1 //===- NVPTXLowerAggrCopies.cpp - ------------------------------*- C++ -*--===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // Lower aggregate copies, memset, memcpy, memmov intrinsics into loops when 10 // the size is large or is not a compile-time constant. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "NVPTXLowerAggrCopies.h" 15 #include "llvm/IR/Constants.h" 16 #include "llvm/IR/DataLayout.h" 17 #include "llvm/IR/Function.h" 18 #include "llvm/IR/IRBuilder.h" 19 #include "llvm/IR/Instructions.h" 20 #include "llvm/IR/IntrinsicInst.h" 21 #include "llvm/IR/Intrinsics.h" 22 #include "llvm/IR/LLVMContext.h" 23 #include "llvm/IR/Module.h" 24 #include "llvm/Support/InstIterator.h" 25 26 using namespace llvm; 27 28 namespace llvm { FunctionPass *createLowerAggrCopies(); } 29 30 char NVPTXLowerAggrCopies::ID = 0; 31 32 // Lower MemTransferInst or load-store pair to loop 33 static void convertTransferToLoop( 34 Instruction *splitAt, Value *srcAddr, Value *dstAddr, Value *len, 35 //unsigned numLoads, 36 bool srcVolatile, bool dstVolatile, LLVMContext &Context, Function &F) { 37 Type *indType = len->getType(); 38 39 BasicBlock *origBB = splitAt->getParent(); 40 BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split"); 41 BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB); 42 43 origBB->getTerminator()->setSuccessor(0, loopBB); 44 IRBuilder<> builder(origBB, origBB->getTerminator()); 45 46 // srcAddr and dstAddr are expected to be pointer types, 47 // so no check is made here. 48 unsigned srcAS = dyn_cast<PointerType>(srcAddr->getType())->getAddressSpace(); 49 unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace(); 50 51 // Cast pointers to (char *) 52 srcAddr = builder.CreateBitCast(srcAddr, Type::getInt8PtrTy(Context, srcAS)); 53 dstAddr = builder.CreateBitCast(dstAddr, Type::getInt8PtrTy(Context, dstAS)); 54 55 IRBuilder<> loop(loopBB); 56 // The loop index (ind) is a phi node. 57 PHINode *ind = loop.CreatePHI(indType, 0); 58 // Incoming value for ind is 0 59 ind->addIncoming(ConstantInt::get(indType, 0), origBB); 60 61 // load from srcAddr+ind 62 Value *val = loop.CreateLoad(loop.CreateGEP(srcAddr, ind), srcVolatile); 63 // store at dstAddr+ind 64 loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), dstVolatile); 65 66 // The value for ind coming from backedge is (ind + 1) 67 Value *newind = loop.CreateAdd(ind, ConstantInt::get(indType, 1)); 68 ind->addIncoming(newind, loopBB); 69 70 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB); 71 } 72 73 // Lower MemSetInst to loop 74 static void convertMemSetToLoop(Instruction *splitAt, Value *dstAddr, 75 Value *len, Value *val, LLVMContext &Context, 76 Function &F) { 77 BasicBlock *origBB = splitAt->getParent(); 78 BasicBlock *newBB = splitAt->getParent()->splitBasicBlock(splitAt, "split"); 79 BasicBlock *loopBB = BasicBlock::Create(Context, "loadstoreloop", &F, newBB); 80 81 origBB->getTerminator()->setSuccessor(0, loopBB); 82 IRBuilder<> builder(origBB, origBB->getTerminator()); 83 84 unsigned dstAS = dyn_cast<PointerType>(dstAddr->getType())->getAddressSpace(); 85 86 // Cast pointer to the type of value getting stored 87 dstAddr = 88 builder.CreateBitCast(dstAddr, PointerType::get(val->getType(), dstAS)); 89 90 IRBuilder<> loop(loopBB); 91 PHINode *ind = loop.CreatePHI(len->getType(), 0); 92 ind->addIncoming(ConstantInt::get(len->getType(), 0), origBB); 93 94 loop.CreateStore(val, loop.CreateGEP(dstAddr, ind), false); 95 96 Value *newind = loop.CreateAdd(ind, ConstantInt::get(len->getType(), 1)); 97 ind->addIncoming(newind, loopBB); 98 99 loop.CreateCondBr(loop.CreateICmpULT(newind, len), loopBB, newBB); 100 } 101 102 bool NVPTXLowerAggrCopies::runOnFunction(Function &F) { 103 SmallVector<LoadInst *, 4> aggrLoads; 104 SmallVector<MemTransferInst *, 4> aggrMemcpys; 105 SmallVector<MemSetInst *, 4> aggrMemsets; 106 107 DataLayout *TD = &getAnalysis<DataLayout>(); 108 LLVMContext &Context = F.getParent()->getContext(); 109 110 // 111 // Collect all the aggrLoads, aggrMemcpys and addrMemsets. 