replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Teach memdep about TBAA tags. 

llvm-svn: 114588
This commit is contained in:
Dan Gohman 2010-09-22 21:41:02 +00:00
parent 0576ce730a
commit 2348393cf5
2 changed files with 123 additions and 77 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
llvm/include/llvm/Analysis/MemoryDependenceAnalysis.h
View File

@ -17,6 +17,7 @@
#include "llvm/BasicBlock.h" #include "llvm/BasicBlock.h"
#include "llvm/Pass.h" #include "llvm/Pass.h"
#include "llvm/Support/ValueHandle.h" #include "llvm/Support/ValueHandle.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/OwningPtr.h" #include "llvm/ADT/OwningPtr.h"
@ -220,11 +221,23 @@ namespace llvm {
/// or not the contents of the block was skipped. /// or not the contents of the block was skipped.
typedef PointerIntPair<BasicBlock*, 1, bool> BBSkipFirstBlockPair; typedef PointerIntPair<BasicBlock*, 1, bool> BBSkipFirstBlockPair;
/// NonLocalPointerInfo - This record is the information kept for each
/// (value, is load) pair.
struct NonLocalPointerInfo {
/// Pair - The pair of the block and the skip-first-block flag.
BBSkipFirstBlockPair Pair;
/// NonLocalDeps - The results of the query for each relevant block.
NonLocalDepInfo NonLocalDeps;
/// TBAATag - The TBAA tag associated with dereferences of the
/// pointer. May be null if there are no tags or conflicting tags.
MDNode *TBAATag;
};
/// CachedNonLocalPointerInfo - This map stores the cached results of doing /// CachedNonLocalPointerInfo - This map stores the cached results of doing
/// a pointer lookup at the bottom of a block. The key of this map is the /// a pointer lookup at the bottom of a block. The key of this map is the
/// pointer+isload bit, the value is a list of <bb->result> mappings. /// pointer+isload bit, the value is a list of <bb->result> mappings.
typedef DenseMap<ValueIsLoadPair, std::pair<BBSkipFirstBlockPair, typedef DenseMap<ValueIsLoadPair,
NonLocalDepInfo> > CachedNonLocalPointerInfo; NonLocalPointerInfo> CachedNonLocalPointerInfo;
CachedNonLocalPointerInfo NonLocalPointerDeps; CachedNonLocalPointerInfo NonLocalPointerDeps;
// A map from instructions to their non-local pointer dependencies. // A map from instructions to their non-local pointer dependencies.
@ -297,9 +310,17 @@ namespace llvm {
/// set of instructions that either define or clobber the value. /// set of instructions that either define or clobber the value.
/// ///
/// This method assumes the pointer has a "NonLocal" dependency within BB. /// This method assumes the pointer has a "NonLocal" dependency within BB.
void getNonLocalPointerDependency(const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *BB,
SmallVectorImpl<NonLocalDepResult> &Result);
/// getNonLocalPointerDependence - A convenience wrapper.
void getNonLocalPointerDependency(Value *Pointer, bool isLoad, void getNonLocalPointerDependency(Value *Pointer, bool isLoad,
BasicBlock *BB, BasicBlock *BB,
SmallVectorImpl<NonLocalDepResult> &Result); SmallVectorImpl<NonLocalDepResult> &Result){
return getNonLocalPointerDependency(AliasAnalysis::Location(Pointer),
isLoad, BB, Result);
}
/// removeInstruction - Remove an instruction from the dependence analysis, /// removeInstruction - Remove an instruction from the dependence analysis,
/// updating the dependence of instructions that previously depended on it. /// updating the dependence of instructions that previously depended on it.
