llvm-6502/lib/Target/NVPTX/NVPTXLowerAggrCopies.cpp

206 lines
7.4 KiB
C++
Raw Normal View History

//===- NVPTXLowerAggrCopies.cpp - ------------------------------*- C++ -*--===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
// Lower aggregate copies, memset, memcpy, memmov intrinsics into loops when
// the size is large or is not a compile-time constant.
//
//===----------------------------------------------------------------------===//
#include "NVPTXLowerAggrCopies.h"
#include "llvm/IR/Constants.h"
#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"
#include "llvm/IR/LLVMContext.h"
#include "llvm/IR/Module.h"
using namespace llvm;
namespace llvm { FunctionPass *createLowerAggrCopies(); }
char NVPTXLowerAggrCopies::ID = 0;
// Lower MemTransferInst or load-store pair to loop
static void convertTransferToLoop(
Instruction *splitAt, Value *srcAddr, Value *dstAddr, Value *len,
//unsigned numLoads,
bool srcVolatile, bool dstVolatile, LLVMContext &Context, Function &F) {
Type *indType = len->getType();
BasicBlock *origBB = splitAt->getParent();
BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
origBB->getTerminator()->setSuccessor(0, loopBB);
IRBuilder<> builder(origBB, origBB->getTerminator());
// srcAddr and dstAddr are expected to be pointer types,
// so no check is made here.
unsigned srcAS = dyn_cast<PointerType>(srcAddr->getType())->getAddressSpace();
unsigned dstAS = dyn_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));
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);
// load from srcAddr+ind
Value *val = loop.CreateLoad(loop.CreateGEP(srcAddr, ind), srcVolatile);
// store at dstAddr+ind
loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), dstVolatile);
// The value for ind coming from backedge is (ind + 1)
Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1));
ind->addIncoming(newind, loopBB);
loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
}
// Lower MemSetInst to loop
static 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);
origBB->getTerminator()->setSuccessor(0, loopBB);
IRBuilder<> builder(origBB, origBB->getTerminator());
unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace();
// Cast pointer to the type of value getting stored
dstAddr =
builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS));
IRBuilder<> loop(loopBB);
PHINode *ind = loop.CreatePHI(len->getType(), 0);
ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB);
loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), false);
Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1));
ind->addIncoming(newind, loopBB);
loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
}
bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
SmallVector<LoadInst *, 4> aggrLoads;
SmallVector<MemTransferInst *, 4> aggrMemcpys;
SmallVector<MemSetInst *, 4> aggrMemsets;
const DataLayout *DL = &getAnalysis<DataLayoutPass>().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() == false)
continue;
if (DL->getTypeStoreSize(load->getType()) < MaxAggrCopySize)
continue;
[C++11] Add range based accessors for the Use-Def chain of a Value. This requires a number of steps. 1) Move value_use_iterator into the Value class as an implementation detail 2) Change it to actually be a *Use* iterator rather than a *User* iterator. 3) Add an adaptor which is a User iterator that always looks through the Use to the User. 4) Wrap these in Value::use_iterator and Value::user_iterator typedefs. 5) Add the range adaptors as Value::uses() and Value::users(). 6) Update *all* of the callers to correctly distinguish between whether they wanted a use_iterator (and to explicitly dig out the User when needed), or a user_iterator which makes the Use itself totally opaque. Because #6 requires churning essentially everything that walked the Use-Def chains, I went ahead and added all of the range adaptors and switched them to range-based loops where appropriate. Also because the renaming requires at least churning every line of code, it didn't make any sense to split these up into multiple commits -- all of which would touch all of the same lies of code. The result is still not quite optimal. The Value::use_iterator is a nice regular iterator, but Value::user_iterator is an iterator over User*s rather than over the User objects themselves. As a consequence, it fits a bit awkwardly into the range-based world and it has the weird extra-dereferencing 'operator->' that so many of our iterators have. I think this could be fixed by providing something which transforms a range of T&s into a range of T*s, but that *can* be separated into another patch, and it isn't yet 100% clear whether this is the right move. However, this change gets us most of the benefit and cleans up a substantial amount of code around Use and User. =] git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203364 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-09 03:16:01 +00:00
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];
[C++11] Add range based accessors for the Use-Def chain of a Value. This requires a number of steps. 1) Move value_use_iterator into the Value class as an implementation detail 2) Change it to actually be a *Use* iterator rather than a *User* iterator. 3) Add an adaptor which is a User iterator that always looks through the Use to the User. 4) Wrap these in Value::use_iterator and Value::user_iterator typedefs. 5) Add the range adaptors as Value::uses() and Value::users(). 6) Update *all* of the callers to correctly distinguish between whether they wanted a use_iterator (and to explicitly dig out the User when needed), or a user_iterator which makes the Use itself totally opaque. Because #6 requires churning essentially everything that walked the Use-Def chains, I went ahead and added all of the range adaptors and switched them to range-based loops where appropriate. Also because the renaming requires at least churning every line of code, it didn't make any sense to split these up into multiple commits -- all of which would touch all of the same lies of code. The result is still not quite optimal. The Value::use_iterator is a nice regular iterator, but Value::user_iterator is an iterator over User*s rather than over the User objects themselves. As a consequence, it fits a bit awkwardly into the range-based world and it has the weird extra-dereferencing 'operator->' that so many of our iterators have. I think this could be fixed by providing something which transforms a range of T&s into a range of T*s, but that *can* be separated into another patch, and it isn't yet 100% clear whether this is the right move. However, this change gets us most of the benefit and cleans up a substantial amount of code around Use and User. =] git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@203364 91177308-0d34-0410-b5e6-96231b3b80d8
2014-03-09 03:16:01 +00:00
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;
}
FunctionPass *llvm::createLowerAggrCopies() {
return new NVPTXLowerAggrCopies();
}