#define DEBUG_TYPE "memdep"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
-#include "llvm/Constants.h"
#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Function.h"
+#include "llvm/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/Target/TargetData.h"
+#include "llvm/DataLayout.h"
using namespace llvm;
STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
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<MemoryDependenceAnalysis> 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() {
}
void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<AliasAnalysis>();
- AU.addRequiredTransitive<TargetData>();
}
bool MemoryDependenceAnalysis::runOnFunction(Function &) {
AA = &getAnalysis<AliasAnalysis>();
- TD = &getAnalysis<TargetData>();
+ TD = getAnalysisIfAvailable<DataLayout>();
+ DT = getAnalysisIfAvailable<DominatorTree>();
if (PredCache == 0)
PredCache.reset(new PredIteratorCache());
return false;
/// 'Inst's set in ReverseMap. If the set becomes empty, remove Inst's entry.
template <typename KeyTy>
static void RemoveFromReverseMap(DenseMap<Instruction*,
- SmallPtrSet<KeyTy*, 4> > &ReverseMap,
- Instruction *Inst, KeyTy *Val) {
- typename DenseMap<Instruction*, SmallPtrSet<KeyTy*, 4> >::iterator
+ SmallPtrSet<KeyTy, 4> > &ReverseMap,
+ Instruction *Inst, KeyTy Val) {
+ typename DenseMap<Instruction*, SmallPtrSet<KeyTy, 4> >::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;
+ 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<LoadInst>(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<StoreInst>(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<VAArgInst>(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<IntrinsicInst>(Inst))
+ switch (II->getIntrinsicID()) {
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ case Intrinsic::invariant_start:
+ Loc = AliasAnalysis::Location(II->getArgOperand(1),
+ cast<ConstantInt>(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<ConstantInt>(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, BasicBlock::iterator ScanIt,
- BasicBlock *BB) {
+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<StoreInst>(Inst)) {
- Pointer = S->getPointerOperand();
- PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
- } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
- Pointer = V->getOperand(0);
- PointerSize = TD->getTypeStoreSize(V->getType());
- } else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
- Pointer = F->getPointerOperand();
-
- // FreeInsts erase the entire structure
- PointerSize = ~0ULL;
- } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
- CallSite InstCS = CallSite::get(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<Value>(Inst)) {
+ // Debug intrinsics don't cause dependences.
+ if (isa<DbgInfoIntrinsic>(Inst)) continue;
// If these two calls do not interfere, look past it.
- if (AA->getModRefInfo(CS, InstCS) == AliasAnalysis::NoModRef)
+ switch (AA->getModRefInfo(CS, InstCS)) {
+ case AliasAnalysis::NoModRef:
+ // If the two calls 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;
-
- // FIXME: If this is a ref/ref result, we should ignore it!
- // X = strlen(P);
- // Y = strlen(Q);
- // Z = strlen(P); // Z = X
-
- // If they interfere, we generally return clobber. However, if they are
- // calls to the same read-only functions we return Def.
- if (!AA->onlyReadsMemory(CS) || CS.getCalledFunction() == 0 ||
- CS.getCalledFunction() != InstCS.getCalledFunction())
+ default:
return MemDepResult::getClobber(Inst);
- return MemDepResult::getDef(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<IntegerType>(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().hasAddressSafetyAttr()){
+ // 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) {
+ 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 (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
+ // Debug intrinsics don't (and can't) cause dependences.
+ if (isa<DbgInfoIntrinsic>(II)) continue;
+
+ // If we reach a lifetime begin or end marker, then the query ends here
+ // because the value is undefined.
+ 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 (the right approach is to use
+ // GetPointerBaseWithConstantOffset).
+ if (AA->isMustAlias(AliasAnalysis::Location(II->getArgOperand(1)),
+ MemLoc))
+ return MemDepResult::getDef(II);
+ 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<LoadInst>(Inst)) {
- Value *Pointer = LI->getPointerOperand();
- uint64_t PointerSize = TD->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<IntegerType>(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<StoreInst>(Inst)) {
- Value *Pointer = SI->getPointerOperand();
- uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
+ // 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, 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.