@ -319,19 +340,20 @@ namespace llvm {
void invalidateCachedPredecessors(); void invalidateCachedPredecessors();
private: private:
MemDepResult getPointerDependencyFrom(Value *Pointer, uint64_t MemSize, MemDepResult getPointerDependencyFrom(const AliasAnalysis::Location &Loc,
bool isLoad, bool isLoad,
BasicBlock::iterator ScanIt, BasicBlock::iterator ScanIt,
BasicBlock *BB); BasicBlock *BB);
MemDepResult getCallSiteDependencyFrom(CallSite C, bool isReadOnlyCall, MemDepResult getCallSiteDependencyFrom(CallSite C, bool isReadOnlyCall,
BasicBlock::iterator ScanIt, BasicBlock::iterator ScanIt,
BasicBlock *BB); BasicBlock *BB);
bool getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t Size, bool getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *BB, bool isLoad, BasicBlock *BB,
SmallVectorImpl<NonLocalDepResult> &Result, SmallVectorImpl<NonLocalDepResult> &Result,
DenseMap<BasicBlock*, Value*> &Visited, DenseMap<BasicBlock*, Value*> &Visited,
bool SkipFirstBlock = false); bool SkipFirstBlock = false);
MemDepResult GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, MemDepResult GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *BB, bool isLoad, BasicBlock *BB,
NonLocalDepInfo *Cache, NonLocalDepInfo *Cache,
unsigned NumSortedEntries); unsigned NumSortedEntries);

164
llvm/lib/Analysis/MemoryDependenceAnalysis.cpp
View File

@ -19,6 +19,7 @@
#include "llvm/Instructions.h" #include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h" #include "llvm/IntrinsicInst.h"
#include "llvm/Function.h" #include "llvm/Function.h"
#include "llvm/LLVMContext.h"
#include "llvm/Analysis/AliasAnalysis.h" #include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/Dominators.h" #include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/InstructionSimplify.h"
@ -108,18 +109,19 @@ getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
Instruction *Inst = --ScanIt; Instruction *Inst = --ScanIt;
// If this inst is a memory op, get the pointer it accessed // If this inst is a memory op, get the pointer it accessed
Value *Pointer = 0; AliasAnalysis::Location Loc;
uint64_t PointerSize = 0;
if (StoreInst *S = dyn_cast<StoreInst>(Inst)) { if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
Pointer = S->getPointerOperand(); Loc = AliasAnalysis::Location(S->getPointerOperand(),
PointerSize = AA->getTypeStoreSize(S->getOperand(0)->getType()); AA->getTypeStoreSize(S->getValueOperand()
->getType()),
S->getMetadata(LLVMContext::MD_tbaa));
} else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) { } else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
Pointer = V->getOperand(0); Loc = AliasAnalysis::Location(V->getPointerOperand(),
PointerSize = AA->getTypeStoreSize(V->getType()); AA->getTypeStoreSize(V->getType()),
V->getMetadata(LLVMContext::MD_tbaa));
} else if (const CallInst *CI = isFreeCall(Inst)) { } else if (const CallInst *CI = isFreeCall(Inst)) {
Pointer = CI->getArgOperand(0);
// calls to free() erase the entire structure // calls to free() erase the entire structure
PointerSize = ~0ULL; Loc = AliasAnalysis::Location(CI->getArgOperand(0));
} else if (CallSite InstCS = cast<Value>(Inst)) { } else if (CallSite InstCS = cast<Value>(Inst)) {
// Debug intrinsics don't cause dependences. // Debug intrinsics don't cause dependences.
if (isa<DbgInfoIntrinsic>(Inst)) continue; if (isa<DbgInfoIntrinsic>(Inst)) continue;
@ -143,7 +145,7 @@ getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
continue; continue;
} }
if (AA->getModRefInfo(CS, Pointer, PointerSize) != AliasAnalysis::NoModRef) if (AA->getModRefInfo(CS, Loc) != AliasAnalysis::NoModRef)
return MemDepResult::getClobber(Inst); return MemDepResult::getClobber(Inst);
} }
@ -158,7 +160,7 @@ getCallSiteDependencyFrom(CallSite CS, bool isReadOnlyCall,
/// location depends. If isLoad is true, this routine ignore may-aliases with /// location depends. If isLoad is true, this routine ignore may-aliases with
/// read-only operations. /// read-only operations.
