X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FMemoryDependenceAnalysis.cpp;h=6af365b76a902e4e1ab61626c7abf323b97b34bf;hb=ef4cfc749a61d0d0252196c957697436ba7ec068;hp=60e9e6b9e4b7e21e586a3ac887806429731b6346;hpb=a8701a6c62158b2b84cc24ed2149e4107d03409a;p=oota-llvm.git diff --git a/lib/Analysis/MemoryDependenceAnalysis.cpp b/lib/Analysis/MemoryDependenceAnalysis.cpp index 60e9e6b9e4b..6af365b76a9 100644 --- a/lib/Analysis/MemoryDependenceAnalysis.cpp +++ b/lib/Analysis/MemoryDependenceAnalysis.cpp @@ -14,56 +14,56 @@ // //===----------------------------------------------------------------------===// +#define DEBUG_TYPE "memdep" #include "llvm/Analysis/MemoryDependenceAnalysis.h" #include "llvm/Constants.h" #include "llvm/Instructions.h" #include "llvm/Function.h" #include "llvm/Analysis/AliasAnalysis.h" -#include "llvm/Support/CFG.h" -#include "llvm/Target/TargetData.h" #include "llvm/ADT/Statistic.h" - -#define DEBUG_TYPE "memdep" - +#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 cached non-local responses"); -STATISTIC(NumUncacheNonlocal, "Number of uncached non-local responses"); +STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); +STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses"); +STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses"); + +STATISTIC(NumCacheNonLocalPtr, + "Number of fully cached non-local ptr responses"); +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; -Instruction* const MemoryDependenceAnalysis::NonLocal = (Instruction*)-3; -Instruction* const MemoryDependenceAnalysis::None = (Instruction*)-4; -Instruction* const MemoryDependenceAnalysis::Dirty = (Instruction*)-5; - // Register this pass... static RegisterPass X("memdep", - "Memory Dependence Analysis"); + "Memory Dependence Analysis", false, true); -void MemoryDependenceAnalysis::ping(Instruction *D) { - for (depMapType::iterator I = depGraphLocal.begin(), E = depGraphLocal.end(); - I != E; ++I) { - assert(I->first != D); - assert(I->second.first != D); - } +MemoryDependenceAnalysis::MemoryDependenceAnalysis() +: FunctionPass(&ID), PredCache(0) { +} +MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { +} - for (nonLocalDepMapType::iterator I = depGraphNonLocal.begin(), E = depGraphNonLocal.end(); - I != E; ++I) { - assert(I->first != D); - } +/// Clean up memory in between runs +void MemoryDependenceAnalysis::releaseMemory() { + LocalDeps.clear(); + NonLocalDeps.clear(); + NonLocalPointerDeps.clear(); + ReverseLocalDeps.clear(); + ReverseNonLocalDeps.clear(); + ReverseNonLocalPtrDeps.clear(); + PredCache->clear(); +} - for (reverseDepMapType::iterator I = reverseDep.begin(), E = reverseDep.end(); - I != E; ++I) - for (SmallPtrSet::iterator II = I->second.begin(), EE = I->second.end(); - II != EE; ++II) - assert(*II != D); - for (reverseDepMapType::iterator I = reverseDepNonLocal.begin(), E = reverseDepNonLocal.end(); - I != E; ++I) - for (SmallPtrSet::iterator II = I->second.begin(), EE = I->second.end(); - II != EE; ++II) - assert(*II != D); -} /// getAnalysisUsage - Does not modify anything. It uses Alias Analysis. /// @@ -73,452 +73,1058 @@ void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequiredTransitive(); } -/// getCallSiteDependency - Private helper for finding the local dependencies -/// of a call site. -Instruction* MemoryDependenceAnalysis::getCallSiteDependency(CallSite C, - Instruction* start, - BasicBlock* block) { - - std::pair& cachedResult = - depGraphLocal[C.getInstruction()]; - AliasAnalysis& AA = getAnalysis(); - TargetData& TD = getAnalysis(); - BasicBlock::iterator blockBegin = C.getInstruction()->getParent()->begin(); - BasicBlock::iterator QI = C.getInstruction(); - - // If the starting point was specifiy, use it - if (start) { - QI = start; - blockBegin = start->getParent()->end(); - // If the starting point wasn't specified, but the block was, use it - } else if (!start && block) { - QI = block->end(); - blockBegin = block->end(); - } - +bool MemoryDependenceAnalysis::runOnFunction(Function &) { + AA = &getAnalysis(); + TD = &getAnalysis(); + if (PredCache == 0) + PredCache.reset(new PredIteratorCache()); + return false; +} + +/// RemoveFromReverseMap - This is a helper function that removes Val from +/// '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 + InstIt = ReverseMap.find(Inst); + assert(InstIt != ReverseMap.end() && "Reverse map out of sync?"); + bool Found = InstIt->second.erase(Val); + assert(Found && "Invalid reverse map!"); Found=Found; + if (InstIt->second.empty()) + ReverseMap.erase(InstIt); +} + + +/// getCallSiteDependencyFrom - Private helper for finding the local +/// dependencies of a call site. +MemDepResult