X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FAnalysis%2FMemoryDependenceAnalysis.cpp;h=9872890494556e61a132acce9e0defcd736c29f9;hb=7372a7d5f87bf1ff65d07f25bae037ddc4df994d;hp=d9d085aee6bc1e6ae3f1d993ed0ad6933f109edb;hpb=4ec2258ffb495d7ce00177e447740ef1123a27db;p=oota-llvm.git diff --git a/lib/Analysis/MemoryDependenceAnalysis.cpp b/lib/Analysis/MemoryDependenceAnalysis.cpp index d9d085aee6b..98728904945 100644 --- a/lib/Analysis/MemoryDependenceAnalysis.cpp +++ b/lib/Analysis/MemoryDependenceAnalysis.cpp @@ -19,15 +19,18 @@ #include "llvm/Instructions.h" #include "llvm/IntrinsicInst.h" #include "llvm/Function.h" +#include "llvm/LLVMContext.h" #include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/MemoryBuiltins.h" #include "llvm/Analysis/PHITransAddr.h" +#include "llvm/Analysis/ValueTracking.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/STLExtras.h" #include "llvm/Support/PredIteratorCache.h" #include "llvm/Support/Debug.h" +#include "llvm/DataLayout.h" using namespace llvm; STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses"); @@ -43,14 +46,23 @@ STATISTIC(NumUncacheNonLocalPtr, STATISTIC(NumCacheCompleteNonLocalPtr, "Number of block queries that were completely cached"); +// Limit for the number of instructions to scan in a block. +// FIXME: Figure out what a sane value is for this. +// (500 is relatively insane.) +static const int BlockScanLimit = 500; + char MemoryDependenceAnalysis::ID = 0; // Register this pass... -static RegisterPass X("memdep", - "Memory Dependence Analysis", false, true); +INITIALIZE_PASS_BEGIN(MemoryDependenceAnalysis, "memdep", + "Memory Dependence Analysis", false, true) +INITIALIZE_AG_DEPENDENCY(AliasAnalysis) +INITIALIZE_PASS_END(MemoryDependenceAnalysis, "memdep", + "Memory Dependence Analysis", false, true) MemoryDependenceAnalysis::MemoryDependenceAnalysis() -: FunctionPass(&ID), PredCache(0) { +: FunctionPass(ID), PredCache(0) { + initializeMemoryDependenceAnalysisPass(*PassRegistry::getPassRegistry()); } MemoryDependenceAnalysis::~MemoryDependenceAnalysis() { } @@ -77,6 +89,8 @@ void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { bool MemoryDependenceAnalysis::runOnFunction(Function &) { AA = &getAnalysis(); + TD = getAnalysisIfAvailable(); + DT = getAnalysisIfAvailable(); if (PredCache == 0) PredCache.reset(new PredIteratorCache()); return false; @@ -92,179 +106,369 @@ static void RemoveFromReverseMap(DenseMapsecond.erase(Val); - assert(Found && "Invalid reverse map!"); Found=Found; + assert(Found && "Invalid reverse map!"); (void)Found; if (InstIt->second.empty()) ReverseMap.erase(InstIt); } +/// GetLocation - If the given instruction references a specific memory +/// location, fill in Loc with the details, otherwise set Loc.Ptr to null. +/// Return a ModRefInfo value describing the general behavior of the +/// instruction. +static +AliasAnalysis::ModRefResult GetLocation(const Instruction *Inst, + AliasAnalysis::Location &Loc, + AliasAnalysis *AA) { + if (const LoadInst *LI = dyn_cast(Inst)) { + if (LI->isUnordered()) { + Loc = AA->getLocation(LI); + return AliasAnalysis::Ref; + } else if (LI->getOrdering() == Monotonic) { + Loc = AA->getLocation(LI); + return AliasAnalysis::ModRef; + } + Loc = AliasAnalysis::Location(); + return AliasAnalysis::ModRef; + } + + if (const StoreInst *SI = dyn_cast(Inst)) { + if (SI->isUnordered()) { + Loc = AA->getLocation(SI); + return AliasAnalysis::Mod; + } else if (SI->getOrdering() == Monotonic) { + Loc = AA->getLocation(SI); + return AliasAnalysis::ModRef; + } + Loc = AliasAnalysis::Location(); + return AliasAnalysis::ModRef; + } + + if (const VAArgInst *V = dyn_cast(Inst)) { + Loc = AA->getLocation(V); + return AliasAnalysis::ModRef; + } + + if (const CallInst *CI = isFreeCall(Inst, AA->getTargetLibraryInfo())) { + // calls to free() deallocate the entire structure + Loc = AliasAnalysis::Location(CI->getArgOperand(0)); + return AliasAnalysis::Mod; + } + + if (const IntrinsicInst *II = dyn_cast(Inst)) + switch (II->getIntrinsicID()) { + case Intrinsic::lifetime_start: + case Intrinsic::lifetime_end: + case Intrinsic::invariant_start: + Loc = AliasAnalysis::Location(II->getArgOperand(1), + cast(II->getArgOperand(0)) + ->getZExtValue(), + II->getMetadata(LLVMContext::MD_tbaa)); + // These intrinsics don't really modify the memory, but returning Mod + // will allow them to be handled conservatively. + return AliasAnalysis::Mod; + case Intrinsic::invariant_end: + Loc = AliasAnalysis::Location(II->getArgOperand(2), + cast(II->getArgOperand(1)) + ->getZExtValue(), + II->getMetadata(LLVMContext::MD_tbaa)); + // These intrinsics