MemDepResult MemoryDependenceAnalysis:: MemDepResult MemoryDependenceAnalysis::
getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad, getPointerDependencyFrom(const AliasAnalysis::Location &MemLoc, bool isLoad,
BasicBlock::iterator ScanIt, BasicBlock *BB) { BasicBlock::iterator ScanIt, BasicBlock *BB) {
Value *InvariantTag = 0; Value *InvariantTag = 0;
@ -184,7 +186,8 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
// FIXME: This only considers queries directly on the invariant-tagged // FIXME: This only considers queries directly on the invariant-tagged
// pointer, not on query pointers that are indexed off of them. It'd // pointer, not on query pointers that are indexed off of them. It'd
// be nice to handle that at some point. // be nice to handle that at some point.
AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(2), MemPtr); AliasAnalysis::AliasResult R =
AA->alias(AliasAnalysis::Location(II->getArgOperand(2)), MemLoc);
if (R == AliasAnalysis::MustAlias) if (R == AliasAnalysis::MustAlias)
InvariantTag = II->getArgOperand(0); InvariantTag = II->getArgOperand(0);
@ -197,7 +200,8 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
// FIXME: This only considers queries directly on the invariant-tagged // FIXME: This only considers queries directly on the invariant-tagged
// pointer, not on query pointers that are indexed off of them. It'd // pointer, not on query pointers that are indexed off of them. It'd
// be nice to handle that at some point. // be nice to handle that at some point.
AliasAnalysis::AliasResult R = AA->alias(II->getArgOperand(1), MemPtr); AliasAnalysis::AliasResult R =
AA->alias(AliasAnalysis::Location(II->getArgOperand(1)), MemLoc);
if (R == AliasAnalysis::MustAlias) if (R == AliasAnalysis::MustAlias)
return MemDepResult::getDef(II); return MemDepResult::getDef(II);
continue; continue;
@ -216,10 +220,12 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
Value *Pointer = LI->getPointerOperand(); Value *Pointer = LI->getPointerOperand();
uint64_t PointerSize = AA->getTypeStoreSize(LI->getType()); uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
// If we found a pointer, check if it could be the same as our pointer. // If we found a pointer, check if it could be the same as our pointer.
AliasAnalysis::AliasResult R = AliasAnalysis::AliasResult R =
AA->alias(Pointer, PointerSize, MemPtr, MemSize); AA->alias(AliasAnalysis::Location(Pointer, PointerSize, TBAATag),
MemLoc);
if (R == AliasAnalysis::NoAlias) if (R == AliasAnalysis::NoAlias)
continue; continue;
@ -239,17 +245,19 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
// If alias analysis can tell that this store is guaranteed to not modify // 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, ignore it. Use getModRefInfo to handle cases where
// the query pointer points to constant memory etc. // the query pointer points to constant memory etc.
if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef) if (AA->getModRefInfo(SI, MemLoc) == AliasAnalysis::NoModRef)
continue; continue;
// Ok, this store might clobber the query pointer. Check to see if it is // 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. // a must alias: in this case, we want to return this as a def.
Value *Pointer = SI->getPointerOperand(); Value *Pointer = SI->getPointerOperand();
uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
MDNode *TBAATag = SI->getMetadata(LLVMContext::MD_tbaa);
// If we found a pointer, check if it could be the same as our pointer. // If we found a pointer, check if it could be the same as our pointer.
