xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp (revision 5f4c09dd85bff675e0ca63c55ea3c517e0fddfcc)
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/IR/Module.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Transforms/Scalar.h"
31 #include "llvm/Transforms/Utils.h"
32 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
33 
34 #define DEBUG_TYPE "loop-data-prefetch"
35 
36 using namespace llvm;
37 
38 // By default, we limit this to creating 16 PHIs (which is a little over half
39 // of the allocatable register set).
40 static cl::opt<bool>
41 PrefetchWrites("loop-prefetch-writes", cl::Hidden, cl::init(false),
42                cl::desc("Prefetch write addresses"));
43 
44 static cl::opt<unsigned>
45     PrefetchDistance("prefetch-distance",
46                      cl::desc("Number of instructions to prefetch ahead"),
47                      cl::Hidden);
48 
49 static cl::opt<unsigned>
50     MinPrefetchStride("min-prefetch-stride",
51                       cl::desc("Min stride to add prefetches"), cl::Hidden);
52 
53 static cl::opt<unsigned> MaxPrefetchIterationsAhead(
54     "max-prefetch-iters-ahead",
55     cl::desc("Max number of iterations to prefetch ahead"), cl::Hidden);
56 
57 STATISTIC(NumPrefetches, "Number of prefetches inserted");
58 
59 namespace {
60 
61 /// Loop prefetch implementation class.
62 class LoopDataPrefetch {
63 public:
64   LoopDataPrefetch(AssumptionCache *AC, DominatorTree *DT, LoopInfo *LI,
65                    ScalarEvolution *SE, const TargetTransformInfo *TTI,
66                    OptimizationRemarkEmitter *ORE)
67       : AC(AC), DT(DT), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
68 
69   bool run();
70 
71 private:
72   bool runOnLoop(Loop *L);
73 
74   /// Check if the stride of the accesses is large enough to
75   /// warrant a prefetch.
76   bool isStrideLargeEnough(const SCEVAddRecExpr *AR, unsigned TargetMinStride);
77 
78   unsigned getMinPrefetchStride(unsigned NumMemAccesses,
79                                 unsigned NumStridedMemAccesses,
80                                 unsigned NumPrefetches,
81                                 bool HasCall) {
82     if (MinPrefetchStride.getNumOccurrences() > 0)
83       return MinPrefetchStride;
84     return TTI->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
85                                      NumPrefetches, HasCall);
86   }
87 
88   unsigned getPrefetchDistance() {
89     if (PrefetchDistance.getNumOccurrences() > 0)
90       return PrefetchDistance;
91     return TTI->getPrefetchDistance();
92   }
93 
94   unsigned getMaxPrefetchIterationsAhead() {
95     if (MaxPrefetchIterationsAhead.getNumOccurrences() > 0)
96       return MaxPrefetchIterationsAhead;
97     return TTI->getMaxPrefetchIterationsAhead();
98   }
99 
100   bool doPrefetchWrites() {
101     if (PrefetchWrites.getNumOccurrences() > 0)
102       return PrefetchWrites;
103     return TTI->enableWritePrefetching();
104   }
105 
106   AssumptionCache *AC;
107   DominatorTree *DT;
108   LoopInfo *LI;
109   ScalarEvolution *SE;
110   const TargetTransformInfo *TTI;
111   OptimizationRemarkEmitter *ORE;
112 };
113 
114 /// Legacy class for inserting loop data prefetches.
