1 //===-------- LoopDataPrefetch.cpp - Loop Data Prefetching Pass -----------===//
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 //===----------------------------------------------------------------------===//
10 // This file implements a Loop Data Prefetching Pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar/LoopDataPrefetch.h"
16 #define DEBUG_TYPE "loop-data-prefetch"
17 #include "llvm/ADT/DepthFirstIterator.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AssumptionCache.h"
20 #include "llvm/Analysis/CodeMetrics.h"
21 #include "llvm/Analysis/InstructionSimplify.h"
22 #include "llvm/Analysis/LoopInfo.h"
23 #include "llvm/Analysis/OptimizationDiagnosticInfo.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
26 #include "llvm/Analysis/ScalarEvolutionExpander.h"
27 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
28 #include "llvm/Analysis/TargetTransformInfo.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/IR/CFG.h"
31 #include "llvm/IR/Dominators.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 #include "llvm/Transforms/Utils/ValueMapper.h"
43 // By default, we limit this to creating 16 PHIs (which is a little over half
44 // of the allocatable register set).
46 PrefetchWrites("loop-prefetch-writes", cl::Hidden, cl::init(false),
47 cl::desc("Prefetch write addresses"));
49 static cl::opt<unsigned>
50 PrefetchDistance("prefetch-distance",
51 cl::desc("Number of instructions to prefetch ahead"),
54 static cl::opt<unsigned>
55 MinPrefetchStride("min-prefetch-stride",
56 cl::desc("Min stride to add prefetches"), cl::Hidden);
58 static cl::opt<unsigned> MaxPrefetchIterationsAhead(
59 "max-prefetch-iters-ahead",
60 cl::desc("Max number of iterations to prefetch ahead"), cl::Hidden);
62 STATISTIC(NumPrefetches, "Number of prefetches inserted");
66 /// Loop prefetch implementation class.
67 class LoopDataPrefetch {
69 LoopDataPrefetch(AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE,
70 const TargetTransformInfo *TTI,
71 OptimizationRemarkEmitter *ORE)
72 : AC(AC), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
77 bool runOnLoop(Loop *L);
79 /// \brief Check if the the stride of the accesses is large enough to
80 /// warrant a prefetch.
81 bool isStrideLargeEnough(const SCEVAddRecExpr *AR);
83 unsigned getMinPrefetchStride() {
84 if (MinPrefetchStride.getNumOccurrences() > 0)
85 return MinPrefetchStride;
86 return TTI->getMinPrefetchStride();
89 unsigned getPrefetchDistance() {
90 if (PrefetchDistance.getNumOccurrences() > 0)
91 return PrefetchDistance;
92 return TTI->getPrefetchDistance();
95 unsigned getMaxPrefetchIterationsAhead() {
96 if (MaxPrefetchIterationsAhead.getNumOccurrences() > 0)
97 return MaxPrefetchIterationsAhead;
98 return TTI->getMaxPrefetchIterationsAhead();
104 const TargetTransformInfo *TTI;
105 OptimizationRemarkEmitter *ORE;
108 /// Legacy class for inserting loop data prefetches.
109 class LoopDataPrefetchLegacyPass : public FunctionPass {
111 static char ID; // Pass ID, replacement for typeid
112 LoopDataPrefetchLegacyPass() : FunctionPass(ID) {
113 initializeLoopDataPrefetchLegacyPassPass(*PassRegistry::getPassRegistry());
116 void getAnalysisUsage(AnalysisUsage &AU) const override {
117 AU.addRequired<AssumptionCacheTracker>();
118 AU.addPreserved<DominatorTreeWrapperPass>();
119 AU.addRequired<LoopInfoWrapperPass>();
120 AU.addPreserved<LoopInfoWrapperPass>();
121 AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
122 AU.addRequired<ScalarEvolutionWrapperPass>();
123 // FIXME: For some reason, preserving SE here breaks LSR (even if
124 // this pass changes nothing).
125 // AU.addPreserved<ScalarEvolutionWrapperPass>();
126 AU.addRequired<TargetTransformInfoWrapperPass>();
129 bool runOnFunction(Function &F) override;
133 char LoopDataPrefetchLegacyPass::ID = 0;
134 INITIALIZE_PASS_BEGIN(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
135 "Loop Data Prefetch", false, false)
136 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
137 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
138 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
139 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
140 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
141 INITIALIZE_PASS_END(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
142 "Loop Data Prefetch", false, false)
144 FunctionPass *llvm::createLoopDataPrefetchPass() {
145 return new LoopDataPrefetchLegacyPass();
148 bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR) {
149 unsigned TargetMinStride = getMinPrefetchStride();
150 // No need to check if any stride goes.
