/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
/// B to be a memcpy from X to Z (or potentially a memmove, depending on
/// circumstances). This allows later passes to remove the first memcpy
/// altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
// We can only optimize statically-sized memcpy's that are non-volatile.
ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
if (CopySize == 0 || M->isVolatile()) return false;
// If the source and destination of the memcpy are the same, then zap it.
if (M->getSource() == M->getDest()) {
MD->removeInstruction(M);
M->eraseFromParent();
return false;
}
// The are two possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
MemDepResult DepInfo = MD->getDependency(M);
if (!DepInfo.isClobber())
return false;
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
CopySize->getZExtValue(), C)) {
M->eraseFromParent();
return true;
}
}
return false;
}
/// HandleFree - Handle frees of entire structures whose dependency is a store
/// to a field of that structure.
bool DSE::HandleFree(CallInst *F) {
MemDepResult Dep = MD->getDependency(F);
do {
if (Dep.isNonLocal()) return false;
Instruction *Dependency = Dep.getInst();
if (!hasMemoryWrite(Dependency) || !isRemovable(Dependency))
return false;
Value *DepPointer =
getStoredPointerOperand(Dependency)->getUnderlyingObject();
// Check for aliasing.
if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
return false;
// DCE instructions only used to calculate that store
DeleteDeadInstruction(Dependency, *MD);
++NumFastStores;
// Inst's old Dependency is now deleted. Compute the next dependency,
// which may also be dead, as in
// s[0] = 0;
// s[1] = 0; // This has just been deleted.
// free(s);
Dep = MD->getDependency(F);
} while (!Dep.isNonLocal());
return true;
}
/// Handle frees of entire structures whose dependency is a store
/// to a field of that structure.
static bool handleFree(CallInst *F, AliasAnalysis *AA,
MemoryDependenceResults *MD, DominatorTree *DT,
const TargetLibraryInfo *TLI,
InstOverlapIntervalsTy &IOL,
DenseMap<Instruction*, size_t> *InstrOrdering) {
bool MadeChange = false;
MemoryLocation Loc = MemoryLocation(F->getOperand(0));
SmallVector<BasicBlock *, 16> Blocks;
Blocks.push_back(F->getParent());
const DataLayout &DL = F->getModule()->getDataLayout();
while (!Blocks.empty()) {
BasicBlock *BB = Blocks.pop_back_val();
Instruction *InstPt = BB->getTerminator();
if (BB == F->getParent()) InstPt = F;
MemDepResult Dep =
MD->getPointerDependencyFrom(Loc, false, InstPt->getIterator(), BB);
while (Dep.isDef() || Dep.isClobber()) {
Instruction *Dependency = Dep.getInst();
if (!hasMemoryWrite(Dependency, *TLI) || !isRemovable(Dependency))
break;
Value *DepPointer =
GetUnderlyingObject(getStoredPointerOperand(Dependency), DL);
// Check for aliasing.
if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
break;
DEBUG(dbgs() << "DSE: Dead Store to soon to be freed memory:\n DEAD: "
<< *Dependency << '\n');
// DCE instructions only used to calculate that store.
BasicBlock::iterator BBI(Dependency);
deleteDeadInstruction(Dependency, &BBI, *MD, *TLI, IOL, InstrOrdering);
++NumFastStores;
MadeChange = true;
// Inst's old Dependency is now deleted. Compute the next dependency,
// which may also be dead, as in
// s[0] = 0;
// s[1] = 0; // This has just been deleted.
// free(s);
Dep = MD->getPointerDependencyFrom(Loc, false, BBI, BB);
}
if (Dep.isNonLocal())
findUnconditionalPreds(Blocks, BB, DT);
}
return MadeChange;
}
DataDependence::DepInfo DataDependence::getDepInfo(MemDepResult dep) {
if (dep.isClobber())
return DepInfo(dep.getInst(), Clobber);
if (dep.isDef())
return DepInfo(dep.getInst(), Def);
if (dep.isNonFuncLocal())
return DepInfo(dep.getInst(), NonFuncLocal);
if (dep.isUnknown())
return DepInfo(dep.getInst(), Unknown);
if (dep.isNonLocal())
return DepInfo(dep.getInst(), NonLocal);
llvm_unreachable("unknown dependence type");
}
/// We've found that the (upward scanning) memory dependence of \p MemCpy is
/// \p MemSet. Try to simplify \p MemSet to only set the trailing bytes that
/// weren't copied over by \p MemCpy.
///
/// In other words, transform:
/// \code
/// memset(dst, c, dst_size);
/// memcpy(dst, src, src_size);
/// \endcode
/// into:
/// \code
/// memcpy(dst, src, src_size);
/// memset(dst + src_size, c, dst_size <= src_size ? 0 : dst_size - src_size);
/// \endcode
bool MemCpyOpt::processMemSetMemCpyDependence(MemCpyInst *MemCpy,
MemSetInst *MemSet) {
// We can only transform memset/memcpy with the same destination.
if (MemSet->getDest() != MemCpy->getDest())
return false;
// Check that there are no other dependencies on the memset destination.
MemDepResult DstDepInfo = MD->getPointerDependencyFrom(
MemoryLocation::getForDest(MemSet), false, MemCpy, MemCpy->getParent());
if (DstDepInfo.getInst() != MemSet)
return false;
// Use the same i8* dest as the memcpy, killing the memset dest if different.
Value *Dest = MemCpy->getRawDest();
Value *DestSize = MemSet->getLength();
Value *SrcSize = MemCpy->getLength();
// By default, create an unaligned memset.
unsigned Align = 1;
// If Dest is aligned, and SrcSize is constant, use the minimum alignment
// of the sum.
const unsigned DestAlign =
std::max(MemSet->getAlignment(), MemCpy->getAlignment());
if (DestAlign > 1)
if (ConstantInt *SrcSizeC = dyn_cast<ConstantInt>(SrcSize))
Align = MinAlign(SrcSizeC->getZExtValue(), DestAlign);
IRBuilder<> Builder(MemCpy);
// If the sizes have different types, zext the smaller one.
if (DestSize->getType() != SrcSize->getType()) {
if (DestSize->getType()->getIntegerBitWidth() >
SrcSize->getType()->getIntegerBitWidth())
SrcSize = Builder.CreateZExt(SrcSize, DestSize->getType());
else
DestSize = Builder.CreateZExt(DestSize, SrcSize->getType());
}
Value *MemsetLen =
Builder.CreateSelect(Builder.CreateICmpULE(DestSize, SrcSize),
ConstantInt::getNullValue(DestSize->getType()),
Builder.CreateSub(DestSize, SrcSize));
const DataLayout &DL = MemCpy->getModule()->getDataLayout();
Builder.CreateMemSet(Builder.CreateGEP(Dest, SrcSize), MemSet->getOperand(1),
MemsetLen, Align,
false, NULL, NULL, NULL, DL.isByteAddressable());
MD->removeInstruction(MemSet);
MemSet->eraseFromParent();
return true;
}
void DeadStoreEliminationPass::runOverwrittenDeadStoreAnalysisOnFn(Function &F) {
MDA = &getAnalysis<MemoryDependenceAnalysis>(F);
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) {
for (BasicBlock::iterator I = BB->begin(), IE = BB->end(); I != IE; ++I) {
Instruction *inst = I;
if (StoreInst* SI = dyn_cast<StoreInst>(inst)) {
Value *ptr = SI->getPointerOperand();
MemDepResult mdr = MDA->getDependency(inst);
Instruction *depInst = mdr.getInst();
if (depInst && (isa<CallInst>(depInst) || isa<InvokeInst>(depInst))) {
Function *calledFn;
if (CallInst* CI = dyn_cast<CallInst>(depInst)) {
calledFn = CI->getCalledFunction();
} else {
InvokeInst *II = dyn_cast<InvokeInst>(depInst);
calledFn = II->getCalledFunction();
}
if (!fnThatStoreOnArgs.count(calledFn)) continue;
CallSite CS(depInst);
CallSite::arg_iterator actualArgIter = CS.arg_begin();
Function::arg_iterator formalArgIter = calledFn->arg_begin();
int size = calledFn->arg_size();
std::set<Value*> storedArgs = fnThatStoreOnArgs[calledFn];
for (int i = 0; i < size; ++i, ++actualArgIter, ++formalArgIter) {
Value *formalArg = formalArgIter;
Value *actualArg = *actualArgIter;
if (ptr == actualArg && storedArgs.count(formalArg)) {
int64_t InstWriteOffset, DepWriteOffset;
DEBUG(errs() << " Verifying if store is completely overwritten.\n");
AliasAnalysis::Location Loc(ptr, getPointerSize(ptr, *AA), NULL);
AliasAnalysis::Location DepLoc(actualArg, getPointerSize(actualArg, *AA), NULL);
OverwriteResult OR = isOverwrite(Loc, DepLoc, *AA, DepWriteOffset, InstWriteOffset);
if (OR == OverwriteComplete) {
DEBUG(errs() << " Store on " << formalArg->getName() << " will be removed with cloning\n");
deadArguments[depInst].insert(formalArg);
}
}
}
if (deadArguments.count(depInst)) {
fn2Clone[calledFn].push_back(depInst);
}
}
}
}
}
}
/// HandleFree - Handle frees of entire structures whose dependency is a store
/// to a field of that structure.
bool DSE::HandleFree(CallInst *F) {
bool MadeChange = false;
MemoryLocation Loc = MemoryLocation(F->getOperand(0));
SmallVector<BasicBlock *, 16> Blocks;
Blocks.push_back(F->getParent());
const DataLayout &DL = F->getModule()->getDataLayout();
while (!Blocks.empty()) {
BasicBlock *BB = Blocks.pop_back_val();
Instruction *InstPt = BB->getTerminator();
if (BB == F->getParent()) InstPt = F;
MemDepResult Dep = MD->getPointerDependencyFrom(Loc, false, InstPt, BB);
while (Dep.isDef() || Dep.isClobber()) {
Instruction *Dependency = Dep.getInst();
if (!hasMemoryWrite(Dependency, *TLI) || !isRemovable(Dependency))
break;
Value *DepPointer =
GetUnderlyingObject(getStoredPointerOperand(Dependency), DL);
// Check for aliasing.
if (!AA->isMustAlias(F->getArgOperand(0), DepPointer))
break;
Instruction *Next = std::next(BasicBlock::iterator(Dependency));
// DCE instructions only used to calculate that store
DeleteDeadInstruction(Dependency, *MD, *TLI);
++NumFastStores;
MadeChange = true;
// Inst's old Dependency is now deleted. Compute the next dependency,
// which may also be dead, as in
// s[0] = 0;
// s[1] = 0; // This has just been deleted.
