#define DEBUG_TYPE "memdep"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
-#include "llvm/Constants.h"
#include "llvm/Instructions.h"
+#include "llvm/IntrinsicInst.h"
#include "llvm/Function.h"
#include "llvm/Analysis/AliasAnalysis.h"
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/InstructionSimplify.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/PHITransAddr.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"
using namespace llvm;
STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
char MemoryDependenceAnalysis::ID = 0;
// Register this pass...
-static RegisterPass<MemoryDependenceAnalysis> X("memdep",
- "Memory Dependence Analysis", false, true);
+INITIALIZE_PASS(MemoryDependenceAnalysis, "memdep",
+ "Memory Dependence Analysis", false, true);
MemoryDependenceAnalysis::MemoryDependenceAnalysis()
: FunctionPass(&ID), PredCache(0) {
void MemoryDependenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
AU.setPreservesAll();
AU.addRequiredTransitive<AliasAnalysis>();
- AU.addRequiredTransitive<TargetData>();
}
bool MemoryDependenceAnalysis::runOnFunction(Function &) {
AA = &getAnalysis<AliasAnalysis>();
- TD = &getAnalysis<TargetData>();
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);
uint64_t PointerSize = 0;
if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
Pointer = S->getPointerOperand();
- PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
+ PointerSize = AA->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 = AA->getTypeStoreSize(V->getType());
+ } else if (const CallInst *CI = isFreeCall(Inst)) {
+ Pointer = CI->getArgOperand(0);
+ // calls to free() erase the entire structure
PointerSize = ~0ULL;
- } else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
- CallSite InstCS = CallSite::get(Inst);
+ } else 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.
switch (AA->getModRefInfo(CS, InstCS)) {
case AliasAnalysis::NoModRef:
/// location depends. If isLoad is true, this routine ignore may-aliases with
/// read-only operations.
MemDepResult MemoryDependenceAnalysis::
-getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
+getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
BasicBlock::iterator ScanIt, BasicBlock *BB) {
+ Value *InvariantTag = 0;
+
// Walk backwards through the basic block, looking for dependencies.
while (ScanIt != BB->begin()) {
Instruction *Inst = --ScanIt;
+ // If we're in an invariant region, no dependencies can be found before
+ // we pass an invariant-begin marker.
+ if (InvariantTag == Inst) {
+ InvariantTag = 0;
+ continue;
+ }
+
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) {
+ // Debug intrinsics don't cause dependences.
+ if (isa<DbgInfoIntrinsic>(Inst)) continue;
+
+ // If we pass an invariant-end marker, then we've just entered an
+ // invariant region and can start ignoring dependencies.
+ if (II->getIntrinsicID() == Intrinsic::invariant_end) {
+ // FIXME: This only considers queries directly on the invariant-tagged
+ // pointer, not on query pointers that are indexed off of them. It'd
+ // be nice to handle that at some point.
+ AliasAnalysis::AliasResult R =
+ AA->alias(II->getArgOperand(2), ~0U, MemPtr, ~0U);
+ if (R == AliasAnalysis::MustAlias) {
+ InvariantTag = II->getArgOperand(0);
+ continue;
+ }
+
+ // If we reach a lifetime begin or end marker, then the query ends here
+ // because the value is undefined.
+ } else if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
+ // FIXME: This only considers queries directly on the invariant-tagged
+ // pointer, not on query pointers that are indexed off of them. It'd
+ // be nice to handle that at some point.
+ AliasAnalysis::AliasResult R =
+ AA->alias(II->getArgOperand(1), ~0U, MemPtr, ~0U);
+ if (R == AliasAnalysis::MustAlias)
+ return MemDepResult::getDef(II);
+ }
+ }
+
+ // If we're querying on a load and we're in an invariant region, we're done
+ // at this point. Nothing a load depends on can live in an invariant region.
+ if (isLoad && InvariantTag) continue;
+
// Values depend on loads if the pointers are must aliased. This means that
// a load depends on another must aliased load from the same value.
if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
Value *Pointer = LI->getPointerOperand();
- uint64_t PointerSize = TD->getTypeStoreSize(LI->getType());
+ uint64_t PointerSize = AA->getTypeStoreSize(LI->getType());
// If we found a pointer, check if it could be the same as our pointer.