+ 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
// 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<AllocationInst>(Inst)) {
- Value *AccessPtr = MemPtr->getUnderlyingObject();
+ // Note: Only determine this to be a malloc if Inst is the malloc call, not
+ // a subsequent bitcast of the malloc call result. There can be stores to
+ // the malloced memory between the malloc call and its bitcast uses, and we
+ // need to continue scanning until the malloc call.
+ const TargetLibraryInfo *TLI = AA->getTargetLibraryInfo();
+ if (isa<AllocaInst>(Inst) || isNoAliasFn(Inst, TLI)) {
+ const Value *AccessPtr = GetUnderlyingObject(MemLoc.Ptr, TD);
- if (AccessPtr == AI ||
- AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
- return MemDepResult::getDef(AI);
- continue;
+ if (AccessPtr == Inst || AA->isMustAlias(Inst, AccessPtr))
+ return MemDepResult::getDef(Inst);
+ // 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<AllocaInst>(Inst) ||
+ isMallocLikeFn(Inst, TLI) || isCallocLikeFn(Inst, TLI))
+ continue;
}
-
+
// See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
- // FIXME: If this is a load, we should ignore readonly calls!
- if (AA->getModRefInfo(Inst, MemPtr, MemSize) == AliasAnalysis::NoModRef)
+ 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;
-
- // Otherwise, there is a dependence.
- return MemDepResult::getClobber(Inst);
+ case AliasAnalysis::Mod:
+ return MemDepResult::getClobber(Inst);
+ case AliasAnalysis::Ref:
+ // If the call is known to never store to the pointer, and if this is a
+ // load query, we can safely ignore it (scan past it).
+ if (isLoad)
+ continue;
+ 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.
+ // 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
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<StoreInst>(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<LoadInst>(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<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
- LocalCache = getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanPos,
- QueryParent);
- } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
- MemPtr = FI->getPointerOperand();
- // FreeInsts erase the entire structure, not just a field.
- MemSize = ~0UL;
+ LocalCache = MemDepResult::getNonFuncLocal();
} else {
- // Non-memory instruction.
- LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
+ 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<IntrinsicInst>(QueryInst))
+ isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_start;
+
+ LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
+ QueryParent);
+ } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
+ CallSite QueryCS(QueryInst);
+ bool isReadOnly = AA->onlyReadsMemory(QueryCS);
+ LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
+ QueryParent);
+ } else
+ // Non-memory instruction.
+ LocalCache = MemDepResult::getUnknown();
}
- // If we need to do a pointer scan, make it happen.
- if (MemPtr)
- LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
- isa<LoadInst>(QueryInst),
- ScanPos, QueryParent);
-
// Remember the result!
if (Instruction *I = LocalCache.getInst())
ReverseLocalDeps[I].insert(QueryInst);
return LocalCache;
}
-/// getNonLocalDependency - Perform a full dependency query for the
-/// specified instruction, returning the set of blocks that the value is
+#ifndef NDEBUG
+/// AssertSorted - This method is used when -debug is specified to verify that
+/// cache arrays are properly kept sorted.
+static void AssertSorted(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
+ int Count = -1) {
+ if (Count == -1) Count = Cache.size();
+ if (Count == 0) return;
+
+ for (unsigned i = 1; i != unsigned(Count); ++i)
+ assert(!(Cache[i] < Cache[i-1]) && "Cache isn't sorted!");
+}
+#endif
+
+/// getNonLocalCallDependency - Perform a full dependency query for the
+/// specified call, returning the set of blocks that the value is
/// potentially live across. The returned set of results will include a
/// "NonLocal" result for all blocks where the value is live across.
///
-/// This method assumes the instruction returns a "nonlocal" dependency
+/// This method assumes the instruction returns a "NonLocal" dependency
/// within its own block.
