#include "llvm/Function.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
#include "llvm/Support/CFG.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/TargetData.h"
using namespace llvm;
-STATISTIC(NumCacheNonLocal, "Number of cached non-local responses");
+STATISTIC(NumCacheNonLocal, "Number of fully cached non-local responses");
+STATISTIC(NumCacheDirtyNonLocal, "Number of dirty cached non-local responses");
STATISTIC(NumUncacheNonLocal, "Number of uncached non-local responses");
+STATISTIC(NumCacheNonLocalPtr,
+ "Number of fully cached non-local ptr responses");
+STATISTIC(NumCacheDirtyNonLocalPtr,
+ "Number of cached, but dirty, non-local ptr responses");
+STATISTIC(NumUncacheNonLocalPtr,
+ "Number of uncached non-local ptr responses");
+
char MemoryDependenceAnalysis::ID = 0;
// Register this pass...
AU.addRequiredTransitive<TargetData>();
}
-/// getCallSiteDependency - Private helper for finding the local dependencies
-/// of a call site.
+bool MemoryDependenceAnalysis::runOnFunction(Function &) {
+ AA = &getAnalysis<AliasAnalysis>();
+ TD = &getAnalysis<TargetData>();
+ return false;
+}
+
+
+/// getCallSiteDependencyFrom - Private helper for finding the local
+/// dependencies of a call site.
MemDepResult MemoryDependenceAnalysis::
-getCallSiteDependency(CallSite C, BasicBlock::iterator ScanIt,
- BasicBlock *BB) {
- AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
- TargetData &TD = getAnalysis<TargetData>();
-
+getCallSiteDependencyFrom(CallSite CS, BasicBlock::iterator ScanIt,
+ BasicBlock *BB) {
// Walk backwards through the block, looking for dependencies
while (ScanIt != BB->begin()) {
Instruction *Inst = --ScanIt;
uint64_t PointerSize = 0;
if (StoreInst *S = dyn_cast<StoreInst>(Inst)) {
Pointer = S->getPointerOperand();
- PointerSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
- } else if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
- Pointer = AI;
- if (ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()))
- // Use ABI size (size between elements), not store size (size of one
- // element without padding).
- PointerSize = C->getZExtValue() *
- TD.getABITypeSize(AI->getAllocatedType());
- else
- PointerSize = ~0UL;
+ PointerSize = TD->getTypeStoreSize(S->getOperand(0)->getType());
} else if (VAArgInst *V = dyn_cast<VAArgInst>(Inst)) {
Pointer = V->getOperand(0);
- PointerSize = TD.getTypeStoreSize(V->getType());
+ PointerSize = TD->getTypeStoreSize(V->getType());
} else if (FreeInst *F = dyn_cast<FreeInst>(Inst)) {
Pointer = F->getPointerOperand();
// FreeInsts erase the entire structure
- PointerSize = ~0UL;
+ PointerSize = ~0ULL;
} else if (isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) {
- if (AA.getModRefBehavior(CallSite::get(Inst)) ==
- AliasAnalysis::DoesNotAccessMemory)
+ CallSite InstCS = CallSite::get(Inst);
+ // If these two calls do not interfere, look past it.
+ if (AA->getModRefInfo(CS, InstCS) == AliasAnalysis::NoModRef)
+ 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())
+ return MemDepResult::getClobber(Inst);
+ return MemDepResult::getDef(Inst);
+ } else {
+ // Non-memory instruction.
+ continue;
+ }
+
+ if (AA->getModRefInfo(CS, Pointer, PointerSize) != 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.
+ if (BB != &BB->getParent()->getEntryBlock())
+ return MemDepResult::getNonLocal();
+ return MemDepResult::getClobber(ScanIt);
+}
+
+/// getPointerDependencyFrom - Return the instruction on which a memory
+/// location depends. If isLoad is true, this routine ignore may-aliases with
+/// read-only operations.
+MemDepResult MemoryDependenceAnalysis::
+getPointerDependencyFrom(Value *MemPtr, uint64_t MemSize, bool isLoad,
+ BasicBlock::iterator ScanIt, BasicBlock *BB) {
+
+ // Walk backwards through the basic block, looking for dependencies.
+ while (ScanIt != BB->begin()) {
+ Instruction *Inst = --ScanIt;
+
+ // Values depend on loads if the pointers are must aliased. This means that
+ // a load depends on another must aliased load from the same value.
+ if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
+ Value *Pointer = LI->getPointerOperand();
+ uint64_t PointerSize = TD->getTypeStoreSize(LI->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);
+ if (R == AliasAnalysis::NoAlias)
continue;
- return MemDepResult::get(Inst);
- } else
+
+ // May-alias loads don't depend on each other without a dependence.
