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 = dyn_cast<PointerType>(srcAddr->getType())->getAddressSpace();
74 unsigned dstAS = dyn_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(srcAddr, ind), srcVolatile);
88 // store at dstAddr+ind
89 loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), dstVolatile);
91 // The value for ind coming from backedge is (ind + 1)
92 Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1));
93 ind->addIncoming(newind, loopBB);
95 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
98 // Lower MemSetInst to loop
99 static void convertMemSetToLoop(Instruction *splitAt, Value *dstAddr,
100 Value *len, Value *val, LLVMContext &Context,
102 BasicBlock *origBB = splitAt->getParent();
103 BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split");
104 BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB);
106 origBB->getTerminator()->setSuccessor(0, loopBB);
107 IRBuilder<> builder(origBB, origBB->getTerminator());
109 unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace();
111 // Cast pointer to the type of value getting stored
113 builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS));
115 IRBuilder<> loop(loopBB);
116 PHINode *ind = loop.CreatePHI(len->getType(), 0);
117 ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB);
119 loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), false);
121 Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1));
122 ind->addIncoming(newind, loopBB);
124 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB);
127 bool NVPTXLowerAggrCopies::runOnFunction(Function &F) {
128 SmallVector<LoadInst *, 4> aggrLoads;
129 SmallVector<MemTransferInst *, 4> aggrMemcpys;
130 SmallVector<MemSetInst *, 4> aggrMemsets;
132 const DataLayout &DL = F.getParent()->getDataLayout();
133 LLVMContext &Context = F.getParent()->getContext();
136 // Collect all the aggrLoads, aggrMemcpys and addrMemsets.
138 //const BasicBlock *firstBB = &F.front(); // first BB in F
139 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
140 //BasicBlock *bb = BI;
141 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE;
143 if (LoadInst *load = dyn_cast<LoadInst>(II)) {
145 if (!load->hasOneUse())
148 if (DL.getTypeStoreSize(load->getType()) < MaxAggrCopySize)
151 User *use = load->user_back();
152 if (StoreInst *store = dyn_cast<StoreInst>(use)) {
153 if (store->getOperand(0) != load) //getValueOperand
155 aggrLoads.push_back(load);
157 } else if (MemTransferInst *intr = dyn_cast<MemTransferInst>(II)) {
158 Value *len = intr->getLength();
159 // If the number of elements being copied is greater
160 // than MaxAggrCopySize, lower it to a loop
161 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
162 if (len_int->getZExtValue() >= MaxAggrCopySize) {
163 aggrMemcpys.push_back(intr);
166 // turn variable length memcpy/memmov into loop
167 aggrMemcpys.push_back(intr);
169 } else if (MemSetInst *memsetintr = dyn_cast<MemSetInst>(II)) {
170 Value *len = memsetintr->getLength();
171 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) {
172 if (len_int->getZExtValue() >= MaxAggrCopySize) {
173 aggrMemsets.push_back(memsetintr);
176 // turn variable length memset into loop
177 aggrMemsets.push_back(memsetintr);
182 if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0) &&
183 (aggrMemsets.size() == 0))
187 // Do the transformation of an aggr load/copy/set to a loop
189 for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) {
190 LoadInst *load = aggrLoads[i];
191 StoreInst *store = dyn_cast<StoreInst>(*load->user_begin());
192 Value *srcAddr = load->getOperand(0);
193 Value *dstAddr = store->getOperand(1);
194 unsigned numLoads = DL.getTypeStoreSize(load->getType());
195 Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads);
197 convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(),
198 store->isVolatile(), Context, F);
200 store->eraseFromParent();
201 load->eraseFromParent();
204 for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) {
205 MemTransferInst *cpy = aggrMemcpys[i];
206 Value *len = cpy->getLength();
207 // llvm 2.7 version of memcpy does not have volatile
208 // operand yet. So always making it non-volatile
209 // optimistically, so that we don't see unnecessary
210 // st.volatile in ptx
211 convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false,
213 cpy->eraseFromParent();
216 for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) {
217 MemSetInst *memsetinst = aggrMemsets[i];
218 Value *len = memsetinst->getLength();
219 Value *val = memsetinst->getValue();
220 convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context,
222 memsetinst->eraseFromParent();
228 FunctionPass *llvm::createLowerAggrCopies() {
229 return new NVPTXLowerAggrCopies();