don't really modify the memory, but returning Mod + // will allow them to be handled conservatively. + return AliasAnalysis::Mod; + default: + break; + } + + // Otherwise, just do the coarse-grained thing that always works. + if (Inst->mayWriteToMemory()) + return AliasAnalysis::ModRef; + if (Inst->mayReadFromMemory()) + return AliasAnalysis::Ref; + return AliasAnalysis::NoModRef; +} /// getCallSiteDependencyFrom - Private helper for finding the local /// dependencies of a call site. MemDepResult MemoryDependenceAnalysis:: getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall, BasicBlock::iterator ScanIt, BasicBlock *BB) { + unsigned Limit = BlockScanLimit; + // Walk backwards through the block, looking for dependencies while (ScanIt != BB->begin()) { + // Limit the amount of scanning we do so we don't end up with quadratic + // running time on extreme testcases. + --Limit; + if (!Limit) + return MemDepResult::getUnknown(); + 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(Inst)) { - Pointer = S->getPointerOperand(); - PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); - } else if (VAArgInst *V = dyn_cast(Inst)) { - Pointer = V->getOperand(0); - PointerSize = AA->getTypeStoreSize(V->getType()); - } else if (isFreeCall(Inst)) { - Pointer = Inst->getOperand(0); - // calls to free() erase the entire structure - PointerSize = ~0ULL; - } else if (isa(Inst) || isa(Inst)) { + AliasAnalysis::Location Loc; + AliasAnalysis::ModRefResult MR = GetLocation(Inst, Loc, AA); + if (Loc.Ptr) { + // A simple instruction. + if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef) + return MemDepResult::getClobber(Inst); + continue; + } + + if (CallSite InstCS = cast(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. switch (AA->getModRefInfo(CS, InstCS)) { case AliasAnalysis::NoModRef: - // If the two calls don't interact (e.g. InstCS is readnone) keep - // scanning. + // If the two calls are the same, return InstCS as a Def, so that + // CS can be found redundant and eliminated. + if (isReadOnlyCall && !(MR & AliasAnalysis::Mod) && + CS.getInstruction()->isIdenticalToWhenDefined(Inst)) + return MemDepResult::getDef(Inst); + + // Otherwise 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 { - // Non-memory instruction. - continue; } - - if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) + + // If we could not obtain a pointer for the instruction and the instruction + // touches memory then assume that this is a dependency. + if (MR != AliasAnalysis::NoModRef) 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. + + // No dependence found. If this is the entry block of the function, it is + // unknown, otherwise it is non-local. if (BB != &BB->getParent()->getEntryBlock()) return MemDepResult::getNonLocal(); - return MemDepResult::getClobber(ScanIt); + return MemDepResult::getNonFuncLocal(); +} + +/// isLoadLoadClobberIfExtendedToFullWidth - Return true if LI is a load that +/// would fully overlap MemLoc if done as a wider legal integer load. +/// +/// MemLocBase, MemLocOffset are lazily computed here the first time the +/// base/offs of memloc is needed. +static bool +isLoadLoadClobberIfExtendedToFullWidth(const AliasAnalysis::Location &MemLoc, + const Value *&MemLocBase, + int64_t &MemLocOffs, + const LoadInst *LI, + const DataLayout *TD) { + // If we have no target data, we can't do this. + if (TD == 0) return false; + + // If we haven't already computed the base/offset of MemLoc, do so now. + if (MemLocBase == 0) + MemLocBase = GetPointerBaseWithConstantOffset(MemLoc.Ptr, MemLocOffs, *TD); + + unsigned Size = MemoryDependenceAnalysis:: + getLoadLoadClobberFullWidthSize(MemLocBase, MemLocOffs, MemLoc.Size, + LI, *TD); + return Size != 0; +} + +/// getLoadLoadClobberFullWidthSize - This is a little bit of analysis that +/// looks at a memory location for a load (specified by MemLocBase, Offs, +/// and Size) and compares it against a load. If the specified load could +/// be safely widened to a larger integer load that is 1) still efficient, +/// 2) safe for the target, and 3) would provide the specified memory +/// location value, then this function returns the size in bytes of the +/// load width to use. If not, this returns zero. +unsigned MemoryDependenceAnalysis:: +getLoadLoadClobberFullWidthSize(const Value *MemLocBase, int64_t MemLocOffs, + unsigned MemLocSize, const LoadInst *LI, + const DataLayout &TD) { + // We can only extend simple integer loads. + if (!isa(LI->getType()) || !LI->isSimple()) return 0; + + // Get the base of this load. + int64_t LIOffs = 0; + const Value *LIBase = + GetPointerBaseWithConstantOffset(LI->getPointerOperand(), LIOffs, TD); + + // If the two pointers are not based on the same pointer, we can't tell that + // they are related. + if (LIBase != MemLocBase) return 0; + + // Okay, the two values are based on the same pointer, but returned as + // no-alias. This happens when we have things like two byte loads at "P+1" + // and "P+3". Check to see if increasing the size of the "LI" load up to its + // alignment (or the largest native integer type) will allow us to load all + // the bits required by MemLoc. + + // If MemLoc is before LI, then no widening of LI will help us out. + if (MemLocOffs < LIOffs) return 0; + + // Get the alignment of the load in bytes. We assume that it is safe to load + // any legal integer up to this size without a problem. For example, if we're + // looking at an i8 load on x86-32 that is known 1024 byte aligned, we can + // widen it up to an i32 load. If it is known 2-byte aligned, we can widen it + // to i16. + unsigned LoadAlign = LI->getAlignment(); + + int64_t MemLocEnd = MemLocOffs+MemLocSize; + + // If no amount of rounding up will let MemLoc fit into LI, then bail out. + if (LIOffs+LoadAlign < MemLocEnd) return 0; + + // This is the size of the load to try. Start with the next larger power of + // two. + unsigned NewLoadByteSize = LI->getType()->getPrimitiveSizeInBits()/8U; + NewLoadByteSize = NextPowerOf2(NewLoadByteSize); + + while (1) { + // If this load size is bigger than our known alignment or would not fit + // into a native integer register, then we fail. + if (NewLoadByteSize > LoadAlign || + !TD.fitsInLegalInteger(NewLoadByteSize*8)) + return 0; + + if (LIOffs+NewLoadByteSize > MemLocEnd && + LI->getParent()->getParent()->getFnAttributes(). + hasAttribute(Attributes::AddressSafety)) + // We will be reading past the location accessed by the original program. + // While this is safe in a regular build, Address Safety analysis tools + // may start reporting false warnings. So, don't do widening. + return 0; + + // If a load of this width would include all of MemLoc, then we succeed. + if (LIOffs+NewLoadByteSize >= MemLocEnd) + return NewLoadByteSize; + + NewLoadByteSize <<= 1; + } } /// getPointerDependencyFrom - Return the instruction on which a memory -/// location depends. If isLoad is true, this routine ignore may-aliases with -/// read-only operations. +/// location depends. If isLoad is true, this routine ignores may-aliases with +/// read-only operations. If isLoad is false, this routine ignores may-aliases +/// with reads from read-only locations. MemDepResult MemoryDependenceAnalysis:: -getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, +getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad, BasicBlock::iterator ScanIt, BasicBlock *BB) { - Value *InvariantTag = 0; + const Value *MemLocBase = 0; + int64_t MemLocOffset = 0; + + unsigned Limit = BlockScanLimit; // Walk backwards through the basic block, looking for dependencies. while (ScanIt != BB->begin()) { + // Limit the amount of scanning we do so we don't end up with quadratic + // running time on extreme testcases. + --Limit; + if (!Limit) + return MemDepResult::getUnknown(); + Instruction *Inst = --ScanIt; - // If we're in an invariant region, no dependencies can be found before - // we pass an invariant-begin marker. - if (InvariantTag == Inst) { - InvariantTag = 0; - continue; - } - if (IntrinsicInst *II = dyn_cast(Inst)) { - // Debug intrinsics don't cause dependences. - if (isa(Inst)) continue; - - // If we pass an invariant-end marker, then we've just entered an - // invariant region and can start ignoring dependencies. - if (II->getIntrinsicID() == Intrinsic::invariant_end) { - // FIXME: This only considers queries directly on the invariant-tagged - // pointer, not on query pointers that are indexed off of them. It'd - // be nice to handle that at some point. - AliasAnalysis::AliasResult R = - AA->alias(II->getOperand(2), ~0U, MemPtr, ~0U); - if (R == AliasAnalysis::MustAlias) { - InvariantTag = II->getOperand(0); - continue; - } + // Debug intrinsics don't (and can't) cause dependences. + if (isa(II)) continue; // If we reach a lifetime begin or end marker, then the query ends here // because the value is undefined. - } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) { + if (II->getIntrinsicID() == Intrinsic::lifetime_start) { // FIXME: This only considers queries directly on the invariant-tagged // pointer, not on query pointers that are indexed off of them. It'd - // be nice to handle that at some point. - AliasAnalysis::AliasResult R = - AA->alias(II->getOperand(1), ~0U, MemPtr, ~0U); - if (R == AliasAnalysis::MustAlias) + // be nice to handle that at some point (the right approach is to use + // GetPointerBaseWithConstantOffset). + if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)), + MemLoc)) return MemDepResult::getDef(II); + continue; } } - // If we're querying on a load and we're in an invariant region, we're done - // at this point. Nothing a load depends on can live in an invariant region. - if (isLoad && InvariantTag) 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 = AA->getTypeStoreSize(LI->getType()); + // Atomic loads have complications involved. + // FIXME: This is overly conservative. + if (!LI->isUnordered()) + return MemDepResult::getClobber(LI); + + AliasAnalysis::Location LoadLoc = AA->getLocation(LI); // If we found a pointer, check if it could be the same as our pointer. - AliasAnalysis::AliasResult R = - AA->alias(Pointer, PointerSize, MemPtr, MemSize); + AliasAnalysis::AliasResult R = AA->alias(LoadLoc, MemLoc); + + if (isLoad) { + if (R == AliasAnalysis::NoAlias) { + // If this is an over-aligned integer load (for example, + // "load i8* %P, align 4") see if it would obviously overlap with the + // queried location if widened to a larger load (e.g. if the queried + // location is 1 byte at P+1). If so, return it as a load/load + // clobber result, allowing the client to decide to widen the load if + // it wants to. + if (IntegerType *ITy = dyn_cast(LI->getType())) + if (LI->getAlignment()*8 > ITy->getPrimitiveSizeInBits() && + isLoadLoadClobberIfExtendedToFullWidth(MemLoc, MemLocBase, + MemLocOffset, LI, TD)) + return MemDepResult::getClobber(Inst); + + continue; + } + + // Must aliased loads are defs of each other. + if (R == AliasAnalysis::MustAlias) + return MemDepResult::getDef(Inst); + +#if 0 // FIXME: Temporarily disabled. GVN is cleverly rewriting loads + // in terms of clobbering loads, but since it does this by looking + // at the clobbering load directly, it doesn't know about any + // phi translation that may have happened along the way. + + // If we have a partial alias, then return this as a clobber for the + // client to handle. + if (R == AliasAnalysis::PartialAlias) + return MemDepResult::getClobber(Inst); +#endif + + // Random may-alias loads don't depend on each other without a + // dependence. + continue; + } + + // Stores don't depend on other no-aliased accesses. if (R == AliasAnalysis::NoAlias) continue; - - // May-alias loads don't depend on each other without a dependence. - if (isLoad && R == AliasAnalysis::MayAlias) + + // Stores don't alias loads from read-only memory. + if (AA->pointsToConstantMemory(LoadLoc)) continue; - // Stores depend on may and must aliased loads, loads depend on must-alias - // loads. + + // Stores depend on may/must aliased loads. return MemDepResult::getDef(Inst); } if (StoreInst *SI = dyn_cast(Inst)) { - // There can't be stores to the value we care about inside an - // invariant region. - if (InvariantTag) continue; - + // Atomic stores have complications involved. + // FIXME: This is overly conservative. + if (!SI->isUnordered()) + return MemDepResult::getClobber(SI); + // 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) + if (AA->getModRefInfo(SI, MemLoc) == 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()); + AliasAnalysis::Location StoreLoc = AA->getLocation(SI); // If we found a pointer, check if it could be the same as our pointer. - AliasAnalysis::AliasResult R = - AA->alias(Pointer, PointerSize, MemPtr, MemSize); + AliasAnalysis::AliasResult R = AA->alias(StoreLoc, MemLoc); if (R == AliasAnalysis::NoAlias) continue; - if (R == AliasAnalysis::MayAlias) - return MemDepResult::getClobber(Inst); - return MemDepResult::getDef(Inst); + if (R == AliasAnalysis::MustAlias) + return MemDepResult::getDef(Inst); + return MemDepResult::getClobber(Inst); } // If this is an allocation, and if we know that the accessed pointer is to @@ -275,25 +479,32 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, // 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) || - (isa(Inst) && extractMallocCall(Inst))) { - Value *AccessPtr = MemPtr->getUnderlyingObject(); + const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo(); + if (isa(Inst) || isNoAliasFn(Inst, TLI)) { + const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD); - if (AccessPtr == Inst || - AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) + if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr)) return MemDepResult::getDef(Inst); - continue; + // Be conservative if the accessed pointer may alias the allocation. + if (AA->alias(Inst, AccessPtr) != AliasAnalysis::NoAlias) + return