AliasAnalysis::AliasResult R =
}
if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
- Value *Pointer = SI->getPointerOperand();
- uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
+ // There can't be stores to the value we care about inside an
+ // invariant region.
+ if (InvariantTag) continue;
+
+ // If alias analysis can tell that this store is guaranteed to not modify
+ // the query pointer, ignore it. Use getModRefInfo to handle cases where
+ // the query pointer points to constant memory etc.
+ if (AA->getModRefInfo(SI, MemPtr, MemSize) == AliasAnalysis::NoModRef)
+ continue;
+ // Ok, this store might clobber the query pointer. Check to see if it is
+ // a must alias: in this case, we want to return this as a def.
+ Value *Pointer = SI->getPointerOperand();
+ uint64_t PointerSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
+
// If we found a pointer, check if it could be the same as our pointer.
AliasAnalysis::AliasResult R =
AA->alias(Pointer, PointerSize, MemPtr, MemSize);
// 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)) {
+ // Note: Only determine this to be a malloc if Inst is the malloc call, not
+ // a subsequent bitcast of the malloc call result. There can be stores to
+ // the malloced memory between the malloc call and its bitcast uses, and we
+ // need to continue scanning until the malloc call.
+ if (isa<AllocaInst>(Inst) ||
+ (isa<CallInst>(Inst) && extractMallocCall(Inst))) {
Value *AccessPtr = MemPtr->getUnderlyingObject();
- if (AccessPtr == AI ||
- AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
- return MemDepResult::getDef(AI);
+ if (AccessPtr == Inst ||
+ AA->alias(Inst, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
+ return MemDepResult::getDef(Inst);
continue;
}
-
+
// See if this instruction (e.g. a call or vaarg) mod/ref's the pointer.
switch (AA->getModRefInfo(Inst, MemPtr, MemSize)) {
case AliasAnalysis::NoModRef:
// If the call has no effect on the queried pointer, just ignore it.
continue;
+ case AliasAnalysis::Mod:
+ // If we're in an invariant region, we can ignore calls that ONLY
+ // modify the pointer.
+ if (InvariantTag) continue;
+ return MemDepResult::getClobber(Inst);
case AliasAnalysis::Ref:
// If the call is known to never store to the pointer, and if this is a
// load query, we can safely ignore it (scan past it).
if (isLoad)
continue;
- // FALL THROUGH.
default:
// Otherwise, there is a potential dependence. Return a clobber.
return MemDepResult::getClobber(Inst);
LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
else {
MemPtr = SI->getPointerOperand();
- MemSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
+ MemSize = AA->getTypeStoreSize(SI->getOperand(0)->getType());
}
} else if (LoadInst *LI = dyn_cast<LoadInst>(QueryInst)) {
// If this is a volatile load, don't mess around with it. Just return the
LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
else {
MemPtr = LI->getPointerOperand();
- MemSize = TD->getTypeStoreSize(LI->getType());
+ MemSize = AA->getTypeStoreSize(LI->getType());
}
- } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
- CallSite QueryCS = CallSite::get(QueryInst);
- bool isReadOnly = AA->onlyReadsMemory(QueryCS);
- LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
- QueryParent);
- } else if (FreeInst *FI = dyn_cast<FreeInst>(QueryInst)) {
- MemPtr = FI->getPointerOperand();
- // FreeInsts erase the entire structure, not just a field.
+ } else if (const CallInst *CI = isFreeCall(QueryInst)) {
+ MemPtr = CI->getArgOperand(0);
+ // calls to free() erase the entire structure, not just a field.
MemSize = ~0UL;
+ } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst)) {
+ int IntrinsicID = 0; // Intrinsic IDs start at 1.
