xref: /llvm-project/llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp (revision 94f9cbbe49b4c836cfbed046637cdc0c63a4a083)
1 //===-------- LoopDataPrefetch.cpp - Loop Data Prefetching Pass -----------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements a Loop Data Prefetching Pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "llvm/Transforms/Scalar/LoopDataPrefetch.h"
14 #include "llvm/InitializePasses.h"
15 
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/LoopInfo.h"
21 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
22 #include "llvm/Analysis/ScalarEvolution.h"
23 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/Support/CommandLine.h"
28 #include "llvm/Support/Debug.h"
29 #include "llvm/Transforms/Scalar.h"
30 #include "llvm/Transforms/Utils.h"
31 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
32 
33 #define DEBUG_TYPE "loop-data-prefetch"
34 
35 using namespace llvm;
36 
37 // By default, we limit this to creating 16 PHIs (which is a little over half
38 // of the allocatable register set).
39 static cl::opt<bool>
40 PrefetchWrites("loop-prefetch-writes", cl::Hidden, cl::init(false),
41                cl::desc("Prefetch write addresses"));
42 
43 static cl::opt<unsigned>
44     PrefetchDistance("prefetch-distance",
45                      cl::desc("Number of instructions to prefetch ahead"),
46                      cl::Hidden);
47 
48 static cl::opt<unsigned>
49     MinPrefetchStride("min-prefetch-stride",
50                       cl::desc("Min stride to add prefetches"), cl::Hidden);
51 
52 static cl::opt<unsigned> MaxPrefetchIterationsAhead(
53     "max-prefetch-iters-ahead",
54     cl::desc("Max number of iterations to prefetch ahead"), cl::Hidden);
55 
56 STATISTIC(NumPrefetches, "Number of prefetches inserted");
57 
58 namespace {
59 
60 /// Loop prefetch implementation class.
61 class LoopDataPrefetch {
62 public:
63   LoopDataPrefetch(AssumptionCache *AC, DominatorTree *DT, LoopInfo *LI,
64                    ScalarEvolution *SE, const TargetTransformInfo *TTI,
65                    OptimizationRemarkEmitter *ORE)
66       : AC(AC), DT(DT), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
67 
68   bool run();
69 
70 private:
71   bool runOnLoop(Loop *L);
72 
73   /// Check if the stride of the accesses is large enough to
74   /// warrant a prefetch.
75   bool isStrideLargeEnough(const SCEVAddRecExpr *AR, unsigned TargetMinStride);
76 
77   unsigned getMinPrefetchStride(unsigned NumMemAccesses,
78                                 unsigned NumStridedMemAccesses,
79                                 unsigned NumPrefetches,
80                                 bool HasCall) {
81     if (MinPrefetchStride.getNumOccurrences() > 0)
82       return MinPrefetchStride;
83     return TTI->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
84                                      NumPrefetches, HasCall);
85   }
86 
87   unsigned getPrefetchDistance() {
88     if (PrefetchDistance.getNumOccurrences() > 0)
89       return PrefetchDistance;
90     return TTI->getPrefetchDistance();
91   }
92 
93   unsigned getMaxPrefetchIterationsAhead() {
94     if (MaxPrefetchIterationsAhead.getNumOccurrences() > 0)
95       return MaxPrefetchIterationsAhead;
96     return TTI->getMaxPrefetchIterationsAhead();
97   }
98 
99   bool doPrefetchWrites() {
100     if (PrefetchWrites.getNumOccurrences() > 0)
101       return PrefetchWrites;
102     return TTI->enableWritePrefetching();
103   }
104 
105   AssumptionCache *AC;
106   DominatorTree *DT;
107   LoopInfo *LI;
108   ScalarEvolution *SE;
109   const TargetTransformInfo *TTI;
110   OptimizationRemarkEmitter *ORE;
111 };
112 
113 /// Legacy class for inserting loop data prefetches.