MemDepResult::getClobber(Inst); + // If the allocation is not aliased and does not read memory (like + // strdup), it is safe to ignore. + if (isa(Inst) || + isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI)) + continue; } // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. - switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { + AliasAnalysis::ModRefResult MR = AA->getModRefInfo(Inst, MemLoc); + // If necessary, perform additional analysis. + if (MR == AliasAnalysis::ModRef) + MR = AA->callCapturesBefore(Inst, MemLoc, DT); + switch (MR) { case AliasAnalysis::NoModRef: // If the call has no effect on the queried pointer, just ignore it. continue; case AliasAnalysis::Mod: - // If we're in an invariant region, we can ignore calls that ONLY - // modify the pointer. - if (InvariantTag) continue; return MemDepResult::getClobber(Inst); case AliasAnalysis::Ref: // If the call is known to never store to the pointer, and if this is a @@ -306,11 +517,11 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, } } - // No dependence found. If this is the entry block of the function, it is a - // clobber, otherwise it is non-local. + // No dependence found. If this is the entry block of the function, it is + // unknown, otherwise it is non-local. if (BB != &BB->getParent()->getEntryBlock()) return MemDepResult::getNonLocal(); - return MemDepResult::getClobber(ScanIt); + return MemDepResult::getNonFuncLocal(); } /// getDependency - Return the instruction on which a memory operation @@ -336,75 +547,33 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *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. + // No dependence found. If this is the entry block of the function, it is + // unknown, 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 = 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 - // previous instruction as a clobber. - if (LI->isVolatile()) - LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); - else { - MemPtr = LI->getPointerOperand(); - 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)) { - int IntrinsicID = 0; // Intrinsic IDs start at 1. - if (IntrinsicInst *II = dyn_cast(QueryInst)) - IntrinsicID = II->getIntrinsicID(); - - switch (IntrinsicID) { - case Intrinsic::lifetime_start: - case Intrinsic::lifetime_end: - case Intrinsic::invariant_start: - MemPtr = QueryInst->getOperand(1); - MemSize = cast(QueryInst->getOperand(0))->getZExtValue(); - break; - case Intrinsic::invariant_end: - MemPtr = QueryInst->getOperand(2); - MemSize = cast(QueryInst->getOperand(1))->getZExtValue(); - break; - default: - CallSite QueryCS = CallSite::get(QueryInst); + LocalCache = MemDepResult::getNonFuncLocal(); + } else { + AliasAnalysis::Location MemLoc; + AliasAnalysis::ModRefResult MR = GetLocation(QueryInst, MemLoc, AA); + if (MemLoc.Ptr) { + // If we can do a pointer scan, make it happen. + bool isLoad = !(MR & AliasAnalysis::Mod); + if (IntrinsicInst *II = dyn_cast(QueryInst)) + isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start; + + LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos, + QueryParent); + } else if (isa(QueryInst) || isa(QueryInst)) { + CallSite QueryCS(QueryInst); bool isReadOnly = AA->onlyReadsMemory(QueryCS); LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos, QueryParent); - break; - } - } else { - // Non-memory instruction. - LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); - } - - // If we need to do a pointer scan, make it happen. - if (MemPtr) { - bool isLoad = !QueryInst->mayWriteToMemory(); - if (IntrinsicInst *II = dyn_cast(QueryInst)) { - isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; - } - LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos, - QueryParent); + } else + // Non-memory instruction. + LocalCache = MemDepResult::getUnknown(); } // Remember the result! @@ -456,7 +625,7 @@ MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { // Okay, we have a cache entry. If we know it is not dirty, just return it // with no computation. if (!CacheP.second) { - NumCacheNonLocal++; + ++NumCacheNonLocal; return Cache; } @@ -478,7 +647,7 @@ MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { BasicBlock *QueryBB = QueryCS.getInstruction()->getParent(); for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI) DirtyBlocks.push_back(*PI); - NumUncacheNonLocal++; + ++NumUncacheNonLocal; } // isReadonlyCall - If this is a read-only call, we can be more aggressive. @@ -538,10 +707,10 @@ MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { 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. + // a clobber, otherwise