1 //===-------- LoopDataPrefetch.cpp - Loop Data Prefetching Pass -----------===// 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 // 10 // This file implements a Loop Data Prefetching Pass. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #define DEBUG_TYPE "loop-data-prefetch" 15 #include "llvm/Transforms/Scalar.h" 16 #include "llvm/ADT/DepthFirstIterator.h" 17 #include "llvm/ADT/Statistic.h" 18 #include "llvm/Analysis/AssumptionCache.h" 19 #include "llvm/Analysis/CodeMetrics.h" 20 #include "llvm/Analysis/InstructionSimplify.h" 21 #include "llvm/Analysis/LoopInfo.h" 22 #include "llvm/Analysis/ScalarEvolution.h" 23 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 24 #include "llvm/Analysis/ScalarEvolutionExpander.h" 25 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 26 #include "llvm/Analysis/TargetTransformInfo.h" 27 #include "llvm/Analysis/ValueTracking.h" 28 #include "llvm/IR/CFG.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/Function.h" 31 #include "llvm/IR/IntrinsicInst.h" 32 #include "llvm/IR/Module.h" 33 #include "llvm/Support/CommandLine.h" 34 #include "llvm/Support/Debug.h" 35 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 36 #include "llvm/Transforms/Utils/Local.h" 37 #include "llvm/Transforms/Utils/ValueMapper.h" 38 using namespace llvm; 39 40 // By default, we limit this to creating 16 PHIs (which is a little over half 41 // of the allocatable register set). 42 static cl::opt<bool> 43 PrefetchWrites("loop-prefetch-writes", cl::Hidden, cl::init(false), 44 cl::desc("Prefetch write addresses")); 45 46 namespace llvm { 47 void initializeLoopDataPrefetchPass(PassRegistry&); 48 } 49 50 namespace { 51 52 class LoopDataPrefetch : public FunctionPass { 53 public: 54 static char ID; // Pass ID, replacement for typeid 55 LoopDataPrefetch() : FunctionPass(ID) { 56 initializeLoopDataPrefetchPass(*PassRegistry::getPassRegistry()); 57 } 58 59 void getAnalysisUsage(AnalysisUsage &AU) const override { 60 AU.addRequired<AssumptionCacheTracker>(); 61 AU.addPreserved<DominatorTreeWrapperPass>(); 62 AU.addRequired<LoopInfoWrapperPass>(); 63 AU.addPreserved<LoopInfoWrapperPass>(); 64 AU.addRequired<ScalarEvolutionWrapperPass>(); 65 // FIXME: For some reason, preserving SE here breaks LSR (even if 66 // this pass changes nothing). 67 // AU.addPreserved<ScalarEvolutionWrapperPass>(); 68 AU.addRequired<TargetTransformInfoWrapperPass>(); 69 } 70 71 bool runOnFunction(Function &F) override; 72 bool runOnLoop(Loop *L); 73 74 private: 75 AssumptionCache *AC; 76 LoopInfo *LI; 77 ScalarEvolution *SE; 78 const TargetTransformInfo *TTI; 79 const DataLayout *DL; 80 }; 81 } 82 83 char LoopDataPrefetch::ID = 0; 84 INITIALIZE_PASS_BEGIN(LoopDataPrefetch, "loop-data-prefetch", 85 "Loop Data Prefetch", false, false) 86 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 87 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 88 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 89 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 90 INITIALIZE_PASS_END(LoopDataPrefetch, "loop-data-prefetch", 91 "Loop Data Prefetch", false, false) 92 93 FunctionPass *llvm::createLoopDataPrefetchPass() { return new LoopDataPrefetch(); } 94 95 bool LoopDataPrefetch::runOnFunction(Function &F) { 96 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 97 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 98 DL = &F.getParent()->getDataLayout(); 99 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 100 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 101 102 // If PrefetchDistance is not set, don't run the pass. This gives an 103 // opportunity for targets to run this pass for selected subtargets only 104 // (whose TTI sets PrefetchDistance). 