///
+/// This returns a reference to an internal data structure that may be
+/// invalidated on the next non-local query or when an instruction is
+/// removed. Clients must copy this data if they want it around longer than
+/// that.
const MemoryDependenceAnalysis::NonLocalDepInfo &
-MemoryDependenceAnalysis::getNonLocalDependency(Instruction *QueryInst) {
- // FIXME: Make this only be for callsites in the future.
- assert(isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst) ||
- isa<LoadInst>(QueryInst) || isa<StoreInst>(QueryInst));
- assert(getDependency(QueryInst).isNonLocal() &&
- "getNonLocalDependency should only be used on insts with non-local deps!");
- PerInstNLInfo &CacheP = NonLocalDeps[QueryInst];
+MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
+ assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
+ "getNonLocalCallDependency should only be used on calls with non-local deps!");
+ PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
NonLocalDepInfo &Cache = CacheP.first;
/// DirtyBlocks - This is the set of blocks that need to be recomputed. In
// 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;
}
// 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);
+ if (I->getResult().isDirty())
+ DirtyBlocks.push_back(I->getBB());
// Sort the cache so that we can do fast binary search lookups below.
std::sort(Cache.begin(), Cache.end());
// << Cache.size() << " cached: " << *QueryInst;
} else {
// Seed DirtyBlocks with each of the preds of QueryInst's block.
- BasicBlock *QueryBB = QueryInst->getParent();
+ BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
DirtyBlocks.push_back(*PI);
- NumUncacheNonLocal++;
+ ++NumUncacheNonLocal;
}
- // Visited checked first, vector in sorted order.
+ // isReadonlyCall - If this is a read-only call, we can be more aggressive.
+ bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
+
SmallPtrSet<BasicBlock*, 64> Visited;
unsigned NumSortedEntries = Cache.size();
+ DEBUG(AssertSorted(Cache));
// Iterate while we still have blocks to update.
while (!DirtyBlocks.empty()) {
// Do a binary search to see if we already have an entry for this block in
// the cache set. If so, find it.
+ DEBUG(AssertSorted(Cache, NumSortedEntries));
NonLocalDepInfo::iterator Entry =
std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
- std::make_pair(DirtyBB, MemDepResult()));
- if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB)
+ NonLocalDepEntry(DirtyBB));
+ if (Entry != Cache.begin() && prior(Entry)->getBB() == DirtyBB)
--Entry;
- MemDepResult *ExistingResult = 0;
+ NonLocalDepEntry *ExistingResult = 0;
if (Entry != Cache.begin()+NumSortedEntries &&
- Entry->first == DirtyBB) {
+ Entry->getBB() == DirtyBB) {
// If we already have an entry, and if it isn't already dirty, the block
// is done.
- if (!Entry->second.isDirty())
+ if (!Entry->getResult().isDirty())
continue;
// Otherwise, remember this slot so we can update the value.
- ExistingResult = &Entry->second;
+ ExistingResult = &*Entry;
}
// 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()) {
+ if (Instruction *Inst = ExistingResult->getResult().getInst()) {
ScanPos = Inst;
// We're removing QueryInst's use of Inst.
- RemoveFromReverseMap(ReverseNonLocalDeps, Inst, QueryInst);
+ RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
+ QueryCS.getInstruction());
}
}
// Find out if this block has a local dependency for QueryInst.
MemDepResult Dep;
- Value *MemPtr = 0;
- uint64_t MemSize = 0;
-
- if (ScanPos == DirtyBB->begin()) {
- // No dependence found. If this is the entry block of the function, it is a
- // clobber, otherwise it is non-local.
- if (DirtyBB != &DirtyBB->getParent()->getEntryBlock())
- Dep = MemDepResult::getNonLocal();
- else
- Dep = MemDepResult::getClobber(ScanPos);
- } else if (StoreInst *SI = dyn_cast<StoreInst>(QueryInst)) {
- // If this is a volatile store, don't mess around with it. Just return the
- // previous instruction as a clobber.