+ if (isLoad && R == AliasAnalysis::MayAlias)
+ continue;
+ // Stores depend on may and must aliased loads, loads depend on must-alias
+ // loads.
+ return MemDepResult::getDef(Inst);
+ }
+
+ if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
+ Value *Pointer = SI->getPointerOperand();
+ uint64_t PointerSize = TD->getTypeStoreSize(SI->getOperand(0)->getType());
+
+ // If we found a pointer, check if it could be the same as our pointer.
+ AliasAnalysis::AliasResult R =
+ AA->alias(Pointer, PointerSize, MemPtr, MemSize);
+
+ if (R == AliasAnalysis::NoAlias)
+ continue;
+ if (R == AliasAnalysis::MayAlias)
+ return MemDepResult::getClobber(Inst);
+ return MemDepResult::getDef(Inst);
+ }
+
+ // If this is an allocation, and if we know that the accessed pointer is to
+ // the allocation, return Def. This means that there is no dependence and
+ // the access can be optimized based on that. For example, a load could
+ // turn into undef.
+ if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
+ Value *AccessPtr = MemPtr->getUnderlyingObject();
+
+ if (AccessPtr == AI ||
+ AA->alias(AI, 1, AccessPtr, 1) == AliasAnalysis::MustAlias)
+ return MemDepResult::getDef(AI);
+ 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)
continue;
- if (AA.getModRefInfo(C, Pointer, PointerSize) != AliasAnalysis::NoModRef)
- return MemDepResult::get(Inst);
+ // Otherwise, there is a dependence.
+ return MemDepResult::getClobber(Inst);
+ }
+
+ // No dependence found. If this is the entry block of the function, it is a
+ // clobber, otherwise it is non-local.
+ if (BB != &BB->getParent()->getEntryBlock())
+ return MemDepResult::getNonLocal();
+ return MemDepResult::getClobber(ScanIt);
+}
+
+/// getDependency - Return the instruction on which a memory operation
+/// depends.
+MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
+ Instruction *ScanPos = QueryInst;
+
+ // Check for a cached result
+ MemDepResult &LocalCache = LocalDeps[QueryInst];
+
+ // If the cached entry is non-dirty, just return it. Note that this depends
+ // on MemDepResult's default constructing to 'dirty'.
+ if (!LocalCache.isDirty())
+ return LocalCache;
+
+ // Otherwise, if we have a dirty entry, we know we can start the scan at that
+ // instruction, which may save us some work.
+ if (Instruction *Inst = LocalCache.getInst()) {
+ ScanPos = Inst;
+
+ SmallPtrSet<Instruction*, 4> &InstMap = ReverseLocalDeps[Inst];
+ bool Found = InstMap.erase(QueryInst);
+ assert(Found && "Invalid reverse map!"); Found=Found;
+ if (InstMap.empty())
+ // FIXME: use an iterator to avoid looking up inst again.
+ ReverseLocalDeps.erase(Inst);
+ }
+
+ BasicBlock *QueryParent = QueryInst->getParent();
+
+ Value *MemPtr = 0;
+ uint64_t MemSize = 0;
+
+ // Do the scan.
+ if (BasicBlock::iterator(QueryInst) == QueryParent->begin()) {
+ // No dependence found. If this is the entry block of the function, it is a
+ // clobber, otherwise it is non-local.
+ if (QueryParent != &QueryParent->getParent()->getEntryBlock())
+ LocalCache = MemDepResult::getNonLocal();
+ else
+ LocalCache = MemDepResult::getClobber(QueryInst);
+ } else if (StoreInst *SI = dyn_cast<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;
+ } else {
+ // Non-memory instruction.
+ LocalCache = MemDepResult::getClobber(--BasicBlock::iterator(ScanPos));
}
- // No dependence found.
- return MemDepResult::getNonLocal();
+ // 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
/// This method assumes the instruction returns a "nonlocal" dependency
/// within its own block.
///
-void MemoryDependenceAnalysis::
-getNonLocalDependency(Instruction *QueryInst,
- SmallVectorImpl<std::pair<BasicBlock*,
- MemDepResult> > &Result) {
+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!");
- DenseMap<BasicBlock*, DepResultTy> &Cache = NonLocalDeps[QueryInst];
+ PerInstNLInfo &CacheP = NonLocalDeps[QueryInst];
+ NonLocalDepInfo &Cache = CacheP.first;
/// DirtyBlocks - This is the set of blocks that need to be recomputed. In
/// the cached case, this can happen due to instructions being deleted etc. In
SmallVector<BasicBlock*, 32> DirtyBlocks;
if (!Cache.empty()) {
+ // Okay, we have a cache entry. If we know it is not dirty, just return it
+ // with no computation.