MemoryDependenceAnalysis:: +getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, + BasicBlock::iterator ScanIt, BasicBlock *BB) { // Walk backwards through the block, looking for dependencies - while (QI != blockBegin) { - --QI; + while (ScanIt != BB->begin()) { + Instruction *Inst = --ScanIt; // If this inst is a memory op, get the pointer it accessed - Value* pointer = 0; - uint64_t pointerSize = 0; - if (StoreInst* S = dyn_cast(QI)) { - pointer = S->getPointerOperand(); - pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType()); - } else if (AllocationInst* AI = dyn_cast(QI)) { - pointer = AI; - if (ConstantInt* C = dyn_cast(AI->getArraySize())) - pointerSize = C->getZExtValue() * \ - TD.getABITypeSize(AI->getAllocatedType()); - else - pointerSize = ~0UL; - } else if (VAArgInst* V = dyn_cast(QI)) { - pointer = V->getOperand(0); - pointerSize = TD.getTypeStoreSize(V->getType()); - } else if (FreeInst* F = dyn_cast(QI)) { - pointer = F->getPointerOperand(); + Value *Pointer = 0; + uint64_t PointerSize = 0; + if (StoreInst *S = dyn_cast(Inst)) { + Pointer = S->getPointerOperand(); + PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType()); + } else if (VAArgInst *V = dyn_cast(Inst)) { + Pointer = V->getOperand(0); + PointerSize = TD->getTypeStoreSize(V->getType()); + } else if (FreeInst *F = dyn_cast(Inst)) { + Pointer = F->getPointerOperand(); // FreeInsts erase the entire structure - pointerSize = ~0UL; - } else if (isa(QI)) { - AliasAnalysis::ModRefBehavior result = - AA.getModRefBehavior(CallSite::get(QI)); - if (result != AliasAnalysis::DoesNotAccessMemory && - result != AliasAnalysis::OnlyReadsMemory) { - if (!start && !block) { - cachedResult.first = QI; - cachedResult.second = true; - reverseDep[QI].insert(C.getInstruction()); - } - return QI; - } else { + PointerSize = ~0ULL; + } else if (isa(Inst) || isa(Inst)) { + CallSite InstCS = CallSite::get(Inst); + // If these two calls do not interfere, look past it. + switch (AA->getModRefInfo(CS, InstCS)) { + case AliasAnalysis::NoModRef: + // If the two calls don't interact (e.g. InstCS is readnone) keep + // scanning. continue; + 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); } - } else + } else { + // Non-memory instruction. continue; - - if (AA.getModRefInfo(C, pointer, pointerSize) != AliasAnalysis::NoModRef) { - if (!start && !block) { - cachedResult.first = QI; - cachedResult.second = true; - reverseDep[QI].insert(C.getInstruction()); - } - return QI; } + + if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) + return MemDepResult::getClobber(Inst); } - // No dependence found - cachedResult.first = NonLocal; - cachedResult.second = true; - reverseDep[NonLocal].insert(C.getInstruction()); - return NonLocal; + // No dependence found. If this is the entry block of the function, it is a + // clobber, otherwise it is non-local. + if (BB != &BB->getParent()->getEntryBlock()) + return MemDepResult::getNonLocal(); + return MemDepResult::getClobber(ScanIt); } -/// nonLocalHelper - Private helper used to calculate non-local dependencies -/// by doing DFS on the predecessors of a block to find its dependencies -void MemoryDependenceAnalysis::nonLocalHelper(Instruction* query, - BasicBlock* block, - DenseMap& resp) { - // Set of blocks that we've already visited in our DFS - SmallPtrSet visited; - // If we're updating a dirtied cache entry, we don't need to reprocess - // already computed entries. - for (DenseMap::iterator I = resp.begin(), - E = resp.end(); I != E; ++I) - if (I->second != Dirty) - visited.insert(I->first); - - // Current stack of the DFS - SmallVector stack; - stack.push_back(block); - - // Do a basic DFS - while (!stack.empty()) { - BasicBlock* BB = stack.back(); - - // If we've already visited this block, no need to revist - if (visited.count(BB)) { - stack.pop_back(); - continue; - } - - // If we find a new block with a local dependency for query, - // then we insert the new dependency and backtrack. - if (BB != block) { - visited.insert(BB); +/// getPointerDependencyFrom - Return the instruction on which a memory +/// location depends. If isLoad is true, this routine ignore may-aliases with +/// read-only operations. +MemDepResult MemoryDependenceAnalysis:: +getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, + BasicBlock::iterator ScanIt, BasicBlock *BB) { + + // Walk backwards through the basic block, looking for dependencies. + while (ScanIt != BB->begin()) { + Instruction *Inst = --ScanIt; + + // 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()); - Instruction* localDep = getDependency(query, 0, BB); - if (localDep != NonLocal) { - resp.insert(std::make_pair(BB, localDep)); - stack.pop_back(); - + // If we found a pointer, check