- if (SI->isVolatile())
- Dep = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
- else {
- MemPtr = SI->getPointerOperand();
- MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
- }
- } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
- // If this is a volatile load, don't mess around with it. Just return the
- // previous instruction as a clobber.
- if (LI->isVolatile())
- Dep = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
- else {
- MemPtr = LI->getPointerOperand();
- MemSize = TD->getTypeStoreSize(LI->getType());
- }
+ if (ScanPos != DirtyBB->begin()) {
+ Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
+ } else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
+ // No dependence found. If this is the entry block of the function, it is
+ // a clobber, otherwise it is unknown.
+ Dep = MemDepResult::getNonLocal();
} else {
- assert(isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst));
- Dep = getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanPos,
- DirtyBB);
+ Dep = MemDepResult::getNonFuncLocal();
}
- if (MemPtr)
- Dep = getPointerDependencyFrom(MemPtr, MemSize, isa<LoadInst>(QueryInst),
- ScanPos, DirtyBB);
-
// If we had a dirty entry for the block, update it. Otherwise, just add
// a new entry.
if (ExistingResult)
- *ExistingResult = Dep;
+ ExistingResult->setResult(Dep);
else
- Cache.push_back(std::make_pair(DirtyBB, Dep));
+ Cache.push_back(NonLocalDepEntry(DirtyBB, Dep));
// If the block has a dependency (i.e. it isn't completely transparent to
// the value), remember the association!
// Keep the ReverseNonLocalDeps map up to date so we can efficiently
// update this when we remove instructions.
if (Instruction *Inst = Dep.getInst())
- ReverseNonLocalDeps[Inst].insert(QueryInst);
+ ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
} else {
// If the block *is* completely transparent to the load, we need to check
/// own block.
///
void MemoryDependenceAnalysis::
-getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
- SmallVectorImpl<NonLocalDepEntry> &Result) {
- assert(isa<PointerType>(Pointer->getType()) &&
+getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
+ BasicBlock *FromBB,
+ SmallVectorImpl<NonLocalDepResult> &Result) {
+ 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<PointerType>(Pointer->getType())->getElementType();
- uint64_t PointeeSize = TD->getTypeStoreSize(EltTy);
+ PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
- // While we have blocks to analyze, get their values.
- SmallPtrSet<BasicBlock*, 64> Visited;
- getNonLocalPointerDepFromBB(Pointer, PointeeSize, isLoad, FromBB,
- Result, Visited);
+ // This is the set of blocks we've inspected, and the pointer we consider in
+ // each block. Because of critical edges, we currently bail out if querying
+ // a block with multiple different pointers. This can happen during PHI
+ // translation.
+ DenseMap<BasicBlock*, Value*> Visited;
+ if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
+ Result, Visited, true))
+ return;
+ Result.clear();
+ Result.push_back(NonLocalDepResult(FromBB,
+ MemDepResult::getUnknown(),
+ const_cast<Value *>(Loc.Ptr)));
}
/// GetNonLocalInfoForBlock - Compute the memdep value for BB with
/// 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) {
// the cache set. If so, find it.
NonLocalDepInfo::iterator Entry =
std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
- std::make_pair(BB, MemDepResult()));
- if (Entry != Cache->begin() && (&*Entry)[-1].first == BB)
+ NonLocalDepEntry(BB));
+ if (Entry != Cache->begin() && (Entry-1)->getBB() == BB)
--Entry;
- MemDepResult *ExistingResult = 0;
- if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
- ExistingResult = &Entry->second;
+ NonLocalDepEntry *ExistingResult = 0;
+ if (Entry != Cache->begin()+NumSortedEntries && Entry->getBB() == BB)
+ ExistingResult = &*Entry;
// If we have a cached entry, and it is non-dirty, use it as the value for
// this dependency.