+ if (!CacheP.second) {
+ NumCacheNonLocal++;
+ return Cache;
+ }
+
// If we already have a partially computed set of results, scan them to
// determine what is dirty, seeding our initial DirtyBlocks worklist.
- // FIXME: In the "don't need to be updated" case, this is expensive, why not
- // have a per-"cache" flag saying it is undirty?
- for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
- E = Cache.end(); I != E; ++I)
- if (I->second.getInt() == Dirty)
+ for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
+ I != E; ++I)
+ if (I->second.isDirty())
DirtyBlocks.push_back(I->first);
- NumCacheNonLocal++;
+ // Sort the cache so that we can do fast binary search lookups below.
+ std::sort(Cache.begin(), Cache.end());
+ ++NumCacheDirtyNonLocal;
//cerr << "CACHED CASE: " << DirtyBlocks.size() << " dirty: "
// << Cache.size() << " cached: " << *QueryInst;
} else {
NumUncacheNonLocal++;
}
+ // Visited checked first, vector in sorted order.
+ SmallPtrSet<BasicBlock*, 64> Visited;
+
+ unsigned NumSortedEntries = Cache.size();
// Iterate while we still have blocks to update.
while (!DirtyBlocks.empty()) {
BasicBlock *DirtyBB = DirtyBlocks.back();
DirtyBlocks.pop_back();
- // Get the entry for this block. Note that this relies on DepResultTy
- // default initializing to Dirty.
- DepResultTy &DirtyBBEntry = Cache[DirtyBB];
+ // Already processed this block?
+ if (!Visited.insert(DirtyBB))
+ continue;
- // If DirtyBBEntry isn't dirty, it ended up on the worklist multiple times.
- if (DirtyBBEntry.getInt() != Dirty) continue;
-
- // Find out if this block has a local dependency for QueryInst.
- // FIXME: If the dirty entry has an instruction pointer, scan from it!
- // FIXME: Don't convert back and forth for MemDepResult <-> DepResultTy.
+ // Do a binary search to see if we already have an entry for this block in
+ // the cache set. If so, find it.
+ NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache.begin(), Cache.begin()+NumSortedEntries,
+ std::make_pair(DirtyBB, MemDepResult()));
+ if (Entry != Cache.begin() && (&*Entry)[-1].first == DirtyBB)
+ --Entry;
+
+ MemDepResult *ExistingResult = 0;
+ if (Entry != Cache.begin()+NumSortedEntries &&
+ Entry->first == DirtyBB) {
+ // If we already have an entry, and if it isn't already dirty, the block
+ // is done.
+ if (!Entry->second.isDirty())
+ continue;
+
+ // Otherwise, remember this slot so we can update the value.
+ ExistingResult = &Entry->second;
+ }
// If the dirty entry has a pointer, start scanning from it so we don't have
// to rescan the entire block.
BasicBlock::iterator ScanPos = DirtyBB->end();
- if (Instruction *Inst = DirtyBBEntry.getPointer())
- ScanPos = Inst;
+ if (ExistingResult) {
+ if (Instruction *Inst = ExistingResult->getInst()) {
+ ScanPos = Inst;
+
+ // We're removing QueryInst's use of Inst.
+ SmallPtrSet<Instruction*, 4> &InstMap = ReverseNonLocalDeps[Inst];
+ bool Found = InstMap.erase(QueryInst);
+ assert(Found && "Invalid reverse map!"); Found=Found;
+ // FIXME: Use an iterator to avoid looking up inst again.
+ if (InstMap.empty()) ReverseNonLocalDeps.erase(Inst);
+ }
+ }
+
+ // 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());
+ }
+ } else {
+ assert(isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst));
+ Dep = getCallSiteDependencyFrom(CallSite::get(QueryInst), ScanPos,
+ DirtyBB);
+ }
+
+ 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;
+ else
+ Cache.push_back(std::make_pair(DirtyBB, Dep));
- DirtyBBEntry = ConvFromResult(getDependencyFrom(QueryInst, ScanPos,
- DirtyBB));
-
// If the block has a dependency (i.e. it isn't completely transparent to
- // the value), remember it!
- if (DirtyBBEntry.getInt() != NonLocal) {
+ // the value), remember the association!
+ if (!Dep.isNonLocal()) {
// Keep the ReverseNonLocalDeps map up to date so we can efficiently
// update this when we remove instructions.