if it could be the same as our pointer. + AliasAnalysis::AliasResult R = + AA->alias(Pointer, PointerSize, MemPtr, MemSize); + if (R == AliasAnalysis::NoAlias) continue; - } - // If we re-encounter the starting block, we still need to search it - // because there might be a dependency in the starting block AFTER - // the position of the query. This is necessary to get loops right. - } else if (BB == block && stack.size() > 1) { - visited.insert(BB); - Instruction* localDep = getDependency(query, 0, BB); - if (localDep != query) - resp.insert(std::make_pair(BB, localDep)); + // May-alias loads don't depend on each other without a dependence. + if (isLoad && R == AliasAnalysis::MayAlias) + continue; + // Stores depend on may and must aliased loads, loads depend on must-alias + // loads. + return MemDepResult::getDef(Inst); + } + + if (StoreInst *SI = dyn_cast(Inst)) { + Value *Pointer = SI->getPointerOperand(); + uint64_t PointerSize = TD->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); - stack.pop_back(); + if (R == AliasAnalysis::NoAlias) + continue; + if (R == AliasAnalysis::MayAlias) + return MemDepResult::getClobber(Inst); + return MemDepResult::getDef(Inst); + } + + // If this is an allocation, and if we know that the accessed pointer is to + // 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)) { + Value *AccessPtr = MemPtr->getUnderlyingObject(); + if (AccessPtr == AI || + AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) + return MemDepResult::getDef(AI); continue; } - // If we didn't find anything, recurse on the precessors of this block - bool predOnStack = false; - bool inserted = false; - for (pred_iterator PI = pred_begin(BB), PE = pred_end(BB); - PI != PE; ++PI) - if (!visited.count(*PI)) { - stack.push_back(*PI); - inserted = true; - } else - predOnStack = true; - - // If we inserted a new predecessor, then we'll come back to this block - if (inserted) + // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. + switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { + case AliasAnalysis::NoModRef: + // If the call has no effect on the queried pointer, just ignore it. continue; - // If we didn't insert because we have no predecessors, then this - // query has no dependency at all. - else if (!inserted && !predOnStack) { - resp.insert(std::make_pair(BB, None)); - // If we didn't insert because our predecessors are already on the stack, - // then we might still have a dependency, but it will be discovered during - // backtracking. - } else if (!inserted && predOnStack){ - resp.insert(std::make_pair(BB, NonLocal)); + 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 + // clobber, otherwise it is non-local. + if (BB != &BB->getParent()->getEntryBlock()) + return MemDepResult::getNonLocal(); + return MemDepResult::getClobber(ScanIt); +} + +/// getDependency - Return the instruction on which a memory operation +/// depends. +MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) { + Instruction *ScanPos = QueryInst; + + // Check for a cached result + MemDepResult &LocalCache = LocalDeps[QueryInst]; + + // If the cached entry is non-dirty, just return it. Note that this depends + // on MemDepResult's default constructing to 'dirty'. + if (!LocalCache.isDirty()) + return LocalCache; - stack.pop_back(); + // Otherwise, if we have a dirty entry, we know we can start the scan at that + // instruction, which may save us some work. + if (Instruction *Inst = LocalCache.getInst()) { + ScanPos = Inst; + + RemoveFromReverseMap(ReverseLocalDeps, Inst, QueryInst); } + + BasicBlock *QueryParent = QueryInst->getParent(); + + Value *MemPtr = 0; + uint64_t MemSize = 0; + + // Do the scan. + if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { + // No dependence found. If this is the entry block of the function, it is a + // clobber, otherwise it is non-local. + if (QueryParent != &QueryParent->getParent()->getEntryBlock()) + LocalCache = MemDepResult::getNonLocal(); + else + LocalCache = MemDepResult::getClobber(QueryInst); + } 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()) + LocalCache = 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()) + LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); + else { + MemPtr = LI->getPointerOperand(); + MemSize = TD->getTypeStoreSize(LI->getType()); + } + } else if (isa(QueryInst) || isa(QueryInst)) { + 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(); + // FreeInsts erase the entire structure, not just a field. + MemSize = ~0UL; + } else { + // Non-memory instruction. + LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); + } + + // If we need to do a pointer scan, make it happen. + if (MemPtr) + LocalCache = getPointerDependencyFrom(MemPtr, MemSize, + isa(QueryInst), + ScanPos, QueryParent); + + // Remember the result! + if (Instruction *I = LocalCache.getInst()) + ReverseLocalDeps[I].insert(QueryInst); + + return LocalCache; } -/// getNonLocalDependency - Fills the passed-in map with the non-local -/// dependencies of the queries. The map will contain NonLocal for -/// blocks between the query and its dependencies. -void MemoryDependenceAnalysis::getNonLocalDependency(Instruction* query, - DenseMap& resp) { - if (depGraphNonLocal.count(query)) { - DenseMap& cached = depGraphNonLocal[query]; - NumCacheNonlocal++; - - SmallVector dirtied; - for (DenseMap::iterator I = cached.begin(), - E = cached.end(); I != E; ++I) - if (I->second == Dirty) - dirtied.push_back(I->first); - - for (SmallVector::iterator I = dirtied.begin(), - E = dirtied.end(); I != E; ++I) { - Instruction* localDep = getDependency(query, 0, *I); - if (localDep != NonLocal) - cached[*I] = localDep; - else { - cached.erase(*I); - nonLocalHelper(query, *I, cached); +#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 +/// 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::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 + /// the cached case, this can happen due to instructions being deleted etc. In + /// the uncached case, this starts out as the set of predecessors we care + /// about. + SmallVector DirtyBlocks; + + if (!Cache.empty()) { + // Okay, we have a cache entry. If we know it is not dirty, just return it + // with no computation. + if (!CacheP.second) { + NumCacheNonLocal++; + return Cache; + } + + // If we already have a partially computed set of results, scan them to + // determine what is dirty, seeding our initial DirtyBlocks worklist. + for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end(); + I != E; ++I) + if (I->second.isDirty()) + DirtyBlocks.push_back(I->first); + + // Sort the cache so that we can do fast binary search lookups below. + std::sort(Cache.begin(), Cache.end()); + + ++NumCacheDirtyNonLocal; + //cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: " + // << Cache.size() << " cached: " << *QueryInst; + } else { + // Seed DirtyBlocks with each of the preds of QueryInst's block. + BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); + for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) + DirtyBlocks.push_back(*PI); + NumUncacheNonLocal++; + } + + // 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()) { + BasicBlock *DirtyBB = DirtyBlocks.back(); + DirtyBlocks.pop_back(); + + // Already processed this block? + if (!Visited.insert(DirtyBB)) + continue; + + // 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() && prior(Entry)->first == DirtyBB) + --Entry; + + MemDepResult *ExistingResult = 0; + if (Entry != Cache.begin()+NumSortedEntries && + Entry->first == DirtyBB) { + // If we already have an entry, and if it isn't already dirty, the block + // is done. + if (!Entry->second.isDirty()) + continue; + + // Otherwise, remember this slot so we can update the value. + ExistingResult = &Entry->second; + } + + // If the dirty entry has a pointer, start scanning from it so we don't have + // to rescan the entire block. + BasicBlock::iterator ScanPos = DirtyBB->end(); + if (ExistingResult) { + if (Instruction *Inst = ExistingResult->getInst()) { + ScanPos = Inst; + // We're removing QueryInst's use of Inst. + RemoveFromReverseMap(ReverseNonLocalDeps, Inst, + QueryCS.getInstruction()); } } - resp = cached; + // Find out if this block has a local dependency for QueryInst. + MemDepResult Dep; + + 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 { + Dep = MemDepResult::getClobber(ScanPos); + } + + // 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(DirtyBB, Dep)); + // If the block has a dependency (i.e. it isn't completely transparent to + // the value), remember the association! + if (!Dep.isNonLocal()) { + // 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(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. + for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI) + DirtyBlocks.push_back(*PI); + } + } + + return Cache; +} + +/// getNonLocalPointerDependency - Perform a full dependency query for an +/// access to the specified (non-volatile) memory location, returning the +/// set of instructions that either define or clobber the value. +/// +/// This method assumes the pointer has a "NonLocal" dependency within its +/// own block. +/// +void MemoryDependenceAnalysis:: +getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, + SmallVectorImpl &Result) { + assert(isa(Pointer->getType()) && + "Can't get pointer deps of a non-pointer!"); + Result.clear(); + + // 