- if (ExistingResult && !ExistingResult->isDirty()) {
+ if (ExistingResult && !ExistingResult->getResult().isDirty()) {
++NumCacheNonLocalPtr;
- return *ExistingResult;
+ return ExistingResult->getResult();
}
// 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 &&
+ if (ExistingResult && ExistingResult->getResult().getInst()) {
+ assert(ExistingResult->getResult().getInst()->getParent() == BB &&
"Instruction invalidated?");
++NumCacheDirtyNonLocalPtr;
- ScanPos = ExistingResult->getInst();
+ ScanPos = ExistingResult->getResult().getInst();
// Eliminating the dirty entry from 'Cache', so update the reverse info.
- ValueIsLoadPair CacheKey(Pointer, isLoad);
- RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos,
- CacheKey.getOpaqueValue());
+ 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.
if (ExistingResult)
- *ExistingResult = Dep;
+ ExistingResult->setResult(Dep);
else
- Cache->push_back(std::make_pair(BB, Dep));
+ Cache->push_back(NonLocalDepEntry(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())
+ 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);
- ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue());
+ ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
+ ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
return Dep;
}
+/// SortNonLocalDepInfoCache - Sort the a NonLocalDepInfo cache, given a certain
+/// number of elements in the array that are already properly ordered. This is
+/// optimized for the case when only a few entries are added.
+static void
+SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
+ unsigned NumSortedEntries) {
+ switch (Cache.size() - NumSortedEntries) {
+ case 0:
+ // done, no new entries.
+ break;
+ case 2: {
+ // Two new entries, insert the last one into place.
+ NonLocalDepEntry Val = Cache.back();
+ Cache.pop_back();
+ MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.end()-1, Val);
+ Cache.insert(Entry, Val);
+ // FALL THROUGH.
+ }
+ case 1:
+ // One new entry, Just insert the new value at the appropriate position.
+ if (Cache.size() != 1) {
+ NonLocalDepEntry Val = Cache.back();
+ Cache.pop_back();
+ MemoryDependenceAnalysis::NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.end(), Val);
+ Cache.insert(Entry, Val);
+ }
+ break;
+ default:
+ // Added many values, do a full scale sort.
+ std::sort(Cache.begin(), Cache.end());
+ break;
+ }
+}
-/// getNonLocalPointerDepFromBB -
-void MemoryDependenceAnalysis::
-getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
+/// 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(const PHITransAddr &Pointer,
+ const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *StartBB,
- SmallVectorImpl<NonLocalDepEntry> &Result,
- SmallPtrSet<BasicBlock*, 64> &Visited) {
- // Look up the cached info for Pointer.
- ValueIsLoadPair CacheKey(Pointer, isLoad);
+ SmallVectorImpl<NonLocalDepResult> &Result,
+ DenseMap<BasicBlock*, Value*> &Visited,
+ bool SkipFirstBlock) {
- std::pair<BasicBlock*, NonLocalDepInfo> &CacheInfo =
- NonLocalPointerDeps[CacheKey];
- NonLocalDepInfo *Cache = &CacheInfo.second;
+ // Look up the cached info for Pointer.
+ ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
+
+ // 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<CachedNonLocalPointerInfo::iterator, bool> 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 == StartBB) {
+ 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
+ // 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<BasicBlock*, Value*>::iterator VI = Visited.find(I->getBB());
+ if (VI == Visited.end() || VI->second == Pointer.getAddr())
+ 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;
+ }
+ }
+
+ Value *Addr = Pointer.getAddr();
for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
- I != E; ++I)
- if (!I->second.isNonLocal())
- Result.push_back(*I);
+ I != E; ++I) {
+ Visited.insert(std::make_pair(I->getBB(), Addr));
+ if (!I->getResult().isNonLocal())
+ Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(), Addr));
+ }
++NumCacheCompleteNonLocalPtr;
- return;
+ return false;
}
// Otherwise, either this is a new block, a block with an invalid cache
// pointer or one that we're about to invalidate by putting more info into it
// than its valid cache info. If empty, the result will be valid cache info,
// otherwise it isn't.
- CacheInfo.first = Cache->empty() ? StartBB : 0;
+ if (Cache->empty())
+ CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
+ else
+ CacheInfo->Pair = BBSkipFirstBlockPair();
SmallVector<BasicBlock*, 32> Worklist;
Worklist.push_back(StartBB);
+ // PredList used inside loop.
