavior for the first block queries (when SkipFirstBlock +/// is true). In this special case, it ignores the contents of the specified +/// block and starts returning dependence info for its predecessors. +/// +/// This function returns false on success, or true to indicate that it could +/// not compute dependence information for some reason. This should be treated +/// as a clobber dependence on the first instruction in the predecessor block. +bool MemoryDependenceAnalysis:: +getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize, + bool isLoad, BasicBlock *StartBB, + SmallVectorImpl &Result, + DenseMap &Visited, + bool SkipFirstBlock) { // Look up the cached info for Pointer. ValueIsLoadPair CacheKey(Pointer, isLoad); - NonLocalDepInfo *Cache = &NonLocalPointerDeps[CacheKey]; + + std::pair *CacheInfo = + &NonLocalPointerDeps[CacheKey]; + NonLocalDepInfo *Cache = &CacheInfo->second; + + // If we have valid cached information for exactly the block we are + // investigating, just return it with no recomputation. + if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) { + // We have a fully cached result for this query then we can just return the + // cached results and populate the visited set. However, we have to verify + // that we don't already have conflicting results for these blocks. Check + // to ensure that if a block in the results set is in the visited set that + // it was for the same pointer query. + if (!Visited.empty()) { + for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); + I != E; ++I) { + DenseMap::iterator VI = Visited.find(I->first); + if (VI == Visited.end() || VI->second == Pointer) continue; + + // We have a pointer mismatch in a block. Just return clobber, saying + // that something was clobbered in this result. We could also do a + // non-fully cached query, but there is little point in doing this. + return true; + } + } + + for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); + I != E; ++I) { + Visited.insert(std::make_pair(I->first, Pointer)); + if (!I->second.isNonLocal()) + Result.push_back(*I); + } + ++NumCacheCompleteNonLocalPtr; + return false; + } + + // Otherwise, either this is a new block, a block with an invalid cache + // pointer or one that we're about to invalidate by putting more info into it + // than its valid cache info. If empty, the result will be valid cache info, + // otherwise it isn't. + if (Cache->empty()) + CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); + else + CacheInfo->first = BBSkipFirstBlockPair(); + + SmallVector Worklist; + Worklist.push_back(StartBB); // Keep track of the entries that we know are sorted. Previously cached // entries will all be sorted. The entries we add we only sort on demand (we @@ -492,93 +720,175 @@ getNonLocalPointerDepInternal(Value *Pointer, uint64_t PointeeSize, // won't get any reuse from currently inserted values, because we don't // revisit blocks after we insert info for them. unsigned NumSortedEntries = Cache->size(); + DEBUG(AssertSorted(*Cache)); while (!Worklist.empty()) { BasicBlock *BB = Worklist.pop_back_val(); - // Analyze the dependency of *Pointer in FromBB. See if we already have - // been here. - if (!Visited.insert(BB)) - continue; + // Skip the first block if we have it. + if (!SkipFirstBlock) { + // Analyze the dependency of *Pointer in FromBB. See if we already have + // been here. + assert(Visited.count(BB) && "Should check 'visited' before adding to WL"); - // Get the dependency info for Pointer in BB. If we have cached - // information, we will use it, otherwise we compute it. - - // Do a binary search to see if we already have an entry for this block in - // the cache set. If so, find it. - NonLocalDepInfo::iterator Entry = - std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries, - std::make_pair(BB, MemDepResult())); - if (Entry != Cache->begin() && (&*Entry)[-1].first == BB) - --Entry; + // Get the dependency info for Pointer in BB. If we have cached + // information, we will use it, otherwise we compute it. + DEBUG(AssertSorted(*Cache, NumSortedEntries)); + MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, + BB, Cache, NumSortedEntries); + + // If we got a Def or Clobber, add this to the list of results. + if (!Dep.isNonLocal()) { + Result.push_back(NonLocalDepEntry(BB, Dep)); + continue; + } + } - MemDepResult *ExistingResult = 0; - if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) - ExistingResult = &Entry->second; + // If 'Pointer' is an instruction defined in this block, then we need to do + // phi translation to change it into a value live in the predecessor block. + // If phi translation fails, then we can't continue dependence analysis. + Instruction *PtrInst = dyn_cast(Pointer); + bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB; - // If we have a cached entry, and it is non-dirty, use it as the value for - // this dependency. - MemDepResult Dep; - if (ExistingResult && !ExistingResult->isDirty()) { - Dep = *ExistingResult; - ++NumCacheNonLocalPtr; - } else { - // Otherwise, we have to scan for the value. If we have a dirty cache - // entry, start scanning from its position, otherwise we scan from the end - // of the block. - BasicBlock::iterator ScanPos = BB->end(); - if (ExistingResult && ExistingResult->getInst()) { - assert(ExistingResult->getInst()->getParent() == BB && - "Instruction invalidated?"); - ++NumCacheDirtyNonLocalPtr; - ScanPos = ExistingResult->getInst(); - - // Eliminating the dirty entry from 'Cache', so update the reverse info. - RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, - CacheKey.getOpaqueValue()); - } else { - ++NumUncacheNonLocalPtr; + // If no PHI translation is needed, just add all the