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;
}
void AMDGPUPromoteAlloca::visitAlloca(AllocaInst &I) {
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I))
return;
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
// FIXME: This is the maximum work group size. We should try to get
// value from the reqd_work_group_size function attribute if it is
// available.
unsigned WorkGroupSize = 256;
int AllocaSize =
WorkGroupSize * Mod->getDataLayout().getTypeAllocSize(AllocaTy);
if (AllocaSize > LocalMemAvailable) {
DEBUG(dbgs() << " Not enough local memory to promote alloca.\n");
return;
}
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
LocalMemAvailable -= AllocaSize;
Type *GVTy = ArrayType::get(I.getAllocatedType(), 256);
GlobalVariable *GV = new GlobalVariable(
*Mod, GVTy, false, GlobalValue::ExternalLinkage, 0, I.getName(), 0,
GlobalVariable::NotThreadLocal, AMDGPUAS::LOCAL_ADDRESS);
FunctionType *FTy = FunctionType::get(
Type::getInt32Ty(Mod->getContext()), false);
AttributeSet AttrSet;
AttrSet.addAttribute(Mod->getContext(), 0, Attribute::ReadNone);
Value *ReadLocalSizeY = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.y", FTy, AttrSet);
Value *ReadLocalSizeZ = Mod->getOrInsertFunction(
"llvm.r600.read.local.size.z", FTy, AttrSet);
Value *ReadTIDIGX = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.x", FTy, AttrSet);
Value *ReadTIDIGY = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.y", FTy, AttrSet);
Value *ReadTIDIGZ = Mod->getOrInsertFunction(
"llvm.r600.read.tidig.z", FTy, AttrSet);
Value *TCntY = Builder.CreateCall(ReadLocalSizeY, {});
Value *TCntZ = Builder.CreateCall(ReadLocalSizeZ, {});
Value *TIdX = Builder.CreateCall(ReadTIDIGX, {});
Value *TIdY = Builder.CreateCall(ReadTIDIGY, {});
Value *TIdZ = Builder.CreateCall(ReadTIDIGZ, {});
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
std::vector<Value*> Indices;
Indices.push_back(Constant::getNullValue(Type::getInt32Ty(Mod->getContext())));
Indices.push_back(TID);
Value *Offset = Builder.CreateGEP(GVTy, GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (std::vector<Value*>::iterator i = WorkList.begin(),
e = WorkList.end(); i != e; ++i) {
Value *V = *i;
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
continue;
}
IntrinsicInst *Intr = dyn_cast<IntrinsicInst>(Call);
if (!Intr) {
std::vector<Type*> ArgTypes;
for (unsigned ArgIdx = 0, ArgEnd = Call->getNumArgOperands();
ArgIdx != ArgEnd; ++ArgIdx) {
ArgTypes.push_back(Call->getArgOperand(ArgIdx)->getType());
}
Function *F = Call->getCalledFunction();
FunctionType *NewType = FunctionType::get(Call->getType(), ArgTypes,
F->isVarArg());
Constant *C = Mod->getOrInsertFunction((F->getName() + ".local").str(),
NewType, F->getAttributes());
Function *NewF = cast<Function>(C);
Call->setCalledFunction(NewF);
continue;
}
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}
/// tryMergingIntoMemset - When 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 consecutive ones, it
/// attempts to merge them together into a memcpy/memset.
Instruction *MemCpyOpt::tryMergingIntoMemset(Instruction *StartInst,
Value *StartPtr, Value *ByteVal) {
if (TD == 0) return 0;
// 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);
BasicBlock::iterator BI = StartInst;
for (++BI; !isa<TerminatorInst>(BI); ++BI) {
if (!isa<StoreInst>(BI) && !isa<MemSetInst>(BI)) {
// If the instruction 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 (BI->mayWriteToMemory() || BI->mayReadFromMemory())
break;
continue;
}
if (StoreInst *NextStore = dyn_cast<StoreInst>(BI)) {
// If this is a store, see if we can merge it in.
if (!NextStore->isSimple()) 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);
} else {
MemSetInst *MSI = cast<MemSetInst>(BI);
if (MSI->isVolatile() || ByteVal != MSI->getValue() ||
!isa<ConstantInt>(MSI->getLength()))
break;
// Check to see if this store is to a constant offset from the start ptr.
int64_t Offset;
if (!IsPointerOffset(StartPtr, MSI->getDest(), Offset, *TD))
break;
Ranges.addMemSet(Offset, MSI);
}
}
// 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 0;
// 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.addInst(0, StartInst);
// If we create any memsets, we put it right before the first instruction that
// isn't part of the memset block. This ensure that the memset is dominated
// by any addressing instruction needed by the start of the block.