- if (Instruction *Inst = DirtyBBEntry.getPointer())
+ if (Instruction *Inst = Dep.getInst())
ReverseNonLocalDeps[Inst].insert(QueryInst);
- continue;
- }
+ } else {
- // If the block *is* completely transparent to the load, we need to check
- // the predecessors of this block. Add them to our worklist.
- DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB));
+ // If the block *is* completely transparent to the load, we need to check
+ // the predecessors of this block. Add them to our worklist.
+ DirtyBlocks.append(pred_begin(DirtyBB), pred_end(DirtyBB));
+ }
}
+ return Cache;
+}
+
+/// getNonLocalPointerDependency - Perform a full dependency query for an
+/// access to the specified (non-volatile) memory location, returning the
+/// set of instructions that either define or clobber the value.
+///
+/// This method assumes the pointer has a "NonLocal" dependency within its
+/// own block.
+///
+void MemoryDependenceAnalysis::
+getNonLocalPointerDependency(Value *Pointer, bool isLoad, BasicBlock *FromBB,
+ SmallVectorImpl<NonLocalDepEntry> &Result) {
+ 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);
- // Copy the result into the output set.
- for (DenseMap<BasicBlock*, DepResultTy>::iterator I = Cache.begin(),
- E = Cache.end(); I != E; ++I)
- Result.push_back(std::make_pair(I->first, ConvToResult(I->second)));
+ // While we have blocks to analyze, get their values.
+ SmallPtrSet<BasicBlock*, 64> Visited;
+
+ for (pred_iterator PI = pred_begin(FromBB), E = pred_end(FromBB); PI != E;
+ ++PI) {
+ // TODO: PHI TRANSLATE.
+ getNonLocalPointerDepInternal(Pointer, PointeeSize, isLoad, *PI,
+ Result, Visited);
+ }
}
-/// getDependency - Return the instruction on which a memory operation
-/// depends. The local parameter indicates if the query should only
-/// evaluate dependencies within the same basic block.
-MemDepResult MemoryDependenceAnalysis::
-getDependencyFrom(Instruction *QueryInst, BasicBlock::iterator ScanIt,
- BasicBlock *BB) {
- AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
- TargetData &TD = getAnalysis<TargetData>();
+void MemoryDependenceAnalysis::
+getNonLocalPointerDepInternal(Value *Pointer, uint64_t PointeeSize,
+ bool isLoad, BasicBlock *StartBB,
+ SmallVectorImpl<NonLocalDepEntry> &Result,
+ SmallPtrSet<BasicBlock*, 64> &Visited) {
+ SmallVector<BasicBlock*, 32> Worklist;
+ Worklist.push_back(StartBB);
- // Get the pointer value for which dependence will be determined
- Value *MemPtr = 0;
- uint64_t MemSize = 0;
- bool MemVolatile = false;
-
- if (StoreInst* S = dyn_cast<StoreInst>(QueryInst)) {
- MemPtr = S->getPointerOperand();
- MemSize = TD.getTypeStoreSize(S->getOperand(0)->getType());
- MemVolatile = S->isVolatile();
- } else if (LoadInst* L = dyn_cast<LoadInst>(QueryInst)) {
- MemPtr = L->getPointerOperand();
- MemSize = TD.getTypeStoreSize(L->getType());
- MemVolatile = L->isVolatile();
- } else if (VAArgInst* V = dyn_cast<VAArgInst>(QueryInst)) {
- MemPtr = V->getOperand(0);
- MemSize = TD.getTypeStoreSize(V->getType());
- } else if (FreeInst* F = dyn_cast<FreeInst>(QueryInst)) {
- MemPtr = F->getPointerOperand();
- // FreeInsts erase the entire structure, not just a field.
- MemSize = ~0UL;
- } else if (isa<CallInst>(QueryInst) || isa<InvokeInst>(QueryInst))
- return getCallSiteDependency(CallSite::get(QueryInst), ScanIt, BB);
- else // Non-memory instructions depend on nothing.
- return MemDepResult::getNone();
+ // Look up the cached info for Pointer.
+ ValueIsLoadPair CacheKey(Pointer, isLoad);
+ NonLocalDepInfo *Cache = &NonLocalPointerDeps[CacheKey];
- // Walk backwards through the basic block, looking for dependencies
- while (ScanIt != BB->begin()) {
- Instruction *Inst = --ScanIt;
+ // 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();
+
+ while (!Worklist.empty()) {
+ BasicBlock *BB = Worklist.pop_back_val();
+
+ // Analyze the dependency of *Pointer in FromBB. See if we already have
+ // been here.
+ if (!Visited.insert(BB))
+ continue;
- // If the access is volatile and this is a volatile load/store, return a
- // dependence.