// free(s);
Dep = MD->getPointerDependencyFrom(Loc, false, Next, BB);
}
if (Dep.isNonLocal())
FindUnconditionalPreds(Blocks, BB, DT);
}
return MadeChange;
}
/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
/// B to be a memcpy from X to Z (or potentially a memmove, depending on
/// circumstances). This allows later passes to remove the first memcpy
/// altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
// We can only optimize statically-sized memcpy's that are non-volatile.
ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
if (CopySize == 0 || M->isVolatile()) return false;
// If the source and destination of the memcpy are the same, then zap it.
if (M->getSource() == M->getDest()) {
MD->removeInstruction(M);
M->eraseFromParent();
return false;
}
// If copying from a constant, try to turn the memcpy into a memset.
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
if (GV->isConstant() && GV->hasDefinitiveInitializer())
if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
IRBuilder<> Builder(M);
Builder.CreateMemSet(M->getRawDest(), ByteVal, CopySize,
M->getAlignment(), false);
MD->removeInstruction(M);
M->eraseFromParent();
++NumCpyToSet;
return true;
}
// The are two possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
MemDepResult DepInfo = MD->getDependency(M);
if (!DepInfo.isClobber())
return false;
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst()))
return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
CopySize->getZExtValue(), C)) {
MD->removeInstruction(M);
M->eraseFromParent();
return true;
}
}
return false;
}
bool MemDepPrinter::runOnFunction(Function &F) {
this->F = &F;
MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
// All this code uses non-const interfaces because MemDep is not
// const-friendly, though nothing is actually modified.
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
Instruction *Inst = &*I;
if (!Inst->mayReadFromMemory() && !Inst->mayWriteToMemory())
continue;
MemDepResult Res = MDA.getDependency(Inst);
if (!Res.isNonLocal()) {
Deps[Inst].insert(std::make_pair(getInstTypePair(Res),
static_cast<BasicBlock *>(nullptr)));
} else if (CallSite CS = cast<Value>(Inst)) {
const MemoryDependenceAnalysis::NonLocalDepInfo &NLDI =
MDA.getNonLocalCallDependency(CS);
DepSet &InstDeps = Deps[Inst];
for (MemoryDependenceAnalysis::NonLocalDepInfo::const_iterator
I = NLDI.begin(), E = NLDI.end(); I != E; ++I) {
const MemDepResult &Res = I->getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I->getBB()));
}
} else {
SmallVector<NonLocalDepResult, 4> NLDI;
assert( (isa<LoadInst>(Inst) || isa<StoreInst>(Inst) ||
isa<VAArgInst>(Inst)) && "Unknown memory instruction!");
MDA.getNonLocalPointerDependency(Inst, NLDI);
DepSet &InstDeps = Deps[Inst];
for (SmallVectorImpl<NonLocalDepResult>::const_iterator
I = NLDI.begin(), E = NLDI.end(); I != E; ++I) {
const MemDepResult &Res = I->getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I->getBB()));
}
}
}
return false;
}
bool MemDepPrinter::runOnFunction(Function &F) {
this->F = &F;
MemoryDependenceResults &MDA = getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
// All this code uses non-const interfaces because MemDep is not
// const-friendly, though nothing is actually modified.
for (auto &I : instructions(F)) {
Instruction *Inst = &I;
if (!Inst->mayReadFromMemory() && !Inst->mayWriteToMemory())
continue;
MemDepResult Res = MDA.getDependency(Inst);
if (!Res.isNonLocal()) {
Deps[Inst].insert(std::make_pair(getInstTypePair(Res),
static_cast<BasicBlock *>(nullptr)));
} else if (auto CS = CallSite(Inst)) {
const MemoryDependenceResults::NonLocalDepInfo &NLDI =
MDA.getNonLocalCallDependency(CS);
DepSet &InstDeps = Deps[Inst];
for (const NonLocalDepEntry &I : NLDI) {
const MemDepResult &Res = I.getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I.getBB()));
}
} else {
SmallVector<NonLocalDepResult, 4> NLDI;
assert( (isa<LoadInst>(Inst) || isa<StoreInst>(Inst) ||
isa<VAArgInst>(Inst)) && "Unknown memory instruction!");
MDA.getNonLocalPointerDependency(Inst, NLDI);
DepSet &InstDeps = Deps[Inst];
for (const NonLocalDepResult &I : NLDI) {
const MemDepResult &Res = I.getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I.getBB()));
}
}
}
return false;
}
/// GetNonLocalInfoForBlock - Compute the memdep value for BB with
/// Pointer/PointeeSize using either cached information in Cache or by doing a
/// lookup (which may use dirty cache info if available). If we do a lookup,
/// add the result to the cache.
MemDepResult MemoryDependenceAnalysis::
GetNonLocalInfoForBlock(Value *Pointer, uint64_t PointeeSize,
bool isLoad, BasicBlock *BB,
NonLocalDepInfo *Cache, unsigned NumSortedEntries) {
// 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() && prior(Entry)->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.
if (ExistingResult && !ExistingResult->isDirty()) {
++NumCacheNonLocalPtr;
return *ExistingResult;
}
// 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();
// Eliminating the dirty entry from 'Cache', so update the reverse info.
ValueIsLoadPair CacheKey(Pointer, isLoad);
RemoveFromReverseMap(ReverseNonLocalPtrDeps, ScanPos,
CacheKey.getOpaqueValue());
} else {
++NumUncacheNonLocalPtr;
}
// Scan the block for the dependency.
MemDepResult Dep = getPointerDependencyFrom(Pointer, PointeeSize, isLoad,
ScanPos, BB);
// 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(BB, Dep));
// If the block has a dependency (i.e. it isn't completely transparent to
// the value), remember the reverse association because we just added it
// to Cache!
if (Dep.isNonLocal())
return Dep;
// Keep the ReverseNonLocalPtrDeps map up to date so we can efficiently
// update MemDep when we remove instructions.
Instruction *Inst = Dep.getInst();
assert(Inst && "Didn't depend on anything?");
ValueIsLoadPair CacheKey(Pointer, isLoad);
ReverseNonLocalPtrDeps[Inst].insert(CacheKey.getOpaqueValue());
return Dep;
}
/// getNonLocalCallDependency - Perform a full dependency query for the
/// specified call, returning the set of blocks that the value is
/// potentially live across. The returned set of results will include a
/// "NonLocal" result for all blocks where the value is live across.
///
/// This method assumes the instruction returns a "NonLocal" dependency
/// within its own block.
///
/// This returns a reference to an internal data structure that may be
/// invalidated on the next non-local query or when an instruction is
/// removed. Clients must copy this data if they want it around longer than
/// that.
const MemoryDependenceAnalysis::NonLocalDepInfo &
MemoryDependenceAnalysis::getNonLocalCallDependency(CallSite QueryCS) {
assert(getDependency(QueryCS.getInstruction()).isNonLocal() &&
"getNonLocalCallDependency should only be used on calls with non-local deps!");
PerInstNLInfo &CacheP = NonLocalDeps[QueryCS.getInstruction()];
NonLocalDepInfo &Cache = CacheP.first;
/// DirtyBlocks - This is the set of blocks that need to be recomputed. In
/// the cached case, this can happen due to instructions being deleted etc. In
/// the uncached case, this starts out as the set of predecessors we care
/// about.
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.
for (NonLocalDepInfo::iterator I = Cache.begin(), E = Cache.end();
I != E; ++I)
if (I->second.isDirty())
DirtyBlocks.push_back(I->first);
// 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 {
// Seed DirtyBlocks with each of the preds of QueryInst's block.
BasicBlock *QueryBB = QueryCS.getInstruction()->getParent();
for (BasicBlock **PI = PredCache->GetPreds(QueryBB); *PI; ++PI)
DirtyBlocks.push_back(*PI);
NumUncacheNonLocal++;
}
// isReadonlyCall - If this is a read-only call, we can be more aggressive.
bool isReadonlyCall = AA->onlyReadsMemory(QueryCS);
SmallPtrSet<BasicBlock*, 64> Visited;
unsigned NumSortedEntries = Cache.size();
DEBUG(AssertSorted(Cache));
// Iterate while we still have blocks to update.
while (!DirtyBlocks.empty()) {
BasicBlock *DirtyBB = DirtyBlocks.back();
DirtyBlocks.pop_back();
// Already processed this block?
if (!Visited.insert(DirtyBB))
continue;
// 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)
--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 (ExistingResult) {
if (Instruction *Inst = ExistingResult->getInst()) {
ScanPos = Inst;
// We're removing QueryInst's use of Inst.
RemoveFromReverseMap(ReverseNonLocalDeps, Inst,
QueryCS.getInstruction());
}
}
// Find out if this block has a local dependency for QueryInst.
MemDepResult Dep;
if (ScanPos != DirtyBB->begin()) {
Dep = getCallSiteDependencyFrom(QueryCS, isReadonlyCall,ScanPos, DirtyBB);
} else if (DirtyBB != &DirtyBB->getParent()->getEntryBlock()) {
// No dependence found. If this is the entry block of the function, it is
// a clobber, otherwise it is non-local.
Dep = MemDepResult::getNonLocal();
} else {
Dep = MemDepResult::getClobber(ScanPos);
}
// 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));
// If the block has a dependency (i.e. it isn't completely transparent to
// 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 = Dep.getInst())
ReverseNonLocalDeps[Inst].insert(QueryCS.getInstruction());
} else {
// If the block *is* completely transparent to the load, we need to check
// the predecessors of this block. Add them to our worklist.
for (BasicBlock **PI = PredCache->GetPreds(DirtyBB); *PI; ++PI)
DirtyBlocks.push_back(*PI);
}
}
return Cache;
}
/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
/// B to be a memcpy from X to Z (or potentially a memmove, depending on
/// circumstances). This allows later passes to remove the first memcpy
/// altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
// We can only optimize non-volatile memcpy's.
if (M->isVolatile()) return false;
// If the source and destination of the memcpy are the same, then zap it.
if (M->getSource() == M->getDest()) {
MD->removeInstruction(M);
M->eraseFromParent();
return false;
}
// If copying from a constant, try to turn the memcpy into a memset.
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
if (GV->isConstant() && GV->hasDefinitiveInitializer())
if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
IRBuilder<> Builder(M);
Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
M->getAlignment(), false);
MD->removeInstruction(M);
M->eraseFromParent();
++NumCpyToSet;
return true;
}
MemDepResult DepInfo = MD->getDependency(M);
// Try to turn a partially redundant memset + memcpy into
// memcpy + smaller memset. We don't need the memcpy size for this.
if (DepInfo.isClobber())
if (MemSetInst *MDep = dyn_cast<MemSetInst>(DepInfo.getInst()))
if (processMemSetMemCpyDependence(M, MDep))
return true;
// The optimizations after this point require the memcpy size.
ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
if (!CopySize) return false;
// There are four possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
// c) memcpy from freshly alloca'd space or space that has just started its
// lifetime copies undefined data, and we can therefore eliminate the
// memcpy in favor of the data that was already at the destination.