AliasAnalysis::AliasResult R = AliasAnalysis::AliasResult R =
AA->alias(Pointer, PointerSize, MemPtr, MemSize); AA->alias(AliasAnalysis::Location(Pointer, PointerSize, TBAATag),
MemLoc);
if (R == AliasAnalysis::NoAlias) if (R == AliasAnalysis::NoAlias)
continue; continue;
@ -268,7 +276,7 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
// need to continue scanning until the malloc call. // need to continue scanning until the malloc call.
if (isa<AllocaInst>(Inst) || if (isa<AllocaInst>(Inst) ||
(isa<CallInst>(Inst) && extractMallocCall(Inst))) { (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
Value *AccessPtr = MemPtr->getUnderlyingObject(); const Value *AccessPtr = MemLoc.Ptr->getUnderlyingObject();
if (AccessPtr == Inst || if (AccessPtr == Inst ||
AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias) AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
@ -277,7 +285,7 @@ getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
} }
// See if this instruction (e.g. a call or vaarg) mod/ref's the pointer. // See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) { switch (AA->getModRefInfo(Inst, MemLoc)) {
case AliasAnalysis::NoModRef: case AliasAnalysis::NoModRef:
// If the call has no effect on the queried pointer, just ignore it. // If the call has no effect on the queried pointer, just ignore it.
continue; continue;
@ -327,8 +335,7 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
BasicBlock *QueryParent = QueryInst->getParent(); BasicBlock *QueryParent = QueryInst->getParent();
Value *MemPtr = 0; AliasAnalysis::Location MemLoc;
uint64_t MemSize = 0;
// Do the scan. // Do the scan.
if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) { if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
@ -343,23 +350,23 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
// previous instruction as a clobber. // previous instruction as a clobber.
if (SI->isVolatile()) if (SI->isVolatile())
LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
else { else
MemPtr = SI->getPointerOperand(); MemLoc = AliasAnalysis::Location(SI->getPointerOperand(),
MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType()); AA->getTypeStoreSize(SI->getOperand(0)
} ->getType()),
SI->getMetadata(LLVMContext::MD_tbaa));
} else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) { } else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
// If this is a volatile load, don't mess around with it. Just return the // If this is a volatile load, don't mess around with it. Just return the
// previous instruction as a clobber. // previous instruction as a clobber.
if (LI->isVolatile()) if (LI->isVolatile())
LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos)); LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
else { else
MemPtr = LI->getPointerOperand(); MemLoc = AliasAnalysis::Location(LI->getPointerOperand(),
MemSize = AA->getTypeStoreSize(LI->getType()); AA->getTypeStoreSize(LI->getType()),
} LI->getMetadata(LLVMContext::MD_tbaa));
} else if (const CallInst *CI = isFreeCall(QueryInst)) { } else if (const CallInst *CI = isFreeCall(QueryInst)) {
MemPtr = CI->getArgOperand(0);
// calls to free() erase the entire structure, not just a field. // calls to free() erase the entire structure, not just a field.
MemSize = ~0UL; MemLoc = AliasAnalysis::Location(CI->getArgOperand(0));
} else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) { } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
int IntrinsicID = 0; // Intrinsic IDs start at 1. int IntrinsicID = 0; // Intrinsic IDs start at 1.
IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst); IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst);
@ -370,12 +377,16 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
case Intrinsic::lifetime_start: case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end: case Intrinsic::lifetime_end:
case Intrinsic::invariant_start: case Intrinsic::invariant_start:
MemPtr = II->getArgOperand(1); MemLoc = AliasAnalysis::Location(II->getArgOperand(1),
MemSize = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue(); cast<ConstantInt>(II->getArgOperand(0))
->getZExtValue(),
II->getMetadata(LLVMContext::MD_tbaa));
break; break;
case Intrinsic::invariant_end: case Intrinsic::invariant_end:
MemPtr = II->getArgOperand(2); MemLoc = AliasAnalysis::Location(II->getArgOperand(2),
MemSize = cast<ConstantInt>(II->getArgOperand(1))->getZExtValue(); cast<ConstantInt>(II->getArgOperand(1))
->getZExtValue(),
II->getMetadata(LLVMContext::MD_tbaa));
break; break;
default: default:
CallSite QueryCS(QueryInst); CallSite QueryCS(QueryInst);
@ -390,12 +401,12 @@ MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
} }
// If we need to do a pointer scan, make it happen. // If we need to do a pointer scan, make it happen.
if (MemPtr) { if (MemLoc.Ptr) {
bool isLoad = !QueryInst->mayWriteToMemory(); bool isLoad = !QueryInst->mayWriteToMemory();
if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) { if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) {
isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end; isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
} }
LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos, LocalCache = getPointerDependencyFrom(MemLoc, isLoad, ScanPos,
QueryParent); QueryParent);
} }
@ -570,31 +581,27 @@ MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
/// own block. /// own block.