+ IntrinsicInst *II = dyn_cast<IntrinsicInst>(QueryInst);
+ if (II)
+ IntrinsicID = II->getIntrinsicID();
+
+ switch (IntrinsicID) {
+ case Intrinsic::lifetime_start:
+ case Intrinsic::lifetime_end:
+ case Intrinsic::invariant_start:
+ MemPtr = II->getArgOperand(1);
+ MemSize = cast<ConstantInt>(II->getArgOperand(0))->getZExtValue();
+ break;
+ case Intrinsic::invariant_end:
+ MemPtr = II->getArgOperand(2);
+ MemSize = cast<ConstantInt>(II->getArgOperand(1))->getZExtValue();
+ break;
+ default:
+ CallSite QueryCS(QueryInst);
+ bool isReadOnly = AA->onlyReadsMemory(QueryCS);
+ LocalCache = getCallSiteDependencyFrom(QueryCS, isReadOnly, ScanPos,
+ QueryParent);
+ break;
+ }
} else {
// Non-memory instruction.
LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
}
// If we need to do a pointer scan, make it happen.
- if (MemPtr)
- LocalCache = getPointerDependencyFrom(MemPtr, MemSize,
- isa<LoadInst>(QueryInst),
- ScanPos, QueryParent);
+ if (MemPtr) {
+ bool isLoad = !QueryInst->mayWriteToMemory();
+ if (IntrinsicInst *II = dyn_cast<MemoryUseIntrinsic>(QueryInst)) {
+ isLoad |= II->getIntrinsicID() == Intrinsic::lifetime_end;
+ }
+ LocalCache = getPointerDependencyFrom(MemPtr, MemSize, isLoad, ScanPos,
+ QueryParent);
+ }
// Remember the result!
if (Instruction *I = LocalCache.getInst())
return LocalCache;
}
+#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
// 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());
BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
DirtyBlocks.push_back(*PI);
- NumUncacheNonLocal++;
+ ++NumUncacheNonLocal;
}
// isReadonlyCall - If this is a read-only call, we can be more aggressive.
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() && prior(Entry)->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,
// 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!
///
void MemoryDependenceAnalysis::
getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
- SmallVectorImpl<NonLocalDepEntry> &Result) {
- assert(isa<PointerType>(Pointer->getType()) &&
+ SmallVectorImpl<NonLocalDepResult> &Result) {
+ assert(Pointer->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);
+ 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
// 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(Pointer, PointeeSize, isLoad, FromBB,
+ if (!getNonLocalPointerDepFromBB(Address, PointeeSize, isLoad, FromBB,
Result, Visited, true))
return;
- Result.push_back(std::make_pair(FromBB,
- MemDepResult::getClobber(FromBB->begin())));
+ Result.clear();
+ Result.push_back(NonLocalDepResult(FromBB,
+ MemDepResult::getClobber(FromBB->begin()),
+ Pointer));
}
/// GetNonLocalInfoForBlock - Compute the memdep value for BB with
// the cache set. If so, find it.
NonLocalDepInfo::iterator Entry =
std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
- std::make_pair(BB, MemDepResult()));
- if (Entry != Cache->begin() && prior(Entry)->first == BB)
+ 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());
+ RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos, CacheKey);
} else {
++NumUncacheNonLocalPtr;
}
// 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
Instruction *Inst = Dep.getInst();
assert(Inst && "Didn't depend on anything?");
ValueIsLoadPair CacheKey(Pointer, isLoad);
- ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue());
+ 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 - Perform a dependency query based on
/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
/// not compute dependence information for some reason. This should be treated
/// as a clobber dependence on the first instruction in the predecessor block.
bool MemoryDependenceAnalysis::
-getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
+getNonLocalPointerDepFromBB(const PHITransAddr &Pointer, uint64_t PointeeSize,
bool isLoad, BasicBlock *StartBB,
- SmallVectorImpl<NonLocalDepEntry> &Result,
+ SmallVectorImpl<NonLocalDepResult> &Result,
DenseMap<BasicBlock*, Value*> &Visited,
bool SkipFirstBlock) {
// Look up the cached info for Pointer.
- ValueIsLoadPair CacheKey(Pointer, isLoad);
+ ValueIsLoadPair CacheKey(Pointer.getAddr(), isLoad);
std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
&NonLocalPointerDeps[CacheKey];
// If we have valid cached information for exactly the block we are
// investigating, just return it with no recomputation.
if (CacheInfo->first == BBSkipFirstBlockPair(StartBB, SkipFirstBlock)) {
+ // We have a fully cached result for this query then we can just return the
+ // cached results and populate the visited set. However, we have to verify
+ // that we don't already have conflicting results for these blocks. Check
+ // to ensure that if a block in the results set is in the visited set that
+ // it was for the same pointer query.