114 class LoopDataPrefetchLegacyPass : public FunctionPass {
115 public:
116   static char ID; // Pass ID, replacement for typeid
117   LoopDataPrefetchLegacyPass() : FunctionPass(ID) {
118     initializeLoopDataPrefetchLegacyPassPass(*PassRegistry::getPassRegistry());
119   }
120 
121   void getAnalysisUsage(AnalysisUsage &AU) const override {
122     AU.addRequired<AssumptionCacheTracker>();
123     AU.addRequired<DominatorTreeWrapperPass>();
124     AU.addPreserved<DominatorTreeWrapperPass>();
125     AU.addRequired<LoopInfoWrapperPass>();
126     AU.addPreserved<LoopInfoWrapperPass>();
127     AU.addRequiredID(LoopSimplifyID);
128     AU.addPreservedID(LoopSimplifyID);
129     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
130     AU.addRequired<ScalarEvolutionWrapperPass>();
131     AU.addPreserved<ScalarEvolutionWrapperPass>();
132     AU.addRequired<TargetTransformInfoWrapperPass>();
133   }
134 
135   bool runOnFunction(Function &F) override;
136   };
137 }
138 
139 char LoopDataPrefetchLegacyPass::ID = 0;
140 INITIALIZE_PASS_BEGIN(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
141                       "Loop Data Prefetch", false, false)
142 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
143 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
144 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
145 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
146 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
147 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
148 INITIALIZE_PASS_END(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
149                     "Loop Data Prefetch", false, false)
150 
151 FunctionPass *llvm::createLoopDataPrefetchPass() {
152   return new LoopDataPrefetchLegacyPass();
153 }
154 
155 bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR,
156                                            unsigned TargetMinStride) {
157   // No need to check if any stride goes.
158   if (TargetMinStride <= 1)
159     return true;
160 
161   const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
162   // If MinStride is set, don't prefetch unless we can ensure that stride is
163   // larger.
164   if (!ConstStride)
165     return false;
166 
167   unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue());
168   return TargetMinStride <= AbsStride;
169 }
170 
171 PreservedAnalyses LoopDataPrefetchPass::run(Function &F,
172                                             FunctionAnalysisManager &AM) {
173   DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
174   LoopInfo *LI = &AM.getResult<LoopAnalysis>(F);
175   ScalarEvolution *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
176   AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F);
177   OptimizationRemarkEmitter *ORE =
178       &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
179   const TargetTransformInfo *TTI = &AM.getResult<TargetIRAnalysis>(F);
180 
181   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
182   bool Changed = LDP.run();
183 
184   if (Changed) {
185     PreservedAnalyses PA;
186     PA.preserve<DominatorTreeAnalysis>();
187     PA.preserve<LoopAnalysis>();
188     return PA;
189   }
190 
191   return PreservedAnalyses::all();
192 }
193 
194 bool LoopDataPrefetchLegacyPass::runOnFunction(Function &F) {
195   if (skipFunction(F))
196     return false;
197 
198   DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
199   LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
200   ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
201   AssumptionCache *AC =
202       &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
203   OptimizationRemarkEmitter *ORE =
204       &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
205   const TargetTransformInfo *TTI =
206       &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
207 
208   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
209   return LDP.run();
210 }
211 
212 bool LoopDataPrefetch::run() {
213   // If PrefetchDistance is not set, don't run the pass.  This gives an
214   // opportunity for targets to run this pass for selected subtargets only
215   // (whose TTI sets PrefetchDistance and CacheLineSize).
216   if (getPrefetchDistance() == 0 || TTI->getCacheLineSize() == 0) {
217     LLVM_DEBUG(dbgs() << "Please set both PrefetchDistance and CacheLineSize "
218                          "for loop data prefetch.\n");
219     return false;
220   }
221 
222   bool MadeChange = false;
223 
224   for (Loop *I : *LI)
225     for (Loop *L : depth_first(I))
226       MadeChange |= runOnLoop(L);
227 
228   return MadeChange;
229 }
230 
231 /// A record for a potential prefetch made during the initial scan of the
232 /// loop. This is used to let a single prefetch target multiple memory accesses.
