1 //===- NVPTXLowerAggrCopies.cpp - ------------------------------*- C++ -*--===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // Lower aggregate copies, memset, memcpy, memmov intrinsics into loops when
10 // the size is large or is not a compile-time constant.
12 //===----------------------------------------------------------------------===//
14 #include "NVPTXLowerAggrCopies.h"
15 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
16 #include "llvm/CodeGen/StackProtector.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/DataLayout.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/InstIterator.h"
22 #include "llvm/IR/Instructions.h"
23 #include "llvm/IR/IntrinsicInst.h"
24 #include "llvm/IR/Intrinsics.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Module.h"
27 #include "llvm/Support/Debug.h"
29 #define DEBUG_TYPE "nvptx"
34 // actual analysis class, which is a functionpass
35 struct NVPTXLowerAggrCopies : public FunctionPass {
38 NVPTXLowerAggrCopies() : FunctionPass(ID) {}
40 void getAnalysisUsage(AnalysisUsage &AU) const override {
41 AU.addPreserved<MachineFunctionAnalysis>();
42 AU.addPreserved<StackProtector>();
45 bool runOnFunction(Function &F) override;
47 static const unsigned MaxAggrCopySize = 128;
49 const char *getPassName() const override {
50 return "Lower aggregate copies/intrinsics into loops";
55 char NVPTXLowerAggrCopies::ID = 0;
57 // Lower MemTransferInst or load-store pair to loop
58 static void convertTransferToLoop(
59 Instruction *splitAt, Value *srcAddr, Value *dstAddr, Value *len,
61 bool srcVolatile, bool dstVolatile, LLVMContext &Context, Function &F) {
62 Type *indType = len->getType();
64 BasicBlock *origBB = splitAt->getParent();
65 BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
66 BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
68 origBB->getTerminator()->setSuccessor(0, loopBB);
69 IRBuilder<> builder(origBB, origBB->getTerminator());
71 // srcAddr and dstAddr are expected to be pointer types,
72 // so no check is made here.
73 unsigned srcAS = cast<PointerType>(srcAddr->getType())->getAddressSpace();
74 unsigned dstAS = cast<PointerType>(dstAddr->getType())->getAddressSpace();
76 // Cast pointers to (char *)
77 srcAddr = builder.CreateBitCast(srcAddr, Type::getInt8PtrTy(Context, srcAS));
78 dstAddr = builder.CreateBitCast(dstAddr, Type::getInt8PtrTy(Context, dstAS));
80 IRBuilder<> loop(loopBB);
81 // The loop index (ind) is a phi node.
82 PHINode *ind = loop.CreatePHI(indType, 0);
83 // Incoming value for ind is 0
84 ind->addIncoming(ConstantInt::get(indType, 0), origBB);
86 // load from srcAddr+ind
87 Value *val = loop.CreateLoad(loop.CreateGEP(loop.getInt8Ty(), srcAddr, ind),
89 // store at dstAddr+ind
90 loop.CreateStore(val, loop.CreateGEP(loop.getInt8Ty(), dstAddr, ind),
93 // The value for ind coming from backedge is (ind + 1)
94 Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1));
95 ind->addIncoming(newind, loopBB);
97 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
100 // Lower MemSetInst to loop
101 static void convertMemSetToLoop(Instruction *splitAt, Value *dstAddr,
102 Value *len, Value *val, LLVMContext &Context,
104 BasicBlock *origBB = splitAt->getParent();
105 BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
106 BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
108 origBB->getTerminator()->setSuccessor(0, loopBB);
109 IRBuilder<> builder(origBB, origBB->getTerminator());
111 unsigned dstAS = cast<PointerType>(dstAddr->getType())->getAddressSpace();
113 // Cast pointer to the type of value getting stored
115 builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS));
117 IRBuilder<> loop(loopBB);
118 PHINode *ind = loop.CreatePHI(len->getType(), 0);
119 ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB);
121 loop.CreateStore(val, loop.CreateGEP(val->getType(), dstAddr, ind), false);
123 Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1));
124 ind->addIncoming(newind, loopBB);
126 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
129 bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
130 SmallVector<LoadInst *, 4> aggrLoads;
131 SmallVector<MemTransferInst *, 4> aggrMemcpys;
132 SmallVector<MemSetInst *, 4> aggrMemsets;
134 const DataLayout &DL = F.getParent()->getDataLayout();
135 LLVMContext &Context = F.getParent()->getContext();
138 // Collect all the aggrLoads, aggrMemcpys and addrMemsets.
140 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
141 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
143 if (LoadInst *load = dyn_cast<LoadInst>(II)) {
144 if (!load->hasOneUse())
147 if (DL.getTypeStoreSize(load->getType()) < MaxAggrCopySize)
150 User *use = load->user_back();
151 if (StoreInst *store = dyn_cast<StoreInst>(use)) {
152 if (store->getOperand(0) != load)
154 aggrLoads.push_back(load);
156 } else if (MemTransferInst *intr = dyn_cast<MemTransferInst>(II)) {
157 Value *len = intr->getLength();
158 // If the number of elements being copied is greater
159 // than MaxAggrCopySize, lower it to a loop
160 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
161 if (len_int->getZExtValue() >= MaxAggrCopySize) {
162 aggrMemcpys.push_back(intr);
165 // turn variable length memcpy/memmov into loop
166 aggrMemcpys.push_back(intr);
168 } else if (MemSetInst *memsetintr = dyn_cast<MemSetInst>(II)) {
169 Value *len = memsetintr->getLength();
170 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
171 if (len_int->getZExtValue() >= MaxAggrCopySize) {
172 aggrMemsets.push_back(memsetintr);
175 // turn variable length memset into loop
176 aggrMemsets.push_back(memsetintr);
181 if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0) &&
182 (aggrMemsets.size() == 0))
186 // Do the transformation of an aggr load/copy/set to a loop
188 for (LoadInst *load : aggrLoads) {
189 StoreInst *store = dyn_cast<StoreInst>(*load->user_begin());
190 Value *srcAddr = load->getOperand(0);
191 Value *dstAddr = store->getOperand(1);
192 unsigned numLoads = DL.getTypeStoreSize(load->getType());
193 Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);
195 convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
196 store->isVolatile(), Context, F);
198 store->eraseFromParent();
199 load->eraseFromParent();
202 for (MemTransferInst *cpy : aggrMemcpys) {
203 Value *len = cpy->getLength();
204 // llvm 2.7 version of memcpy does not have volatile
205 // operand yet. So always making it non-volatile
206 // optimistically, so that we don't see unnecessary
207 // st.volatile in ptx
208 convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
210 cpy->eraseFromParent();
213 for (MemSetInst *memsetinst : aggrMemsets) {
214 Value *len = memsetinst->getLength();
215 Value *val = memsetinst->getValue();
216 convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
218 memsetinst->eraseFromParent();
224 FunctionPass *llvm::createLowerAggrCopies() {
225 return new NVPTXLowerAggrCopies();