// d) memcpy from a just-memset'd source can be turned into memset.
if (DepInfo.isClobber()) {
if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
CopySize->getZExtValue(), M->getAlignment(),
C)) {
MD->removeInstruction(M);
M->eraseFromParent();
return true;
}
}
}
AliasAnalysis::Location SrcLoc = AliasAnalysis::getLocationForSource(M);
MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true,
M, M->getParent());
if (SrcDepInfo.isClobber()) {
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
return processMemCpyMemCpyDependence(M, MDep);
} else if (SrcDepInfo.isDef()) {
Instruction *I = SrcDepInfo.getInst();
bool hasUndefContents = false;
if (isa<AllocaInst>(I)) {
hasUndefContents = true;
} else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
if (II->getIntrinsicID() == Intrinsic::lifetime_start)
if (ConstantInt *LTSize = dyn_cast<ConstantInt>(II->getArgOperand(0)))
if (LTSize->getZExtValue() >= CopySize->getZExtValue())
hasUndefContents = true;
}
if (hasUndefContents) {
MD->removeInstruction(M);
M->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
if (SrcDepInfo.isClobber())
if (MemSetInst *MDep = dyn_cast<MemSetInst>(SrcDepInfo.getInst()))
if (performMemCpyToMemSetOptzn(M, MDep)) {
MD->removeInstruction(M);
M->eraseFromParent();
++NumCpyToSet;
return true;
}
return false;
}
bool DSE::runOnBasicBlock(BasicBlock &BB) {
bool MadeChange = false;
// Do a top-down walk on the BB.
for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
Instruction *Inst = BBI++;
// Handle 'free' calls specially.
if (CallInst *F = isFreeCall(Inst, TLI)) {
MadeChange |= HandleFree(F);
continue;
}
// If we find something that writes memory, get its memory dependence.
if (!hasMemoryWrite(Inst, TLI))
continue;
MemDepResult InstDep = MD->getDependency(Inst);
// Ignore any store where we can't find a local dependence.
// FIXME: cross-block DSE would be fun. :)
if (!InstDep.isDef() && !InstDep.isClobber())
continue;
// If we're storing the same value back to a pointer that we just
// loaded from, then the store can be removed.
if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
SI->getOperand(0) == DepLoad && isRemovable(SI)) {
DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
<< "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
// DeleteDeadInstruction can delete the current instruction. Save BBI
// in case we need it.
WeakVH NextInst(BBI);
DeleteDeadInstruction(SI, *MD, TLI);
if (!NextInst) // Next instruction deleted.
BBI = BB.begin();
else if (BBI != BB.begin()) // Revisit this instruction if possible.
--BBI;
++NumFastStores;
MadeChange = true;
continue;
}
}
}
// Figure out what location is being stored to.
AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
// If we didn't get a useful location, fail.
if (!Loc.Ptr)
continue;
while (InstDep.isDef() || InstDep.isClobber()) {
// Get the memory clobbered by the instruction we depend on. MemDep will
// skip any instructions that 'Loc' clearly doesn't interact with. If we
// end up depending on a may- or must-aliased load, then we can't optimize
// away the store and we bail out. However, if we depend on on something
// that overwrites the memory location we *can* potentially optimize it.
//
// Find out what memory location the dependent instruction stores.
Instruction *DepWrite = InstDep.getInst();
AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
// If we didn't get a useful location, or if it isn't a size, bail out.
if (!DepLoc.Ptr)
break;
// If we find a write that is a) removable (i.e., non-volatile), b) is
// completely obliterated by the store to 'Loc', and c) which we know that
// 'Inst' doesn't load from, then we can remove it.
if (isRemovable(DepWrite) &&
!isPossibleSelfRead(Inst, Loc, DepWrite, *AA)) {
int64_t InstWriteOffset, DepWriteOffset;
OverwriteResult OR = isOverwrite(Loc, DepLoc, *AA,
DepWriteOffset, InstWriteOffset);
if (OR == OverwriteComplete) {
DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: "
<< *DepWrite << "\n KILLER: " << *Inst << '\n');
// Delete the store and now-dead instructions that feed it.
DeleteDeadInstruction(DepWrite, *MD, TLI);
++NumFastStores;
MadeChange = true;
// DeleteDeadInstruction can delete the current instruction in loop
// cases, reset BBI.
BBI = Inst;
if (BBI != BB.begin())
--BBI;
break;
} else if (OR == OverwriteEnd && isShortenable(DepWrite)) {
// TODO: base this on the target vector size so that if the earlier
// store was too small to get vector writes anyway then its likely
// a good idea to shorten it
// Power of 2 vector writes are probably always a bad idea to optimize
// as any store/memset/memcpy is likely using vector instructions so
// shortening it to not vector size is likely to be slower
MemIntrinsic* DepIntrinsic = cast<MemIntrinsic>(DepWrite);
unsigned DepWriteAlign = DepIntrinsic->getAlignment();
if (llvm::isPowerOf2_64(InstWriteOffset) ||
((DepWriteAlign != 0) && InstWriteOffset % DepWriteAlign == 0)) {
DEBUG(dbgs() << "DSE: Remove Dead Store:\n OW END: "
<< *DepWrite << "\n KILLER (offset "
<< InstWriteOffset << ", "
<< DepLoc.Size << ")"
<< *Inst << '\n');
Value* DepWriteLength = DepIntrinsic->getLength();
Value* TrimmedLength = ConstantInt::get(DepWriteLength->getType(),
InstWriteOffset -
DepWriteOffset);
DepIntrinsic->setLength(TrimmedLength);
MadeChange = true;
}
}
}
// If this is a may-aliased store that is clobbering the store value, we
// can keep searching past it for another must-aliased pointer that stores
// to the same location. For example, in:
// store -> P
// store -> Q
// store -> P
// we can remove the first store to P even though we don't know if P and Q
// alias.
if (DepWrite == &BB.front()) break;
// Can't look past this instruction if it might read 'Loc'.
if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
break;
InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
}
}
// If this block ends in a return, unwind, or unreachable, all allocas are
// dead at its end, which means stores to them are also dead.
if (BB.getTerminator()->getNumSuccessors() == 0)
MadeChange |= handleEndBlock(BB);
return MadeChange;
}
void DSWP::buildPDG(Loop *L) {
//Initialize PDG
for (Loop::block_iterator bi = L->getBlocks().begin(); bi != L->getBlocks().end(); bi++) {
BasicBlock *BB = *bi;
for (BasicBlock::iterator ui = BB->begin(); ui != BB->end(); ui++) {
Instruction *inst = &(*ui);
//standardlize the name for all expr
if (util.hasNewDef(inst)) {
inst->setName(util.genId());
dname[inst] = inst->getNameStr();
} else {
dname[inst] = util.genId();
}
pdg[inst] = new vector<Edge>();
rev[inst] = new vector<Edge>();
}
}
//LoopInfo &li = getAnalysis<LoopInfo>();
/*
* Memory dependency analysis
*/
MemoryDependenceAnalysis &mda = getAnalysis<MemoryDependenceAnalysis>();
for (Loop::block_iterator bi = L->getBlocks().begin(); bi != L->getBlocks().end(); bi++) {
BasicBlock *BB = *bi;
for (BasicBlock::iterator ii = BB->begin(); ii != BB->end(); ii++) {
Instruction *inst = &(*ii);
//data dependence = register dependence + memory dependence
//begin register dependence
for (Value::use_iterator ui = ii->use_begin(); ui != ii->use_end(); ui++) {
if (Instruction *user = dyn_cast<Instruction>(*ui)) {
addEdge(inst, user, REG);
}
}
//finish register dependence
//begin memory dependence
MemDepResult mdr = mda.getDependency(inst);
//TODO not sure clobbers mean!!
if (mdr.isDef()) {
Instruction *dep = mdr.getInst();
if (isa<LoadInst>(inst)) {
if (isa<StoreInst>(dep)) {
addEdge(dep, inst, DTRUE); //READ AFTER WRITE
}
}
if (isa<StoreInst>(inst)) {
if (isa<LoadInst>(dep)) {
addEdge(dep, inst, DANTI); //WRITE AFTER READ
}
if (isa<StoreInst>(dep)) {
addEdge(dep, inst, DOUT); //WRITE AFTER WRITE
}
}
//READ AFTER READ IS INSERT AFTER PDG BUILD
}
//end memory dependence
}//for ii
}//for bi
/*
* begin control dependence
*/
PostDominatorTree &pdt = getAnalysis<PostDominatorTree>();
//cout << pdt.getRootNode()->getBlock()->getNameStr() << endl;
/*
* alien code part 1
*/
LoopInfo *LI = &getAnalysis<LoopInfo>();
std::set<BranchInst*> backedgeParents;
for (Loop::block_iterator bi = L->getBlocks().begin(); bi
!= L->getBlocks().end(); bi++) {
BasicBlock *BB = *bi;
for (BasicBlock::iterator ii = BB->begin(); ii != BB->end(); ii++) {
Instruction *inst = ii;
if (BranchInst *brInst = dyn_cast<BranchInst>(inst)) {
// get the loop this instruction (and therefore basic block) belongs to
Loop *instLoop = LI->getLoopFor(BB);
bool branchesToHeader = false;
for (int i = brInst->getNumSuccessors() - 1; i >= 0
&& !branchesToHeader; i--) {
// if the branch could exit, store it
if (LI->getLoopFor(brInst->getSuccessor(i)) != instLoop) {
branchesToHeader = true;
}
}
if (branchesToHeader) {
backedgeParents.insert(brInst);
}
}
}
}
//build information for predecessor of blocks in post dominator tree
for (Function::iterator bi = func->begin(); bi != func->end(); bi++) {
BasicBlock *BB = bi;
DomTreeNode *dn = pdt.getNode(BB);
for (DomTreeNode::iterator di = dn->begin(); di != dn->end(); di++) {
BasicBlock *CB = (*di)->getBlock();
pre[CB] = BB;
}
}
//
// //add dependency within a basicblock
// for (Loop::block_iterator bi = L->getBlocks().begin(); bi != L->getBlocks().end(); bi++) {
// BasicBlock *BB = *bi;
// Instruction *pre = NULL;
// for (BasicBlock::iterator ui = BB->begin(); ui != BB->end(); ui++) {
// Instruction *inst = &(*ui);
// if (pre != NULL) {
// addEdge(pre, inst, CONTROL);
// }
// pre = inst;
// }
// }
// //the special kind of dependence need loop peeling ? I don't know whether this is needed
// for (Loop::block_iterator bi = L->getBlocks().begin(); bi != L->getBlocks().end(); bi++) {
// BasicBlock *BB = *bi;
// for (succ_iterator PI = succ_begin(BB); PI != succ_end(BB); ++PI) {
// BasicBlock *succ = *PI;
//
// checkControlDependence(BB, succ, pdt);
// }
// }
/*
* alien code part 2
*/
// add normal control dependencies
// loop through each instruction
for (Loop::block_iterator bbIter = L->block_begin(); bbIter
!= L->block_end(); ++bbIter) {
BasicBlock *bb = *bbIter;
// check the successors of this basic block
if (BranchInst *branchInst = dyn_cast<BranchInst>(bb->getTerminator())) {
if (branchInst->getNumSuccessors() > 1) {
BasicBlock * succ = branchInst->getSuccessor(0);
// if the successor is nested shallower than the current basic block, continue
if (LI->getLoopDepth(bb) < LI->getLoopDepth(succ)) {
continue;
}
// otherwise, add all instructions to graph as control dependence
while (succ != NULL && succ != bb && LI->getLoopDepth(succ)
>= LI->getLoopDepth(bb)) {
Instruction *terminator = bb->getTerminator();
for (BasicBlock::iterator succInstIter = succ->begin(); &(*succInstIter)
!= succ->getTerminator(); ++succInstIter) {
addEdge(terminator, &(*succInstIter), CONTROL);
}
if (BranchInst *succBrInst = dyn_cast<BranchInst>(succ->getTerminator())) {
if (succBrInst->getNumSuccessors() > 1) {
addEdge(terminator, succ->getTerminator(),
CONTROL);
}
}
if (BranchInst *br = dyn_cast<BranchInst>(succ->getTerminator())) {
if (br->getNumSuccessors() == 1) {
succ = br->getSuccessor(0);
} else {
succ = NULL;
}
} else {
succ = NULL;
}
}
}
}
}
/*
* alien code part 3
*/
for (std::set<BranchInst*>::iterator exitIter = backedgeParents.begin(); exitIter != backedgeParents.end(); ++exitIter) {
BranchInst *exitBranch = *exitIter;
if (exitBranch->isConditional()) {
BasicBlock *header = LI->getLoopFor(exitBranch->getParent())->getHeader();
for (BasicBlock::iterator ctrlIter = header->begin(); ctrlIter != header->end(); ++ctrlIter) {
addEdge(exitBranch, &(*ctrlIter), CONTROL);
}
}
}
//end control dependence
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (!SI->isSimple()) return false;
// Avoid merging nontemporal stores since the resulting
// memcpy/memset would not be able to preserve the nontemporal hint.
