xref: /llvm-project/llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp (revision fd3e8044cd0594ca2903bb0f0264d27691885a93) 
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/CFG.h"
26 #include "llvm/IR/Dominators.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Transforms/Scalar.h"
32 #include "llvm/Transforms/Utils.h"
33 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
34 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
35 #include "llvm/Transforms/Utils/ValueMapper.h"
36 
37 #define DEBUG_TYPE "loop-data-prefetch"
38 
39 using namespace llvm;
40 
41 // By default, we limit this to creating 16 PHIs (which is a little over half
42 // of the allocatable register set).
43 static cl::opt<bool>
44 PrefetchWrites("loop-prefetch-writes", cl::Hidden, cl::init(false),
45                cl::desc("Prefetch write addresses"));
46 
47 static cl::opt<unsigned>
48     PrefetchDistance("prefetch-distance",
49                      cl::desc("Number of instructions to prefetch ahead"),
50                      cl::Hidden);
51 
52 static cl::opt<unsigned>
53     MinPrefetchStride("min-prefetch-stride",
54                       cl::desc("Min stride to add prefetches"), cl::Hidden);
55 
56 static cl::opt<unsigned> MaxPrefetchIterationsAhead(
57     "max-prefetch-iters-ahead",
58     cl::desc("Max number of iterations to prefetch ahead"), cl::Hidden);
59 
60 STATISTIC(NumPrefetches, "Number of prefetches inserted");
61 
62 namespace {
63 
64 /// Loop prefetch implementation class.
65 class LoopDataPrefetch {
66 public:
67   LoopDataPrefetch(AssumptionCache *AC, DominatorTree *DT, LoopInfo *LI,
68                    ScalarEvolution *SE, const TargetTransformInfo *TTI,
69                    OptimizationRemarkEmitter *ORE)
70       : AC(AC), DT(DT), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
71 
72   bool run();
73 
74 private:
75   bool runOnLoop(Loop *L);
76 
77   /// Check if the stride of the accesses is large enough to
78   /// warrant a prefetch.
79   bool isStrideLargeEnough(const SCEVAddRecExpr *AR, unsigned TargetMinStride);
80 
81   unsigned getMinPrefetchStride(unsigned NumMemAccesses,
82                                 unsigned NumStridedMemAccesses,
83                                 unsigned NumPrefetches,
84                                 bool HasCall) {
85     if (MinPrefetchStride.getNumOccurrences() > 0)
86       return MinPrefetchStride;
87     return TTI->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
88                                      NumPrefetches, HasCall);
89   }
90 
91   unsigned getPrefetchDistance() {
92     if (PrefetchDistance.getNumOccurrences() > 0)
93       return PrefetchDistance;
94     return TTI->getPrefetchDistance();
95   }
96 
97   unsigned getMaxPrefetchIterationsAhead() {
98     if (MaxPrefetchIterationsAhead.getNumOccurrences() > 0)
99       return MaxPrefetchIterationsAhead;
100     return TTI->getMaxPrefetchIterationsAhead();
101   }
102 
103   bool doPrefetchWrites() {
104     if (PrefetchWrites.getNumOccurrences() > 0)
105       return PrefetchWrites;
106     return TTI->enableWritePrefetching();
107   }
108 
109   AssumptionCache *AC;
110   DominatorTree *DT;
111   LoopInfo *LI;
112   ScalarEvolution *SE;
113   const TargetTransformInfo *TTI;
114   OptimizationRemarkEmitter *ORE;
115 };
116 
117 /// Legacy class for inserting loop data prefetches.
118 class LoopDataPrefetchLegacyPass : public FunctionPass {
119 public:
120   static char ID; // Pass ID, replacement for typeid
121   LoopDataPrefetchLegacyPass() : FunctionPass(ID) {
122     initializeLoopDataPrefetchLegacyPassPass(*PassRegistry::getPassRegistry());
123   }
124 
125   void getAnalysisUsage(AnalysisUsage &AU) const override {
126     AU.addRequired<AssumptionCacheTracker>();
127     AU.addRequired<DominatorTreeWrapperPass>();
128     AU.addPreserved<DominatorTreeWrapperPass>();
129     AU.addRequired<LoopInfoWrapperPass>();
130     AU.addPreserved<LoopInfoWrapperPass>();
131     AU.addRequiredID(LoopSimplifyID);
132     AU.addPreservedID(LoopSimplifyID);
133     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
134     AU.addRequired<ScalarEvolutionWrapperPass>();
135     AU.addPreserved<ScalarEvolutionWrapperPass>();
136     AU.addRequired<TargetTransformInfoWrapperPass>();
137   }
138 
139   bool runOnFunction(Function &F) override;
140   };
141 }
142 
143 char LoopDataPrefetchLegacyPass::ID = 0;
144 INITIALIZE_PASS_BEGIN(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
145                       "Loop Data Prefetch", false, false)
146 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
147 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
148 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
149 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
150 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
151 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
152 INITIALIZE_PASS_END(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
153                     "Loop Data Prefetch", false, false)
154 
155 FunctionPass *llvm::createLoopDataPrefetchPass() {
156   return new LoopDataPrefetchLegacyPass();
157 }
158 
159 bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR,
160                                            unsigned TargetMinStride) {
161   // No need to check if any stride goes.