- if (MemVolatile &&
- ((isa<LoadInst>(Inst) && cast<LoadInst>(Inst)->isVolatile()) ||
- (isa<StoreInst>(Inst) && cast<StoreInst>(Inst)->isVolatile())))
- return MemDepResult::get(Inst);
+ // Get the dependency info for Pointer in BB. If we have cached
+ // information, we will use it, otherwise we compute it.
+
+ // Do a binary search to see if we already have an entry for this block in
+ // the cache set. If so, find it.
+ NonLocalDepInfo::iterator Entry =
+ std::upper_bound(Cache->begin(), Cache->begin()+NumSortedEntries,
+ std::make_pair(BB, MemDepResult()));
+ if (Entry != Cache->begin() && (&*Entry)[-1].first == BB)
+ --Entry;
+
+ MemDepResult *ExistingResult = 0;
+ if (Entry != Cache->begin()+NumSortedEntries && Entry->first == BB)
+ ExistingResult = &Entry->second;
+
+ // If we have a cached entry, and it is non-dirty, use it as the value for
+ // this dependency.
+ MemDepResult Dep;
+ if (ExistingResult && !ExistingResult->isDirty()) {
+ Dep = *ExistingResult;
+ ++NumCacheNonLocalPtr;
+ } else {
+ // Otherwise, we have to scan for the value. If we have a dirty cache
+ // entry, start scanning from its position, otherwise we scan from the end
+ // of the block.
+ BasicBlock::iterator ScanPos = BB->end();
+ if (ExistingResult && ExistingResult->getInst()) {
+ assert(ExistingResult->getInst()->getParent() == BB &&
+ "Instruction invalidated?");
+ ++NumCacheDirtyNonLocalPtr;
+ ScanPos = ExistingResult->getInst();
- // MemDep is broken w.r.t. loads: it says that two loads of the same pointer
- // depend on each other. :(
- // FIXME: ELIMINATE THIS!
- if (LoadInst *L = dyn_cast<LoadInst>(Inst)) {
- Value *Pointer = L->getPointerOperand();
- uint64_t PointerSize = TD.getTypeStoreSize(L->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);
+ // Eliminating the dirty entry from 'Cache', so update the reverse info.
+ SmallPtrSet<void *, 4> &InstMap = ReverseNonLocalPtrDeps[ScanPos];
+ bool Contained = InstMap.erase(CacheKey.getOpaqueValue());
+ assert(Contained && "Invalid cache entry"); Contained=Contained;
+ // FIXME: Use an iterator to avoid a repeated lookup in ".erase".
+ if (InstMap.empty()) ReverseNonLocalPtrDeps.erase(ScanPos);
+ } else {
+ ++NumUncacheNonLocalPtr;
+ }
- if (R == AliasAnalysis::NoAlias)
- continue;
+ // Scan the block for the dependency.
+ Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad, ScanPos, BB);
- // May-alias loads don't depend on each other without a dependence.
- if (isa<LoadInst>(QueryInst) && R == AliasAnalysis::MayAlias)
- continue;
- return MemDepResult::get(Inst);
- }
-
- // FIXME: This claims that an access depends on the allocation. This may
- // make sense, but is dubious at best. It would be better to fix GVN to
- // handle a 'None' Query.
- if (AllocationInst *AI = dyn_cast<AllocationInst>(Inst)) {
- Value *Pointer = AI;
- uint64_t PointerSize;
- if (ConstantInt *C = dyn_cast<ConstantInt>(AI->getArraySize()))
- // Use ABI size (size between elements), not store size (size of one
- // element without padding).
- PointerSize = C->getZExtValue() *
- TD.getABITypeSize(AI->getAllocatedType());
+ // If we had a dirty entry for the block, update it. Otherwise, just add
+ // a new entry.
+ if (ExistingResult)
+ *ExistingResult = Dep;
else
- PointerSize = ~0UL;
-
- AliasAnalysis::AliasResult R =
- AA.alias(Pointer, PointerSize, MemPtr, MemSize);
+ Cache->push_back(std::make_pair(BB, Dep));
- if (R == AliasAnalysis::NoAlias)
- continue;
- return MemDepResult::get(Inst);
+ // 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()) {
+ // 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?");
+ ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue());
+ }
}
-
-
- // See if this instruction mod/ref's the pointer.
- AliasAnalysis::ModRefResult MRR = AA.getModRefInfo(Inst, MemPtr, MemSize);
-
- if (MRR == AliasAnalysis::NoModRef)
- continue;
- // Loads don't depend on read-only instructions.
- if (isa<LoadInst>(QueryInst) && MRR == AliasAnalysis::Ref)
+ // If we got a Def or Clobber, add this to the list of results.