// In theory we could teach how to propagate the !nontemporal metadata to
// memset calls. However, that change would force the backend to
// conservatively expand !nontemporal memset calls back to sequences of
// store instructions (effectively undoing the merging).
if (SI->getMetadata(LLVMContext::MD_nontemporal))
return false;
const DataLayout &DL = SI->getModule()->getDataLayout();
// Load to store forwarding can be interpreted as memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (LI->isSimple() && LI->hasOneUse() &&
LI->getParent() == SI->getParent()) {
auto *T = LI->getType();
if (T->isAggregateType()) {
AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
MemoryLocation LoadLoc = MemoryLocation::get(LI);
// We use alias analysis to check if an instruction may store to
// the memory we load from in between the load and the store. If
// such an instruction is found, we store it in AI.
Instruction *AI = nullptr;
for (BasicBlock::iterator I = ++LI->getIterator(), E = SI->getIterator();
I != E; ++I) {
if (AA.getModRefInfo(&*I, LoadLoc) & MRI_Mod) {
AI = &*I;
break;
}
}
// If no aliasing instruction is found, then we can promote the
// load/store pair to a memcpy at the store loaction.
if (!AI) {
// If we load from memory that may alias the memory we store to,
// memmove must be used to preserve semantic. If not, memcpy can
// be used.
bool UseMemMove = false;
if (!AA.isNoAlias(MemoryLocation::get(SI), LoadLoc))
UseMemMove = true;
unsigned Align = findCommonAlignment(DL, SI, LI);
uint64_t Size = DL.getTypeStoreSize(T);
IRBuilder<> Builder(SI);
Instruction *M;
if (UseMemMove)
M = Builder.CreateMemMove(SI->getPointerOperand(),
LI->getPointerOperand(), Size,
Align, SI->isVolatile());
else
M = Builder.CreateMemCpy(SI->getPointerOperand(),
LI->getPointerOperand(), Size,
Align, SI->isVolatile());
DEBUG(dbgs() << "Promoting " << *LI << " to " << *SI
<< " => " << *M << "\n");
MD->removeInstruction(SI);
SI->eraseFromParent();
MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
// Make sure we do not invalidate the iterator.
BBI = M->getIterator();
return true;
}
}
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
MemDepResult ldep = MD->getDependency(LI);
CallInst *C = nullptr;
if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
C = dyn_cast<CallInst>(ldep.getInst());
if (C) {
// Check that nothing touches the dest of the "copy" between
// the call and the store.
AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
MemoryLocation StoreLoc = MemoryLocation::get(SI);
for (BasicBlock::iterator I = --SI->getIterator(), E = C->getIterator();
I != E; --I) {
if (AA.getModRefInfo(&*I, StoreLoc) != MRI_NoModRef) {
C = nullptr;
break;
}
}
}
if (C) {
bool changed = performCallSlotOptzn(
LI, SI->getPointerOperand()->stripPointerCasts(),
LI->getPointerOperand()->stripPointerCasts(),
DL.getTypeStoreSize(SI->getOperand(0)->getType()),
findCommonAlignment(DL, SI, LI), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
}
}
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
ByteVal)) {
BBI = I->getIterator(); // Don't invalidate iterator.
return true;
}
return false;
}
/// processMemCpy - perform simplication of memcpy's. If we have memcpy A which
/// copies X to Y, and memcpy B which copies Y to Z, then we can rewrite B to be
/// a memcpy from X to Z (or potentially a memmove, depending on circumstances).
/// This allows later passes to remove the first memcpy altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
MemoryDependenceAnalysis &MD = getAnalysis<MemoryDependenceAnalysis>();
// The are two possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
MemDepResult dep = MD.getDependency(M);
if (!dep.isClobber())
return false;
if (!isa<MemCpyInst>(dep.getInst())) {
if (CallInst *C = dyn_cast<CallInst>(dep.getInst()))
return performCallSlotOptzn(M, C);
return false;
}
MemCpyInst *MDep = cast<MemCpyInst>(dep.getInst());
// We can only transforms memcpy's where the dest of one is the source of the
// other
if (M->getSource() != MDep->getDest())
return false;
// Second, the length of the memcpy's must be the same, or the preceeding one
// must be larger than the following one.
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
ConstantInt *C2 = dyn_cast<ConstantInt>(M->getLength());
if (!C1 || !C2)
return false;
uint64_t DepSize = C1->getValue().getZExtValue();
uint64_t CpySize = C2->getValue().getZExtValue();
if (DepSize < CpySize)
return false;
// Finally, we have to make sure that the dest of the second does not
// alias the source of the first
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
if (AA.alias(M->getRawDest(), CpySize, MDep->getRawSource(), DepSize) !=
AliasAnalysis::NoAlias)
return false;
else if (AA.alias(M->getRawDest(), CpySize, M->getRawSource(), CpySize) !=
AliasAnalysis::NoAlias)
return false;
else if (AA.alias(MDep->getRawDest(), DepSize, MDep->getRawSource(), DepSize)
!= AliasAnalysis::NoAlias)
return false;
// If all checks passed, then we can transform these memcpy's
const Type *ArgTys[3] = { M->getRawDest()->getType(),
MDep->getRawSource()->getType(),
M->getLength()->getType() };
Function *MemCpyFun = Intrinsic::getDeclaration(
M->getParent()->getParent()->getParent(),
M->getIntrinsicID(), ArgTys, 3);
Value *Args[5] = {
M->getRawDest(), MDep->getRawSource(), M->getLength(),
M->getAlignmentCst(), M->getVolatileCst()
};
CallInst *C = CallInst::Create(MemCpyFun, Args, Args+5, "", M);
// If C and M don't interfere, then this is a valid transformation. If they
// did, this would mean that the two sources overlap, which would be bad.
if (MD.getDependency(C) == dep) {
MD.removeInstruction(M);
M->eraseFromParent();
++NumMemCpyInstr;
return true;
}
// Otherwise, there was no point in doing this, so we remove the call we
// inserted and act like nothing happened.
MD.removeInstruction(C);
C->eraseFromParent();
return false;
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (!SI->isSimple()) return false;
// Avoid merging nontemporal stores since the resulting
// memcpy/memset would not be able to preserve the nontemporal hint.
// In theory we could teach how to propagate the !nontemporal metadata to
// memset calls. However, that change would force the backend to
// conservatively expand !nontemporal memset calls back to sequences of
// store instructions (effectively undoing the merging).
if (SI->getMetadata(LLVMContext::MD_nontemporal))
return false;
const DataLayout &DL = SI->getModule()->getDataLayout();
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (LI->isSimple() && LI->hasOneUse() &&
LI->getParent() == SI->getParent()) {
MemDepResult ldep = MD->getDependency(LI);
CallInst *C = nullptr;
if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
C = dyn_cast<CallInst>(ldep.getInst());
if (C) {
// Check that nothing touches the dest of the "copy" between
// the call and the store.
AliasAnalysis &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
MemoryLocation StoreLoc = MemoryLocation::get(SI);
for (BasicBlock::iterator I = --SI->getIterator(), E = C->getIterator();
I != E; --I) {
if (AA.getModRefInfo(&*I, StoreLoc) != MRI_NoModRef) {
C = nullptr;
break;
}
}
}
if (C) {
unsigned storeAlign = SI->getAlignment();
if (!storeAlign)
storeAlign = DL.getABITypeAlignment(SI->getOperand(0)->getType());
unsigned loadAlign = LI->getAlignment();
if (!loadAlign)
loadAlign = DL.getABITypeAlignment(LI->getType());
bool changed = performCallSlotOptzn(
LI, SI->getPointerOperand()->stripPointerCasts(),
LI->getPointerOperand()->stripPointerCasts(),
DL.getTypeStoreSize(SI->getOperand(0)->getType()),
std::min(storeAlign, loadAlign), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
}
}
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
ByteVal)) {
BBI = I->getIterator(); // Don't invalidate iterator.
return true;
}
return false;
}
/// processStore - When GVN is scanning forward over instructions, we look for
/// some other patterns to fold away. In particular, this looks for stores to
/// neighboring locations of memory. If it sees enough consequtive ones
/// (currently 4) it attempts to merge them together into a memcpy/memset.