115 class LoopDataPrefetchLegacyPass : public FunctionPass {
116 public:
117   static char ID; // Pass ID, replacement for typeid
118   LoopDataPrefetchLegacyPass() : FunctionPass(ID) {
119     initializeLoopDataPrefetchLegacyPassPass(*PassRegistry::getPassRegistry());
120   }
121 
122   void getAnalysisUsage(AnalysisUsage &AU) const override {
123     AU.addRequired<AssumptionCacheTracker>();
124     AU.addRequired<DominatorTreeWrapperPass>();
125     AU.addPreserved<DominatorTreeWrapperPass>();
126     AU.addRequired<LoopInfoWrapperPass>();
127     AU.addPreserved<LoopInfoWrapperPass>();
128     AU.addRequiredID(LoopSimplifyID);
129     AU.addPreservedID(LoopSimplifyID);
130     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
131     AU.addRequired<ScalarEvolutionWrapperPass>();
132     AU.addPreserved<ScalarEvolutionWrapperPass>();
133     AU.addRequired<TargetTransformInfoWrapperPass>();
134   }
135 
136   bool runOnFunction(Function &F) override;
137   };
138 }
139 
140 char LoopDataPrefetchLegacyPass::ID = 0;
141 INITIALIZE_PASS_BEGIN(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
142                       "Loop Data Prefetch", false, false)
143 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
144 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
145 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
146 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
147 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
148 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
149 INITIALIZE_PASS_END(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
150                     "Loop Data Prefetch", false, false)
151 
152 FunctionPass *llvm::createLoopDataPrefetchPass() {
153   return new LoopDataPrefetchLegacyPass();
154 }
155 
156 bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR,
157                                            unsigned TargetMinStride) {
158   // No need to check if any stride goes.
159   if (TargetMinStride <= 1)
160     return true;
161 
162   const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
163   // If MinStride is set, don't prefetch unless we can ensure that stride is
164   // larger.
165   if (!ConstStride)
166     return false;
167 
168   unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue());
169   return TargetMinStride <= AbsStride;
170 }
171 
172 PreservedAnalyses LoopDataPrefetchPass::run(Function &F,
173                                             FunctionAnalysisManager &AM) {
174   DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
175   LoopInfo *LI = &AM.getResult<LoopAnalysis>(F);
176   ScalarEvolution *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
177   AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F);
178   OptimizationRemarkEmitter *ORE =
179       &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
180   const TargetTransformInfo *TTI = &AM.getResult<TargetIRAnalysis>(F);
181 
182   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
183   bool Changed = LDP.run();
184 
185   if (Changed) {
186     PreservedAnalyses PA;
187     PA.preserve<DominatorTreeAnalysis>();
188     PA.preserve<LoopAnalysis>();
189     return PA;
190   }
191 
192   return PreservedAnalyses::all();
193 }
194 
195 bool LoopDataPrefetchLegacyPass::runOnFunction(Function &F) {
196   if (skipFunction(F))
197     return false;
198 
199   DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
200   LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
201   ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
202   AssumptionCache *AC =
203       &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
204   OptimizationRemarkEmitter *ORE =
205       &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
206   const TargetTransformInfo *TTI =
207       &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
208 
209   LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
210   return LDP.run();
211 }
212 
213 bool LoopDataPrefetch::run() {
214   // If PrefetchDistance is not set, don't run the pass.  This gives an
215   // opportunity for targets to run this pass for selected subtargets only
216   // (whose TTI sets PrefetchDistance and CacheLineSize).
217   if (getPrefetchDistance() == 0 || TTI->getCacheLineSize() == 0) {
218     LLVM_DEBUG(dbgs() << "Please set both PrefetchDistance and CacheLineSize "
219                          "for loop data prefetch.\n");
220     return false;
221   }
222 
223   bool MadeChange = false;
224 
225   for (Loop *I : *LI)
226     for (Loop *L : depth_first(I))
227       MadeChange |= runOnLoop(L);
228 
229   return MadeChange;
230 }
231 
232 /// A record for a potential prefetch made during the initial scan of the
233 /// loop. This is used to let a single prefetch target multiple memory accesses.
234 struct Prefetch {
235   /// The address formula for this prefetch as returned by ScalarEvolution.
236   const SCEVAddRecExpr *LSCEVAddRec;
237   /// The point of insertion for the prefetch instruction.
238   Instruction *InsertPt = nullptr;
239   /// True if targeting a write memory access.
