1 //===-------- PPCLoopDataPrefetch.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 #define DEBUG_TYPE "ppc-loop-data-prefetch"
16 #include "llvm/Transforms/Scalar.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/ScalarEvolutionExpander.h"
24 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
25 #include "llvm/Analysis/TargetTransformInfo.h"
26 #include "llvm/Analysis/ValueTracking.h"
27 #include "llvm/IR/CFG.h"
28 #include "llvm/IR/Dominators.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/Local.h"
36 #include "llvm/Transforms/Utils/ValueMapper.h"
39 // By default, we limit this to creating 16 PHIs (which is a little over half
40 // of the allocatable register set).
42 PrefetchWrites("ppc-loop-prefetch-writes", cl::Hidden, cl::init(false),
43 cl::desc("Prefetch write addresses"));
45 // This seems like a reasonable default for the BG/Q (this pass is enabled, by
46 // default, only on the BG/Q).
47 static cl::opt<unsigned>
48 PrefDist("ppc-loop-prefetch-distance", cl::Hidden, cl::init(300),
49 cl::desc("The loop prefetch distance"));
51 static cl::opt<unsigned>
52 CacheLineSize("ppc-loop-prefetch-cache-line", cl::Hidden, cl::init(64),
53 cl::desc("The loop prefetch cache line size"));
56 void initializePPCLoopDataPrefetchPass(PassRegistry&);
61 class PPCLoopDataPrefetch : public FunctionPass {
63 static char ID; // Pass ID, replacement for typeid
64 PPCLoopDataPrefetch() : FunctionPass(ID) {
65 initializePPCLoopDataPrefetchPass(*PassRegistry::getPassRegistry());
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 AU.addRequired<AssumptionCacheTracker>();
70 AU.addPreserved<DominatorTreeWrapperPass>();
71 AU.addRequired<LoopInfoWrapperPass>();
72 AU.addPreserved<LoopInfoWrapperPass>();
73 AU.addRequired<ScalarEvolution>();
74 // FIXME: For some reason, preserving SE here breaks LSR (even if
75 // this pass changes nothing).
76 // AU.addPreserved<ScalarEvolution>();
77 AU.addRequired<TargetTransformInfoWrapperPass>();
80 bool runOnFunction(Function &F) override;
81 bool runOnLoop(Loop *L);
87 const TargetTransformInfo *TTI;
92 char PPCLoopDataPrefetch::ID = 0;
93 INITIALIZE_PASS_BEGIN(PPCLoopDataPrefetch, "ppc-loop-data-prefetch",
94 "PPC Loop Data Prefetch", false, false)
95 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
96 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
97 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
98 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
99 INITIALIZE_PASS_END(PPCLoopDataPrefetch, "ppc-loop-data-prefetch",
100 "PPC Loop Data Prefetch", false, false)
102 FunctionPass *llvm::createPPCLoopDataPrefetchPass() { return new PPCLoopDataPrefetch(); }
104 bool PPCLoopDataPrefetch::runOnFunction(Function &F) {
105 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
106 SE = &getAnalysis<ScalarEvolution>();
107 DL = &F.getParent()->getDataLayout();
108 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
109 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
111 bool MadeChange = false;
113 for (LoopInfo::iterator I = LI->begin(), E = LI->end();
116 MadeChange |= runOnLoop(L);
122 bool PPCLoopDataPrefetch::runOnLoop(Loop *L) {
123 bool MadeChange = false;
125 // Only prefetch in the inner-most loop
129 SmallPtrSet<const Value *, 32> EphValues;
130 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
132 // Calculate the number of iterations ahead to prefetch
134 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
137 // If the loop already has prefetches, then assume that the user knows
138 // what he or she is doing and don't add any more.
139 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end();
141 if (CallInst *CI = dyn_cast<CallInst>(J))
142 if (Function *F = CI->getCalledFunction())
143 if (F->getIntrinsicID() == Intrinsic::prefetch)
146 Metrics.analyzeBasicBlock(*I, *TTI, EphValues);
148 unsigned LoopSize = Metrics.NumInsts;
152 unsigned ItersAhead = PrefDist/LoopSize;
156 SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads;
157 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
159 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end();
164 if (LoadInst *LMemI = dyn_cast<LoadInst>(J)) {
166 PtrValue = LMemI->getPointerOperand();
167 } else if (StoreInst *SMemI = dyn_cast<StoreInst>(J)) {
168 if (!PrefetchWrites) continue;
170 PtrValue = SMemI->getPointerOperand();
173 unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
177 if (L->isLoopInvariant(PtrValue))
180 const SCEV *LSCEV = SE->getSCEV(PtrValue);
181 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
185 // We don't want to double prefetch individual cache lines. If this load
186 // is known to be within one cache line of some other load that has
187 // already been prefetched, then don't prefetch this one as well.
188 bool DupPref = false;
189 for (SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>,
190 16>::iterator K = PrefLoads.begin(), KE = PrefLoads.end();
192 const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, K->second);
193 if (const SCEVConstant *ConstPtrDiff =
194 dyn_cast<SCEVConstant>(PtrDiff)) {
195 int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
196 if (PD < (int64_t) CacheLineSize) {
205 const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr(
206 SE->getConstant(LSCEVAddRec->getType(), ItersAhead),
207 LSCEVAddRec->getStepRecurrence(*SE)));
208 if (!isSafeToExpand(NextLSCEV, *SE))
211 PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec));
213 Type *I8Ptr = Type::getInt8PtrTy((*I)->getContext(), PtrAddrSpace);
214 SCEVExpander SCEVE(*SE, J->getModule()->getDataLayout(), "prefaddr");
215 Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI);
217 IRBuilder<> Builder(MemI);
218 Module *M = (*I)->getParent()->getParent();
219 Type *I32 = Type::getInt32Ty((*I)->getContext());
220 Value *PrefetchFunc = Intrinsic::getDeclaration(M, Intrinsic::prefetch);
221 Builder.CreateCall4(PrefetchFunc, PrefPtrValue,
222 ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1),
223 ConstantInt::get(I32, 3), ConstantInt::get(I32, 1));