+ if (!Dep.isNonLocal()) {
+ Result.push_back(NonLocalDepEntry(BB, Dep));
continue;
+ }
- // Otherwise, there is a dependence.
- return MemDepResult::get(Inst);
+ // Otherwise, we have to process all the predecessors of this block to scan
+ // them as well.
+ for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
+ // TODO: PHI TRANSLATE.
+ Worklist.push_back(*PI);
+ }
}
- // If we found nothing, return the non-local flag.
- return MemDepResult::getNonLocal();
+ // If we computed new values, re-sort Cache.
+ if (NumSortedEntries != Cache->size())
+ std::sort(Cache->begin(), Cache->end());
}
-/// getDependency - Return the instruction on which a memory operation
-/// depends.
-MemDepResult MemoryDependenceAnalysis::getDependency(Instruction *QueryInst) {
- Instruction *ScanPos = QueryInst;
+/// RemoveCachedNonLocalPointerDependencies - If P exists in
+/// CachedNonLocalPointerInfo, remove it.
+void MemoryDependenceAnalysis::
+RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair P) {
+ CachedNonLocalPointerInfo::iterator It =
+ NonLocalPointerDeps.find(P);
+ if (It == NonLocalPointerDeps.end()) return;
- // Check for a cached result
- DepResultTy &LocalCache = LocalDeps[QueryInst];
+ // Remove all of the entries in the BB->val map. This involves removing
+ // instructions from the reverse map.
+ NonLocalDepInfo &PInfo = It->second;
- // If the cached entry is non-dirty, just return it. Note that this depends
- // on DepResultTy's default constructing to 'dirty'.
- if (LocalCache.getInt() != Dirty)
- return ConvToResult(LocalCache);
+ for (unsigned i = 0, e = PInfo.size(); i != e; ++i) {
+ Instruction *Target = PInfo[i].second.getInst();
+ if (Target == 0) continue; // Ignore non-local dep results.
+ assert(Target->getParent() == PInfo[i].first && Target != P.getPointer());
- // Otherwise, if we have a dirty entry, we know we can start the scan at that
- // instruction, which may save us some work.
- if (Instruction *Inst = LocalCache.getPointer())
- ScanPos = Inst;
-
- // Do the scan.
- MemDepResult Res =
- getDependencyFrom(QueryInst, ScanPos, QueryInst->getParent());
-
- // Remember the result!
- // FIXME: Don't convert back and forth! Make a shared helper function.
- LocalCache = ConvFromResult(Res);
- if (Instruction *I = Res.getInst())
- ReverseLocalDeps[I].insert(QueryInst);
+ // Eliminating the dirty entry from 'Cache', so update the reverse info.
+ SmallPtrSet<void *, 4> &InstMap = ReverseNonLocalPtrDeps[Target];
+ bool Contained = InstMap.erase(P.getOpaqueValue());
+ assert(Contained && "Invalid cache entry"); Contained=Contained;
+
+ // FIXME: Use an iterator to avoid a repeated lookup in ".erase".
+ if (InstMap.empty()) ReverseNonLocalPtrDeps.erase(Target);
+ }
- return Res;
+ // Remove P from NonLocalPointerDeps (which deletes NonLocalDepInfo).
+ NonLocalPointerDeps.erase(It);
}
+
/// removeInstruction - Remove an instruction from the dependence analysis,
/// updating the dependence of instructions that previously depended on it.
/// This method attempts to keep the cache coherent using the reverse map.
void MemoryDependenceAnalysis::removeInstruction(Instruction *RemInst) {
// Walk through the Non-local dependencies, removing this one as the value
// for any cached queries.
- for (DenseMap<BasicBlock*, DepResultTy>::iterator DI =
- NonLocalDeps[RemInst].begin(), DE = NonLocalDeps[RemInst].end();
- DI != DE; ++DI)
- if (Instruction *Inst = DI->second.getPointer())
- ReverseNonLocalDeps[Inst].erase(RemInst);
+ NonLocalDepMapType::iterator NLDI = NonLocalDeps.find(RemInst);
+ if (NLDI != NonLocalDeps.end()) {
+ NonLocalDepInfo &BlockMap = NLDI->second.first;
+ for (NonLocalDepInfo::iterator DI = BlockMap.begin(), DE = BlockMap.end();
+ DI != DE; ++DI)
+ if (Instruction *Inst = DI->second.getInst())
+ ReverseNonLocalDeps[Inst].erase(RemInst);
+ NonLocalDeps.erase(NLDI);
+ }
// If we have a cached local dependence query for this instruction, remove it.
