StoreInst *
StackAllocationPromoter::promoteAllocationInBlock(SILBasicBlock *BB) {
LLVM_DEBUG(llvm::dbgs() << "*** Promoting ASI in block: " << *ASI);
// We don't know the value of the alloca until we find the first store.
SILValue RunningVal = SILValue();
// Keep track of the last StoreInst that we found.
StoreInst *LastStore = nullptr;
// For all instructions in the block.
for (auto BBI = BB->begin(), E = BB->end(); BBI != E;) {
SILInstruction *Inst = &*BBI;
++BBI;
if (isLoadFromStack(Inst, ASI)) {
auto Load = cast<LoadInst>(Inst);
if (RunningVal) {
// If we are loading from the AllocStackInst and we already know the
// content of the Alloca then use it.
LLVM_DEBUG(llvm::dbgs() << "*** Promoting load: " << *Load);
replaceLoad(Load, RunningVal, ASI);
NumInstRemoved++;
} else if (Load->getOperand() == ASI) {
// If we don't know the content of the AllocStack then the loaded
// value *is* the new value;
LLVM_DEBUG(llvm::dbgs() << "*** First load: " << *Load);
RunningVal = Load;
}
continue;
}
// Remove stores and record the value that we are saving as the running
// value.
if (auto *SI = dyn_cast<StoreInst>(Inst)) {
if (SI->getDest() != ASI)
continue;
// The stored value is the new running value.
RunningVal = SI->getSrc();
// If we met a store before this one, delete it.
if (LastStore) {
NumInstRemoved++;
LLVM_DEBUG(llvm::dbgs() << "*** Removing redundant store: "
<< *LastStore);
LastStore->eraseFromParent();
}
LastStore = SI;
continue;
}
// Replace debug_value_addr with debug_value of the promoted value
// if we have a valid value to use at this point. Otherwise we'll
// promote this when we deal with hooking up phis.
if (auto *DVAI = dyn_cast<DebugValueAddrInst>(Inst)) {
if (DVAI->getOperand() == ASI &&
RunningVal)
promoteDebugValueAddr(DVAI, RunningVal, B);
continue;
}
// Replace destroys with a release of the value.
if (auto *DAI = dyn_cast<DestroyAddrInst>(Inst)) {
if (DAI->getOperand() == ASI &&
RunningVal) {
replaceDestroy(DAI, RunningVal);
}
continue;
}
// Stop on deallocation.
if (auto *DSI = dyn_cast<DeallocStackInst>(Inst)) {
if (DSI->getOperand() == ASI)
break;
}
}
if (LastStore) {
LLVM_DEBUG(llvm::dbgs() << "*** Finished promotion. Last store: "
<< *LastStore);
} else {
LLVM_DEBUG(llvm::dbgs() << "*** Finished promotion with no stores.\n");
}
return LastStore;
}
/// Attempt to merge an objc_release with a store, load, and objc_retain to form
/// an objc_storeStrong. This can be a little tricky because the instructions
/// don't always appear in order, and there may be unrelated intervening
/// instructions.
void ObjCARCContract::ContractRelease(Instruction *Release,
inst_iterator &Iter) {
LoadInst *Load = dyn_cast<LoadInst>(GetObjCArg(Release));
if (!Load || !Load->isSimple()) return;
// For now, require everything to be in one basic block.
BasicBlock *BB = Release->getParent();
if (Load->getParent() != BB) return;
// Walk down to find the store and the release, which may be in either order.
BasicBlock::iterator I = Load, End = BB->end();
++I;
AliasAnalysis::Location Loc = AA->getLocation(Load);
StoreInst *Store = 0;
bool SawRelease = false;
for (; !Store || !SawRelease; ++I) {
if (I == End)
return;
Instruction *Inst = I;
if (Inst == Release) {
SawRelease = true;
continue;
}
InstructionClass Class = GetBasicInstructionClass(Inst);
// Unrelated retains are harmless.
if (IsRetain(Class))
continue;
if (Store) {
// The store is the point where we're going to put the objc_storeStrong,
// so make sure there are no uses after it.
if (CanUse(Inst, Load, PA, Class))
return;
} else if (AA->getModRefInfo(Inst, Loc) & AliasAnalysis::Mod) {
// We are moving the load down to the store, so check for anything
// else which writes to the memory between the load and the store.