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); + + // 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; - } else - NumUncacheNonlocal++; + 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) { - // If not, go ahead and search for non-local deps. - nonLocalHelper(query, query->getParent(), resp); + // 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; - // Update the non-local dependency cache - for (DenseMap::iterator I = resp.begin(), E = resp.end(); - I != E; ++I) { - depGraphNonLocal[query].insert(*I); - reverseDepNonLocal[I->second].insert(query); + MemDepResult *ExistingResult = 0; + if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB) + ExistingResult = &Entry->second; + + // 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.getOpaqueValue()); + } 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.getOpaqueValue()); + return Dep; } -/// getDependency - Return the instruction on which a memory operation -/// depends. The local paramter indicates if the query should only -/// evaluate dependencies within the same basic block. -Instruction* MemoryDependenceAnalysis::getDependency(Instruction* query, - Instruction* start, - BasicBlock* block) { - // Start looking for dependencies with the queried inst - BasicBlock::iterator QI = query; + +/// 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) { - // Check for a cached result - std::pair& cachedResult = depGraphLocal[query]; - // If we have a _confirmed_ cached entry, return it - if (!block && !start) { - if (cachedResult.second) - return cachedResult.first; - else if (cachedResult.first && cachedResult.first != NonLocal) - // If we have an unconfirmed cached entry, we can start our search from there - QI = cachedResult.first; + // Look up the cached info for Pointer. + ValueIsLoadPair CacheKey(Pointer, isLoad); + + 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; } - if (start) - QI = start; - else if (!start && block) - QI = block->end(); - - AliasAnalysis& AA = getAnalysis(); - TargetData& TD = getAnalysis(); - - // Get the pointer value for which dependence will be determined - Value* dependee = 0; - uint64_t dependeeSize = 0; - bool queryIsVolatile = false; - if (StoreInst* S = dyn_cast(query)) { - dependee = S->getPointerOperand(); - dependeeSize = TD.getTypeStoreSize(S->getOperand(0)->getType()); - queryIsVolatile = S->isVolatile(); - } else if (LoadInst* L = dyn_cast(query)) { - dependee = L->getPointerOperand(); - dependeeSize = TD.getTypeStoreSize(L->getType()); - queryIsVolatile = L->isVolatile(); - } else if (VAArgInst* V = dyn_cast(query)) { - dependee = V->getOperand(0); - dependeeSize = TD.getTypeStoreSize(V->getType()); - } else if (FreeInst* F = dyn_cast(query)) { - dependee = F->getPointerOperand(); - - // FreeInsts erase the entire structure, not just a field - dependeeSize = ~0UL; - } else if (CallSite::get(query).getInstruction() != 0) - return getCallSiteDependency(CallSite::get(query), start, block); - else if (isa(query)) - return None; + // 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 - return None; + CacheInfo->first = BBSkipFirstBlockPair(); + + SmallVector Worklist; + Worklist.push_back(StartBB); - BasicBlock::iterator blockBegin = block ? block->begin() - : query->getParent()->begin(); + // 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 + // don't insert every element into its sorted position). We know that we + // 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)); - // Walk backwards through the basic block, looking for dependencies - while (QI != blockBegin) { - --QI; + while (!Worklist.empty()) { + BasicBlock *BB = Worklist.pop_back_val(); - // If this inst is a memory op, get the pointer it accessed - Value* pointer = 0; - uint64_t pointerSize = 0; - if (StoreInst* S = dyn_cast(QI)) { - // All volatile loads/stores depend on each other - if (queryIsVolatile && S->isVolatile()) { - if (!start && !block) { - cachedResult.first = S; - cachedResult.second = true; - reverseDep[S].insert(query); - } - - return S; - } + // 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. + DEBUG(AssertSorted(*Cache, NumSortedEntries)); + MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad, + BB, Cache, NumSortedEntries); - pointer = S->getPointerOperand(); - pointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType()); - } else if (LoadInst* L = dyn_cast(QI)) { - // All volatile loads/stores depend on each other - if (queryIsVolatile && L->isVolatile()) { - if (!start && !block) { - cachedResult.first = L; - cachedResult.second = true; - reverseDep[L].insert(query); - } - - return L; + // If we got a Def or Clobber, add this to the list of results. + if (!Dep.isNonLocal()) { + Result.push_back(NonLocalDepEntry(BB, Dep)); + continue; } - - pointer = L->getPointerOperand(); - pointerSize = TD.getTypeStoreSize(L->getType()); - } else if (AllocationInst* AI = dyn_cast(QI)) { - pointer = AI; - if (ConstantInt* C = dyn_cast(AI->getArraySize())) - pointerSize = C->getZExtValue() * \ - TD.getABITypeSize(AI->getAllocatedType()); - else - pointerSize = ~0UL; - } else if (VAArgInst* V = dyn_cast(QI)) { - pointer = V->getOperand(0); - pointerSize = TD.getTypeStoreSize(V->getType()); - } else if (FreeInst* F = dyn_cast(QI)) { - pointer = F->getPointerOperand(); - - // FreeInsts erase the entire structure - pointerSize = ~0UL; - } else if (CallSite::get(QI).getInstruction() != 0) { - // Call insts need special handling. Check if they can modify our pointer - AliasAnalysis::ModRefResult MR = AA.getModRefInfo(CallSite::get(QI), - dependee, dependeeSize); - - if (MR != AliasAnalysis::NoModRef) { - // Loads don't depend on read-only calls - if (isa(query) && MR == AliasAnalysis::Ref) + } + + // 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 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 (!start && !block) { - cachedResult.first = QI; - cachedResult.second = true; - reverseDep[QI].insert(query); } - return QI; - } else { - 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 found a pointer, check if it could be the same as our pointer - if (pointer) { - AliasAnalysis::AliasResult R = AA.alias(pointer, pointerSize, - dependee, dependeeSize); - - if (R != AliasAnalysis::NoAlias) { - // May-alias loads don't depend on each other - if (isa(query) && isa(QI) && - R == AliasAnalysis::MayAlias) - 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. + + // 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)) { + for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { + BasicBlock *Pred = *PI; + Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred); - if (!start && !block) { - cachedResult.first = QI; - cachedResult.second = true; - reverseDep[QI].insert(query); + // 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; } + + // 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 that could reuse the cache + // value will only see properly sorted cache arrays. + if (Cache && NumSortedEntries != Cache->size()) + std::sort(Cache->begin(), Cache->end()); + Cache = 0; - return QI; + // 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; } + + // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. + CacheInfo = &NonLocalPointerDeps[CacheKey]; + Cache = &CacheInfo->second; + NumSortedEntries = Cache->size(); + + // 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; + } + + // 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(); + } else if (NumSortedEntries != Cache->size()) { + std::sort(Cache->begin(), Cache->end()); + NumSortedEntries = Cache->size(); + } + + // 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.getOpaqueValue()); + Result.push_back(*I); + break; } } + + // Okay, we're done now. If we added new values to the cache, re-sort it. + switch (Cache->size()-NumSortedEntries) { + case 0: + // done, no new entries. + break; + case 2: { + // Two new entries, insert the last one into place. + NonLocalDepEntry Val = Cache->back(); + Cache->pop_back(); + 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) { + NonLocalDepEntry Val = Cache->back(); + Cache->pop_back(); + 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()); + } + DEBUG(AssertSorted(*Cache)); + return false; +} + +/// RemoveCachedNonLocalPointerDependencies - If P exists in +/// CachedNonLocalPointerInfo, remove it. +void MemoryDependenceAnalysis:: +RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) { + CachedNonLocalPointerInfo::iterator It = + NonLocalPointerDeps.find(P); + if (It == NonLocalPointerDeps.end()) return; + + // Remove all of the entries in the BB->val map. This involves removing + // instructions from the reverse map. + NonLocalDepInfo &PInfo = It->second.second; - // If we found nothing, return the non-local flag - if (!start && !block) { - cachedResult.first = NonLocal; - cachedResult.second = true; - reverseDep[NonLocal].insert(query); + 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); + + // Eliminating the dirty entry from 'Cache', so update the reverse info. + RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue()); } - return NonLocal; + // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo). + NonLocalPointerDeps.erase(It); +} + + +/// 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. -void MemoryDependenceAnalysis::removeInstruction(Instruction* rem) { - // Figure out the new dep for things that currently depend on rem - Instruction* newDep = NonLocal; - - for (DenseMap::iterator