240   bool Writes = false;
241   /// The (first seen) prefetched instruction.
242   Instruction *MemI = nullptr;
243 
244   /// Constructor to create a new Prefetch for \p I.
245   Prefetch(const SCEVAddRecExpr *L, Instruction *I) : LSCEVAddRec(L) {
246     addInstruction(I);
247   };
248 
249   /// Add the instruction \param I to this prefetch. If it's not the first
250   /// one, 'InsertPt' and 'Writes' will be updated as required.
251   /// \param PtrDiff the known constant address difference to the first added
252   /// instruction.
253   void addInstruction(Instruction *I, DominatorTree *DT = nullptr,
254                       int64_t PtrDiff = 0) {
255     if (!InsertPt) {
256       MemI = I;
257       InsertPt = I;
258       Writes = isa<StoreInst>(I);
259     } else {
260       BasicBlock *PrefBB = InsertPt->getParent();
261       BasicBlock *InsBB = I->getParent();
262       if (PrefBB != InsBB) {
263         BasicBlock *DomBB = DT->findNearestCommonDominator(PrefBB, InsBB);
264         if (DomBB != PrefBB)
265           InsertPt = DomBB->getTerminator();
266       }
267 
268       if (isa<StoreInst>(I) && PtrDiff == 0)
269         Writes = true;
270     }
271   }
272 };
273 
274 bool LoopDataPrefetch::runOnLoop(Loop *L) {
275   bool MadeChange = false;
276 
277   // Only prefetch in the inner-most loop
278   if (!L->isInnermost())
279     return MadeChange;
280 
281   SmallPtrSet<const Value *, 32> EphValues;
282   CodeMetrics::collectEphemeralValues(L, AC, EphValues);
283 
284   // Calculate the number of iterations ahead to prefetch
285   CodeMetrics Metrics;
286   bool HasCall = false;
287   for (const auto BB : L->blocks()) {
288     // If the loop already has prefetches, then assume that the user knows
289     // what they are doing and don't add any more.
290     for (auto &I : *BB) {
291       if (isa<CallInst>(&I) || isa<InvokeInst>(&I)) {
292         if (const Function *F = cast<CallBase>(I).getCalledFunction()) {
293           if (F->getIntrinsicID() == Intrinsic::prefetch)
294             return MadeChange;
295           if (TTI->isLoweredToCall(F))
296             HasCall = true;
297         } else { // indirect call.
298           HasCall = true;
299         }
300       }
301     }
302     Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
303   }
304 
305   if (!Metrics.NumInsts.isValid())
306     return MadeChange;
307 
308   unsigned LoopSize = *Metrics.NumInsts.getValue();
309   if (!LoopSize)
310     LoopSize = 1;
311 
312   unsigned ItersAhead = getPrefetchDistance() / LoopSize;
313   if (!ItersAhead)
314     ItersAhead = 1;
315 
316   if (ItersAhead > getMaxPrefetchIterationsAhead())
317     return MadeChange;
318 
319   unsigned ConstantMaxTripCount = SE->getSmallConstantMaxTripCount(L);
320   if (ConstantMaxTripCount && ConstantMaxTripCount < ItersAhead + 1)
321     return MadeChange;
322 
323   unsigned NumMemAccesses = 0;
324   unsigned NumStridedMemAccesses = 0;
325   SmallVector<Prefetch, 16> Prefetches;
326   for (const auto BB : L->blocks())
327     for (auto &I : *BB) {
328       Value *PtrValue;
329       Instruction *MemI;
330 
331       if (LoadInst *LMemI = dyn_cast<LoadInst>(&I)) {
332         MemI = LMemI;
333         PtrValue = LMemI->getPointerOperand();
334       } else if (StoreInst *SMemI = dyn_cast<StoreInst>(&I)) {
335         if (!doPrefetchWrites()) continue;
336         MemI = SMemI;
337         PtrValue = SMemI->getPointerOperand();
338       } else continue;
339 
340       unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
341       if (!TTI->shouldPrefetchAddressSpace(PtrAddrSpace))
342         continue;
343       NumMemAccesses++;
344       if (L->isLoopInvariant(PtrValue))
345         continue;
346 
347       const SCEV *LSCEV = SE->getSCEV(PtrValue);
348       const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
349       if (!LSCEVAddRec)
350         continue;
351       NumStridedMemAccesses++;
352 
353       // We don't want to double prefetch individual cache lines. If this
354       // access is known to be within one cache line of some other one that
355       // has already been prefetched, then don't prefetch this one as well.