112 // 113 //const BasicBlock *firstBB = &F.front(); // first BB in F 114 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 115 //BasicBlock *bb = BI; 116 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; 117 ++II) { 118 if (LoadInst *load = dyn_cast<LoadInst>(II)) { 119 120 if (load->hasOneUse() == false) 121 continue; 122 123 if (TD->getTypeStoreSize(load->getType()) < MaxAggrCopySize) 124 continue; 125 126 User *use = *(load->use_begin()); 127 if (StoreInst *store = dyn_cast<StoreInst>(use)) { 128 if (store->getOperand(0) != load) //getValueOperand 129 continue; 130 aggrLoads.push_back(load); 131 } 132 } else if (MemTransferInst *intr = dyn_cast<MemTransferInst>(II)) { 133 Value *len = intr->getLength(); 134 // If the number of elements being copied is greater 135 // than MaxAggrCopySize, lower it to a loop 136 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) { 137 if (len_int->getZExtValue() >= MaxAggrCopySize) { 138 aggrMemcpys.push_back(intr); 139 } 140 } else { 141 // turn variable length memcpy/memmov into loop 142 aggrMemcpys.push_back(intr); 143 } 144 } else if (MemSetInst *memsetintr = dyn_cast<MemSetInst>(II)) { 145 Value *len = memsetintr->getLength(); 146 if (ConstantInt *len_int = dyn_cast<ConstantInt>(len)) { 147 if (len_int->getZExtValue() >= MaxAggrCopySize) { 148 aggrMemsets.push_back(memsetintr); 149 } 150 } else { 151 // turn variable length memset into loop 152 aggrMemsets.push_back(memsetintr); 153 } 154 } 155 } 156 } 157 if ((aggrLoads.size() == 0) && (aggrMemcpys.size() == 0) && 158 (aggrMemsets.size() == 0)) 159 return false; 160 161 // 162 // Do the transformation of an aggr load/copy/set to a loop 163 // 164 for (unsigned i = 0, e = aggrLoads.size(); i != e; ++i) { 165 LoadInst *load = aggrLoads[i]; 166 StoreInst *store = dyn_cast<StoreInst>(*load->use_begin()); 167 Value *srcAddr = load->getOperand(0); 168 Value *dstAddr = store->getOperand(1); 169 unsigned numLoads = TD->getTypeStoreSize(load->getType()); 170 Value *len = ConstantInt::get(Type::getInt32Ty(Context), numLoads); 171 172 convertTransferToLoop(store, srcAddr, dstAddr, len, load->isVolatile(), 173 store->isVolatile(), Context, F); 174 175 store->eraseFromParent(); 176 load->eraseFromParent(); 177 } 178 179 for (unsigned i = 0, e = aggrMemcpys.size(); i != e; ++i) { 180 MemTransferInst *cpy = aggrMemcpys[i]; 181 Value *len = cpy->getLength(); 182 // llvm 2.7 version of memcpy does not have volatile 183 // operand yet. So always making it non-volatile 184 // optimistically, so that we don't see unnecessary 185 // st.volatile in ptx 186 convertTransferToLoop(cpy, cpy->getSource(), cpy->getDest(), len, false, 187 false, Context, F); 188 cpy->eraseFromParent(); 189 } 190 191 for (unsigned i = 0, e = aggrMemsets.size(); i != e; ++i) { 192 MemSetInst *memsetinst = aggrMemsets[i]; 193 Value *len = memsetinst->getLength(); 194 Value *val = memsetinst->getValue(); 195 convertMemSetToLoop(memsetinst, memsetinst->getDest(), len, val, Context, 196 F); 197 memsetinst->eraseFromParent(); 198 } 199 200 return true; 201 } 202 203 FunctionPass *llvm::createLowerAggrCopies() { 204 return new NVPTXLowerAggrCopies(); 205 } 206