+ if (!Visited.empty()) {
+ for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
+ I != E; ++I) {
+ DenseMap<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 false;
}
// won't get any reuse from currently inserted values, because we don't
// revisit blocks after we insert info for them.
unsigned NumSortedEntries = Cache->size();
+ DEBUG(AssertSorted(*Cache));
while (!Worklist.empty()) {
BasicBlock *BB = Worklist.pop_back_val();
// 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(Pointer.getAddr(), PointeeSize,
+ isLoad, BB, Cache,
+ NumSortedEntries);
// If we got a Def or Clobber, add this to the list of results.
if (!Dep.isNonLocal()) {
- Result.push_back(NonLocalDepEntry(BB, Dep));
+ Result.push_back(NonLocalDepResult(BB, Dep, Pointer.getAddr()));
continue;
}
}
// If 'Pointer' is an instruction defined in this block, then we need to do
// phi translation to change it into a value live in the predecessor block.
- // If phi translation fails, then we can't continue dependence analysis.
- Instruction *PtrInst = dyn_cast<Instruction>(Pointer);
- bool NeedsPHITranslation = PtrInst && PtrInst->getParent() == BB;
-
- // If no PHI translation is needed, just add all the predecessors of this
- // block to scan them as well.
- if (!NeedsPHITranslation) {
+ // If not, we just add the predecessors to the worklist and scan them with
+ // the same Pointer.
+ if (!Pointer.NeedsPHITranslationFromBlock(BB)) {
SkipFirstBlock = false;
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));
+ InsertRes = Visited.insert(std::make_pair(*PI, Pointer.getAddr()));
if (InsertRes.second) {
// First time we've looked at *PI.
Worklist.push_back(*PI);
// 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)
+ if (InsertRes.first->second != Pointer.getAddr())
goto PredTranslationFailure;
}
continue;
}
- // If we do need to do phi translation, then there are a bunch of different
- // cases, because we have to find a Value* live in the predecessor block. We
- // know that PtrInst is defined in this block at least.
+ // We 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;
- // If this is directly a PHI node, just use the incoming values for each
- // pred as the phi translated version.
- if (PHINode *PtrPHI = dyn_cast<PHINode>(PtrInst)) {
- for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI){
- BasicBlock *Pred = *PI;
- Value *PredPtr = PtrPHI->getIncomingValueForBlock(Pred);
-
- // Check to see if we have already visited this pred block with another
- // pointer. If so, we can't do this lookup. This failure can occur
- // with PHI translation when a critical edge exists and the PHI node in
- // the successor translates to a pointer value different than the
- // pointer the block was first analyzed with.
- std::pair<DenseMap<BasicBlock*,Value*>::iterator, bool>
- InsertRes = Visited.insert(std::make_pair(Pred, PredPtr));
+ // 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;
+
+ for (BasicBlock **PI = PredCache->GetPreds(BB); *PI; ++PI) {
+ BasicBlock *Pred = *PI;
+
+ // Get the PHI translated pointer in this predecessor. This can fail if
+ // not translatable, in which case the getAddr() returns null.
+ PHITransAddr PredPointer(Pointer);
+ PredPointer.PHITranslateValue(BB, Pred, 0);
- if (!InsertRes.second) {
- // If the predecessor was visited with PredPtr, then we already did
- // the analysis and can ignore it.
- if (InsertRes.first->second == PredPtr)
- continue;
-
- // Otherwise, the block was previously analyzed with a different
- // pointer. We can't represent the result of this case, so we just
- // treat this as a phi translation failure.
- goto PredTranslationFailure;
- }
+ 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) {
+ // If the predecessor was visited with PredPtr, then we already did
+ // the analysis and can ignore it.
+ if (InsertRes.first->second == PredPtrVal)
+ continue;
- // If we have a problem phi translating, fall through to the code below
- // to handle the failure condition.
- if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
- Result, Visited))
- goto PredTranslationFailure;
+ // Otherwise, the block was previously analyzed with a different
+ // pointer. We can't represent the result of this case, so we just
+ // treat this as a phi translation failure.