DI = - depGraphNonLocal[rem].begin(), DE = depGraphNonLocal[rem].end(); - DI != DE; ++DI) - if (DI->second != None) - reverseDepNonLocal[DI->second].erase(rem); - - depMapType::iterator depGraphEntry = depGraphLocal.find(rem); - - if (depGraphEntry != depGraphLocal.end()) { - reverseDep[depGraphLocal[rem].first].erase(rem); - - if (depGraphEntry->second.first != NonLocal && - depGraphEntry->second.first != None && - depGraphEntry->second.second) { - // If we have dep info for rem, set them to it - BasicBlock::iterator RI = depGraphEntry->second.first; - RI++; - newDep = RI; - } else if ( (depGraphEntry->second.first == NonLocal || - depGraphEntry->second.first == None ) && - depGraphEntry->second.second ) { - // If we have a confirmed non-local flag, use it - newDep = depGraphEntry->second.first; - } else { - // Otherwise, use the immediate successor of rem - // NOTE: This is because, when getDependence is called, it will first - // check the immediate predecessor of what is in the cache. - BasicBlock::iterator RI = rem; - RI++; - newDep = RI; - } - } else { - // Otherwise, use the immediate successor of rem - // NOTE: This is because, when getDependence is called, it will first - // check the immediate predecessor of what is in the cache. - BasicBlock::iterator RI = rem; - RI++; - newDep = RI; +void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { + // Walk through the Non-local dependencies, removing this one as the value + // for any cached queries. + NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst); + if (NLDI != NonLocalDeps.end()) { + NonLocalDepInfo &BlockMap = NLDI->second.first; + for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end(); + DI != DE; ++DI) + if (Instruction *Inst = DI->second.getInst()) + RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst); + NonLocalDeps.erase(NLDI); + } + + // If we have a cached local dependence query for this instruction, remove it. + // + LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst); + if (LocalDepEntry != LocalDeps.end()) { + // Remove us from DepInst's reverse set now that the local dep info is gone. + if (Instruction *Inst = LocalDepEntry->second.getInst()) + RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst); + + // Remove this local dependency info. + LocalDeps.erase(LocalDepEntry); } - SmallPtrSet& set = reverseDep[rem]; - for (SmallPtrSet::iterator I = set.begin(), E = set.end(); - I != E; ++I) { - // Insert the new dependencies - // Mark it as unconfirmed as long as it is not the non-local flag - depGraphLocal[*I] = std::make_pair(newDep, (newDep == NonLocal || - newDep == None)); + // If we have any cached pointer dependencies on this instruction, remove + // them. If the instruction has non-pointer type, then it can't be a pointer + // base. + + // Remove it from both the load info and the store info. The instruction + // can't be in either of these maps if it is non-pointer. + if (isa(RemInst->getType())) { + RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false)); + RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true)); } - depGraphLocal.erase(rem); - reverseDep.erase(rem); + // Loop over all of the things that depend on the instruction we're removing. + // + SmallVector, 8> ReverseDepsToAdd; + + // If we find RemInst as a clobber or Def in any of the maps for other values, + // we need to replace its entry with a dirty version of the instruction after + // it. If RemInst is a terminator, we use a null dirty value. + // + // Using a dirty version of the instruction after RemInst saves having to scan + // the entire block to get to this point. + MemDepResult NewDirtyVal; + if (!RemInst->isTerminator()) + NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst)); - if (reverseDepNonLocal.count(rem)) { - SmallPtrSet& set = reverseDepNonLocal[rem]; - for (SmallPtrSet::iterator I = set.begin(), E = set.end(); - I != E; ++I) - for (DenseMap::iterator DI = - depGraphNonLocal[*I].begin(), DE = depGraphNonLocal[*I].end(); - DI != DE; ++DI) - if (DI->second == rem) - DI->second = Dirty; + ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst); + if (ReverseDepIt != ReverseLocalDeps.end()) { + SmallPtrSet &ReverseDeps = ReverseDepIt->second; + // RemInst can't be the terminator if it has local stuff depending on it. + assert(!ReverseDeps.empty() && !isa(RemInst) && + "Nothing can locally depend on a terminator"); + for (SmallPtrSet::iterator I = ReverseDeps.begin(), + E = ReverseDeps.end(); I != E; ++I) { + Instruction *InstDependingOnRemInst = *I; + assert(InstDependingOnRemInst != RemInst && + "Already removed our local dep info"); + + LocalDeps[InstDependingOnRemInst] = NewDirtyVal; + + // Make