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (SI->isVolatile()) return false;
TargetData *TD = getAnalysisIfAvailable<TargetData>();
if (!TD) return false;
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (!LI->isVolatile() && LI->hasOneUse()) {
MemDepResult dep = MD->getDependency(LI);
CallInst *C = 0;
if (dep.isClobber() && !isa<MemCpyInst>(dep.getInst()))
C = dyn_cast<CallInst>(dep.getInst());
if (C) {
bool changed = performCallSlotOptzn(LI,
SI->getPointerOperand()->stripPointerCasts(),
LI->getPointerOperand()->stripPointerCasts(),
TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
}
}
LLVMContext &Context = SI->getContext();
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
Value *ByteVal = isBytewiseValue(SI->getOperand(0));
if (!ByteVal)
return false;
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
Module *M = SI->getParent()->getParent()->getParent();
// Okay, so we now have a single store that can be splatable. Scan to find
// all subsequent stores of the same value to offset from the same pointer.
// Join these together into ranges, so we can decide whether contiguous blocks
// are stored.
MemsetRanges Ranges(*TD);
Value *StartPtr = SI->getPointerOperand();
BasicBlock::iterator BI = SI;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
if (isa<CallInst>(BI) || isa<InvokeInst>(BI)) {
// If the call is readnone, ignore it, otherwise bail out. We don't even
// allow readonly here because we don't want something like:
// A[1] = 2; strlen(A); A[2] = 2; -> memcpy(A, ...); strlen(A).
if (AA.getModRefBehavior(CallSite(BI)) ==
AliasAnalysis::DoesNotAccessMemory)
continue;
// TODO: If this is a memset, try to join it in.
break;
} else if (isa<VAArgInst>(BI) || isa<LoadInst>(BI))
break;
// If this is a non-store instruction it is fine, ignore it.
StoreInst *NextStore = dyn_cast<StoreInst>(BI);
if (NextStore == 0) continue;
// If this is a store, see if we can merge it in.
if (NextStore->isVolatile()) break;
// Check to see if this stored value is of the same byte-splattable value.
if (ByteVal != isBytewiseValue(NextStore->getOperand(0)))
break;
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, NextStore->getPointerOperand(), Offset, *TD))
break;
Ranges.addStore(Offset, NextStore);
}
// If we have no ranges, then we just had a single store with nothing that
// could be merged in. This is a very common case of course.
if (Ranges.empty())
return false;
// If we had at least one store that could be merged in, add the starting
// store as well. We try to avoid this unless there is at least something
// interesting as a small compile-time optimization.
Ranges.addStore(0, SI);
// Now that we have full information about ranges, loop over the ranges and
// emit memset's for anything big enough to be worthwhile.
bool MadeChange = false;
for (MemsetRanges::const_iterator I = Ranges.begin(), E = Ranges.end();
I != E; ++I) {
const MemsetRange &Range = *I;
if (Range.TheStores.size() == 1) continue;
// If it is profitable to lower this range to memset, do so now.
if (!Range.isProfitableToUseMemset(*TD))
continue;
// Otherwise, we do want to transform this! Create a new memset. We put
// the memset right before the first instruction that isn't part of this
// memset block. This ensure that the memset is dominated by any addressing
// instruction needed by the start of the block.
BasicBlock::iterator InsertPt = BI;
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
const Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
// Cast the start ptr to be i8* as memset requires.
const PointerType* StartPTy = cast<PointerType>(StartPtr->getType());
const PointerType *i8Ptr = Type::getInt8PtrTy(Context,
StartPTy->getAddressSpace());
if (StartPTy!= i8Ptr)
StartPtr = new BitCastInst(StartPtr, i8Ptr, StartPtr->getName(),
InsertPt);
Value *Ops[] = {
StartPtr, ByteVal, // Start, value
// size
ConstantInt::get(Type::getInt64Ty(Context), Range.End-Range.Start),
// align
ConstantInt::get(Type::getInt32Ty(Context), Alignment),
// volatile
ConstantInt::getFalse(Context),
};
const Type *Tys[] = { Ops[0]->getType(), Ops[2]->getType() };
Function *MemSetF = Intrinsic::getDeclaration(M, Intrinsic::memset, Tys, 2);
Value *C = CallInst::Create(MemSetF, Ops, Ops+5, "", InsertPt);
DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
dbgs() << "With: " << *C << '\n'); C=C;
// Don't invalidate the iterator
BBI = BI;
// Zap all the stores.
for (SmallVector<StoreInst*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI)
(*SI)->eraseFromParent();
++NumMemSetInfer;
MadeChange = true;
}
return MadeChange;
}
/// processMemCpyMemCpyDependence - We've found that the (upward scanning)
/// memory dependence of memcpy 'M' is the memcpy 'MDep'. Try to simplify M to
/// copy from MDep's input if we can. MSize is the size of M's copy.
///
bool MemCpyOpt::processMemCpyMemCpyDependence(MemCpyInst *M, MemCpyInst *MDep,
uint64_t MSize) {
// We can only transforms memcpy's where the dest of one is the source of the
// other.
if (M->getSource() != MDep->getDest() || MDep->isVolatile())
return false;
// If dep instruction is reading from our current input, then it is a noop
// transfer and substituting the input won't change this instruction. Just
// ignore the input and let someone else zap MDep. This handles cases like:
// memcpy(a <- a)
// memcpy(b <- a)
if (M->getSource() == MDep->getSource())
return false;
// Second, the length of the memcpy's must be the same, or the preceding one
// must be larger than the following one.
ConstantInt *MDepLen = dyn_cast<ConstantInt>(MDep->getLength());
ConstantInt *MLen = dyn_cast<ConstantInt>(M->getLength());
if (!MDepLen || !MLen || MDepLen->getZExtValue() < MLen->getZExtValue())
return false;
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
// Verify that the copied-from memory doesn't change in between the two
// transfers. For example, in:
// memcpy(a <- b)
// *b = 42;
// memcpy(c <- a)
// It would be invalid to transform the second memcpy into memcpy(c <- b).
//
// TODO: If the code between M and MDep is transparent to the destination "c",
// then we could still perform the xform by moving M up to the first memcpy.
//
// NOTE: This is conservative, it will stop on any read from the source loc,
// not just the defining memcpy.
MemDepResult SourceDep =
MD->getPointerDependencyFrom(AA.getLocationForSource(MDep),
false, M, M->getParent());
if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
return false;
// If the dest of the second might alias the source of the first, then the
// source and dest might overlap. We still want to eliminate the intermediate
// value, but we have to generate a memmove instead of memcpy.
bool UseMemMove = false;
if (!AA.isNoAlias(AA.getLocationForDest(M), AA.getLocationForSource(MDep)))
UseMemMove = true;
// If all checks passed, then we can transform M.
// Make sure to use the lesser of the alignment of the source and the dest
// since we're changing where we're reading from, but don't want to increase
// the alignment past what can be read from or written to.
// TODO: Is this worth it if we're creating a less aligned memcpy? For
// example we could be moving from movaps -> movq on x86.
unsigned Align = std::min(MDep->getAlignment(), M->getAlignment());
IRBuilder<> Builder(M);
if (UseMemMove)
Builder.CreateMemMove(M->getRawDest(), MDep->getRawSource(), M->getLength(),
Align, M->isVolatile());
else
Builder.CreateMemCpy(M->getRawDest(), MDep->getRawSource(), M->getLength(),
Align, M->isVolatile());
// Remove the instruction we're replacing.
MD->removeInstruction(M);
M->eraseFromParent();
++NumMemCpyInstr;
return true;
}
bool DSE::runOnBasicBlock(BasicBlock &BB) {
bool MadeChange = false;
// Do a top-down walk on the BB.
for (BasicBlock::iterator BBI = BB.begin(), BBE = BB.end(); BBI != BBE; ) {
Instruction *Inst = BBI++;
// Handle 'free' calls specially.
if (CallInst *F = isFreeCall(Inst)) {
MadeChange |= HandleFree(F);
continue;
}
// If we find something that writes memory, get its memory dependence.
if (!hasMemoryWrite(Inst))
continue;
MemDepResult InstDep = MD->getDependency(Inst);
// Ignore non-local store liveness.
// FIXME: cross-block DSE would be fun. :)
if (InstDep.isNonLocal() ||
// Ignore self dependence, which happens in the entry block of the
// function.
InstDep.getInst() == Inst)
continue;
// If we're storing the same value back to a pointer that we just
// loaded from, then the store can be removed.
if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
if (LoadInst *DepLoad = dyn_cast<LoadInst>(InstDep.getInst())) {
if (SI->getPointerOperand() == DepLoad->getPointerOperand() &&
SI->getOperand(0) == DepLoad && !SI->isVolatile()) {
DEBUG(dbgs() << "DSE: Remove Store Of Load from same pointer:\n "
<< "LOAD: " << *DepLoad << "\n STORE: " << *SI << '\n');
// DeleteDeadInstruction can delete the current instruction. Save BBI
// in case we need it.
WeakVH NextInst(BBI);
DeleteDeadInstruction(SI, *MD);
if (NextInst == 0) // Next instruction deleted.
BBI = BB.begin();
else if (BBI != BB.begin()) // Revisit this instruction if possible.
--BBI;
++NumFastStores;
MadeChange = true;
continue;
}
}
}
// Figure out what location is being stored to.
AliasAnalysis::Location Loc = getLocForWrite(Inst, *AA);
// If we didn't get a useful location, fail.
if (Loc.Ptr == 0)
continue;
while (!InstDep.isNonLocal()) {
// Get the memory clobbered by the instruction we depend on. MemDep will
// skip any instructions that 'Loc' clearly doesn't interact with. If we
// end up depending on a may- or must-aliased load, then we can't optimize
// away the store and we bail out. However, if we depend on on something
// that overwrites the memory location we *can* potentially optimize it.
//
// Find out what memory location the dependant instruction stores.
Instruction *DepWrite = InstDep.getInst();
AliasAnalysis::Location DepLoc = getLocForWrite(DepWrite, *AA);
// If we didn't get a useful location, or if it isn't a size, bail out.
if (DepLoc.Ptr == 0)
break;
// If we find a write that is a) removable (i.e., non-volatile), b) is
// completely obliterated by the store to 'Loc', and c) which we know that
// 'Inst' doesn't load from, then we can remove it.
if (isRemovable(DepWrite) && isCompleteOverwrite(Loc, DepLoc, *AA) &&
!isPossibleSelfRead(Inst, Loc, DepWrite, *AA)) {
DEBUG(dbgs() << "DSE: Remove Dead Store:\n DEAD: "
<< *DepWrite << "\n KILLER: " << *Inst << '\n');
// Delete the store and now-dead instructions that feed it.
DeleteDeadInstruction(DepWrite, *MD);
++NumFastStores;
MadeChange = true;
// DeleteDeadInstruction can delete the current instruction in loop
// cases, reset BBI.