IRBuilder<> Builder(BI);
// Now that we have full information about ranges, loop over the ranges and
// emit memset's for anything big enough to be worthwhile.
Instruction *AMemSet = 0;
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.
// Get the starting pointer of the block.
StartPtr = Range.StartPtr;
// Determine alignment
unsigned Alignment = Range.Alignment;
if (Alignment == 0) {
Type *EltType =
cast<PointerType>(StartPtr->getType())->getElementType();
Alignment = TD->getABITypeAlignment(EltType);
}
AMemSet =
Builder.CreateMemSet(StartPtr, ByteVal, Range.End-Range.Start, Alignment);
DEBUG(dbgs() << "Replace stores:\n";
for (unsigned i = 0, e = Range.TheStores.size(); i != e; ++i)
dbgs() << *Range.TheStores[i] << '\n';
dbgs() << "With: " << *AMemSet << '\n');
if (!Range.TheStores.empty())
AMemSet->setDebugLoc(Range.TheStores[0]->getDebugLoc());
// Zap all the stores.
for (SmallVector<Instruction*, 16>::const_iterator
SI = Range.TheStores.begin(),
SE = Range.TheStores.end(); SI != SE; ++SI) {
MD->removeInstruction(*SI);
(*SI)->eraseFromParent();
}
++NumMemSetInfer;
}
return AMemSet;
}
// FIXME: Should try to pick the most likely to be profitable allocas first.
void AMDGPUPromoteAlloca::handleAlloca(AllocaInst &I) {
// Array allocations are probably not worth handling, since an allocation of
// the array type is the canonical form.
if (!I.isStaticAlloca() || I.isArrayAllocation())
return;
IRBuilder<> Builder(&I);
// First try to replace the alloca with a vector
Type *AllocaTy = I.getAllocatedType();
DEBUG(dbgs() << "Trying to promote " << I << '\n');
if (tryPromoteAllocaToVector(&I)) {
DEBUG(dbgs() << " alloca is not a candidate for vectorization.\n");
return;
}
const Function &ContainingFunction = *I.getParent()->getParent();
// Don't promote the alloca to LDS for shader calling conventions as the work
// item ID intrinsics are not supported for these calling conventions.
// Furthermore not all LDS is available for some of the stages.
if (AMDGPU::isShader(ContainingFunction.getCallingConv()))
return;
// FIXME: We should also try to get this value from the reqd_work_group_size
// function attribute if it is available.
unsigned WorkGroupSize = AMDGPU::getMaximumWorkGroupSize(ContainingFunction);
const DataLayout &DL = Mod->getDataLayout();
unsigned Align = I.getAlignment();
if (Align == 0)
Align = DL.getABITypeAlignment(I.getAllocatedType());
// FIXME: This computed padding is likely wrong since it depends on inverse
// usage order.
//
// FIXME: It is also possible that if we're allowed to use all of the memory
// could could end up using more than the maximum due to alignment padding.
uint32_t NewSize = alignTo(CurrentLocalMemUsage, Align);
uint32_t AllocSize = WorkGroupSize * DL.getTypeAllocSize(AllocaTy);
NewSize += AllocSize;
if (NewSize > LocalMemLimit) {
DEBUG(dbgs() << " " << AllocSize
<< " bytes of local memory not available to promote\n");
return;
}
CurrentLocalMemUsage = NewSize;
std::vector<Value*> WorkList;
if (!collectUsesWithPtrTypes(&I, &I, WorkList)) {
DEBUG(dbgs() << " Do not know how to convert all uses\n");
return;
}
DEBUG(dbgs() << "Promoting alloca to local memory\n");
Function *F = I.getParent()->getParent();
Type *GVTy = ArrayType::get(I.getAllocatedType(), WorkGroupSize);
GlobalVariable *GV = new GlobalVariable(
*Mod, GVTy, false, GlobalValue::InternalLinkage,
UndefValue::get(GVTy),
Twine(F->getName()) + Twine('.') + I.getName(),
nullptr,
GlobalVariable::NotThreadLocal,
AMDGPUAS::LOCAL_ADDRESS);
GV->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
GV->setAlignment(I.getAlignment());
Value *TCntY, *TCntZ;
std::tie(TCntY, TCntZ) = getLocalSizeYZ(Builder);
Value *TIdX = getWorkitemID(Builder, 0);
Value *TIdY = getWorkitemID(Builder, 1);
Value *TIdZ = getWorkitemID(Builder, 2);
Value *Tmp0 = Builder.CreateMul(TCntY, TCntZ, "", true, true);
Tmp0 = Builder.CreateMul(Tmp0, TIdX);
Value *Tmp1 = Builder.CreateMul(TIdY, TCntZ, "", true, true);
Value *TID = Builder.CreateAdd(Tmp0, Tmp1);
TID = Builder.CreateAdd(TID, TIdZ);
Value *Indices[] = {
Constant::getNullValue(Type::getInt32Ty(Mod->getContext())),
TID
};
Value *Offset = Builder.CreateInBoundsGEP(GVTy, GV, Indices);
I.mutateType(Offset->getType());
I.replaceAllUsesWith(Offset);
I.eraseFromParent();
for (Value *V : WorkList) {
CallInst *Call = dyn_cast<CallInst>(V);
if (!Call) {
if (ICmpInst *CI = dyn_cast<ICmpInst>(V)) {
Value *Src0 = CI->getOperand(0);
Type *EltTy = Src0->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
if (isa<ConstantPointerNull>(CI->getOperand(0)))
CI->setOperand(0, ConstantPointerNull::get(NewTy));
if (isa<ConstantPointerNull>(CI->getOperand(1)))
CI->setOperand(1, ConstantPointerNull::get(NewTy));
continue;
}
// The operand's value should be corrected on its own.