/// ///
void MemoryDependenceAnalysis:: void MemoryDependenceAnalysis::
getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB, getNonLocalPointerDependency(const AliasAnalysis::Location &Loc, bool isLoad,
BasicBlock *FromBB,
SmallVectorImpl<NonLocalDepResult> &Result) { SmallVectorImpl<NonLocalDepResult> &Result) {
assert(Pointer->getType()->isPointerTy() && assert(Loc.Ptr->getType()->isPointerTy() &&
"Can't get pointer deps of a non-pointer!"); "Can't get pointer deps of a non-pointer!");
Result.clear(); Result.clear();
// We know that the pointer value is live into FromBB find the def/clobbers PHITransAddr Address(const_cast<Value *>(Loc.Ptr), TD);
// from presecessors.
const Type *EltTy = cast<PointerType>(Pointer->getType())->getElementType();
uint64_t PointeeSize = AA->getTypeStoreSize(EltTy);
PHITransAddr Address(Pointer, TD);
// This is the set of blocks we've inspected, and the pointer we consider in // 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 // each block. Because of critical edges, we currently bail out if querying
// a block with multiple different pointers. This can happen during PHI // a block with multiple different pointers. This can happen during PHI
// translation. // translation.
DenseMap<BasicBlock*, Value*> Visited; DenseMap<BasicBlock*, Value*> Visited;
if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB, if (!getNonLocalPointerDepFromBB(Address, Loc, isLoad, FromBB,
Result, Visited, true)) Result, Visited, true))
return; return;
Result.clear(); Result.clear();
Result.push_back(NonLocalDepResult(FromBB, Result.push_back(NonLocalDepResult(FromBB,
MemDepResult::getClobber(FromBB->begin()), MemDepResult::getClobber(FromBB->begin()),
Pointer)); const_cast<Value *>(Loc.Ptr)));
} }
/// GetNonLocalInfoForBlock - Compute the memdep value for BB with /// GetNonLocalInfoForBlock - Compute the memdep value for BB with
@ -602,7 +609,7 @@ getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
/// lookup (which may use dirty cache info if available). If we do a lookup, /// lookup (which may use dirty cache info if available). If we do a lookup,
/// add the result to the cache. /// add the result to the cache.
MemDepResult MemoryDependenceAnalysis:: MemDepResult MemoryDependenceAnalysis::
GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize, GetNonLocalInfoForBlock(const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *BB, bool isLoad, BasicBlock *BB,
NonLocalDepInfo *Cache, unsigned NumSortedEntries) { NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
@ -636,15 +643,14 @@ GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
ScanPos = ExistingResult->getResult().getInst(); ScanPos = ExistingResult->getResult().getInst();
// Eliminating the dirty entry from 'Cache', so update the reverse info. // Eliminating the dirty entry from 'Cache', so update the reverse info.
ValueIsLoadPair CacheKey(Pointer, isLoad); ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey); RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
} else { } else {
++NumUncacheNonLocalPtr; ++NumUncacheNonLocalPtr;
} }
// Scan the block for the dependency. // Scan the block for the dependency.
MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, MemDepResult Dep = getPointerDependencyFrom(Loc, isLoad, ScanPos, BB);
ScanPos, BB);
// If we had a dirty entry for the block, update it. Otherwise, just add // If we had a dirty entry for the block, update it. Otherwise, just add
// a new entry. // a new entry.
@ -663,7 +669,7 @@ GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
// update MemDep when we remove instructions. // update MemDep when we remove instructions.