it is unknown. Dep = MemDepResult::getNonLocal(); } else { - Dep = MemDepResult::getClobber(ScanPos); + Dep = MemDepResult::getNonFuncLocal(); } // If we had a dirty entry for the block, update it. Otherwise, just add @@ -578,31 +747,27 @@ MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) { /// own block. /// void MemoryDependenceAnalysis:: -getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, +getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad, + BasicBlock *FromBB, SmallVectorImpl &Result) { - assert(Pointer->getType()->isPointerTy() && + assert(Loc.Ptr->getType()->isPointerTy() && "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 = AA->getTypeStoreSize(EltTy); - - PHITransAddr Address(Pointer, TD); + PHITransAddr Address(const_cast(Loc.Ptr), TD); // 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(Address, PointeeSize, isLoad, FromBB, + if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB, Result, Visited, true)) return; Result.clear(); Result.push_back(NonLocalDepResult(FromBB, - MemDepResult::getClobber(FromBB->begin()), - Pointer)); + MemDepResult::getUnknown(), + const_cast(Loc.Ptr))); } /// GetNonLocalInfoForBlock - Compute the memdep value for BB with @@ -610,7 +775,7 @@ getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, /// 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, +GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc, bool isLoad, BasicBlock *BB, NonLocalDepInfo *Cache, unsigned NumSortedEntries) { @@ -644,15 +809,14 @@ GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, ScanPos = ExistingResult->getResult().getInst(); // Eliminating the dirty entry from 'Cache', so update the reverse info. - ValueIsLoadPair CacheKey(Pointer, isLoad); + ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); } else { ++NumUncacheNonLocalPtr; } // Scan the block for the dependency. - MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, - ScanPos, BB); + MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB); // If we had a dirty entry for the block, update it. Otherwise, just add // a new entry. @@ -664,14 +828,14 @@ GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, // 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()) + if (!Dep.isDef() && !Dep.isClobber()) 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); + ValueIsLoadPair CacheKey(Loc.Ptr, isLoad); ReverseNonLocalPtrDeps[Inst].insert(CacheKey); return Dep; } @@ -725,7 +889,8 @@ SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache, /// 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(const PHITransAddr &Pointer, uint64_t PointeeSize, +getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, + const AliasAnalysis::Location &Loc, bool isLoad, BasicBlock *StartBB, SmallVectorImpl &Result, DenseMap &Visited, @@ -733,14 +898,68 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // Look up the cached info for Pointer. ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); - - std::pair *CacheInfo = - &NonLocalPointerDeps[CacheKey]; - NonLocalDepInfo *Cache = &CacheInfo->second; + + // Set up a temporary NLPI value. If the map doesn't yet have an entry for + // CacheKey, this value will be inserted as the associated value. Otherwise, + // it'll be ignored, and we'll have to check to see if the cached size and + // tbaa tag are consistent with the current query. + NonLocalPointerInfo InitialNLPI; + InitialNLPI.Size = Loc.Size; + InitialNLPI.TBAATag = Loc.TBAATag; + + // Get the NLPI for CacheKey, inserting one into the map if it doesn't + // already have one. + std::pair Pair = + NonLocalPointerDeps.insert(std::make_pair(CacheKey, InitialNLPI)); + NonLocalPointerInfo *CacheInfo = &Pair.first->second; + + // If we already have a cache entry for this CacheKey, we may need to do some + // work to reconcile the cache entry and the current query. + if (!Pair.second) { + if (CacheInfo->Size < Loc.Size) { + // The query's Size is greater than the cached one. Throw out the + // cached data and proceed with the query at the greater size. + CacheInfo->Pair = BBSkipFirstBlockPair(); + CacheInfo->Size = Loc.Size; + for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), + DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) + if (Instruction *Inst = DI->getResult().getInst()) + RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); + CacheInfo->NonLocalDeps.clear(); + } else if (CacheInfo->Size > Loc.Size) { + // This query's Size is less than the cached one. Conservatively restart + // the query using the greater size. + return getNonLocalPointerDepFromBB(Pointer, + Loc.getWithNewSize(CacheInfo->Size), + isLoad, StartBB, Result, Visited, + SkipFirstBlock); + } + + // If