+ SmallVector<std::pair<BasicBlock*, PHITransAddr>, 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
// 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();
-
- // SkipFirstBlock - If this is the very first block that we're processing, we
- // don't want to scan or think about its body, because the client was supposed
- // to do a local dependence query. Instead, just start processing it by
- // adding its predecessors to the worklist and iterating.
- bool SkipFirstBlock = Visited.empty();
+ DEBUG(AssertSorted(*Cache));
while (!Worklist.empty()) {
BasicBlock *BB = Worklist.pop_back_val();
// Skip the first block if we have it.
- if (SkipFirstBlock) {
- SkipFirstBlock = false;
- } else {
+ if (!SkipFirstBlock) {
// Analyze the dependency of *Pointer in FromBB. See if we already have
// been here.
- if (!Visited.insert(BB))
- continue;
+ 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.
- MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, PointeeSize, isLoad,
- BB, Cache, NumSortedEntries);
+ DEBUG(AssertSorted(*Cache, NumSortedEntries));
+ 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()) {
- Result.push_back(NonLocalDepEntry(BB, Dep));
+ Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
continue;
}
}
- // Otherwise, we have to process all the predecessors of this block to scan
- // them as well.
+ // 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 not, we just add the predecessors to the worklist and scan them with
+ // the same Pointer.
+ if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
+ SkipFirstBlock = false;
+ SmallVector<BasicBlock*, 16> NewBlocks;
+ for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
+ // Verify that we haven't looked at this block yet.
+ std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
+ InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
+ if (InsertRes.second) {
+ // First time we've looked at *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()) {
+ // 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;
+ }
+
+ // We do need to do phi translation, if we know ahead of time we can't phi
+ // translate this value, don't even try.
+ if (!Pointer.IsPotentiallyPHITranslatable())
+ 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 and other routines that could reuse the cache
+ // value will only see properly sorted cache arrays.
+ if (Cache && NumSortedEntries != Cache->size()) {
+ SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
+ NumSortedEntries = Cache->size();
+ }
+ Cache = 0;
+
+ PredList.clear();
for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
- // TODO: PHI TRANSLATE.
- Worklist.push_back(*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 = PredList.back().second;
+ PredPointer.PHITranslateValue(BB, Pred, 0);
+
+ Value *PredPtrVal = PredPointer.getAddr();
+
+ // 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<DenseMap<BasicBlock*,Value*>::iterator, bool>
+ 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)
+ 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.
+
+ // 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)
+ 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::getUnknown(), PredPtrVal);
+ Result.push_back(Entry);
+
+ // Since we had a phi translation failure, the cache for CacheKey won't
+ // include all of the entries that we need to immediately satisfy future
+ // 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.
+ NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
+ NLPI.Pair = BBSkipFirstBlockPair();
+ continue;
+ }
+ }
+
+ // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
+ CacheInfo = &NonLocalPointerDeps[CacheKey];
+ 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->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->NonLocalDeps;
+ NumSortedEntries = Cache->size();
+ }
+
+ // Since we failed 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->Pair = BBSkipFirstBlockPair();
+
+ // 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,
+ // 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->getBB() != BB)
+ continue;
+
+ assert(I->getResult().isNonLocal() &&
+ "Should only be here with transparent block");
+ I->setResult(MemDepResult::getUnknown());
+ Result.push_back(NonLocalDepResult(I->getBB(), I->getResult(),
+ Pointer.getAddr()));
+ 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.
- 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());
- }
+ SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
+ DEBUG(AssertSorted(*Cache));
+ return false;
}
/// RemoveCachedNonLocalPointerDependencies - If P exists in
// 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].second.getInst();
+ Instruction *Target = PInfo[i].getResult().getInst();
if (Target == 0) continue; // Ignore non-local dep results.