predecessors of this + // block to scan them as well. + if (!NeedsPHITranslation) { + SkipFirstBlock = false; + for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { + // Verify that we haven't looked at this block yet. + std::pair::iterator, bool> + InsertRes = Visited.insert(std::make_pair(*PI, Pointer)); + if (InsertRes.second) { + // First time we've looked at *PI. + Worklist.push_back(*PI); + continue; + } + + // If we have seen this block before, but it was with a different + // pointer then we have a phi translation failure and we have to treat + // this as a clobber. + if (InsertRes.first->second != Pointer) + goto PredTranslationFailure; } + continue; + } + + // If we do need to do phi translation, then there are a bunch of different + // cases, because we have to find a Value* live in the predecessor block. We + // know that PtrInst is defined in this block at least. + + // We may have added values to the cache list before this PHI translation. + // If so, we haven't done anything to ensure that the cache remains sorted. + // Sort it now (if needed) so that recursive invocations of + // getNonLocalPointerDepFromBB and other routines that could reuse the cache + // value will only see properly sorted cache arrays. + if (Cache && NumSortedEntries != Cache->size()) { + SortNonLocalDepInfoCache(*Cache, NumSortedEntries); + NumSortedEntries = Cache->size(); + } + + // If this is directly a PHI node, just use the incoming values for each + // pred as the phi translated version. + if (PHINode *PtrPHI = dyn_cast(PtrInst)) { + Cache = 0; - // Scan the block for the dependency. - Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, ScanPos, BB); + for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { + BasicBlock *Pred = *PI; + Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred); + + // Check to see if we have already visited this pred block with another + // pointer. If so, we can't do this lookup. This failure can occur + // with PHI translation when a critical edge exists and the PHI node in + // the successor translates to a pointer value different than the + // pointer the block was first analyzed with. + std::pair::iterator, bool> + InsertRes = Visited.insert(std::make_pair(Pred, PredPtr)); + + if (!InsertRes.second) { + // If the predecessor was visited with PredPtr, then we already did + // the analysis and can ignore it. + if (InsertRes.first->second == PredPtr) + continue; + + // Otherwise, the block was previously analyzed with a different + // pointer. We can't represent the result of this case, so we just + // treat this as a phi translation failure. + goto PredTranslationFailure; + } + + // FIXME: it is entirely possible that PHI translating will end up with + // the same value. Consider PHI translating something like: + // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need* + // to recurse here, pedantically speaking. + + // If we have a problem phi translating, fall through to the code below + // to handle the failure condition. + if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred, + Result, Visited)) + goto PredTranslationFailure; + } - // If we had a dirty entry for the block, update it. Otherwise, just add - // a new entry. - if (ExistingResult) - *ExistingResult = Dep; - else - Cache->push_back(std::make_pair(BB, Dep)); + // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. + CacheInfo = &NonLocalPointerDeps[CacheKey]; + Cache = &CacheInfo->second; + NumSortedEntries = Cache->size(); - // If the block has a dependency (i.e. it isn't completely transparent to - // the value), remember the reverse association because we just added it - // to Cache! - if (!Dep.isNonLocal()) { - // Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently - // update MemDep when we remove instructions. - Instruction *Inst = Dep.getInst(); - assert(Inst && "Didn't depend on anything?"); - ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue()); - } + // Since we did phi translation, the "Cache" set won't contain all of the + // results for the query. This is ok (we can still use it to accelerate + // specific block queries) but we can't do the fastpath "return all + // results from the set" Clear out the indicator for this. + CacheInfo->first = BBSkipFirstBlockPair(); + SkipFirstBlock = false; + continue; } - // If we got a Def or Clobber, add this to the list of results. - if (!Dep.isNonLocal()) { - Result.push_back(NonLocalDepEntry(BB, Dep)); - continue; + // TODO: BITCAST, GEP. + + // cerr << "MEMDEP: Could not PHI translate: " << *Pointer; + // if (isa(PtrInst) || isa(PtrInst)) + // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0); + PredTranslationFailure: + + if (Cache == 0) { + // Refresh the CacheInfo/Cache pointer if it got invalidated. + CacheInfo = &NonLocalPointerDeps[CacheKey]; + Cache = &CacheInfo->second; + NumSortedEntries = Cache->size(); } - // Otherwise, we have to process all the predecessors of this block to scan - // them as well. - for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { - // TODO: PHI TRANSLATE. - Worklist.push_back(*PI); + // Since we did phi translation, the "Cache" set won't contain all of the + // results for the query. This is ok (we can still use it to accelerate + // specific block queries) but we can't do the fastpath "return all + // results from the set" Clear out the indicator for this. + CacheInfo->first = BBSkipFirstBlockPair(); + + // If *nothing* works, mark the pointer as being clobbered by the first + // instruction in this block. + // + // If this is the magic first block, return this as a clobber of the whole + // incoming value. Since we can't phi translate to one of the predecessors, + // we have to