the query's TBAATag is inconsistent with the cached one, + // conservatively throw out the cached data and restart the query with + // no tag if needed. + if (CacheInfo->TBAATag != Loc.TBAATag) { + if (CacheInfo->TBAATag) { + CacheInfo->Pair = BBSkipFirstBlockPair(); + CacheInfo->TBAATag = 0; + for (NonLocalDepInfo::iterator DI = CacheInfo->NonLocalDeps.begin(), + DE = CacheInfo->NonLocalDeps.end(); DI != DE; ++DI) + if (Instruction *Inst = DI->getResult().getInst()) + RemoveFromReverseMap(ReverseNonLocalPtrDeps, Inst, CacheKey); + CacheInfo->NonLocalDeps.clear(); + } + if (Loc.TBAATag) + return getNonLocalPointerDepFromBB(Pointer, Loc.getWithoutTBAATag(), + isLoad, StartBB, Result, Visited, + SkipFirstBlock); + } + } + + NonLocalDepInfo *Cache = &CacheInfo->NonLocalDeps; // 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)) { + if (CacheInfo->Pair == 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 @@ -764,7 +983,7 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end(); I != E; ++I) { Visited.insert(std::make_pair(I->getBB(), Addr)); - if (!I->getResult().isNonLocal()) + if (!I->getResult().isNonLocal() && DT->isReachableFromEntry(I->getBB())) Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr)); } ++NumCacheCompleteNonLocalPtr; @@ -776,13 +995,16 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // 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); + CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); else - CacheInfo->first = BBSkipFirstBlockPair(); + CacheInfo->Pair = BBSkipFirstBlockPair(); SmallVector Worklist; Worklist.push_back(StartBB); + // PredList used inside loop. + SmallVector, 16> PredList; + // 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 @@ -803,12 +1025,11 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // 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.getAddr(), PointeeSize, - isLoad, BB, Cache, + MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache, NumSortedEntries); // If we got a Def or Clobber, add this to the list of results. - if (!Dep.isNonLocal()) { + if (!Dep.isNonLocal() && DT->isReachableFromEntry(BB)) { Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr())); continue; } @@ -820,22 +1041,29 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // the same Pointer. if (!Pointer.NeedsPHITranslationFromBlock(BB)) { SkipFirstBlock = false; + SmallVector NewBlocks; 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.getAddr())); if (InsertRes.second) { // First time we've looked at *PI. - Worklist.push_back(*PI); + NewBlocks.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.getAddr()) + if (InsertRes.first->second != Pointer.getAddr()) { + // Make sure to clean up the Visited map before continuing on to + // PredTranslationFailure. + for (unsigned i = 0; i < NewBlocks.size(); i++) + Visited.erase(NewBlocks[i]); goto PredTranslationFailure; + } } + Worklist.append(NewBlocks.begin(), NewBlocks.end()); continue; } @@ -854,13 +1082,15 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, NumSortedEntries = Cache->size(); } Cache = 0; - + + PredList.clear(); for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) { BasicBlock *Pred = *PI; - + PredList.push_back(std::make_pair(Pred, Pointer)); + // Get the PHI translated pointer in this predecessor. This can fail if // not translatable, in which case the getAddr() returns null. - PHITransAddr PredPointer(Pointer); + PHITransAddr &PredPointer = PredList.back().second; PredPointer.PHITranslateValue(BB, Pred, 0); Value *PredPtrVal = PredPointer.getAddr(); @@ -874,6 +1104,9 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, InsertRes = Visited.insert(std::make_pair(Pred, PredPtrVal)); if (!InsertRes.second) { + // We found the pred; take it off the list of preds to visit. + PredList.pop_back(); + // If the predecessor was visited with PredPtr, then we already did // the analysis and can ignore it. if (InsertRes.first->second == PredPtrVal) @@ -882,18 +1115,49 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // 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. + + // Make sure to clean up the Visited map before continuing on to + // PredTranslationFailure. + for (unsigned i = 0; i < PredList.size(); i++) + Visited.erase(PredList[i].first); + goto PredTranslationFailure; } - + } + + // Actually process results here; this need to be a separate loop to avoid + // calling getNonLocalPointerDepFromBB for blocks we don't want to return + // any results for. (getNonLocalPointerDepFromBB will modify our + // datastructures in ways