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/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/ScalarEvolution.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"
40 // By default, we limit this to creating 16 PHIs (which is a little over half
41 // of the allocatable register set).
43 PrefetchWrites("ppc-loop-prefetch-writes", cl::Hidden, cl::init(false),
44 cl::desc("Prefetch write addresses"));
46 // This seems like a reasonable default for the BG/Q (this pass is enabled, by
47 // default, only on the BG/Q).
48 static cl::opt<unsigned>
49 PrefDist("ppc-loop-prefetch-distance", cl::Hidden, cl::init(300),
50 cl::desc("The loop prefetch distance"));
52 static cl::opt<unsigned>
53 CacheLineSize("ppc-loop-prefetch-cache-line", cl::Hidden, cl::init(64),
54 cl::desc("The loop prefetch cache line size"));
57 void initializePPCLoopDataPrefetchPass(PassRegistry&);
62 class PPCLoopDataPrefetch : public FunctionPass {
64 static char ID; // Pass ID, replacement for typeid
65 PPCLoopDataPrefetch() : FunctionPass(ID) {
66 initializePPCLoopDataPrefetchPass(*PassRegistry::getPassRegistry());
69 void getAnalysisUsage(AnalysisUsage &AU) const override {
70 AU.addRequired<AssumptionCacheTracker>();
71 AU.addPreserved<DominatorTreeWrapperPass>();
72 AU.addRequired<LoopInfoWrapperPass>();
73 AU.addPreserved<LoopInfoWrapperPass>();
74 AU.addRequired<ScalarEvolution>();
75 // FIXME: For some reason, preserving SE here breaks LSR (even if
76 // this pass changes nothing).
77 // AU.addPreserved<ScalarEvolution>();
78 AU.addRequired<TargetTransformInfoWrapperPass>();
81 bool runOnFunction(Function &F) override;
82 bool runOnLoop(Loop *L);
88 const TargetTransformInfo *TTI;
93 char PPCLoopDataPrefetch::ID = 0;
94 INITIALIZE_PASS_BEGIN(PPCLoopDataPrefetch, "ppc-loop-data-prefetch",
95 "PPC Loop Data Prefetch", false, false)
96 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
97 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
98 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
99 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
100 INITIALIZE_PASS_END(PPCLoopDataPrefetch, "ppc-loop-data-prefetch",
101 "PPC Loop Data Prefetch", false, false)
103 FunctionPass *llvm::createPPCLoopDataPrefetchPass() { return new PPCLoopDataPrefetch(); }
105 bool PPCLoopDataPrefetch::runOnFunction(Function &F) {
106 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
107 SE = &getAnalysis<ScalarEvolution>();
108 DL = &F.getParent()->getDataLayout();
109 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
110 TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
112 bool MadeChange = false;
117 for (auto I = df_begin(*LI->begin()), E = df_end(*LI->begin()); I != E; ++I) {
119 MadeChange |= runOnLoop(L);
125 bool PPCLoopDataPrefetch::runOnLoop(Loop *L) {
126 bool MadeChange = false;
128 // Only prefetch in the inner-most loop
132 SmallPtrSet<const Value *, 32> EphValues;
133 CodeMetrics::collectEphemeralValues(L, AC, EphValues);
135 // Calculate the number of iterations ahead to prefetch
137 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
140 // If the loop already has prefetches, then assume that the user knows
141 // what he or she is doing and don't add any more.
142 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end();
144 if (CallInst *CI = dyn_cast<CallInst>(J))
145 if (Function *F = CI->getCalledFunction())
146 if (F->getIntrinsicID() == Intrinsic::prefetch)
149 Metrics.analyzeBasicBlock(*I, *TTI, EphValues);
151 unsigned LoopSize = Metrics.NumInsts;
155 unsigned ItersAhead = PrefDist/LoopSize;
159 SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads;
160 for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
162 for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end();
167 if (LoadInst *LMemI = dyn_cast<LoadInst>(J)) {
169 PtrValue = LMemI->getPointerOperand();
170 } else if (StoreInst *SMemI = dyn_cast<StoreInst>(J)) {
171 if (!PrefetchWrites) continue;
173 PtrValue = SMemI->getPointerOperand();
176 unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
180 if (L->isLoopInvariant(PtrValue))
183 const SCEV *LSCEV = SE->getSCEV(PtrValue);
184 const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
188 // We don't want to double prefetch individual cache lines. If this load
189 // is known to be within one cache line of some other load that has
190 // already been prefetched, then don't prefetch this one as well.
191 bool DupPref = false;
192 for (SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>,
193 16>::iterator K = PrefLoads.begin(), KE = PrefLoads.end();
195 const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, K->second);
196 if (const SCEVConstant *ConstPtrDiff =
197 dyn_cast<SCEVConstant>(PtrDiff)) {
198 int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
199 if (PD < (int64_t) CacheLineSize) {
208 const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr(
209 SE->getConstant(LSCEVAddRec->getType(), ItersAhead),
210 LSCEVAddRec->getStepRecurrence(*SE)));
211 if (!isSafeToExpand(NextLSCEV, *SE))
214 PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec));
216 Type *I8Ptr = Type::getInt8PtrTy((*I)->getContext(), PtrAddrSpace);
217 SCEVExpander SCEVE(*SE, J->getModule()->getDataLayout(), "prefaddr");
218 Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI);
220 IRBuilder<> Builder(MemI);
221 Module *M = (*I)->getParent()->getParent();
222 Type *I32 = Type::getInt32Ty((*I)->getContext());
223 Value *PrefetchFunc = Intrinsic::getDeclaration(M, Intrinsic::prefetch);
224 Builder.CreateCall4(PrefetchFunc, PrefPtrValue,
225 ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1),
226 ConstantInt::get(I32, 3), ConstantInt::get(I32, 1));