//
LocalDepMapType::iterator LocalDepEntry = LocalDeps.find(RemInst);
if (LocalDepEntry != LocalDeps.end()) {
// Remove us from DepInst's reverse set now that the local dep info is gone.
- if (Instruction *Inst = LocalDepEntry->second.getPointer()) {
+ if (Instruction *Inst = LocalDepEntry->second.getInst()) {
SmallPtrSet<Instruction*, 4> &RLD = ReverseLocalDeps[Inst];
- RLD.erase(RemInst);
+ bool Found = RLD.erase(RemInst);
+ assert(Found && "Invalid reverse map!"); Found=Found;
+ // FIXME: Use an iterator to avoid looking up Inst again.
if (RLD.empty())
ReverseLocalDeps.erase(Inst);
}
// Remove this local dependency info.
LocalDeps.erase(LocalDepEntry);
- }
+ }
+
+ // If we have any cached pointer dependencies on this instruction, remove
+ // them. If the instruction has non-pointer type, then it can't be a pointer
+ // base.
+
+ // Remove it from both the load info and the store info. The instruction
+ // can't be in either of these maps if it is non-pointer.
+ if (isa<PointerType>(RemInst->getType())) {
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, false));
+ RemoveCachedNonLocalPointerDependencies(ValueIsLoadPair(RemInst, true));
+ }
// Loop over all of the things that depend on the instruction we're removing.
//
ReverseDepMapType::iterator ReverseDepIt = ReverseLocalDeps.find(RemInst);
if (ReverseDepIt != ReverseLocalDeps.end()) {
SmallPtrSet<Instruction*, 4> &ReverseDeps = ReverseDepIt->second;
- // RemInst can't be the terminator if it has stuff depending on it.
+ // RemInst can't be the terminator if it has local stuff depending on it.
assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
"Nothing can locally depend on a terminator");
// dependent on the instruction after RemInst. It will have the dirty flag
// set so it will rescan. This saves having to scan the entire block to get
// to this point.
- Instruction *NewDepInst = next(BasicBlock::iterator(RemInst));
+ Instruction *NewDepInst = ++BasicBlock::iterator(RemInst);
for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
E = ReverseDeps.end(); I != E; ++I) {
Instruction *InstDependingOnRemInst = *I;
-
- // If we thought the instruction depended on itself (possible for
- // unconfirmed dependencies) ignore the update.
- if (InstDependingOnRemInst == RemInst) continue;
+ assert(InstDependingOnRemInst != RemInst &&
+ "Already removed our local dep info");
- LocalDeps[InstDependingOnRemInst] = DepResultTy(NewDepInst, Dirty);
+ LocalDeps[InstDependingOnRemInst] = MemDepResult::getDirty(NewDepInst);
// Make sure to remember that new things depend on NewDepInst.
ReverseDepsToAdd.push_back(std::make_pair(NewDepInst,
ReverseDepIt = ReverseNonLocalDeps.find(RemInst);
if (ReverseDepIt != ReverseNonLocalDeps.end()) {
- SmallPtrSet<Instruction*, 4>& set = ReverseDepIt->second;
- for (SmallPtrSet<Instruction*, 4>::iterator I = set.begin(), E = set.end();
- I != E; ++I)
- for (DenseMap<BasicBlock*, DepResultTy>::iterator
- DI = NonLocalDeps[*I].begin(), DE = NonLocalDeps[*I].end();
- DI != DE; ++DI)
- if (DI->second.getPointer() == RemInst) {
- // Convert to a dirty entry for the subsequent instruction.
- DI->second.setInt(Dirty);
- if (RemInst->isTerminator())
- DI->second.setPointer(0);
- else {
- Instruction *NextI = next(BasicBlock::iterator(RemInst));
- DI->second.setPointer(NextI);
- assert(NextI != RemInst);
- ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
- }
+ SmallPtrSet<Instruction*, 4> &Set = ReverseDepIt->second;
+ for (SmallPtrSet<Instruction*, 4>::iterator I = Set.begin(), E = Set.end();
+ I != E; ++I) {
+ assert(*I != RemInst && "Already removed NonLocalDep info for RemInst");
+
+ PerInstNLInfo &INLD = NonLocalDeps[*I];
+ // The information is now dirty!
+ INLD.second = true;
+
+ for (NonLocalDepInfo::iterator DI = INLD.first.begin(),
+ DE = INLD.first.end(); DI != DE; ++DI) {
+ if (DI->second.getInst() != RemInst) continue;
+
+ // Convert to a dirty entry for the subsequent instruction.