the code after the PredTranslationFailure label + // doesn't expect.) + for (unsigned i = 0; i < PredList.size(); i++) { + BasicBlock *Pred = PredList[i].first; + PHITransAddr &PredPointer = PredList[i].second; + Value *PredPtrVal = PredPointer.getAddr(); + + bool CanTranslate = true; // If PHI translation was unable to find an available pointer in this // predecessor, then we have to assume that the pointer is clobbered in // that predecessor. We can still do PRE of the load, which would insert // a computation of the pointer in this predecessor. - if (PredPtrVal == 0) { + if (PredPtrVal == 0) + CanTranslate = false; + + // 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 getNonLocalPointerDepFromBB fails here, that means the cached + // result conflicted with the Visited list; we have to conservatively + // assume it is unknown, but this also does not block PRE of the load. + if (!CanTranslate || + getNonLocalPointerDepFromBB(PredPointer, + Loc.getWithNewPtr(PredPtrVal), + isLoad, Pred, + Result, Visited)) { // Add the entry to the Result list. - NonLocalDepResult Entry(Pred, - MemDepResult::getClobber(Pred->getTerminator()), - PredPtrVal); + NonLocalDepResult Entry(Pred, MemDepResult::getUnknown(), PredPtrVal); Result.push_back(Entry); // Since we had a phi translation failure, the cache for CacheKey won't @@ -901,41 +1165,34 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // queries. Mark this in NonLocalPointerDeps by setting the // BBSkipFirstBlockPair pointer to null. This requires reuse of the // cached value to do more work but not miss the phi trans failure. - NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair(); + NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey]; + NLPI.Pair = BBSkipFirstBlockPair(); continue; } - - // 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(PredPointer, PointeeSize, isLoad, Pred, - Result, Visited)) - goto PredTranslationFailure; } // Refresh the CacheInfo/Cache pointer so that it isn't invalidated. CacheInfo = &NonLocalPointerDeps[CacheKey]; - Cache = &CacheInfo->second; + Cache = &CacheInfo->NonLocalDeps; 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(); + CacheInfo->Pair = BBSkipFirstBlockPair(); SkipFirstBlock = false; continue; PredTranslationFailure: + // The following code is "failure"; we can't produce a sane translation + // for the given block. It assumes that we haven't modified any of + // our datastructures while processing the current block. if (Cache == 0) { // Refresh the CacheInfo/Cache pointer if it got invalidated. CacheInfo = &NonLocalPointerDeps[CacheKey]; - Cache = &CacheInfo->second; + Cache = &CacheInfo->NonLocalDeps; NumSortedEntries = Cache->size(); } @@ -943,10 +1200,9 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, // 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(); + CacheInfo->Pair = BBSkipFirstBlockPair(); - // If *nothing* works, mark the pointer as being clobbered by the first - // instruction in this block. + // If *nothing* works, mark the pointer as unknown. // // 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, @@ -961,8 +1217,7 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, assert(I->getResult().isNonLocal() && "Should only be here with transparent block"); - I->setResult(MemDepResult::getClobber(BB->begin())); - ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey); + I->setResult(MemDepResult::getUnknown()); Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Pointer.getAddr())); break; @@ -985,7 +1240,7 @@ 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.second; + NonLocalDepInfo &PInfo = It->second.NonLocalDeps; for (unsigned i = 0, e = PInfo.size(); i != e; ++i) { Instruction *Target = PInfo[i].getResult().getInst(); @@ -1156,10 +1411,10 @@ void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) { assert(P.getPointer() != RemInst && "Already removed NonLocalPointerDeps info for RemInst"); - NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second; + NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps; // The cache is not valid for any specific block anymore. - NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); + NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair(); // Update any entries for RemInst to use the instruction after it. for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); @@ -1205,7 +1460,7 @@ 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.second; + const NonLocalDepInfo &Val = I->second.NonLocalDeps; for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end(); II != E; ++II) assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");