356       bool DupPref = false;
357       for (auto &Pref : Prefetches) {
358         const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, Pref.LSCEVAddRec);
359         if (const SCEVConstant *ConstPtrDiff =
360             dyn_cast<SCEVConstant>(PtrDiff)) {
361           int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
362           if (PD < (int64_t) TTI->getCacheLineSize()) {
363             Pref.addInstruction(MemI, DT, PD);
364             DupPref = true;
365             break;
366           }
367         }
368       }
369       if (!DupPref)
370         Prefetches.push_back(Prefetch(LSCEVAddRec, MemI));
371     }
372 
373   unsigned TargetMinStride =
374     getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
375                          Prefetches.size(), HasCall);
376 
377   LLVM_DEBUG(dbgs() << "Prefetching " << ItersAhead
378              << " iterations ahead (loop size: " << LoopSize << ") in "
379              << L->getHeader()->getParent()->getName() << ": " << *L);
380   LLVM_DEBUG(dbgs() << "Loop has: "
381              << NumMemAccesses << " memory accesses, "
382              << NumStridedMemAccesses << " strided memory accesses, "
383              << Prefetches.size() << " potential prefetch(es), "
384              << "a minimum stride of " << TargetMinStride << ", "
385              << (HasCall ? "calls" : "no calls") << ".\n");
386 
387   for (auto &P : Prefetches) {
388     // Check if the stride of the accesses is large enough to warrant a
389     // prefetch.
390     if (!isStrideLargeEnough(P.LSCEVAddRec, TargetMinStride))
391       continue;
392 
393     BasicBlock *BB = P.InsertPt->getParent();
394     SCEVExpander SCEVE(*SE, BB->getModule()->getDataLayout(), "prefaddr");
395     const SCEV *NextLSCEV = SE->getAddExpr(P.LSCEVAddRec, SE->getMulExpr(
396       SE->getConstant(P.LSCEVAddRec->getType(), ItersAhead),
397       P.LSCEVAddRec->getStepRecurrence(*SE)));
398     if (!SCEVE.isSafeToExpand(NextLSCEV))
399       continue;
400 
401     unsigned PtrAddrSpace = NextLSCEV->getType()->getPointerAddressSpace();
402     Type *I8Ptr = Type::getInt8PtrTy(BB->getContext(), PtrAddrSpace);
403     Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, P.InsertPt);
404 
405     IRBuilder<> Builder(P.InsertPt);
406     Module *M = BB->getParent()->getParent();
407     Type *I32 = Type::getInt32Ty(BB->getContext());
408     Function *PrefetchFunc = Intrinsic::getDeclaration(
409         M, Intrinsic::prefetch, PrefPtrValue->getType());
410     Builder.CreateCall(
411         PrefetchFunc,
412         {PrefPtrValue,
413          ConstantInt::get(I32, P.Writes),
414          ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)});
415     ++NumPrefetches;
416     LLVM_DEBUG(dbgs() << "  Access: "
417                << *P.MemI->getOperand(isa<LoadInst>(P.MemI) ? 0 : 1)
418                << ", SCEV: " << *P.LSCEVAddRec << "\n");
419     ORE->emit([&]() {
420         return OptimizationRemark(DEBUG_TYPE, "Prefetched", P.MemI)
421           << "prefetched memory access";
422       });
423 
424     MadeChange = true;
425   }
426 
427   return MadeChange;
428 }
429