BBI = Inst;
if (BBI != BB.begin())
--BBI;
break;
}
// If this is a may-aliased store that is clobbering the store value, we
// can keep searching past it for another must-aliased pointer that stores
// to the same location. For example, in:
// store -> P
// store -> Q
// store -> P
// we can remove the first store to P even though we don't know if P and Q
// alias.
if (DepWrite == &BB.front()) break;
// Can't look past this instruction if it might read 'Loc'.
if (AA->getModRefInfo(DepWrite, Loc) & AliasAnalysis::Ref)
break;
InstDep = MD->getPointerDependencyFrom(Loc, false, DepWrite, &BB);
}
}
// If this block ends in a return, unwind, or unreachable, all allocas are
// dead at its end, which means stores to them are also dead.
if (BB.getTerminator()->getNumSuccessors() == 0)
MadeChange |= handleEndBlock(BB);
return MadeChange;
}
bool MemDepPrinter::runOnFunction(Function &F) {
this->F = &F;
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
MemoryDependenceAnalysis &MDA = getAnalysis<MemoryDependenceAnalysis>();
// All this code uses non-const interfaces because MemDep is not
// const-friendly, though nothing is actually modified.
for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) {
Instruction *Inst = &*I;
if (!Inst->mayReadFromMemory() && !Inst->mayWriteToMemory())
continue;
MemDepResult Res = MDA.getDependency(Inst);
if (!Res.isNonLocal()) {
Deps[Inst].insert(std::make_pair(getInstTypePair(Res),
static_cast<BasicBlock *>(nullptr)));
} else if (CallSite CS = cast<Value>(Inst)) {
const MemoryDependenceAnalysis::NonLocalDepInfo &NLDI =
MDA.getNonLocalCallDependency(CS);
DepSet &InstDeps = Deps[Inst];
for (MemoryDependenceAnalysis::NonLocalDepInfo::const_iterator
I = NLDI.begin(), E = NLDI.end(); I != E; ++I) {
const MemDepResult &Res = I->getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I->getBB()));
}
} else {
SmallVector<NonLocalDepResult, 4> NLDI;
if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) {
if (!LI->isUnordered()) {
// FIXME: Handle atomic/volatile loads.
Deps[Inst].insert(std::make_pair(getInstTypePair(nullptr, Unknown),
static_cast<BasicBlock *>(nullptr)));
continue;
}
AliasAnalysis::Location Loc = AA.getLocation(LI);
MDA.getNonLocalPointerDependency(Loc, true, LI->getParent(), NLDI);
} else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) {
if (!SI->isUnordered()) {
// FIXME: Handle atomic/volatile stores.
Deps[Inst].insert(std::make_pair(getInstTypePair(nullptr, Unknown),
static_cast<BasicBlock *>(nullptr)));
continue;
}
AliasAnalysis::Location Loc = AA.getLocation(SI);
MDA.getNonLocalPointerDependency(Loc, false, SI->getParent(), NLDI);
} else if (VAArgInst *VI = dyn_cast<VAArgInst>(Inst)) {
AliasAnalysis::Location Loc = AA.getLocation(VI);
MDA.getNonLocalPointerDependency(Loc, false, VI->getParent(), NLDI);
} else {
llvm_unreachable("Unknown memory instruction!");
}
DepSet &InstDeps = Deps[Inst];
for (SmallVectorImpl<NonLocalDepResult>::const_iterator
I = NLDI.begin(), E = NLDI.end(); I != E; ++I) {
const MemDepResult &Res = I->getResult();
InstDeps.insert(std::make_pair(getInstTypePair(Res), I->getBB()));
}
}
}
return false;
}
/// getNonLocalPointerDepFromBB - Perform a dependency query based on
/// pointer/pointeesize starting at the end of StartBB. Add any clobber/def
/// results to the results vector and keep track of which blocks are visited in
/// 'Visited'.
///
/// This has special behavior for the first block queries (when SkipFirstBlock
/// is true). In this special case, it ignores the contents of the specified
/// block and starts returning dependence info for its predecessors.
///
/// This function returns false on success, or true to indicate that it could
/// not compute dependence information for some reason. This should be treated
/// as a clobber dependence on the first instruction in the predecessor block.
bool MemoryDependenceAnalysis::
getNonLocalPointerDepFromBB(Value *Pointer, uint64_t PointeeSize,
bool isLoad, BasicBlock *StartBB,
SmallVectorImpl<NonLocalDepEntry> &Result,
DenseMap<BasicBlock*, Value*> &Visited,
bool SkipFirstBlock) {
// Look up the cached info for Pointer.
ValueIsLoadPair CacheKey(Pointer, isLoad);
std::pair<BBSkipFirstBlockPair, NonLocalDepInfo> *CacheInfo =
&NonLocalPointerDeps[CacheKey];
NonLocalDepInfo *Cache = &CacheInfo->second;
// 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->first);
if (VI == Visited.end() || VI->second == Pointer) 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;
}
}
for (NonLocalDepInfo::iterator I = Cache->begin(), E = Cache->end();
I != E; ++I) {
Visited.insert(std::make_pair(I->first, Pointer));
if (!I->second.isNonLocal())
Result.push_back(*I);
}
++NumCacheCompleteNonLocalPtr;
return false;
}
// Otherwise, either this is a new block, a block with an invalid cache
// pointer or one that we're about to invalidate by putting more info into it
// than its valid cache info. If empty, the result will be valid cache info,
// otherwise it isn't.
if (Cache->empty())
CacheInfo->first = BBSkipFirstBlockPair(StartBB, SkipFirstBlock);
else
CacheInfo->first = BBSkipFirstBlockPair();
SmallVector<BasicBlock*, 32> Worklist;
Worklist.push_back(StartBB);
// 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();
DEBUG(AssertSorted(*Cache));
while (!Worklist.empty()) {
BasicBlock *BB = Worklist.pop_back_val();
// Skip the first block if we have it.
if (!SkipFirstBlock) {
// Analyze the dependency of *Pointer in FromBB. See if we already have
// been here.
assert(Visited.count(BB) && "Should check 'visited' before adding to WL");
// Get the dependency info for Pointer in BB. If we have cached
// information, we will use it, otherwise we compute it.
DEBUG(AssertSorted(*Cache, NumSortedEntries));
MemDepResult Dep = GetNonLocalInfoForBlock(Pointer, 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));
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) {
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));
if (InsertRes.second) {
// First time we've looked at *PI.
Worklist.push_back(*PI);
continue;
}
// If we have seen this block before, but it was with a different
// pointer then we have a phi translation failure and we have to treat
// this as a clobber.
if (InsertRes.first->second != Pointer)
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.
// 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));
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;
}
// 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 that could reuse the cache
// value will only see properly sorted cache arrays.
if (Cache && NumSortedEntries != Cache->size())
std::sort(Cache->begin(), Cache->end());
Cache = 0;
// FIXME: it is entirely possible that PHI translating will end up with
// the same value. Consider PHI translating something like:
// X = phi [x, bb1], [y, bb2]. PHI translating for bb1 doesn't *need*
// to recurse here, pedantically speaking.
// If we have a problem phi translating, fall through to the code below
// to handle the failure condition.
if (getNonLocalPointerDepFromBB(PredPtr, PointeeSize, isLoad, Pred,
Result, Visited))
goto 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;
}
// 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:
if (Cache == 0) {
// Refresh the CacheInfo/Cache pointer if it got invalidated.
CacheInfo = &NonLocalPointerDeps[CacheKey];
Cache = &CacheInfo->second;
NumSortedEntries = Cache->size();
} else if (NumSortedEntries != Cache->size()) {
std::sort(Cache->begin(), Cache->end());
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();
// If *nothing* works, mark the pointer as being clobbered by the first
// instruction in this block.
//
// If this is the magic first block, return this as a clobber of the whole
// incoming value. Since we can't phi translate to one of the predecessors,
// we have to bail out.
if (SkipFirstBlock)
return true;
for (NonLocalDepInfo::reverse_iterator I = Cache->rbegin(); ; ++I) {
assert(I != Cache->rend() && "Didn't find current block??");
if (I->first != BB)
continue;
assert(I->second.isNonLocal() &&
"Should only be here with transparent block");
I->second = MemDepResult::getClobber(BB->begin());
ReverseNonLocalPtrDeps[BB->begin()].insert(CacheKey.getOpaqueValue());
Result.push_back(*I);
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());
}
DEBUG(AssertSorted(*Cache));
return false;
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (SI->isVolatile()) return false;
if (TD == 0) return false;
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (!LI->isVolatile() && LI->hasOneUse()) {
MemDepResult ldep = MD->getDependency(LI);
CallInst *C = 0;
if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
C = dyn_cast<CallInst>(ldep.getInst());
if (C) {
// Check that nothing touches the dest of the "copy" between
// the call and the store.
MemDepResult sdep = MD->getDependency(SI);
if (!sdep.isNonLocal()) {
bool FoundCall = false;
for (BasicBlock::iterator I = SI, E = sdep.getInst(); I != E; --I) {
if (&*I == C) {
FoundCall = true;
break;
}
}
if (!FoundCall)
C = 0;
}
}
if (C) {
bool changed = performCallSlotOptzn(LI,
SI->getPointerOperand()->stripPointerCasts(),
LI->getPointerOperand()->stripPointerCasts(),
TD->getTypeStoreSize(SI->getOperand(0)->getType()), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
}
}
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
ByteVal)) {
BBI = I; // Don't invalidate iterator.
return true;
}
return false;
}
/// 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.
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())
RemoveFromReverseMap(ReverseNonLocalDeps, Inst, 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.getInst())
RemoveFromReverseMap(ReverseLocalDeps, Inst, RemInst);
// 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.
//
SmallVector<std::pair<Instruction*, Instruction*>, 8> ReverseDepsToAdd;
// If we find RemInst as a clobber or Def in any of the maps for other values,
// we need to replace its entry with a dirty version of the instruction after
// it. If RemInst is a terminator, we use a null dirty value.
//
// Using a dirty version of the instruction after RemInst saves having to scan
// the entire block to get to this point.
MemDepResult NewDirtyVal;
if (!RemInst->isTerminator())
NewDirtyVal = MemDepResult::getDirty(++BasicBlock::iterator(RemInst));
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 local stuff depending on it.
assert(!ReverseDeps.empty() && !isa<TerminatorInst>(RemInst) &&
"Nothing can locally depend on a terminator");
for (SmallPtrSet<Instruction*, 4>::iterator I = ReverseDeps.begin(),
E = ReverseDeps.end(); I != E; ++I) {
Instruction *InstDependingOnRemInst = *I;
assert(InstDependingOnRemInst != RemInst &&
"Already removed our local dep info");
LocalDeps[InstDependingOnRemInst] = NewDirtyVal;
// Make sure to remember that new things depend on NewDepInst.
assert(NewDirtyVal.getInst() && "There is no way something else can have "
"a local dep on this if it is a terminator!");
ReverseDepsToAdd.push_back(std::make_pair(NewDirtyVal.getInst(),
InstDependingOnRemInst));
}
ReverseLocalDeps.erase(ReverseDepIt);
// Add new reverse deps after scanning the set, to avoid invalidating the
// 'ReverseDeps' reference.
while (!ReverseDepsToAdd.empty()) {
ReverseLocalDeps[ReverseDepsToAdd.back().first]
.insert(ReverseDepsToAdd.back().second);
ReverseDepsToAdd.pop_back();
}
}
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) {
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.