if (isa<AddrSpaceCastInst>(V))
continue;
Type *EltTy = V->getType()->getPointerElementType();
PointerType *NewTy = PointerType::get(EltTy, AMDGPUAS::LOCAL_ADDRESS);
// FIXME: It doesn't really make sense to try to do this for all
// instructions.
V->mutateType(NewTy);
// Adjust the types of any constant operands.
if (SelectInst *SI = dyn_cast<SelectInst>(V)) {
if (isa<ConstantPointerNull>(SI->getOperand(1)))
SI->setOperand(1, ConstantPointerNull::get(NewTy));
if (isa<ConstantPointerNull>(SI->getOperand(2)))
SI->setOperand(2, ConstantPointerNull::get(NewTy));
} else if (PHINode *Phi = dyn_cast<PHINode>(V)) {
for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I) {
if (isa<ConstantPointerNull>(Phi->getIncomingValue(I)))
Phi->setIncomingValue(I, ConstantPointerNull::get(NewTy));
}
}
continue;
}
IntrinsicInst *Intr = cast<IntrinsicInst>(Call);
Builder.SetInsertPoint(Intr);
switch (Intr->getIntrinsicID()) {
case Intrinsic::lifetime_start:
case Intrinsic::lifetime_end:
// These intrinsics are for address space 0 only
Intr->eraseFromParent();
continue;
case Intrinsic::memcpy: {
MemCpyInst *MemCpy = cast<MemCpyInst>(Intr);
Builder.CreateMemCpy(MemCpy->getRawDest(), MemCpy->getRawSource(),
MemCpy->getLength(), MemCpy->getAlignment(),
MemCpy->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memmove: {
MemMoveInst *MemMove = cast<MemMoveInst>(Intr);
Builder.CreateMemMove(MemMove->getRawDest(), MemMove->getRawSource(),
MemMove->getLength(), MemMove->getAlignment(),
MemMove->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::memset: {
MemSetInst *MemSet = cast<MemSetInst>(Intr);
Builder.CreateMemSet(MemSet->getRawDest(), MemSet->getValue(),
MemSet->getLength(), MemSet->getAlignment(),
MemSet->isVolatile());
Intr->eraseFromParent();
continue;
}
case Intrinsic::invariant_start:
case Intrinsic::invariant_end:
case Intrinsic::invariant_group_barrier:
Intr->eraseFromParent();
// FIXME: I think the invariant marker should still theoretically apply,
// but the intrinsics need to be changed to accept pointers with any
// address space.
continue;
case Intrinsic::objectsize: {
Value *Src = Intr->getOperand(0);
Type *SrcTy = Src->getType()->getPointerElementType();
Function *ObjectSize = Intrinsic::getDeclaration(Mod,
Intrinsic::objectsize,
{ Intr->getType(), PointerType::get(SrcTy, AMDGPUAS::LOCAL_ADDRESS) }
);
CallInst *NewCall
= Builder.CreateCall(ObjectSize, { Src, Intr->getOperand(1) });
Intr->replaceAllUsesWith(NewCall);
Intr->eraseFromParent();
continue;
}
default:
Intr->dump();
llvm_unreachable("Don't know how to promote alloca intrinsic use.");
}
}
}