sure to remember that new things depend on NewDepInst. + assert(NewDirtyVal.getInst() && "There is no way something else can have " + "a local dep on this if it is a terminator!"); + ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(), + InstDependingOnRemInst)); + } + + ReverseLocalDeps.erase(ReverseDepIt); + + // Add new reverse deps after scanning the set, to avoid invalidating the + // 'ReverseDeps' reference. + while (!ReverseDepsToAdd.empty()) { + ReverseLocalDeps[ReverseDepsToAdd.back().first] + .insert(ReverseDepsToAdd.back().second); + ReverseDepsToAdd.pop_back(); + } } - reverseDepNonLocal.erase(rem); - nonLocalDepMapType::iterator I = depGraphNonLocal.find(rem); - if (I != depGraphNonLocal.end()) - depGraphNonLocal.erase(I); + ReverseDepIt = ReverseNonLocalDeps.find(RemInst); + if (ReverseDepIt != ReverseNonLocalDeps.end()) { + SmallPtrSet &Set = ReverseDepIt->second; + for (SmallPtrSet::iterator I = Set.begin(), E = Set.end(); + I != E; ++I) { + assert(*I != RemInst && "Already removed NonLocalDep info for RemInst"); + + PerInstNLInfo &INLD = NonLocalDeps[*I]; + // The information is now dirty! + INLD.second = true; + + for (NonLocalDepInfo::iterator DI = INLD.first.begin(), + DE = INLD.first.end(); DI != DE; ++DI) { + if (DI->second.getInst() != RemInst) continue; + + // Convert to a dirty entry for the subsequent instruction. + DI->second = NewDirtyVal; + + if (Instruction *NextI = NewDirtyVal.getInst()) + ReverseDepsToAdd.push_back(std::make_pair(NextI, *I)); + } + } - getAnalysis().deleteValue(rem); + ReverseNonLocalDeps.erase(ReverseDepIt); + + // Add new reverse deps after scanning the set, to avoid invalidating 'Set' + while (!ReverseDepsToAdd.empty()) { + ReverseNonLocalDeps[ReverseDepsToAdd.back().first] + .insert(ReverseDepsToAdd.back().second); + ReverseDepsToAdd.pop_back(); + } + } + + // If the instruction is in ReverseNonLocalPtrDeps then it appears as a + // value in the NonLocalPointerDeps info. + ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt = + ReverseNonLocalPtrDeps.find(RemInst); + if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) { + SmallPtrSet &Set = ReversePtrDepIt->second; + SmallVector,8> ReversePtrDepsToAdd; + + for (SmallPtrSet::iterator I = Set.begin(), E = Set.end(); + I != E; ++I) { + ValueIsLoadPair P; + P.setFromOpaqueValue(*I); + assert(P.getPointer() != RemInst && + "Already removed NonLocalPointerDeps info for RemInst"); + + 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(); + DI != DE; ++DI) { + if (DI->second.getInst() != RemInst) continue; + + // Convert to a dirty entry for the subsequent instruction. + DI->second = NewDirtyVal; + + 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()); + ReversePtrDepsToAdd.pop_back(); + } + } + + + assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?"); + 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 { + for (LocalDepMapType::const_iterator I = LocalDeps.begin(), + E = LocalDeps.end(); I != E; ++I) { + assert(I->first != D && "Inst occurs in data structures"); + assert(I->second.getInst() != D && + "Inst occurs in data structures"); + } + + 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.second; + for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); + II != E; ++II) + assert(II->second.getInst() != D && "Inst occurs as NLPD value"); + } + + for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(), + E = NonLocalDeps.end(); I != E; ++I) { + assert(I->first != D && "Inst occurs in data structures"); + const PerInstNLInfo &INLD = I->second; + for (NonLocalDepInfo::const_iterator II = INLD.first.begin(), + EE = INLD.first.end(); II != EE; ++II) + assert(II->second.getInst() != D && "Inst occurs in data structures"); + } + + for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(), + E = ReverseLocalDeps.end(); I != E; ++I) { + assert(I->first != D && "Inst occurs in data structures"); + for (SmallPtrSet::const_iterator II = I->second.begin(), + EE = I->second.end(); II != EE; ++II) + assert(*II != D && "Inst occurs in data structures"); + } + + for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(), + E = ReverseNonLocalDeps.end(); + I != E; ++I) { + assert(I->first != D && "Inst occurs in data structures"); + for (SmallPtrSet::const_iterator II = I->second.begin(), + EE = I->second.end(); II != EE; ++II) + assert(*II != D && "Inst occurs in data structures"); + } + + for (ReverseNonLocalPtrDepTy::const_iterator + I = ReverseNonLocalPtrDeps.begin(), + E = ReverseNonLocalPtrDeps.end(); I != E; ++I) { + assert(I->first != D && "Inst occurs in rev NLPD map"); + + 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() && + "Inst occurs in ReverseNonLocalPtrDeps map"); + } + }