+ Instruction *NextI = 0;
+ if (!RemInst->isTerminator()) {
+ NextI = ++BasicBlock::iterator(RemInst);
+ ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
}
+ DI->second = MemDepResult::getDirty(NextI);
+ }
+ }
ReverseNonLocalDeps.erase(ReverseDepIt);
}
}
- NonLocalDeps.erase(RemInst);
- getAnalysis<AliasAnalysis>().deleteValue(RemInst);
+ // If the instruction is in ReverseNonLocalPtrDeps then it appears as a
+ // value in the NonLocalPointerDeps info.
+ ReverseNonLocalPtrDepTy::iterator ReversePtrDepIt =
+ ReverseNonLocalPtrDeps.find(RemInst);
+ if (ReversePtrDepIt != ReverseNonLocalPtrDeps.end()) {
+ SmallPtrSet<void*, 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);
+ assert(P.getPointer() != RemInst &&
+ "Already removed NonLocalPointerDeps info for RemInst");
+
+ NonLocalDepInfo &NLPDI = NonLocalPointerDeps[P];
+
+ MemDepResult NewDirtyVal;
+ if (!RemInst->isTerminator())
+ NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(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;
+
+ // Convert to a dirty entry for the subsequent instruction.
+ DI->second = NewDirtyVal;
+
+ if (Instruction *NewDirtyInst = NewDirtyVal.getInst())
+ ReversePtrDepsToAdd.push_back(std::make_pair(NewDirtyInst, P));
+ }
+ }
+
+ ReverseNonLocalPtrDeps.erase(ReversePtrDepIt);
+
+ while (!ReversePtrDepsToAdd.empty()) {
+ ReverseNonLocalPtrDeps[ReversePtrDepsToAdd.back().first]
+ .insert(ReversePtrDepsToAdd.back().second.getOpaqueValue());
+ ReversePtrDepsToAdd.pop_back();
+ }
+ }
+
+
+ assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
+ AA->deleteValue(RemInst);
DEBUG(verifyRemoved(RemInst));
}
for (LocalDepMapType::const_iterator I = LocalDeps.begin(),
E = LocalDeps.end(); I != E; ++I) {
assert(I->first != D && "Inst occurs in data structures");
- assert(I->second.getPointer() != D &&
+ assert(I->second.getInst() != D &&
"Inst occurs in data structures");
}
+ for (CachedNonLocalPointerInfo::const_iterator I =NonLocalPointerDeps.begin(),
+ E = NonLocalPointerDeps.end(); I != E; ++I) {
+ assert(I->first.getPointer() != D && "Inst occurs in NLPD map key");
+ const NonLocalDepInfo &Val = I->second;
+ for (NonLocalDepInfo::const_iterator II = Val.begin(), E = Val.end();
+ II != E; ++II)
+ assert(II->second.getInst() != D && "Inst occurs as NLPD value");
+ }
+
for (NonLocalDepMapType::const_iterator I = NonLocalDeps.begin(),
E = NonLocalDeps.end(); I != E; ++I) {
assert(I->first != D && "Inst occurs in data structures");
- for (DenseMap<BasicBlock*, DepResultTy>::iterator II = I->second.begin(),
- EE = I->second.end(); II != EE; ++II)
- assert(II->second.getPointer() != D && "Inst occurs in data structures");
+ const PerInstNLInfo &INLD = I->second;
+ for (NonLocalDepInfo::const_iterator II = INLD.first.begin(),
+ EE = INLD.first.end(); II != EE; ++II)
+ assert(II->second.getInst() != D && "Inst occurs in data structures");
}
for (ReverseDepMapType::const_iterator I = ReverseLocalDeps.begin(),
- E = ReverseLocalDeps.end(); I != E; ++I)
+ E = ReverseLocalDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
EE = I->second.end(); II != EE; ++II)
assert(*II != D && "Inst occurs in data structures");
+ }
for (ReverseDepMapType::const_iterator I = ReverseNonLocalDeps.begin(),
E = ReverseNonLocalDeps.end();
- I != E; ++I)
+ I != E; ++I) {
+ assert(I->first != D && "Inst occurs in data structures");
for (SmallPtrSet<Instruction*, 4>::const_iterator II = I->second.begin(),
EE = I->second.end(); II != EE; ++II)
assert(*II != D && "Inst occurs in data structures");
+ }
+
+ for (ReverseNonLocalPtrDepTy::const_iterator
+ I = ReverseNonLocalPtrDeps.begin(),
+ E = ReverseNonLocalPtrDeps.end(); I != E; ++I) {
+ assert(I->first != D && "Inst occurs in rev NLPD map");
+
+ for (SmallPtrSet<void*, 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() &&
+ "Inst occurs in ReverseNonLocalPtrDeps map");
+ }
+
}