Instruction *Inst = Dep.getInst(); Instruction *Inst = Dep.getInst();
assert(Inst && "Didn't depend on anything?"); assert(Inst && "Didn't depend on anything?");
ValueIsLoadPair CacheKey(Pointer, isLoad); ValueIsLoadPair CacheKey(Loc.Ptr, isLoad);
ReverseNonLocalPtrDeps[Inst].insert(CacheKey); ReverseNonLocalPtrDeps[Inst].insert(CacheKey);
return Dep; return Dep;
} }
@ -717,7 +723,8 @@ SortNonLocalDepInfoCache(MemoryDependenceAnalysis::NonLocalDepInfo &Cache,
/// not compute dependence information for some reason. This should be treated /// not compute dependence information for some reason. This should be treated
/// as a clobber dependence on the first instruction in the predecessor block. /// as a clobber dependence on the first instruction in the predecessor block.
bool MemoryDependenceAnalysis:: bool MemoryDependenceAnalysis::
getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize, getNonLocalPointerDepFromBB(const PHITransAddr &Pointer,
const AliasAnalysis::Location &Loc,
bool isLoad, BasicBlock *StartBB, bool isLoad, BasicBlock *StartBB,
SmallVectorImpl<NonLocalDepResult> &Result, SmallVectorImpl<NonLocalDepResult> &Result,
DenseMap<BasicBlock*, Value*> &Visited, DenseMap<BasicBlock*, Value*> &Visited,
@ -725,14 +732,23 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// Look up the cached info for Pointer. // Look up the cached info for Pointer.
ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad); ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
NonLocalPointerInfo *CacheInfo = &NonLocalPointerDeps[CacheKey];
std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo = // If this query's TBAATag is inconsistent with the cached one, discard the
&NonLocalPointerDeps[CacheKey]; // tag and restart the query.
NonLocalDepInfo *Cache = &CacheInfo->second; if (CacheInfo->TBAATag != Loc.TBAATag) {
CacheInfo->TBAATag = 0;
NonLocalPointerDeps.erase(CacheKey);
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 // If we have valid cached information for exactly the block we are
// investigating, just return it with no recomputation. // 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 // 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 // cached results and populate the visited set. However, we have to verify
// that we don't already have conflicting results for these blocks. Check // that we don't already have conflicting results for these blocks. Check
@ -768,9 +784,11 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// than its valid cache info. If empty, the result will be valid cache info, // than its valid cache info. If empty, the result will be valid cache info,
// otherwise it isn't. // otherwise it isn't.
if (Cache->empty()) if (Cache->empty())
CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock); CacheInfo->Pair = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
else else {
CacheInfo->first = BBSkipFirstBlockPair(); CacheInfo->Pair = BBSkipFirstBlockPair();
CacheInfo->TBAATag = 0;
}
SmallVector<BasicBlock*, 32> Worklist; SmallVector<BasicBlock*, 32> Worklist;
Worklist.push_back(StartBB); Worklist.push_back(StartBB);
@ -795,8 +813,7 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// Get the dependency info for Pointer in BB. If we have cached // Get the dependency info for Pointer in BB. If we have cached
// information, we will use it, otherwise we compute it. // information, we will use it, otherwise we compute it.
DEBUG(AssertSorted(*Cache, NumSortedEntries)); DEBUG(AssertSorted(*Cache, NumSortedEntries));
MemDepResult Dep = GetNonLocalInfoForBlock(Pointer.getAddr(), PointeeSize, MemDepResult Dep = GetNonLocalInfoForBlock(Loc, isLoad, BB, Cache,
isLoad, BB, Cache,
NumSortedEntries); NumSortedEntries);
// If we got a Def or Clobber, add this to the list of results. // If we got a Def or Clobber, add this to the list of results.
@ -893,7 +910,9 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// queries. Mark this in NonLocalPointerDeps by setting the // queries. Mark this in NonLocalPointerDeps by setting the
// BBSkipFirstBlockPair pointer to null. This requires reuse of the // BBSkipFirstBlockPair pointer to null. This requires reuse of the
// cached value to do more work but not miss the phi trans failure. // cached value to do more work but not miss the phi trans failure.
NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair(); NonLocalPointerInfo &NLPI = NonLocalPointerDeps[CacheKey];
NLPI.Pair = BBSkipFirstBlockPair();
NLPI.TBAATag = 0;
continue; continue;
} }
@ -904,21 +923,24 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// If we have a problem phi translating, fall through to the code below // If we have a problem phi translating, fall through to the code below
// to handle the failure condition. // to handle the failure condition.
if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred, if (getNonLocalPointerDepFromBB(PredPointer,
Loc.getWithNewPtr(PredPointer.getAddr()),
isLoad, Pred,
Result, Visited)) Result, Visited))
goto PredTranslationFailure; goto PredTranslationFailure;
} }
// Refresh the CacheInfo/Cache pointer so that it isn't invalidated. // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
CacheInfo = &NonLocalPointerDeps[CacheKey]; CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second; Cache = &CacheInfo->NonLocalDeps;
NumSortedEntries = Cache->size(); NumSortedEntries = Cache->size();
// Since we did phi translation, the "Cache" set won't contain all of the // 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 // 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 // specific block queries) but we can't do the fastpath "return all
// results from the set" Clear out the indicator for this. // results from the set" Clear out the indicator for this.
CacheInfo->first = BBSkipFirstBlockPair(); CacheInfo->Pair = BBSkipFirstBlockPair();
CacheInfo->TBAATag = 0;
SkipFirstBlock = false; SkipFirstBlock = false;
continue; continue;
@ -927,7 +949,7 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
if (Cache == 0) { if (Cache == 0) {
// Refresh the CacheInfo/Cache pointer if it got invalidated. // Refresh the CacheInfo/Cache pointer if it got invalidated.
CacheInfo = &NonLocalPointerDeps[CacheKey]; CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second; Cache = &CacheInfo->NonLocalDeps;
NumSortedEntries = Cache->size(); NumSortedEntries = Cache->size();
} }
@ -935,7 +957,8 @@ getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
// results for the query. This is ok (we can still use it to accelerate // 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 // specific block queries) but we can't do the fastpath "return all
// results from the set". Clear out the indicator for this. // results from the set". Clear out the indicator for this.
CacheInfo->first = BBSkipFirstBlockPair(); CacheInfo->Pair = BBSkipFirstBlockPair();
CacheInfo->TBAATag = 0;
// If *nothing* works, mark the pointer as being clobbered by the first // If *nothing* works, mark the pointer as being clobbered by the first
// instruction in this block. // instruction in this block.
@ -977,7 +1000,7 @@ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
// Remove all of the entries in the BB->val map. This involves removing // Remove all of the entries in the BB->val map. This involves removing
// instructions from the reverse map. // 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) { for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
Instruction *Target = PInfo[i].getResult().getInst(); Instruction *Target = PInfo[i].getResult().getInst();
@ -1148,10 +1171,11 @@ void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
assert(P.getPointer() != RemInst && assert(P.getPointer() != RemInst &&
"Already removed NonLocalPointerDeps info for 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. // The cache is not valid for any specific block anymore.
NonLocalPointerDeps[P].first = BBSkipFirstBlockPair(); NonLocalPointerDeps[P].Pair = BBSkipFirstBlockPair();
NonLocalPointerDeps[P].TBAATag = 0;
// Update any entries for RemInst to use the instruction after it. // Update any entries for RemInst to use the instruction after it.
for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end(); for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
@ -1197,7 +1221,7 @@ void MemoryDependenceAnalysis::verifyRemoved(Instruction *D) const {
for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(), for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
E = NonLocalPointerDeps.end(); I != E; ++I) { E = NonLocalPointerDeps.end(); I != E; ++I) {
assert(I->first.getPointer() != D && "Inst occurs in NLPD map key"); 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(); for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
II != E; ++II) II != E; ++II)
assert(II->getResult().getInst() != D && "Inst occurs as NLPD value"); assert(II->getResult().getInst() != D && "Inst occurs as NLPD value");