162   if (TargetMinStride <= 1)
163     return true;
164 
165   const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
166   // If MinStride is set, don't prefetch unless we can ensure that stride is
167   // larger.
168   if (!ConstStride)
169     return false;
170 
171   unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue());
172   return TargetMinStride <= AbsStride;
173 }
174 
175 PreservedAnalyses LoopDataPrefetchPass::run(Function &F,
176                                             FunctionAnalysisManager &AM) {
177   DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
178   LoopInfo *LI = &AM.getResult<LoopAnalysis>(F);
179   ScalarEvolution *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
180   AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F);
181   OptimizationRemarkEmitter *ORE =
182       &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
183   const TargetTransformInfo *TTI = &AM.getResult<TargetIRAnalysis>(F);
184 
185   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
186   bool Changed = LDP.run();
187 
188   if (Changed) {
189     PreservedAnalyses PA;
190     PA.preserve<DominatorTreeAnalysis>();
191     PA.preserve<LoopAnalysis>();
192     return PA;
193   }
194 
195   return PreservedAnalyses::all();
196 }
197 
198 bool LoopDataPrefetchLegacyPass::runOnFunction(Function &F) {
199   if (skipFunction(F))
200     return false;
201 
202   DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
203   LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
204   ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
205   AssumptionCache *AC =
206       &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
207   OptimizationRemarkEmitter *ORE =
208       &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
209   const TargetTransformInfo *TTI =
210       &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
211 
212   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
213   return LDP.run();
214 }
215 
216 bool LoopDataPrefetch::run() {
217   // If PrefetchDistance is not set, don't run the pass.  This gives an
218   // opportunity for targets to run this pass for selected subtargets only
219   // (whose TTI sets PrefetchDistance).
220   if (getPrefetchDistance() == 0)
221     return false;
222   assert(TTI->getCacheLineSize() && "Cache line size is not set for target");
223 
224   bool MadeChange = false;
225 
226   for (Loop *I : *LI)
227     for (Loop *L : depth_first(I))
228       MadeChange |= runOnLoop(L);
229 
230   return MadeChange;
231 }
232 
233 /// A record for a potential prefetch made during the initial scan of the
234 /// loop. This is used to let a single prefetch target multiple memory accesses.
235 struct Prefetch {
236   /// The address formula for this prefetch as returned by ScalarEvolution.
237   const SCEVAddRecExpr *LSCEVAddRec;
238   /// The point of insertion for the prefetch instruction.
239   Instruction *InsertPt = nullptr;
240   /// True if targeting a write memory access.
241   bool Writes = false;
242   /// The (first seen) prefetched instruction.
243   Instruction *MemI = nullptr;
244 
245   /// Constructor to create a new Prefetch for \p I.
246   Prefetch(const SCEVAddRecExpr *L, Instruction *I) : LSCEVAddRec(L) {
247     addInstruction(I);
248   };
249 
250   /// Add the instruction \param I to this prefetch. If it's not the first
251   /// one, 'InsertPt' and 'Writes' will be updated as required.
252   /// \param PtrDiff the known constant address difference to the first added
253   /// instruction.
254   void addInstruction(Instruction *I, DominatorTree *DT = nullptr,
255                       int64_t PtrDiff = 0) {
256     if (!InsertPt) {
257       MemI = I;
258       InsertPt = I;
259       Writes = isa<StoreInst>(I);
260     } else {
261       BasicBlock *PrefBB = InsertPt->getParent();
262       BasicBlock *InsBB = I->getParent();
263       if (PrefBB != InsBB) {
264         BasicBlock *DomBB = DT->findNearestCommonDominator(PrefBB, InsBB);
265         if (DomBB != PrefBB)
266           InsertPt = DomBB->getTerminator();
267       }
268 
269       if (isa<StoreInst>(I) && PtrDiff == 0)
270         Writes = true;
271     }
272   }
273 };
274 
275 bool LoopDataPrefetch::runOnLoop(Loop *L) {
276   bool MadeChange = false;
277 
278   // Only prefetch in the inner-most loop
279   if (!L->isInnermost())
280     return MadeChange;
281 
282   SmallPtrSet<const Value *, 32> EphValues;
283   CodeMetrics::collectEphemeralValues(L, AC, EphValues);
284 
285   // Calculate the number of iterations ahead to prefetch
286   CodeMetrics Metrics;
287   bool HasCall = false;
288   for (const auto BB : L->blocks()) {
289     // If the loop already has prefetches, then assume that the user knows
290     // what they are doing and don't add any more.
291     for (auto &I : *BB) {
292       if (isa<CallInst>(&I) || isa<InvokeInst>(&I)) {
293         if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
294           if (F->getIntrinsicID() == Intrinsic::prefetch)
295             return MadeChange;
296           if (TTI->isLoweredToCall(F))
297             HasCall = true;
298         } else { // indirect call.