151 if (TargetMinStride <= 1)
154 const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
155 // If MinStride is set, don't prefetch unless we can ensure that stride is
160 unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue());
161 return TargetMinStride <= AbsStride;
164 PreservedAnalyses LoopDataPrefetchPass::run(Function &F,
165 FunctionAnalysisManager &AM) {
166 LoopInfo *LI = &AM.getResult<LoopAnalysis>(F);
167 ScalarEvolution *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
168 AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F);
169 OptimizationRemarkEmitter *ORE =
170 &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
171 const TargetTransformInfo *TTI = &AM.getResult<TargetIRAnalysis>(F);
173 LoopDataPrefetch LDP(AC, LI, SE, TTI, ORE);
174 bool Changed = LDP.run();
177 PreservedAnalyses PA;
178 PA.preserve<DominatorTreeAnalysis>();
179 PA.preserve<LoopAnalysis>();
183 return PreservedAnalyses::all();
186 bool LoopDataPrefetchLegacyPass::runOnFunction(Function &F) {
190 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
191 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
192 AssumptionCache *AC =
193 &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
194 OptimizationRemarkEmitter *ORE =
195 &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
196 const TargetTransformInfo *TTI =
197 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
199 LoopDataPrefetch LDP(AC, LI, SE, TTI, ORE);
203 bool LoopDataPrefetch::run() {
204 // If PrefetchDistance is not set, don't run the pass. This gives an
205 // opportunity for targets to run this pass for selected subtargets only
206 // (whose TTI sets PrefetchDistance).
207 if (getPrefetchDistance() == 0)
209 assert(TTI->getCacheLineSize() && "Cache line size is not set for target");
211 bool MadeChange = false;
214 for (auto L = df_begin(I), LE = df_end(I); L != LE; ++L)
215 MadeChange |= runOnLoop(*L);
220 bool LoopDataPrefetch::runOnLoop(Loop *L) {
221 bool MadeChange = false;
223 // Only prefetch in the inner-most loop
227 SmallPtrSet<const Value *, 32> EphValues;
228 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
230 // Calculate the number of iterations ahead to prefetch
232 for (const auto BB : L->blocks()) {
233 // If the loop already has prefetches, then assume that the user knows
234 // what they are doing and don't add any more.
236 if (CallInst *CI = dyn_cast<CallInst>(&I))
237 if (Function *F = CI->getCalledFunction())
238 if (F->getIntrinsicID() == Intrinsic::prefetch)
241 Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
243 unsigned LoopSize = Metrics.NumInsts;
247 unsigned ItersAhead = getPrefetchDistance() / LoopSize;
251 if (ItersAhead > getMaxPrefetchIterationsAhead())
254 DEBUG(dbgs() << "Prefetching " << ItersAhead
255 << " iterations ahead (loop size: " << LoopSize << ") in "
256 << L->getHeader()->getParent()->getName() << ": " << *L);
258 SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads;
259 for (const auto BB : L->blocks()) {
260 for (auto &I : *BB) {
264 if (LoadInst *LMemI = dyn_cast<LoadInst>(&I)) {
266 PtrValue = LMemI->getPointerOperand();
267 } else if (StoreInst *SMemI = dyn_cast<StoreInst>(&I)) {
268 if (!PrefetchWrites) continue;
270 PtrValue = SMemI->getPointerOperand();
273 unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
277 if (L->isLoopInvariant(PtrValue))
280 const SCEV *LSCEV = SE->getSCEV(PtrValue);
281 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
285 // Check if the the stride of the accesses is large enough to warrant a
287 if (!isStrideLargeEnough(LSCEVAddRec))
290 // We don't want to double prefetch individual cache lines. If this load
291 // is known to be within one cache line of some other load that has
292 // already been prefetched, then don't prefetch this one as well.
293 bool DupPref = false;
294 for (const auto &PrefLoad : PrefLoads) {
295 const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, PrefLoad.second);
296 if (const SCEVConstant *ConstPtrDiff =
297 dyn_cast<SCEVConstant>(PtrDiff)) {
298 int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
299 if (PD < (int64_t) TTI->getCacheLineSize()) {
308 const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr(
309 SE->getConstant(LSCEVAddRec->getType(), ItersAhead),
310 LSCEVAddRec->getStepRecurrence(*SE)));
311 if (!isSafeToExpand(NextLSCEV, *SE))
314 PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec));
316 Type *I8Ptr = Type::getInt8PtrTy(BB->getContext(), PtrAddrSpace);
317 SCEVExpander SCEVE(*SE, I.getModule()->getDataLayout(), "prefaddr");
318 Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI);
320 IRBuilder<> Builder(MemI);
321 Module *M = BB->getParent()->getParent();
322 Type *I32 = Type::getInt32Ty(BB->getContext());
323 Value *PrefetchFunc = Intrinsic::getDeclaration(M, Intrinsic::prefetch);
327 ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1),
328 ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)});
330 DEBUG(dbgs() << " Access: " << *PtrValue << ", SCEV: " << *LSCEV
332 ORE->emit(OptimizationRemark(DEBUG_TYPE, "Prefetched", MemI)
333 << "prefetched memory access");