DI->second = NewDirtyVal;
if (Instruction *NextI = NewDirtyVal.getInst())
ReverseDepsToAdd.push_back(std::make_pair(NextI, *I));
}
}
ReverseNonLocalDeps.erase(ReverseDepIt);
// Add new reverse deps after scanning the set, to avoid invalidating 'Set'
while (!ReverseDepsToAdd.empty()) {
ReverseNonLocalDeps[ReverseDepsToAdd.back().first]
.insert(ReverseDepsToAdd.back().second);
ReverseDepsToAdd.pop_back();
}
}
// 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].second;
// The cache is not valid for any specific block anymore.
NonLocalPointerDeps[P].first = BBSkipFirstBlockPair();
// Update any entries for RemInst to use the instruction after it.
for (NonLocalDepInfo::iterator DI = NLPDI.begin(), DE = NLPDI.end();
DI != DE; ++DI) {
if (DI->second.getInst() != RemInst) continue;
// 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));
}
// 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());
ReversePtrDepsToAdd.pop_back();
}
}
assert(!NonLocalDeps.count(RemInst) && "RemInst got reinserted?");
AA->deleteValue(RemInst);
DEBUG(verifyRemoved(RemInst));
}
void DataDependence::processDepResult(Instruction *inst,
MemoryDependenceAnalysis &MDA, AliasAnalysis &AA) {
// TODO: This is probably a good place to check of the dependency
// information is calculated on-demand
MemDepResult Res = MDA.getDependency(inst);
if (!Res.isNonLocal()) {
// local results (not-non-local) can be simply handled. They are just
// a pair of insturctions and a dependency type
// Get dependency information
DepInfo newInfo;
newInfo = getDepInfo(Res);
#ifdef MK_DEBUG
//errs() << "[DEBUG] newInfo depInst == " << Res.getInst() << '\n';
if (Res.getInst() == NULL) {
errs() << "[DEBUG] NULL dependency found, dep type: "
<< depTypeToString(newInfo.Type_) << '\n';
}
#endif
// Save into map
assert(newInfo.valid());
LocalDeps_[inst] = newInfo;
}
else {
// Handle NonLocal dependencies. The function call
// getNonLocalPointerDependency() assumes that a result of NonLocal
// has already been encountered
// Get dependency information
DepInfo newInfo;
newInfo = getDepInfo(Res);
assert(newInfo.Type_ == NonLocal);
assert(Res.isNonLocal());
SmallVector<NonLocalDepResult, 4> NLDep;
if (LoadInst *LI = dyn_cast<LoadInst>(inst)) {
if (!LI->isUnordered()) {
// FIXME: Handle atomic/volatile loads.
errs() << "[WARNING] atomic/volatile loads are not handled\n";
assert(false && "atomic/volatile loads not handled");
//Deps[Inst].insert(std::make_pair(getInstTypePair(0, Unknown),
//static_cast<BasicBlock *>(0)));
return;
}
AliasAnalysis::Location Loc = AA.getLocation(LI);
MDA.getNonLocalPointerDependency(Loc, true, LI->getParent(), NLDep);
}
else if (StoreInst *SI = dyn_cast<StoreInst>(inst)) {
if (!SI->isUnordered()) {
// FIXME: Handle atomic/volatile stores.
errs() << "[WARNING] atomic/volatile stores are not handled\n";
assert(false && "atomic/volatile stores not handled");
//Deps[Inst].insert(std::make_pair(getInstTypePair(0, Unknown),
//static_cast<BasicBlock *>(0)));
return;
}
AliasAnalysis::Location Loc = AA.getLocation(SI);
MDA.getNonLocalPointerDependency(Loc, false, SI->getParent(),
NLDep);
}
else if (VAArgInst *VI = dyn_cast<VAArgInst>(inst)) {
AliasAnalysis::Location Loc = AA.getLocation(VI);
MDA.getNonLocalPointerDependency(Loc, false, VI->getParent(),
NLDep);
}
else {
llvm_unreachable("Unknown memory instruction!");
}
#ifdef MK_DEBUG
errs() << "[DEBUG] NLDep.size() == " << NLDep.size() << '\n';
#endif
for (auto I = NLDep.begin(), E = NLDep.end(); I != E; ++I) {
NonLocalDeps_[inst].push_back(*I);
}
} // end else
}
bool MemCpyOpt::processStore(StoreInst *SI, BasicBlock::iterator &BBI) {
if (!SI->isSimple()) return false;
if (TD == 0) return false;
// Detect cases where we're performing call slot forwarding, but
// happen to be using a load-store pair to implement it, rather than
// a memcpy.
if (LoadInst *LI = dyn_cast<LoadInst>(SI->getOperand(0))) {
if (LI->isSimple() && LI->hasOneUse() &&
LI->getParent() == SI->getParent()) {
MemDepResult ldep = MD->getDependency(LI);
CallInst *C = 0;
if (ldep.isClobber() && !isa<MemCpyInst>(ldep.getInst()))
C = dyn_cast<CallInst>(ldep.getInst());
if (C) {
// Check that nothing touches the dest of the "copy" between
// the call and the store.
AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
AliasAnalysis::Location StoreLoc = AA.getLocation(SI);
for (BasicBlock::iterator I = --BasicBlock::iterator(SI),
E = C; I != E; --I) {
if (AA.getModRefInfo(&*I, StoreLoc) != AliasAnalysis::NoModRef) {
C = 0;
break;
}
}
}
if (C) {
unsigned storeAlign = SI->getAlignment();
if (!storeAlign)
storeAlign = TD->getABITypeAlignment(SI->getOperand(0)->getType());
unsigned loadAlign = LI->getAlignment();
if (!loadAlign)
loadAlign = TD->getABITypeAlignment(LI->getType());
bool changed = performCallSlotOptzn(LI,
SI->getPointerOperand()->stripPointerCasts(),
LI->getPointerOperand()->stripPointerCasts(),
TD->getTypeStoreSize(SI->getOperand(0)->getType()),
std::min(storeAlign, loadAlign), C);
if (changed) {
MD->removeInstruction(SI);
SI->eraseFromParent();
MD->removeInstruction(LI);
LI->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
}
}
// There are two cases that are interesting for this code to handle: memcpy
// and memset. Right now we only handle memset.
// Ensure that the value being stored is something that can be memset'able a
// byte at a time like "0" or "-1" or any width, as well as things like
// 0xA0A0A0A0 and 0.0.
if (Value *ByteVal = isBytewiseValue(SI->getOperand(0)))
if (Instruction *I = tryMergingIntoMemset(SI, SI->getPointerOperand(),
ByteVal)) {
BBI = I; // Don't invalidate iterator.
return true;
}
return false;
}
/// processMemCpy - perform simplification of memcpy's. If we have memcpy A
/// which copies X to Y, and memcpy B which copies Y to Z, then we can rewrite
/// B to be a memcpy from X to Z (or potentially a memmove, depending on
/// circumstances). This allows later passes to remove the first memcpy
/// altogether.
bool MemCpyOpt::processMemCpy(MemCpyInst *M) {
// We can only optimize non-volatile memcpy's.
if (M->isVolatile()) return false;
// If the source and destination of the memcpy are the same, then zap it.
if (M->getSource() == M->getDest()) {
MD->removeInstruction(M);
M->eraseFromParent();
return false;
}
// If copying from a constant, try to turn the memcpy into a memset.
if (GlobalVariable *GV = dyn_cast<GlobalVariable>(M->getSource()))
if (GV->isConstant() && GV->hasDefinitiveInitializer())
if (Value *ByteVal = isBytewiseValue(GV->getInitializer())) {
IRBuilder<> Builder(M);
Builder.CreateMemSet(M->getRawDest(), ByteVal, M->getLength(),
M->getAlignment(), false);
MD->removeInstruction(M);
M->eraseFromParent();
++NumCpyToSet;
return true;
}
// The optimizations after this point require the memcpy size.
ConstantInt *CopySize = dyn_cast<ConstantInt>(M->getLength());
if (CopySize == 0) return false;
// The are three possible optimizations we can do for memcpy:
// a) memcpy-memcpy xform which exposes redundance for DSE.
// b) call-memcpy xform for return slot optimization.
// c) memcpy from freshly alloca'd space copies undefined data, and we can
// therefore eliminate the memcpy in favor of the data that was already
// at the destination.
MemDepResult DepInfo = MD->getDependency(M);
if (DepInfo.isClobber()) {
if (CallInst *C = dyn_cast<CallInst>(DepInfo.getInst())) {
if (performCallSlotOptzn(M, M->getDest(), M->getSource(),
CopySize->getZExtValue(), M->getAlignment(),
C)) {
MD->removeInstruction(M);
M->eraseFromParent();
return true;
}
}
}
AliasAnalysis::Location SrcLoc = AliasAnalysis::getLocationForSource(M);
MemDepResult SrcDepInfo = MD->getPointerDependencyFrom(SrcLoc, true,
M, M->getParent());
if (SrcDepInfo.isClobber()) {
if (MemCpyInst *MDep = dyn_cast<MemCpyInst>(SrcDepInfo.getInst()))
return processMemCpyMemCpyDependence(M, MDep, CopySize->getZExtValue());
} else if (SrcDepInfo.isDef()) {
if (isa<AllocaInst>(SrcDepInfo.getInst())) {
MD->removeInstruction(M);
M->eraseFromParent();
++NumMemCpyInstr;
return true;
}
}
return false;
}
/// processByValArgument - This is called on every byval argument in call sites.
bool MemCpyOpt::processByValArgument(CallSite CS, unsigned ArgNo) {
if (TD == 0) return false;
// Find out what feeds this byval argument.
Value *ByValArg = CS.getArgument(ArgNo);
Type *ByValTy = cast<PointerType>(ByValArg->getType())->getElementType();
uint64_t ByValSize = TD->getTypeAllocSize(ByValTy);
MemDepResult DepInfo =
MD->getPointerDependencyFrom(AliasAnalysis::Location(ByValArg, ByValSize),
true, CS.getInstruction(),
CS.getInstruction()->getParent());
if (!DepInfo.isClobber())
return false;
// If the byval argument isn't fed by a memcpy, ignore it. If it is fed by
// a memcpy, see if we can byval from the source of the memcpy instead of the
// result.
MemCpyInst *MDep = dyn_cast<MemCpyInst>(DepInfo.getInst());
if (MDep == 0 || MDep->isVolatile() ||
ByValArg->stripPointerCasts() != MDep->getDest())
return false;
// The length of the memcpy must be larger or equal to the size of the byval.
