maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Streamline the coding style in NVPTXLowerAggrCopies 

Make the style consistent with LLVM style throughout and clang-format.

llvm-svn: 242439
This commit is contained in:
Eli Bendersky 2015-07-16 20:42:38 +00:00
parent 0693b145ec
commit f871e09d2d
1 changed files with 123 additions and 107 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

230
llvm/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp
View File

@ -7,6 +7,7 @@
//
//===----------------------------------------------------------------------===//
//
// \file
// Lower aggregate copies, memset, memcpy, memmov intrinsics into loops when
// the size is large or is not a compile-time constant.
//
@ -19,7 +20,6 @@
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstIterator.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
@ -57,47 +57,52 @@ struct NVPTXLowerAggrCopies : public FunctionPass {
char NVPTXLowerAggrCopies::ID = 0;
// Lower memcpy to loop.
void convertMemCpyToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
Value *len, bool srcVolatile, bool dstVolatile,
LLVMContext &Context, Function &F) {
Type *indType = len->getType();
void convertMemCpyToLoop(Instruction *ConvertedInst, Value *SrcAddr,
Value *DstAddr, Value *CopyLen, bool SrcIsVolatile,
bool DstIsVolatile, LLVMContext &Context,
Function &F) {
Type *TypeOfCopyLen = CopyLen->getType();
BasicBlock *origBB = splitAt->getParent();
BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
BasicBlock *OrigBB = ConvertedInst->getParent();
BasicBlock *NewBB =
ConvertedInst->getParent()->splitBasicBlock(ConvertedInst, "split");
BasicBlock *LoopBB = BasicBlock::Create(Context, "loadstoreloop", &F, NewBB);
origBB->getTerminator()->setSuccessor(0, loopBB);
IRBuilder<> builder(origBB, origBB->getTerminator());
OrigBB->getTerminator()->setSuccessor(0, LoopBB);
IRBuilder<> Builder(OrigBB, OrigBB->getTerminator());
// srcAddr and dstAddr are expected to be pointer types,
// SrcAddr and DstAddr are expected to be pointer types,
// so no check is made here.
unsigned srcAS = cast<PointerType>(srcAddr->getType())->getAddressSpace();
unsigned dstAS = cast<PointerType>(dstAddr->getType())->getAddressSpace();
unsigned SrcAS = cast<PointerType>(SrcAddr->getType())->getAddressSpace();
unsigned DstAS = cast<PointerType>(DstAddr->getType())->getAddressSpace();
// Cast pointers to (char *)
srcAddr = builder.CreateBitCast(srcAddr, Type::getInt8PtrTy(Context, srcAS));
dstAddr = builder.CreateBitCast(dstAddr, Type::getInt8PtrTy(Context, dstAS));
SrcAddr = Builder.CreateBitCast(SrcAddr, Builder.getInt8PtrTy(SrcAS));
DstAddr = Builder.CreateBitCast(DstAddr, Builder.getInt8PtrTy(DstAS));
IRBuilder<> loop(loopBB);
// The loop index (ind) is a phi node.
PHINode *ind = loop.CreatePHI(indType, 0);
// Incoming value for ind is 0
ind->addIncoming(ConstantInt::get(indType, 0), origBB);
IRBuilder<> LoopBuilder(LoopBB);
PHINode *LoopIndex = LoopBuilder.CreatePHI(TypeOfCopyLen, 0);
LoopIndex->addIncoming(ConstantInt::get(TypeOfCopyLen, 0), OrigBB);
// load from srcAddr+ind
// load from SrcAddr+LoopIndex
// TODO: we can leverage the align parameter of llvm.memcpy for more efficient
// word-sized loads and stores.