105 if (TTI->getPrefetchDistance() == 0) 106 return false; 107 assert(TTI->getCacheLineSize() && "Cache line size is not set for target"); 108 109 bool MadeChange = false; 110 111 for (auto I = LI->begin(), IE = LI->end(); I != IE; ++I) 112 for (auto L = df_begin(*I), LE = df_end(*I); L != LE; ++L) 113 MadeChange |= runOnLoop(*L); 114 115 return MadeChange; 116 } 117 118 bool LoopDataPrefetch::runOnLoop(Loop *L) { 119 bool MadeChange = false; 120 121 // Only prefetch in the inner-most loop 122 if (!L->empty()) 123 return MadeChange; 124 125 SmallPtrSet<const Value *, 32> EphValues; 126 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 127 128 // Calculate the number of iterations ahead to prefetch 129 CodeMetrics Metrics; 130 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end(); 131 I != IE; ++I) { 132 133 // If the loop already has prefetches, then assume that the user knows 134 // what he or she is doing and don't add any more. 135 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end(); 136 J != JE; ++J) 137 if (CallInst *CI = dyn_cast<CallInst>(J)) 138 if (Function *F = CI->getCalledFunction()) 139 if (F->getIntrinsicID() == Intrinsic::prefetch) 140 return MadeChange; 141 142 Metrics.analyzeBasicBlock(*I, *TTI, EphValues); 143 } 144 unsigned LoopSize = Metrics.NumInsts; 145 if (!LoopSize) 146 LoopSize = 1; 147 148 unsigned ItersAhead = TTI->getPrefetchDistance() / LoopSize; 149 if (!ItersAhead) 150 ItersAhead = 1; 151 152 SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads; 153 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end(); 154 I != IE; ++I) { 155 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end(); 156 J != JE; ++J) { 157 Value *PtrValue; 158 Instruction *MemI; 159 160 if (LoadInst *LMemI = dyn_cast<LoadInst>(J)) { 161 MemI = LMemI; 162 PtrValue = LMemI->getPointerOperand(); 163 } else if (StoreInst *SMemI = dyn_cast<StoreInst>(J)) { 164 if (!PrefetchWrites) continue; 165 MemI = SMemI; 166 PtrValue = SMemI->getPointerOperand(); 167 } else continue; 168 169 unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace(); 170 if (PtrAddrSpace) 171 continue; 172 173 if (L->isLoopInvariant(PtrValue)) 174 continue; 175 176 const SCEV *LSCEV = SE->getSCEV(PtrValue); 177 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV); 178 if (!LSCEVAddRec) 179 continue; 180 181 // We don't want to double prefetch individual cache lines. If this load 182 // is known to be within one cache line of some other load that has 183 // already been prefetched, then don't prefetch this one as well. 184 bool DupPref = false; 185 for (SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 186 16>::iterator K = PrefLoads.begin(), KE = PrefLoads.end(); 187 K != KE; ++K) { 188 const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, K->second); 189 if (const SCEVConstant *ConstPtrDiff = 190 dyn_cast<SCEVConstant>(PtrDiff)) { 191 int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue()); 192 if (PD < (int64_t) TTI->getCacheLineSize()) { 193 DupPref = true; 194 break; 195 } 196 } 197 } 198 if (DupPref) 199 continue; 200 201 const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr( 202 SE->getConstant(LSCEVAddRec->getType(), ItersAhead), 203 LSCEVAddRec->getStepRecurrence(*SE))); 204 if (!isSafeToExpand(NextLSCEV, *SE)) 205 continue; 206 207 PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec)); 208 209 Type *I8Ptr = Type::getInt8PtrTy((*I)->getContext(), PtrAddrSpace); 210 SCEVExpander SCEVE(*SE, J->getModule()->getDataLayout(), "prefaddr"); 211 Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI); 212 213 IRBuilder<> Builder(MemI); 214 Module *M = (*I)->getParent()->getParent(); 215 Type *I32 = Type::getInt32Ty((*I)->getContext()); 216 Value *PrefetchFunc = Intrinsic::getDeclaration(M, Intrinsic::prefetch); 217 Builder.CreateCall( 218 PrefetchFunc, 219 {PrefPtrValue, 220 ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1), 221 ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)}); 222 223 MadeChange = true; 224 } 225 } 226 227 return MadeChange; 228 } 229 230