ConstantInt *C1 = dyn_cast<ConstantInt>(MDep->getLength());
if (C1 == 0 || C1->getValue().getZExtValue() < ByValSize)
return false;
// Get the alignment of the byval. If the call doesn't specify the alignment,
// then it is some target specific value that we can't know.
unsigned ByValAlign = CS.getParamAlignment(ArgNo+1);
if (ByValAlign == 0) return false;
// If it is greater than the memcpy, then we check to see if we can force the
// source of the memcpy to the alignment we need. If we fail, we bail out.
if (MDep->getAlignment() < ByValAlign &&
getOrEnforceKnownAlignment(MDep->getSource(),ByValAlign, TD) < ByValAlign)
return false;
// Verify that the copied-from memory doesn't change in between the memcpy and
// the byval call.
// memcpy(a <- b)
// *b = 42;
// foo(*a)
// It would be invalid to transform the second memcpy into foo(*b).
//
// NOTE: This is conservative, it will stop on any read from the source loc,
// not just the defining memcpy.
MemDepResult SourceDep =
MD->getPointerDependencyFrom(AliasAnalysis::getLocationForSource(MDep),
false, CS.getInstruction(), MDep->getParent());
if (!SourceDep.isClobber() || SourceDep.getInst() != MDep)
return false;
Value *TmpCast = MDep->getSource();
if (MDep->getSource()->getType() != ByValArg->getType())
TmpCast = new BitCastInst(MDep->getSource(), ByValArg->getType(),
"tmpcast", CS.getInstruction());
DEBUG(dbgs() << "MemCpyOpt: Forwarding memcpy to byval:\n"
<< " " << *MDep << "\n"
<< " " << *CS.getInstruction() << "\n");
// Otherwise we're good! Update the byval argument.
CS.setArgument(ArgNo, TmpCast);
++NumMemCpyInstr;
return true;
}
bool AMDGPURewriteOutArguments::runOnFunction(Function &F) {
if (skipFunction(F))
return false;
// TODO: Could probably handle variadic functions.
if (F.isVarArg() || F.hasStructRetAttr() ||
AMDGPU::isEntryFunctionCC(F.getCallingConv()))
return false;
MDA = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
unsigned ReturnNumRegs = 0;
SmallSet<int, 4> OutArgIndexes;
SmallVector<Type *, 4> ReturnTypes;
Type *RetTy = F.getReturnType();
if (!RetTy->isVoidTy()) {
ReturnNumRegs = DL->getTypeStoreSize(RetTy) / 4;
if (ReturnNumRegs >= MaxNumRetRegs)
return false;
ReturnTypes.push_back(RetTy);
}
SmallVector<Argument *, 4> OutArgs;
for (Argument &Arg : F.args()) {
if (isOutArgumentCandidate(Arg)) {
LLVM_DEBUG(dbgs() << "Found possible out argument " << Arg
<< " in function " << F.getName() << '\n');
OutArgs.push_back(&Arg);
}
}
if (OutArgs.empty())
return false;
using ReplacementVec = SmallVector<std::pair<Argument *, Value *>, 4>;
DenseMap<ReturnInst *, ReplacementVec> Replacements;
SmallVector<ReturnInst *, 4> Returns;
for (BasicBlock &BB : F) {
if (ReturnInst *RI = dyn_cast<ReturnInst>(&BB.back()))
Returns.push_back(RI);
}
if (Returns.empty())
return false;
bool Changing;
do {
Changing = false;
// Keep retrying if we are able to successfully eliminate an argument. This
// helps with cases with multiple arguments which may alias, such as in a
// sincos implemntation. If we have 2 stores to arguments, on the first
// attempt the MDA query will succeed for the second store but not the
// first. On the second iteration we've removed that out clobbering argument
// (by effectively moving it into another function) and will find the second
// argument is OK to move.
for (Argument *OutArg : OutArgs) {
bool ThisReplaceable = true;
SmallVector<std::pair<ReturnInst *, StoreInst *>, 4> ReplaceableStores;
Type *ArgTy = OutArg->getType()->getPointerElementType();
// Skip this argument if converting it will push us over the register
// count to return limit.
// TODO: This is an approximation. When legalized this could be more. We
// can ask TLI for exactly how many.
unsigned ArgNumRegs = DL->getTypeStoreSize(ArgTy) / 4;
if (ArgNumRegs + ReturnNumRegs > MaxNumRetRegs)
continue;
// An argument is convertible only if all exit blocks are able to replace
// it.
for (ReturnInst *RI : Returns) {
BasicBlock *BB = RI->getParent();
MemDepResult Q = MDA->getPointerDependencyFrom(MemoryLocation(OutArg),
true, BB->end(), BB, RI);
StoreInst *SI = nullptr;
if (Q.isDef())
SI = dyn_cast<StoreInst>(Q.getInst());
if (SI) {
LLVM_DEBUG(dbgs() << "Found out argument store: " << *SI << '\n');
ReplaceableStores.emplace_back(RI, SI);
} else {
ThisReplaceable = false;
break;
}
}
if (!ThisReplaceable)
continue; // Try the next argument candidate.
for (std::pair<ReturnInst *, StoreInst *> Store : ReplaceableStores) {
Value *ReplVal = Store.second->getValueOperand();
auto &ValVec = Replacements[Store.first];
if (llvm::find_if(ValVec,
[OutArg](const std::pair<Argument *, Value *> &Entry) {
return Entry.first == OutArg;}) != ValVec.end()) {
LLVM_DEBUG(dbgs()
<< "Saw multiple out arg stores" << *OutArg << '\n');
// It is possible to see stores to the same argument multiple times,
// but we expect these would have been optimized out already.
ThisReplaceable = false;
break;
}
ValVec.emplace_back(OutArg, ReplVal);
Store.second->eraseFromParent();
}
if (ThisReplaceable) {
ReturnTypes.push_back(ArgTy);
OutArgIndexes.insert(OutArg->getArgNo());
++NumOutArgumentsReplaced;
Changing = true;
}
}
} while (Changing);
if (Replacements.empty())
return false;
LLVMContext &Ctx = F.getParent()->getContext();
StructType *NewRetTy = StructType::create(Ctx, ReturnTypes, F.getName());
FunctionType *NewFuncTy = FunctionType::get(NewRetTy,
F.getFunctionType()->params(),
F.isVarArg());
LLVM_DEBUG(dbgs() << "Computed new return type: " << *NewRetTy << '\n');
Function *NewFunc = Function::Create(NewFuncTy, Function::PrivateLinkage,
F.getName() + ".body");
F.getParent()->getFunctionList().insert(F.getIterator(), NewFunc);
NewFunc->copyAttributesFrom(&F);
NewFunc->setComdat(F.getComdat());
// We want to preserve the function and param attributes, but need to strip
// off any return attributes, e.g. zeroext doesn't make sense with a struct.
NewFunc->stealArgumentListFrom(F);
AttrBuilder RetAttrs;
RetAttrs.addAttribute(Attribute::SExt);
RetAttrs.addAttribute(Attribute::ZExt);
RetAttrs.addAttribute(Attribute::NoAlias);
NewFunc->removeAttributes(AttributeList::ReturnIndex, RetAttrs);
// TODO: How to preserve metadata?
// Move the body of the function into the new rewritten function, and replace
// this function with a stub.
NewFunc->getBasicBlockList().splice(NewFunc->begin(), F.getBasicBlockList());
for (std::pair<ReturnInst *, ReplacementVec> &Replacement : Replacements) {
ReturnInst *RI = Replacement.first;
IRBuilder<> B(RI);
B.SetCurrentDebugLocation(RI->getDebugLoc());
int RetIdx = 0;
Value *NewRetVal = UndefValue::get(NewRetTy);
Value *RetVal = RI->getReturnValue();
if (RetVal)
NewRetVal = B.CreateInsertValue(NewRetVal, RetVal, RetIdx++);
for (std::pair<Argument *, Value *> ReturnPoint : Replacement.second) {
Argument *Arg = ReturnPoint.first;
Value *Val = ReturnPoint.second;
Type *EltTy = Arg->getType()->getPointerElementType();
if (Val->getType() != EltTy) {
Type *EffectiveEltTy = EltTy;
if (StructType *CT = dyn_cast<StructType>(EltTy)) {
assert(CT->getNumElements() == 1);
EffectiveEltTy = CT->getElementType(0);
}
if (DL->getTypeSizeInBits(EffectiveEltTy) !=
DL->getTypeSizeInBits(Val->getType())) {
assert(isVec3ToVec4Shuffle(EffectiveEltTy, Val->getType()));
Val = B.CreateShuffleVector(Val, UndefValue::get(Val->getType()),
{ 0, 1, 2 });
}
Val = B.CreateBitCast(Val, EffectiveEltTy);
// Re-create single element composite.
if (EltTy != EffectiveEltTy)
Val = B.CreateInsertValue(UndefValue::get(EltTy), Val, 0);
}
NewRetVal = B.CreateInsertValue(NewRetVal, Val, RetIdx++);
}
if (RetVal)
RI->setOperand(0, NewRetVal);
else {
B.CreateRet(NewRetVal);
RI->eraseFromParent();
}
}
SmallVector<Value *, 16> StubCallArgs;
for (Argument &Arg : F.args()) {
if (OutArgIndexes.count(Arg.getArgNo())) {
// It's easier to preserve the type of the argument list. We rely on
// DeadArgumentElimination to take care of these.
StubCallArgs.push_back(UndefValue::get(Arg.getType()));
} else {
StubCallArgs.push_back(&Arg);
}
}
BasicBlock *StubBB = BasicBlock::Create(Ctx, "", &F);
IRBuilder<> B(StubBB);
CallInst *StubCall = B.CreateCall(NewFunc, StubCallArgs);
int RetIdx = RetTy->isVoidTy() ? 0 : 1;
for (Argument &Arg : F.args()) {
if (!OutArgIndexes.count(Arg.getArgNo()))
continue;
PointerType *ArgType = cast<PointerType>(Arg.getType());
auto *EltTy = ArgType->getElementType();
unsigned Align = Arg.getParamAlignment();
if (Align == 0)
Align = DL->getABITypeAlignment(EltTy);
Value *Val = B.CreateExtractValue(StubCall, RetIdx++);
Type *PtrTy = Val->getType()->getPointerTo(ArgType->getAddressSpace());
// We can peek through bitcasts, so the type may not match.
Value *PtrVal = B.CreateBitCast(&Arg, PtrTy);
B.CreateAlignedStore(Val, PtrVal, Align);
}
if (!RetTy->isVoidTy()) {
B.CreateRet(B.CreateExtractValue(StubCall, 0));
} else {
B.CreateRetVoid();
}
// The function is now a stub we want to inline.
F.addFnAttr(Attribute::AlwaysInline);
++NumOutArgumentFunctionsReplaced;
return true;
}
