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 STATISTIC(NumPrefetches, "Number of prefetches inserted"); 47 48 namespace llvm { 49 void initializeLoopDataPrefetchPass(PassRegistry&); 50 } 51 52 namespace { 53 54 class LoopDataPrefetch : public FunctionPass { 55 public: 56 static char ID; // Pass ID, replacement for typeid 57 LoopDataPrefetch() : FunctionPass(ID) { 58 initializeLoopDataPrefetchPass(*PassRegistry::getPassRegistry()); 59 } 60 61 void getAnalysisUsage(AnalysisUsage &AU) const override { 62 AU.addRequired<AssumptionCacheTracker>(); 63 AU.addPreserved<DominatorTreeWrapperPass>(); 64 AU.addRequired<LoopInfoWrapperPass>(); 65 AU.addPreserved<LoopInfoWrapperPass>(); 66 AU.addRequired<ScalarEvolutionWrapperPass>(); 67 // FIXME: For some reason, preserving SE here breaks LSR (even if 68 // this pass changes nothing). 69 // AU.addPreserved<ScalarEvolutionWrapperPass>(); 70 AU.addRequired<TargetTransformInfoWrapperPass>(); 71 } 72 73 bool runOnFunction(Function &F) override; 74 bool runOnLoop(Loop *L); 75 76 /// \brief Check if the the stride of the accesses is large enough to 77 /// warrant a prefetch. 78 bool isStrideLargeEnough(const SCEVAddRecExpr *AR); 79 80 private: 81 AssumptionCache *AC; 82 LoopInfo *LI; 83 ScalarEvolution *SE; 84 const TargetTransformInfo *TTI; 85 const DataLayout *DL; 86 }; 87 } 88 89 char LoopDataPrefetch::ID = 0; 90 INITIALIZE_PASS_BEGIN(LoopDataPrefetch, "loop-data-prefetch", 91 "Loop Data Prefetch", false, false) 92 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 93 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 94 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 95 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass) 96 INITIALIZE_PASS_END(LoopDataPrefetch, "loop-data-prefetch", 97 "Loop Data Prefetch", false, false) 98 99 FunctionPass *llvm::createLoopDataPrefetchPass() { return new LoopDataPrefetch(); } 100 101 bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR) { 102 unsigned TargetMinStride = TTI->getMinPrefetchStride(); 103 // No need to check if any stride goes. 104 if (TargetMinStride <= 1) 105 return true; 106 107 const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 108 // If MinStride is set, don't prefetch unless we can ensure that stride is 109 // larger. 110 if (!ConstStride) 111 return false; 112 113 unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue()); 114 return TargetMinStride <= AbsStride; 115 } 116 117 bool LoopDataPrefetch::runOnFunction(Function &F) { 118 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 119 SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 120 DL = &F.getParent()->getDataLayout(); 121 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 122 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 123 124 // If PrefetchDistance is not set, don't run the pass. This gives an 125 // opportunity for targets to run this pass for selected subtargets only 126 // (whose TTI sets PrefetchDistance). 127 if (TTI->getPrefetchDistance() == 0) 128 return false; 129 assert(TTI->getCacheLineSize() && "Cache line size is not set for target"); 130 131 bool MadeChange = false; 132 133 for (auto I = LI->begin(), IE = LI->end(); I != IE; ++I) 134 for (auto L = df_begin(*I), LE = df_end(*I); L != LE; ++L) 135 MadeChange |= runOnLoop(*L); 136 137 return MadeChange; 138 } 139 140 bool LoopDataPrefetch::runOnLoop(Loop *L) { 141 bool MadeChange = false; 142 143 // Only prefetch in the inner-most loop 144 if (!L->empty()) 145 return MadeChange; 146 147 SmallPtrSet<const Value *, 32> EphValues; 148 CodeMetrics::collectEphemeralValues(L, AC, EphValues); 149 150 // Calculate the number of iterations ahead to prefetch 151 CodeMetrics Metrics; 152 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end(); 153 I != IE; ++I) { 154 155 // If the loop already has prefetches, then assume that the user knows 156 // what he or she is doing and don't add any more. 157 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end(); 158 J != JE; ++J) 159 if (CallInst *CI = dyn_cast<CallInst>(J)) 160 if (Function *F = CI->getCalledFunction()) 161 if (F->getIntrinsicID() == Intrinsic::prefetch) 162 return MadeChange; 163 164 Metrics.analyzeBasicBlock(*I, *TTI, EphValues); 165 } 166 unsigned LoopSize = Metrics.NumInsts; 167 if (!LoopSize) 168 LoopSize = 1; 169 170 unsigned ItersAhead = TTI->getPrefetchDistance() / LoopSize; 171 if (!ItersAhead) 172 ItersAhead = 1; 173 174 if (ItersAhead > TTI->getMaxPrefetchIterationsAhead()) 175 return MadeChange; 176 177 DEBUG(dbgs() << "Prefetching " << ItersAhead 178 << " iterations ahead (loop size: " << LoopSize << ") in " 179 << L->getHeader()->getParent()->getName() << ": " << *L); 180 181 SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads; 182 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end(); 183 I != IE; ++I) { 184 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end(); 185 J != JE; ++J) { 186 Value *PtrValue; 187 Instruction *MemI; 188 189 if (LoadInst *LMemI = dyn_cast<LoadInst>(J)) { 190 MemI = LMemI; 191 PtrValue = LMemI->getPointerOperand(); 192 } else if (StoreInst *SMemI = dyn_cast<StoreInst>(J)) { 193 if (!PrefetchWrites) continue; 194 MemI = SMemI; 195 PtrValue = SMemI->getPointerOperand(); 196 } else continue; 197 198 unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace(); 199 if (PtrAddrSpace) 200 continue; 201 202 if (L->isLoopInvariant(PtrValue)) 203 continue; 204 205 const SCEV *LSCEV = SE->getSCEV(PtrValue); 206 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV); 207 if (!LSCEVAddRec) 208 continue; 209 210 // Check if the the stride of the accesses is large enough to warrant a 211 // prefetch. 212 if (!isStrideLargeEnough(LSCEVAddRec)) 213 continue; 214 215 // We don't want to double prefetch individual cache lines. If this load 216 // is known to be within one cache line of some other load that has 217 // already been prefetched, then don't prefetch this one as well. 218 bool DupPref = false; 219 for (SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 220 16>::iterator K = PrefLoads.begin(), KE = PrefLoads.end(); 221 K != KE; ++K) { 222 const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, K->second); 223 if (const SCEVConstant *ConstPtrDiff = 224 dyn_cast<SCEVConstant>(PtrDiff)) { 225 int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue()); 226 if (PD < (int64_t) TTI->getCacheLineSize()) { 227 DupPref = true; 228 break; 229 } 230 } 231 } 232 if (DupPref) 233 continue; 234 235 const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr( 236 SE->getConstant(LSCEVAddRec->getType(), ItersAhead), 237 LSCEVAddRec->getStepRecurrence(*SE))); 238 if (!isSafeToExpand(NextLSCEV, *SE)) 239 continue; 240 241 PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec)); 242 243 Type *I8Ptr = Type::getInt8PtrTy((*I)->getContext(), PtrAddrSpace); 244 SCEVExpander SCEVE(*SE, J->getModule()->getDataLayout(), "prefaddr"); 245 Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI); 246 247 IRBuilder<> Builder(MemI); 248 Module *M = (*I)->getParent()->getParent(); 249 Type *I32 = Type::getInt32Ty((*I)->getContext()); 250 Value *PrefetchFunc = Intrinsic::getDeclaration(M, Intrinsic::prefetch); 251 Builder.CreateCall( 252 PrefetchFunc, 253 {PrefPtrValue, 254 ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1), 255 ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)}); 256 ++NumPrefetches; 257 DEBUG(dbgs() << " Access: " << *PtrValue << ", SCEV: " << *LSCEV 258 << "\n"); 259 260 MadeChange = true; 261 } 262 } 263 264 return MadeChange; 265 } 266 267