+ goto PredTranslationFailure;
}
- // Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
- CacheInfo = &NonLocalPointerDeps[CacheKey];
- Cache = &CacheInfo->second;
+ // 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) {
+ // Add the entry to the Result list.
+ NonLocalDepResult Entry(Pred,
+ MemDepResult::getClobber(Pred->getTerminator()),
+ 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.
+ NonLocalPointerDeps[CacheKey].first = BBSkipFirstBlockPair();
+ continue;
+ }
+
+ // FIXME: it is entirely possible that PHI translating will end up with
+ // the same value. Consider PHI translating something like:
+ // X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
+ // to recurse here, pedantically speaking.
- // Since we did phi translation, the "Cache" set won't contain all of the
- // results for the query. This is ok (we can still use it to accelerate
- // specific block queries) but we can't do the fastpath "return all
- // results from the set" Clear out the indicator for this.
- CacheInfo->first = BBSkipFirstBlockPair();
- SkipFirstBlock = false;
- continue;
+ // If we have a problem phi translating, fall through to the code below
+ // to handle the failure condition.
+ if (getNonLocalPointerDepFromBB(PredPointer, PointeeSize, isLoad, Pred,
+ Result, Visited))
+ goto PredTranslationFailure;
}
- // TODO: BITCAST, GEP.
-
- // cerr << "MEMDEP: Could not PHI translate: " << *Pointer;
- // if (isa<BitCastInst>(PtrInst) || isa<GetElementPtrInst>(PtrInst))
- // cerr << "OP:\t\t\t\t" << *PtrInst->getOperand(0);
- PredTranslationFailure:
-
// Refresh the CacheInfo/Cache pointer so that it isn't invalidated.
CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second;
+ NumSortedEntries = Cache->size();
// Since we did phi translation, the "Cache" set won't contain all of the
// results for the query. This is ok (we can still use it to accelerate
// specific block queries) but we can't do the fastpath "return all
// results from the set" Clear out the indicator for this.
CacheInfo->first = BBSkipFirstBlockPair();
+ SkipFirstBlock = false;
+ continue;
+
+ PredTranslationFailure:
+
+ if (Cache == 0) {
+ // Refresh the CacheInfo/Cache pointer if it got invalidated.
+ CacheInfo = &NonLocalPointerDeps[CacheKey];
+ Cache = &CacheInfo->second;
+ 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->first = BBSkipFirstBlockPair();
// If *nothing* works, mark the pointer as being clobbered by the first
// instruction in this block.
for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
assert(I != Cache->rend() && "Didn't find current block??");
- if (I->first != BB)
+ if (I->getBB() != BB)
continue;
- assert(I->second.isNonLocal() &&
+ assert(I->getResult().isNonLocal() &&
"Should only be here with transparent block");
- I->second = MemDepResult::getClobber(BB->begin());
- ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey.getOpaqueValue());
- Result.push_back(*I);
+ I->setResult(MemDepResult::getClobber(BB->begin()));
+ ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey);
+ 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.
- if (Cache->size() != 1) {
- NonLocalDepEntry Val = Cache->back();
- Cache->pop_back();
- NonLocalDepInfo::iterator Entry =
- std::upper_bound(Cache->begin(), Cache->end(), Val);
- Cache->insert(Entry, Val);
- }
- break;
- default:
- // Added many values, do a full scale sort.
- std::sort(Cache->begin(), Cache->end());
- }
-
+ SortNonLocalDepInfoCache(*Cache, NumSortedEntries);
+ DEBUG(AssertSorted(*Cache));
return false;
}
NonLocalDepInfo &PInfo = It->second.second;
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);
+ 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).
/// in more places that cached info does not necessarily keep.
void MemoryDependenceAnalysis::invalidateCachedPointerInfo(Value *Ptr) {
// If Ptr isn't really a pointer, just ignore it.
- if (!isa<PointerType>(Ptr->getType())) return;
+ 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");
// 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();
}
}
const NonLocalDepInfo &Val = I->second.second;
for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
II != E; ++II)
- assert(II->second.getInst() != D && "Inst occurs as NLPD value");
+ 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");
}