Store = dyn_cast<StoreInst>(Inst);
if (!Store || !Store->isSimple()) return;
if (Store->getPointerOperand() != Loc.Ptr) return;
}
}
Value *New = StripPointerCastsAndObjCCalls(Store->getValueOperand());
// Walk up to find the retain.
I = Store;
BasicBlock::iterator Begin = BB->begin();
while (I != Begin && GetBasicInstructionClass(I) != IC_Retain)
--I;
Instruction *Retain = I;
if (GetBasicInstructionClass(Retain) != IC_Retain) return;
if (GetObjCArg(Retain) != New) return;
Changed = true;
++NumStoreStrongs;
LLVMContext &C = Release->getContext();
Type *I8X = PointerType::getUnqual(Type::getInt8Ty(C));
Type *I8XX = PointerType::getUnqual(I8X);
Value *Args[] = { Load->getPointerOperand(), New };
if (Args[0]->getType() != I8XX)
Args[0] = new BitCastInst(Args[0], I8XX, "", Store);
if (Args[1]->getType() != I8X)
Args[1] = new BitCastInst(Args[1], I8X, "", Store);
CallInst *StoreStrong =
CallInst::Create(getStoreStrongCallee(BB->getParent()->getParent()),
Args, "", Store);
StoreStrong->setDoesNotThrow();
StoreStrong->setDebugLoc(Store->getDebugLoc());
// We can't set the tail flag yet, because we haven't yet determined
// whether there are any escaping allocas. Remember this call, so that
// we can set the tail flag once we know it's safe.
StoreStrongCalls.insert(StoreStrong);
if (&*Iter == Store) ++Iter;
Store->eraseFromParent();
Release->eraseFromParent();
EraseInstruction(Retain);
if (Load->use_empty())
Load->eraseFromParent();
}
/// \brief Removes instructions that create the callee value if they are no
/// longer necessary after inlining.
static void
cleanupCalleeValue(SILValue CalleeValue, ArrayRef<SILValue> CaptureArgs,
ArrayRef<SILValue> FullArgs) {
SmallVector<SILInstruction*, 16> InstsToDelete;
for (SILValue V : FullArgs) {
if (SILInstruction *I = dyn_cast<SILInstruction>(V))
if (I != CalleeValue &&
isInstructionTriviallyDead(I))
InstsToDelete.push_back(I);
}
recursivelyDeleteTriviallyDeadInstructions(InstsToDelete, true);
// Handle the case where the callee of the apply is a load instruction.
if (LoadInst *LI = dyn_cast<LoadInst>(CalleeValue)) {
auto *PBI = cast<ProjectBoxInst>(LI->getOperand());
auto *ABI = cast<AllocBoxInst>(PBI->getOperand());
// The load instruction must have no more uses left to erase it.
if (!LI->use_empty())
return;
LI->eraseFromParent();
// Look through uses of the alloc box the load is loading from to find up to
// one store and up to one strong release.
StrongReleaseInst *SRI = nullptr;
for (Operand *ABIUse : ABI->getUses()) {
if (SRI == nullptr && isa<StrongReleaseInst>(ABIUse->getUser())) {
SRI = cast<StrongReleaseInst>(ABIUse->getUser());
} else if (ABIUse->getUser() != PBI)
return;
}
StoreInst *SI = nullptr;
for (Operand *PBIUse : PBI->getUses()) {
if (SI == nullptr && isa<StoreInst>(PBIUse->getUser())) {
SI = cast<StoreInst>(PBIUse->getUser());
} else
return;
}
// If we found a store, record its source and erase it.