233 struct Prefetch {
234   /// The address formula for this prefetch as returned by ScalarEvolution.
235   const SCEVAddRecExpr *LSCEVAddRec;
236   /// The point of insertion for the prefetch instruction.
237   Instruction *InsertPt = nullptr;
238   /// True if targeting a write memory access.
239   bool Writes = false;
240   /// The (first seen) prefetched instruction.
241   Instruction *MemI = nullptr;
242 
243   /// Constructor to create a new Prefetch for \p I.
244   Prefetch(const SCEVAddRecExpr *L, Instruction *I) : LSCEVAddRec(L) {
245     addInstruction(I);
246   };
247 
248   /// Add the instruction \param I to this prefetch. If it's not the first
249   /// one, 'InsertPt' and 'Writes' will be updated as required.
250   /// \param PtrDiff the known constant address difference to the first added
251   /// instruction.
252   void addInstruction(Instruction *I, DominatorTree *DT = nullptr,
253                       int64_t PtrDiff = 0) {
254     if (!InsertPt) {
255       MemI = I;
256       InsertPt = I;
257       Writes = isa<StoreInst>(I);
258     } else {
259       BasicBlock *PrefBB = InsertPt->getParent();
260       BasicBlock *InsBB = I->getParent();
261       if (PrefBB != InsBB) {
262         BasicBlock *DomBB = DT->findNearestCommonDominator(PrefBB, InsBB);
263         if (DomBB != PrefBB)
264           InsertPt = DomBB->getTerminator();
265       }
266 
267       if (isa<StoreInst>(I) && PtrDiff == 0)
268         Writes = true;
269     }
270   }
271 };
272 
273 bool LoopDataPrefetch::runOnLoop(Loop *L) {
274   bool MadeChange = false;
275 
276   // Only prefetch in the inner-most loop
277   if (!L->isInnermost())
278     return MadeChange;
279 
280   SmallPtrSet<const Value *, 32> EphValues;
281   CodeMetrics::collectEphemeralValues(L, AC, EphValues);
282 
283   // Calculate the number of iterations ahead to prefetch
284   CodeMetrics Metrics;
285   bool HasCall = false;
286   for (const auto BB : L->blocks()) {
287     // If the loop already has prefetches, then assume that the user knows
288     // what they are doing and don't add any more.
289     for (auto &I : *BB) {
290       if (isa<CallInst>(&I) || isa<InvokeInst>(&I)) {
291         if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
292           if (F->getIntrinsicID() == Intrinsic::prefetch)
293             return MadeChange;
294           if (TTI->isLoweredToCall(F))
295             HasCall = true;
296         } else { // indirect call.
297           HasCall = true;
298         }
299       }
300     }
301     Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
302   }
303 
304   if (!Metrics.NumInsts.isValid())
305     return MadeChange;
306 
307   unsigned LoopSize = *Metrics.NumInsts.getValue();
308   if (!LoopSize)
309     LoopSize = 1;
310 
311   unsigned ItersAhead = getPrefetchDistance() / LoopSize;
312   if (!ItersAhead)
313     ItersAhead = 1;
314 
315   if (ItersAhead > getMaxPrefetchIterationsAhead())
316     return MadeChange;
317 
318   unsigned ConstantMaxTripCount = SE->getSmallConstantMaxTripCount(L);
319   if (ConstantMaxTripCount && ConstantMaxTripCount < ItersAhead + 1)
320     return MadeChange;
321 
322   unsigned NumMemAccesses = 0;
323   unsigned NumStridedMemAccesses = 0;
324   SmallVector<Prefetch, 16> Prefetches;
325   for (const auto BB : L->blocks())
326     for (auto &I : *BB) {
327       Value *PtrValue;
328       Instruction *MemI;
329 
330       if (LoadInst *LMemI = dyn_cast<LoadInst>(&I)) {
331         MemI = LMemI;
332         PtrValue = LMemI->getPointerOperand();
333       } else if (StoreInst *SMemI = dyn_cast<StoreInst>(&I)) {
334         if (!doPrefetchWrites()) continue;
335         MemI = SMemI;
336         PtrValue = SMemI->getPointerOperand();
337       } else continue;
338 
339       unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
340       if (!TTI->shouldPrefetchAddressSpace(PtrAddrSpace))
341         continue;
342       NumMemAccesses++;
343       if (L->isLoopInvariant(PtrValue))
344         continue;
345 
346       const SCEV *LSCEV = SE->getSCEV(PtrValue);
347       const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
348       if (!LSCEVAddRec)
349         continue;
350       NumStridedMemAccesses++;
351 
352       // We don't want to double prefetch individual cache lines. If this
353       // access is known to be within one cache line of some other one that
354       // has already been prefetched, then don't prefetch this one as well.