- assert(Target->getParent() == PInfo[i].first && Target != P.getPointer());
+ assert(Target->getParent() == PInfo[i].getBB());
// Eliminating the dirty entry from 'Cache', so update the reverse info.
- RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P.getOpaqueValue());
+ RemoveFromReverseMap(ReverseNonLocalPtrDeps, Target, P);
}
// Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
}
+/// 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 (!Ptr->getType()->isPointerTy()) return;
+ // Flush store info for the pointer.
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, false));
+ // Flush load info for the pointer.
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(Ptr, true));
+}
+
+/// invalidateCachedPredecessors - Clear the PredIteratorCache info.
+/// This needs to be done when the CFG changes, e.g., due to splitting
+/// critical edges.
+void MemoryDependenceAnalysis::invalidateCachedPredecessors() {
+ PredCache->clear();
+}
+
/// 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.
NonLocalDepInfo &BlockMap = NLDI->second.first;
for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
DI != DE; ++DI)
- if (Instruction *Inst = DI->second.getInst())
+ if (Instruction *Inst = DI->getResult().getInst())
RemoveFromReverseMap(ReverseNonLocalDeps, Inst, RemInst);
NonLocalDeps.erase(NLDI);
}
// 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<PointerType>(RemInst->getType())) {
+ if (RemInst->getType()->isPointerTy()) {
RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
}
for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
DE = INLD.first.end(); DI != DE; ++DI) {
- if (DI->second.getInst() != RemInst) continue;
+ if (DI->getResult().getInst() != RemInst) continue;
// Convert to a dirty entry for the subsequent instruction.
- DI->second = NewDirtyVal;
+ DI->setResult(NewDirtyVal);
if (Instruction *NextI = NewDirtyVal.getInst())
ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
ReverseNonLocalPtrDeps.find(RemInst);
if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
- SmallPtrSet<void*, 4> &Set = ReversePtrDepIt->second;
+ SmallPtrSet<ValueIsLoadPair, 4> &Set = ReversePtrDepIt->second;
SmallVector<std::pair<Instruction*, ValueIsLoadPair>,8> ReversePtrDepsToAdd;
- for (SmallPtrSet<void*, 4>::iterator I = Set.begin(), E = Set.end();
- I != E; ++I) {
- ValueIsLoadPair P;
- P.setFromOpaqueValue(*I);
+ for (SmallPtrSet<ValueIsLoadPair, 4>::iterator I = Set.begin(),
+ E = Set.end(); I != E; ++I) {
+ ValueIsLoadPair P = *I;
assert(P.getPointer() != RemInst &&
"Already removed NonLocalPointerDeps info for RemInst");
- NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].second;
+ NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P].NonLocalDeps;
// The cache is not valid for any specific block anymore.
- NonLocalPointerDeps[P].first = 0;
+ NonLocalPointerDeps[P].Pair = 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;
+ if (DI->getResult().getInst() != RemInst) continue;
// Convert to a dirty entry for the subsequent instruction.
- DI->second = NewDirtyVal;
+ DI->setResult(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());
+ .insert(ReversePtrDepsToAdd.back().second);
ReversePtrDepsToAdd.pop_back();
}
}
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 (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->second.getInst() != D && "Inst occurs as NLPD value");
+ assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");
}
for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
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");
+ assert(II->getResult().getInst() != D && "Inst occurs in data structures");
}
for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
assert(I->first != D && "Inst occurs in rev NLPD map");
- for (SmallPtrSet<void*, 4>::const_iterator II = I->second.begin(),
+ for (SmallPtrSet<ValueIsLoadPair, 4>::const_iterator II = I->second.begin(),
E = I->second.end(); II != E; ++II)
- assert(*II != ValueIsLoadPair(D, false).getOpaqueValue() &&
- *II != ValueIsLoadPair(D, true).getOpaqueValue() &&
+ assert(*II != ValueIsLoadPair(D, false) &&
+ *II != ValueIsLoadPair(D, true) &&
"Inst occurs in ReverseNonLocalPtrDeps map");
}
projects / oota-llvm.git / blobdiff