if (SI) {
CalleeValue = SI->getSrc();
SI->eraseFromParent();
} else {
CalleeValue = SILValue();
}
// If we found a strong release, replace it with a strong release of the
// source of the store and erase it.
if (SRI) {
if (CalleeValue)
SILBuilderWithScope(SRI)
.emitStrongReleaseAndFold(SRI->getLoc(), CalleeValue);
SRI->eraseFromParent();
}
assert(PBI->use_empty());
PBI->eraseFromParent();
assert(ABI->use_empty());
ABI->eraseFromParent();
if (!CalleeValue)
return;
}
if (auto *PAI = dyn_cast<PartialApplyInst>(CalleeValue)) {
SILValue Callee = PAI->getCallee();
if (!tryDeleteDeadClosure(PAI))
return;
CalleeValue = Callee;
}
if (auto *TTTFI = dyn_cast<ThinToThickFunctionInst>(CalleeValue)) {
SILValue Callee = TTTFI->getCallee();
if (!tryDeleteDeadClosure(TTTFI))
return;
CalleeValue = Callee;
}
if (FunctionRefInst *FRI = dyn_cast<FunctionRefInst>(CalleeValue)) {
if (!FRI->use_empty())
return;
FRI->eraseFromParent();
}
}
bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
SmallVector<LoadInst *, 4> aggrLoads;
SmallVector<MemTransferInst *, 4> aggrMemcpys;
SmallVector<MemSetInst *, 4> aggrMemsets;
const DataLayout &DL = F.getParent()->getDataLayout();
LLVMContext &Context = F.getParent()->getContext();
//
// Collect all the aggrLoads, aggrMemcpys and addrMemsets.
//
//const BasicBlock *firstBB = &F.front(); // first BB in F
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
//BasicBlock *bb = BI;
for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
++II) {
if (LoadInst *load = dyn_cast<LoadInst>(II)) {
if (!load->hasOneUse())
continue;
if (DL.getTypeStoreSize(load->getType()) < MaxAggrCopySize)
continue;
User *use = load->user_back();
if (StoreInst *store = dyn_cast<StoreInst>(use)) {
if (store->getOperand(0) != load) //getValueOperand
continue;
aggrLoads.push_back(load);
}
} else if (MemTransferInst *intr = dyn_cast<MemTransferInst>(II)) {
Value *len = intr->getLength();
// If the number of elements being copied is greater
// than MaxAggrCopySize, lower it to a loop
if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemcpys.push_back(intr);
}
} else {
// turn variable length memcpy/memmov into loop
aggrMemcpys.push_back(intr);
}
} else if (MemSetInst *memsetintr = dyn_cast<MemSetInst>(II)) {
Value *len = memsetintr->getLength();
if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
if (len_int->getZExtValue() >= MaxAggrCopySize) {
aggrMemsets.push_back(memsetintr);
}
} else {
// turn variable length memset into loop
aggrMemsets.push_back(memsetintr);
}
}
}
}
if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0) &&
(aggrMemsets.size() == 0))
return false;
//
// Do the transformation of an aggr load/copy/set to a loop
//
for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) {
LoadInst *load = aggrLoads[i];
StoreInst *store = dyn_cast<StoreInst>(*load->user_begin());
Value *srcAddr = load->getOperand(0);
Value *dstAddr = store->getOperand(1);
unsigned numLoads = DL.getTypeStoreSize(load->getType());
Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);
convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
store->isVolatile(), Context, F);
store->eraseFromParent();
load->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) {
MemTransferInst *cpy = aggrMemcpys[i];
Value *len = cpy->getLength();
// llvm 2.7 version of memcpy does not have volatile
// operand yet. So always making it non-volatile
// optimistically, so that we don't see unnecessary
// st.volatile in ptx
convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
false, Context, F);
cpy->eraseFromParent();
}
for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) {
MemSetInst *memsetinst = aggrMemsets[i];
Value *len = memsetinst->getLength();
Value *val = memsetinst->getValue();
convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
F);
memsetinst->eraseFromParent();
}
return true;
}