355       bool DupPref = false;
356       for (auto &Pref : Prefetches) {
357         const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, Pref.LSCEVAddRec);
358         if (const SCEVConstant *ConstPtrDiff =
359             dyn_cast<SCEVConstant>(PtrDiff)) {
360           int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
361           if (PD < (int64_t) TTI->getCacheLineSize()) {
362             Pref.addInstruction(MemI, DT, PD);
363             DupPref = true;
364             break;
365           }
366         }
367       }
368       if (!DupPref)
369         Prefetches.push_back(Prefetch(LSCEVAddRec, MemI));
370     }
371 
372   unsigned TargetMinStride =
373     getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
374                          Prefetches.size(), HasCall);
375 
376   LLVM_DEBUG(dbgs() << "Prefetching " << ItersAhead
377              << " iterations ahead (loop size: " << LoopSize << ") in "
378              << L->getHeader()->getParent()->getName() << ": " << *L);
379   LLVM_DEBUG(dbgs() << "Loop has: "
380              << NumMemAccesses << " memory accesses, "
381              << NumStridedMemAccesses << " strided memory accesses, "
382              << Prefetches.size() << " potential prefetch(es), "
383              << "a minimum stride of " << TargetMinStride << ", "
384              << (HasCall ? "calls" : "no calls") << ".\n");
385 
386   for (auto &P : Prefetches) {
387     // Check if the stride of the accesses is large enough to warrant a
388     // prefetch.
389     if (!isStrideLargeEnough(P.LSCEVAddRec, TargetMinStride))
390       continue;
391 
392     BasicBlock *BB = P.InsertPt->getParent();
393     SCEVExpander SCEVE(*SE, BB->getDataLayout(), "prefaddr");
394     const SCEV *NextLSCEV = SE->getAddExpr(P.LSCEVAddRec, SE->getMulExpr(
395       SE->getConstant(P.LSCEVAddRec->getType(), ItersAhead),
396       P.LSCEVAddRec->getStepRecurrence(*SE)));
397     if (!SCEVE.isSafeToExpand(NextLSCEV))
398       continue;
399 
400     unsigned PtrAddrSpace = NextLSCEV->getType()->getPointerAddressSpace();
401     Type *I8Ptr = PointerType::get(BB->getContext(), PtrAddrSpace);
402     Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, P.InsertPt);
403 
404     IRBuilder<> Builder(P.InsertPt);
405     Type *I32 = Type::getInt32Ty(BB->getContext());
406     Builder.CreateIntrinsic(Intrinsic::prefetch, PrefPtrValue->getType(),
407                             {PrefPtrValue, ConstantInt::get(I32, P.Writes),
408                              ConstantInt::get(I32, 3),
409                              ConstantInt::get(I32, 1)});
410     ++NumPrefetches;
411     LLVM_DEBUG(dbgs() << "  Access: "
412                << *P.MemI->getOperand(isa<LoadInst>(P.MemI) ? 0 : 1)
413                << ", SCEV: " << *P.LSCEVAddRec << "\n");
414     ORE->emit([&]() {
415         return OptimizationRemark(DEBUG_TYPE, "Prefetched", P.MemI)
416           << "prefetched memory access";
417       });
418 
419     MadeChange = true;
420   }
421 
422   return MadeChange;
423 }
424