Value *val = loop.CreateLoad(loop.CreateGEP(loop.getInt8Ty(), srcAddr, ind),
srcVolatile);
// store at dstAddr+ind
loop.CreateStore(val, loop.CreateGEP(loop.getInt8Ty(), dstAddr, ind),
dstVolatile);
Value *Element = LoopBuilder.CreateLoad(
LoopBuilder.CreateGEP(LoopBuilder.getInt8Ty(), SrcAddr, LoopIndex),
SrcIsVolatile);
// store at DstAddr+LoopIndex
LoopBuilder.CreateStore(
Element,
LoopBuilder.CreateGEP(LoopBuilder.getInt8Ty(), DstAddr, LoopIndex),
DstIsVolatile);
// The value for ind coming from backedge is (ind + 1)
Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1));
ind->addIncoming(newind, loopBB);
// The value for LoopIndex coming from backedge is (LoopIndex + 1)
Value *NewIndex =
LoopBuilder.CreateAdd(LoopIndex, ConstantInt::get(TypeOfCopyLen, 1));
LoopIndex->addIncoming(NewIndex, LoopBB);
loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
LoopBuilder.CreateCondBr(LoopBuilder.CreateICmpULT(NewIndex, CopyLen), LoopBB,
NewBB);
}
// Lower memmove to IR. memmove is required to correctly copy overlapping memory
@ -122,11 +127,12 @@ void convertMemCpyToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
// }
// return dst;
// }
void convertMemMoveToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
Value *len, bool srcVolatile, bool dstVolatile,
LLVMContext &Context, Function &F) {
Type *TypeOfLen = len->getType();
BasicBlock *OrigBB = splitAt->getParent();
void convertMemMoveToLoop(Instruction *ConvertedInst, Value *SrcAddr,
Value *DstAddr, Value *CopyLen, bool SrcIsVolatile,
bool DstIsVolatile, LLVMContext &Context,
Function &F) {
Type *TypeOfCopyLen = CopyLen->getType();
BasicBlock *OrigBB = ConvertedInst->getParent();
// Create the a comparison of src and dst, based on which we jump to either
// the forward-copy part of the function (if src >= dst) or the backwards-copy
@ -134,10 +140,11 @@ void convertMemMoveToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
// SplitBlockAndInsertIfThenElse conveniently creates the basic if-then-else
// structure. Its block terminators (unconditional branches) are replaced by
// the appropriate conditional branches when the loop is built.
ICmpInst *PtrCompare = new ICmpInst(splitAt, ICmpInst::ICMP_ULT, srcAddr,
dstAddr, "compare_src_dst");
ICmpInst *PtrCompare = new ICmpInst(ConvertedInst, ICmpInst::ICMP_ULT,
SrcAddr, DstAddr, "compare_src_dst");
TerminatorInst *ThenTerm, *ElseTerm;
SplitBlockAndInsertIfThenElse(PtrCompare, splitAt, &ThenTerm, &ElseTerm);
SplitBlockAndInsertIfThenElse(PtrCompare, ConvertedInst, &ThenTerm,
&ElseTerm);
// Each part of the function consists of two blocks:
// copy_backwards: used to skip the loop when n == 0
@ -148,31 +155,31 @@ void convertMemMoveToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
CopyBackwardsBB->setName("copy_backwards");
BasicBlock *CopyForwardBB = ElseTerm->getParent();
CopyForwardBB->setName("copy_forward");
BasicBlock *ExitBB = splitAt->getParent();
BasicBlock *ExitBB = ConvertedInst->getParent();
ExitBB->setName("memmove_done");
// Initial comparison of n == 0 that lets us skip the loops altogether. Shared
// between both backwards and forward copy clauses.
ICmpInst *CompareN =
new ICmpInst(OrigBB->getTerminator(), ICmpInst::ICMP_EQ, len,
ConstantInt::get(TypeOfLen, 0), "compare_n_to_0");
new ICmpInst(OrigBB->getTerminator(), ICmpInst::ICMP_EQ, CopyLen,
ConstantInt::get(TypeOfCopyLen, 0), "compare_n_to_0");
// Copying backwards.
BasicBlock *LoopBB =
BasicBlock::Create(Context, "copy_backwards_loop", &F, CopyForwardBB);
IRBuilder<> LoopBuilder(LoopBB);
PHINode *LoopPhi = LoopBuilder.CreatePHI(TypeOfLen, 0);
PHINode *LoopPhi = LoopBuilder.CreatePHI(TypeOfCopyLen, 0);
Value *IndexPtr = LoopBuilder.CreateSub(
LoopPhi, ConstantInt::get(TypeOfLen, 1), "index_ptr");
LoopPhi, ConstantInt::get(TypeOfCopyLen, 1), "index_ptr");
Value *Element = LoopBuilder.CreateLoad(
LoopBuilder.CreateInBoundsGEP(srcAddr, IndexPtr), "element");