bail out. + if (SkipFirstBlock) + return true; + + for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) { + assert(I != Cache->rend() && "Didn't find current block??"); + if (I->first != BB) + continue; + + assert(I->second.isNonLocal() && + "Should only be here with transparent block"); + I->second = MemDepResult::getClobber(BB->begin()); + ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); + Result.push_back(*I); + break; } } - - // If we computed new values, re-sort Cache. - if (NumSortedEntries != Cache->size()) - std::sort(Cache->begin(), Cache->end()); + + // Okay, we're done now. If we added new values to the cache, re-sort it. + SortNonLocalDepInfoCache(*Cache, NumSortedEntries); + DEBUG(AssertSorted(*Cache)); + return false; } /// RemoveCachedNonLocalPointerDependencies - If P exists in @@ -591,15 +901,15 @@ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { // Remove all of the entries in the BB->val map. This involves removing // instructions from the reverse map. - NonLocalDepInfo &PInfo = It->second; + NonLocalDepInfo &PInfo = It->second.second; for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { Instruction *Target = PInfo[i].second.getInst(); if (Target == 0) continue; // Ignore non-local dep results. - assert(Target->getParent() == PInfo[i].first && Target != P.getPointer()); + assert(Target->getParent() == PInfo[i].first); // Eliminating the dirty entry from 'Cache', so update the reverse info. - RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue()); + RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P); } // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). @@ -607,6 +917,21 @@ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { } +/// invalidateCachedPointerInfo - This method is used to invalidate cached +/// information about the specified pointer, because it may be too +/// conservative in memdep. This is an optional call that can be used when +/// the client detects an equivalence between the pointer and some other +/// value and replaces the other value with ptr. This can make Ptr available +/// in more places that cached info does not necessarily keep. +void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) { + // If Ptr isn't really a pointer, just ignore it. + if (!isa(Ptr->getType())) return; + // Flush store info for the pointer. + RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false)); + // Flush load info for the pointer. + RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true)); +} + /// removeInstruction - Remove an instruction from the dependence analysis, /// updating the dependence of instructions that previously depended on it. /// This method attempts to keep the cache coherent using the reverse map. @@ -731,17 +1056,19 @@ void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = ReverseNonLocalPtrDeps.find(RemInst); if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { - SmallPtrSet &Set = ReversePtrDepIt->second; + SmallPtrSet &Set = ReversePtrDepIt->second; SmallVector,8> ReversePtrDepsToAdd; - for (SmallPtrSet::iterator I = Set.begin(), E = Set.end(); - I != E; ++I) { - ValueIsLoadPair P; - P.setFromOpaqueValue(*I); + for (SmallPtrSet::iterator I = Set.begin(), + E = Set.end(); I != E; ++I) { + ValueIsLoadPair P = *I; assert(P.getPointer() != RemInst && "Already removed NonLocalPointerDeps info for RemInst"); - NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P]; + NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; + + // The cache is not valid for any specific block anymore. + NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); // Update any entries for RemInst to use the instruction after it. for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); @@ -754,13 +1081,17 @@ void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { if (Instruction *NewDirtyInst = NewDirtyVal.getInst()) ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P)); } + + // Re-sort the NonLocalDepInfo. Changing the dirty entry to its + // subsequent value may invalidate the sortedness. + std::sort(NLPDI.begin(), NLPDI.end()); } ReverseNonLocalPtrDeps.erase(ReversePtrDepIt); while (!ReversePtrDepsToAdd.empty()) { ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first] - .insert(ReversePtrDepsToAdd.back().second.getOpaqueValue()); + .insert(ReversePtrDepsToAdd.back().second); ReversePtrDepsToAdd.pop_back(); } } @@ -770,7 +1101,6 @@ void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { AA->deleteValue(RemInst); DEBUG(verifyRemoved(RemInst)); } - /// verifyRemoved - Verify that the specified instruction does not occur /// in our internal data structures. void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { @@ -784,7 +1114,7 @@ void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), E = NonLocalPointerDeps.end(); I != E; ++I) { assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); - const NonLocalDepInfo &Val = I->second; + const NonLocalDepInfo &Val = I->second.second; for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); II != E; ++II) assert(II->second.getInst() != D && "Inst occurs as NLPD value"); @@ -821,10 +1151,10 @@ void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const { E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { assert(I->first != D && "Inst occurs in rev NLPD map"); - for (SmallPtrSet::const_iterator II = I->second.begin(), + for (SmallPtrSet::const_iterator II = I->second.begin(), E = I->second.end(); II != E; ++II) - assert(*II != ValueIsLoadPair(D, false).getOpaqueValue() && - *II != ValueIsLoadPair(D, true).getOpaqueValue() && + assert(*II != ValueIsLoadPair(D, false) && + *II != ValueIsLoadPair(D, true) && "Inst occurs in ReverseNonLocalPtrDeps map"); }