299           HasCall = true;
300         }
301       }
302     }
303     Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
304   }
305   unsigned LoopSize = Metrics.NumInsts;
306   if (!LoopSize)
307     LoopSize = 1;
308 
309   unsigned ItersAhead = getPrefetchDistance() / LoopSize;
310   if (!ItersAhead)
311     ItersAhead = 1;
312 
313   if (ItersAhead > getMaxPrefetchIterationsAhead())
314     return MadeChange;
315 
316   unsigned ConstantMaxTripCount = SE->getSmallConstantMaxTripCount(L);
317   if (ConstantMaxTripCount && ConstantMaxTripCount < ItersAhead + 1)
318     return MadeChange;
319 
320   unsigned NumMemAccesses = 0;
321   unsigned NumStridedMemAccesses = 0;
322   SmallVector<Prefetch, 16> Prefetches;
323   for (const auto BB : L->blocks())
324     for (auto &I : *BB) {
325       Value *PtrValue;
326       Instruction *MemI;
327 
328       if (LoadInst *LMemI = dyn_cast<LoadInst>(&I)) {
329         MemI = LMemI;
330         PtrValue = LMemI->getPointerOperand();
331       } else if (StoreInst *SMemI = dyn_cast<StoreInst>(&I)) {
332         if (!doPrefetchWrites()) continue;
333         MemI = SMemI;
334         PtrValue = SMemI->getPointerOperand();
335       } else continue;
336 
337       unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
338       if (PtrAddrSpace)
339         continue;
340       NumMemAccesses++;
341       if (L->isLoopInvariant(PtrValue))
342         continue;
343 
344       const SCEV *LSCEV = SE->getSCEV(PtrValue);
345       const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
346       if (!LSCEVAddRec)
347         continue;
348       NumStridedMemAccesses++;
349 
350       // We don't want to double prefetch individual cache lines. If this
351       // access is known to be within one cache line of some other one that
352       // has already been prefetched, then don't prefetch this one as well.
353       bool DupPref = false;
354       for (auto &Pref : Prefetches) {
355         const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, Pref.LSCEVAddRec);
356         if (const SCEVConstant *ConstPtrDiff =
357             dyn_cast<SCEVConstant>(PtrDiff)) {
358           int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
359           if (PD < (int64_t) TTI->getCacheLineSize()) {
360             Pref.addInstruction(MemI, DT, PD);
361             DupPref = true;
362             break;
363           }
364         }
365       }
366       if (!DupPref)
367         Prefetches.push_back(Prefetch(LSCEVAddRec, MemI));
368     }
369 
370   unsigned TargetMinStride =
371     getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
372                          Prefetches.size(), HasCall);
373 
374   LLVM_DEBUG(dbgs() << "Prefetching " << ItersAhead
375              << " iterations ahead (loop size: " << LoopSize << ") in "
376              << L->getHeader()->getParent()->getName() << ": " << *L);
377   LLVM_DEBUG(dbgs() << "Loop has: "
378              << NumMemAccesses << " memory accesses, "
379              << NumStridedMemAccesses << " strided memory accesses, "
380              << Prefetches.size() << " potential prefetch(es), "
381              << "a minimum stride of " << TargetMinStride << ", "
382              << (HasCall ? "calls" : "no calls") << ".\n");
383 
384   for (auto &P : Prefetches) {
385     // Check if the stride of the accesses is large enough to warrant a
386     // prefetch.
387     if (!isStrideLargeEnough(P.LSCEVAddRec, TargetMinStride))
388       continue;
389 
390     const SCEV *NextLSCEV = SE->getAddExpr(P.LSCEVAddRec, SE->getMulExpr(
391       SE->getConstant(P.LSCEVAddRec->getType(), ItersAhead),
392       P.LSCEVAddRec->getStepRecurrence(*SE)));
393     if (!isSafeToExpand(NextLSCEV, *SE))
394       continue;
395 
396     BasicBlock *BB = P.InsertPt->getParent();
397     Type *I8Ptr = Type::getInt8PtrTy(BB->getContext(), 0/*PtrAddrSpace*/);
398     SCEVExpander SCEVE(*SE, BB->getModule()->getDataLayout(), "prefaddr");
399     Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, P.InsertPt);
400 
401     IRBuilder<> Builder(P.InsertPt);
402     Module *M = BB->getParent()->getParent();
403     Type *I32 = Type::getInt32Ty(BB->getContext());
404     Function *PrefetchFunc = Intrinsic::getDeclaration(
405         M, Intrinsic::prefetch, PrefPtrValue->getType());
406     Builder.CreateCall(
407         PrefetchFunc,
408         {PrefPtrValue,
409          ConstantInt::get(I32, P.Writes),
410          ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)});
411     ++NumPrefetches;
412     LLVM_DEBUG(dbgs() << "  Access: "
413                << *P.MemI->getOperand(isa<LoadInst>(P.MemI) ? 0 : 1)
414                << ", SCEV: " << *P.LSCEVAddRec << "\n");
415     ORE->emit([&]() {
416         return OptimizationRemark(DEBUG_TYPE, "Prefetched", P.MemI)
417           << "prefetched memory access";
418       });
419 
420     MadeChange = true;
421   }
422 
423   return MadeChange;
424 }
425