LoopBuilder.CreateInBoundsGEP(SrcAddr, IndexPtr), "element");
LoopBuilder.CreateStore(Element,
LoopBuilder.CreateInBoundsGEP(dstAddr, IndexPtr));
LoopBuilder.CreateInBoundsGEP(DstAddr, IndexPtr));
LoopBuilder.CreateCondBr(
LoopBuilder.CreateICmpEQ(IndexPtr, ConstantInt::get(TypeOfLen, 0)),
LoopBuilder.CreateICmpEQ(IndexPtr, ConstantInt::get(TypeOfCopyLen, 0)),
ExitBB, LoopBB);
LoopPhi->addIncoming(IndexPtr, LoopBB);
LoopPhi->addIncoming(len, CopyBackwardsBB);
LoopPhi->addIncoming(CopyLen, CopyBackwardsBB);
BranchInst::Create(ExitBB, LoopBB, CompareN, ThenTerm);
ThenTerm->eraseFromParent();
@ -180,52 +187,57 @@ void convertMemMoveToLoop(Instruction *splitAt, Value *srcAddr, Value *dstAddr,
BasicBlock *FwdLoopBB =
BasicBlock::Create(Context, "copy_forward_loop", &F, ExitBB);
IRBuilder<> FwdLoopBuilder(FwdLoopBB);
PHINode *FwdCopyPhi = FwdLoopBuilder.CreatePHI(TypeOfLen, 0, "index_ptr");
PHINode *FwdCopyPhi = FwdLoopBuilder.CreatePHI(TypeOfCopyLen, 0, "index_ptr");
Value *FwdElement = FwdLoopBuilder.CreateLoad(
FwdLoopBuilder.CreateInBoundsGEP(srcAddr, FwdCopyPhi), "element");
FwdLoopBuilder.CreateInBoundsGEP(SrcAddr, FwdCopyPhi), "element");
FwdLoopBuilder.CreateStore(
FwdElement, FwdLoopBuilder.CreateInBoundsGEP(dstAddr, FwdCopyPhi));
FwdElement, FwdLoopBuilder.CreateInBoundsGEP(DstAddr, FwdCopyPhi));
Value *FwdIndexPtr = FwdLoopBuilder.CreateAdd(
FwdCopyPhi, ConstantInt::get(TypeOfLen, 1), "index_increment");
FwdLoopBuilder.CreateCondBr(FwdLoopBuilder.CreateICmpEQ(FwdIndexPtr, len),
FwdCopyPhi, ConstantInt::get(TypeOfCopyLen, 1), "index_increment");
FwdLoopBuilder.CreateCondBr(FwdLoopBuilder.CreateICmpEQ(FwdIndexPtr, CopyLen),
ExitBB, FwdLoopBB);
FwdCopyPhi->addIncoming(FwdIndexPtr, FwdLoopBB);
FwdCopyPhi->addIncoming(ConstantInt::get(TypeOfLen, 0), CopyForwardBB);
FwdCopyPhi->addIncoming(ConstantInt::get(TypeOfCopyLen, 0), CopyForwardBB);
BranchInst::Create(ExitBB, FwdLoopBB, CompareN, ElseTerm);
ElseTerm->eraseFromParent();
}
// Lower memset to loop.
void convertMemSetToLoop(Instruction *splitAt, Value *dstAddr, Value *len,
Value *val, LLVMContext &Context, Function &F) {
BasicBlock *origBB = splitAt->getParent();
BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
void convertMemSetToLoop(Instruction *ConvertedInst, Value *DstAddr,
Value *CopyLen, Value *SetValue, LLVMContext &Context,
Function &F) {
BasicBlock *OrigBB = ConvertedInst->getParent();
BasicBlock *NewBB =
ConvertedInst->getParent()->splitBasicBlock(ConvertedInst, "split");
BasicBlock *LoopBB = BasicBlock::Create(Context, "loadstoreloop", &F, NewBB);
origBB->getTerminator()->setSuccessor(0, loopBB);
IRBuilder<> builder(origBB, origBB->getTerminator());
unsigned dstAS = cast<PointerType>(dstAddr->getType())->getAddressSpace();
OrigBB->getTerminator()->setSuccessor(0, LoopBB);
IRBuilder<> Builder(OrigBB, OrigBB->getTerminator());
// Cast pointer to the type of value getting stored
dstAddr =
builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS));
unsigned dstAS = cast<PointerType>(DstAddr->getType())->getAddressSpace();
DstAddr = Builder.CreateBitCast(DstAddr,
PointerType::get(SetValue->getType(), dstAS));
IRBuilder<> loop(loopBB);
PHINode *ind = loop.CreatePHI(len->getType(), 0);
ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB);
IRBuilder<> LoopBuilder(LoopBB);
PHINode *LoopIndex = LoopBuilder.CreatePHI(CopyLen->getType(), 0);
LoopIndex->addIncoming(ConstantInt::get(CopyLen->getType(), 0), OrigBB);
loop.CreateStore(val, loop.CreateGEP(val->getType(), dstAddr, ind), false);
LoopBuilder.CreateStore(
SetValue, LoopBuilder.CreateGEP(SetValue->getType(), DstAddr, LoopIndex),
false);
Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1));
ind->addIncoming(newind, loopBB);
Value *NewIndex =
LoopBuilder.CreateAdd(LoopIndex, ConstantInt::get(CopyLen->getType(), 1));
LoopIndex->addIncoming(NewIndex, LoopBB);
loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
LoopBuilder.CreateCondBr(LoopBuilder.CreateICmpULT(NewIndex, CopyLen), LoopBB,
NewBB);
}
bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
SmallVector<LoadInst *, 4> aggrLoads;
SmallVector<LoadInst *, 4> AggrLoads;
SmallVector<MemIntrinsic *, 4> MemCalls;
const DataLayout &DL = F.getParent()->getDataLayout();
@ -235,18 +247,17 @@ bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
++II) {
if (LoadInst *load = dyn_cast<LoadInst>(II)) {
if (!load->hasOneUse())
if (LoadInst *LI = dyn_cast<LoadInst>(II)) {
if (!LI->hasOneUse())
continue;
if (DL.getTypeStoreSize(load->getType()) < MaxAggrCopySize)
if (DL.getTypeStoreSize(LI->getType()) < MaxAggrCopySize)
continue;
User *use = load->user_back();
if (StoreInst *store = dyn_cast<StoreInst>(use)) {
if (store->getOperand(0) != load)
if (StoreInst *SI = dyn_cast<StoreInst>(LI->user_back())) {
if (SI->getOperand(0) != LI)
continue;
aggrLoads.push_back(load);
AggrLoads.push_back(LI);
}
} else if (MemIntrinsic *IntrCall = dyn_cast<MemIntrinsic>(II)) {
// Convert intrinsic calls with variable size or with constant size
@ -262,55 +273,60 @@ bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
}
}
if (aggrLoads.size() == 0 && MemCalls.size() == 0) {
if (AggrLoads.size() == 0 && MemCalls.size() == 0) {
return false;
}
//
// Do the transformation of an aggr load/copy/set to a loop
//
for (LoadInst *load : aggrLoads) {
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);
for (LoadInst *LI : AggrLoads) {
StoreInst *SI = dyn_cast<StoreInst>(*LI->user_begin());
Value *SrcAddr = LI->getOperand(0);
Value *DstAddr = SI->getOperand(1);
unsigned NumLoads = DL.getTypeStoreSize(LI->getType());
Value *CopyLen = ConstantInt::get(Type::getInt32Ty(Context), NumLoads);
convertMemCpyToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
store->isVolatile(), Context, F);
convertMemCpyToLoop(/* ConvertedInst */ SI,
/* SrcAddr */ SrcAddr, /* DstAddr */ DstAddr,
/* CopyLen */ CopyLen,
/* SrcIsVolatile */ LI->isVolatile(),
/* DstIsVolatile */ SI->isVolatile(),
/* Context */ Context,
/* Function F */ F);
store->eraseFromParent();
load->eraseFromParent();
SI->eraseFromParent();
LI->eraseFromParent();
}
// Transform mem* intrinsic calls.
for (MemIntrinsic *MemCall : MemCalls) {
if (MemCpyInst *Memcpy = dyn_cast<MemCpyInst>(MemCall)) {
convertMemCpyToLoop(/* splitAt */ Memcpy,
/* srcAddr */ Memcpy->getRawSource(),
/* dstAddr */ Memcpy->getRawDest(),
/* len */ Memcpy->getLength(),
/* srcVolatile */ Memcpy->isVolatile(),
/* dstVolatile */ Memcpy->isVolatile(),
convertMemCpyToLoop(/* ConvertedInst */ Memcpy,
/* SrcAddr */ Memcpy->getRawSource(),
/* DstAddr */ Memcpy->getRawDest(),
/* CopyLen */ Memcpy->getLength(),
/* SrcIsVolatile */ Memcpy->isVolatile(),
/* DstIsVolatile */ Memcpy->isVolatile(),
/* Context */ Context,
/* Function F */ F);
} else if (MemMoveInst *Memmove = dyn_cast<MemMoveInst>(MemCall)) {
convertMemMoveToLoop(/* splitAt */ Memmove,
/* srcAddr */ Memmove->getRawSource(),
/* dstAddr */ Memmove->getRawDest(),
/* len */ Memmove->getLength(),
/* srcVolatile */ Memmove->isVolatile(),
/* dstVolatile */ Memmove->isVolatile(),
convertMemMoveToLoop(/* ConvertedInst */ Memmove,
/* SrcAddr */ Memmove->getRawSource(),
/* DstAddr */ Memmove->getRawDest(),
/* CopyLen */ Memmove->getLength(),
/* SrcIsVolatile */ Memmove->isVolatile(),
/* DstIsVolatile */ Memmove->isVolatile(),
/* Context */ Context,
/* Function F */ F);
} else if (MemSetInst *Memset = dyn_cast<MemSetInst>(MemCall)) {
convertMemSetToLoop(/* splitAt */ Memset,
/* dstAddr */ Memset->getRawDest(),
/* len */ Memset->getLength(),
/* val */ Memset->getValue(),
convertMemSetToLoop(/* ConvertedInst */ Memset,
/* DstAddr */ Memset->getRawDest(),
/* CopyLen */ Memset->getLength(),
/* SetValue */ Memset->getValue(),
/* Context */ Context,
/* F */ F);
/* Function F */ F);
}
MemCall->eraseFromParent();
}