1 //===- LoopCacheAnalysis.cpp - Loop Cache Analysis -------------------------==//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
6 // See https://llvm.org/LICENSE.txt for license information.
7 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
8 //
9 //===----------------------------------------------------------------------===//
10 ///
11 /// \file
12 /// This file defines the implementation for the loop cache analysis.
13 /// The implementation is largely based on the following paper:
14 ///
15 /// Compiler Optimizations for Improving Data Locality
16 /// By: Steve Carr, Katherine S. McKinley, Chau-Wen Tseng
17 /// http://www.cs.utexas.edu/users/mckinley/papers/asplos-1994.pdf
18 ///
19 /// The general approach taken to estimate the number of cache lines used by the
20 /// memory references in an inner loop is:
21 /// 1. Partition memory references that exhibit temporal or spacial reuse
22 /// into reference groups.
23 /// 2. For each loop L in the a loop nest LN:
24 /// a. Compute the cost of the reference group
25 /// b. Compute the loop cost by summing up the reference groups costs
26 //===----------------------------------------------------------------------===//
27
28 #include "llvm/Analysis/LoopCacheAnalysis.h"
29 #include "llvm/ADT/BreadthFirstIterator.h"
30 #include "llvm/ADT/Sequence.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/Analysis/AliasAnalysis.h"
33 #include "llvm/Analysis/DependenceAnalysis.h"
34 #include "llvm/Analysis/LoopInfo.h"
35 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
36 #include "llvm/Analysis/TargetTransformInfo.h"
37 #include "llvm/Support/CommandLine.h"
38 #include "llvm/Support/Debug.h"
39
40 using namespace llvm;
41
42 #define DEBUG_TYPE "loop-cache-cost"
43
44 static cl::opt<unsigned> DefaultTripCount(
45 "default-trip-count", cl::init(100), cl::Hidden,
46 cl::desc("Use this to specify the default trip count of a loop"));
47
48 // In this analysis two array references are considered to exhibit temporal
49 // reuse if they access either the same memory location, or a memory location
50 // with distance smaller than a configurable threshold.
51 static cl::opt<unsigned> TemporalReuseThreshold(
52 "temporal-reuse-threshold", cl::init(2), cl::Hidden,
53 cl::desc("Use this to specify the max. distance between array elements "
54 "accessed in a loop so that the elements are classified to have "
55 "temporal reuse"));
56
57 /// Retrieve the innermost loop in the given loop nest \p Loops. It returns a
58 /// nullptr if any loops in the loop vector supplied has more than one sibling.
59 /// The loop vector is expected to contain loops collected in breadth-first
60 /// order.
getInnerMostLoop(const LoopVectorTy & Loops)61 static Loop *getInnerMostLoop(const LoopVectorTy &Loops) {
62 assert(!Loops.empty() && "Expecting a non-empy loop vector");
63
64 Loop *LastLoop = Loops.back();
65 Loop *ParentLoop = LastLoop->getParentLoop();
66
67 if (ParentLoop == nullptr) {
68 assert(Loops.size() == 1 && "Expecting a single loop");
69 return LastLoop;
70 }
71
72 return (llvm::is_sorted(Loops,
73 [](const Loop *L1, const Loop *L2) {
74 return L1->getLoopDepth() < L2->getLoopDepth();
75 }))
76 ? LastLoop
77 : nullptr;
78 }
79
isOneDimensionalArray(const SCEV & AccessFn,const SCEV & ElemSize,const Loop & L,ScalarEvolution & SE)80 static bool isOneDimensionalArray(const SCEV &AccessFn, const SCEV &ElemSize,
81 const Loop &L, ScalarEvolution &SE) {
82 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&AccessFn);
83 if (!AR || !AR->isAffine())
84 return false;
85
86 assert(AR->getLoop() && "AR should have a loop");
87
88 // Check that start and increment are not add recurrences.
89 const SCEV *Start = AR->getStart();
90 const SCEV *Step = AR->getStepRecurrence(SE);
91 if (isa<SCEVAddRecExpr>(Start) || isa<SCEVAddRecExpr>(Step))
92 return false;
93
94 // Check that start and increment are both invariant in the loop.
95 if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
96 return false;
97
98 const SCEV *StepRec = AR->getStepRecurrence(SE);
99 if (StepRec && SE.isKnownNegative(StepRec))
100 StepRec = SE.getNegativeSCEV(StepRec);
101
102 return StepRec == &ElemSize;
103 }
104
105 /// Compute the trip count for the given loop \p L. Return the SCEV expression
106 /// for the trip count or nullptr if it cannot be computed.
computeTripCount(const Loop & L,ScalarEvolution & SE)107 static const SCEV *computeTripCount(const Loop &L, ScalarEvolution &SE) {
108 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(&L);
109 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
110 !isa<SCEVConstant>(BackedgeTakenCount))
111 return nullptr;
112 return SE.getTripCountFromExitCount(BackedgeTakenCount);
113 }
114
115 //===----------------------------------------------------------------------===//
116 // IndexedReference implementation
117 //
operator <<(raw_ostream & OS,const IndexedReference & R)118 raw_ostream &llvm::operator<<(raw_ostream &OS, const IndexedReference &R) {
119 if (!R.IsValid) {
120 OS << R.StoreOrLoadInst;
121 OS << ", IsValid=false.";
122 return OS;
123 }
124
125 OS << *R.BasePointer;
126 for (const SCEV *Subscript : R.Subscripts)
127 OS << "[" << *Subscript << "]";
128
129 OS << ", Sizes: ";
130 for (const SCEV *Size : R.Sizes)
131 OS << "[" << *Size << "]";
132
133 return OS;
134 }
135
IndexedReference(Instruction & StoreOrLoadInst,const LoopInfo & LI,ScalarEvolution & SE)136 IndexedReference::IndexedReference(Instruction &StoreOrLoadInst,
137 const LoopInfo &LI, ScalarEvolution &SE)
138 : StoreOrLoadInst(StoreOrLoadInst), SE(SE) {
139 assert((isa<StoreInst>(StoreOrLoadInst) || isa<LoadInst>(StoreOrLoadInst)) &&
140 "Expecting a load or store instruction");
141
142 IsValid = delinearize(LI);
143 if (IsValid)
144 LLVM_DEBUG(dbgs().indent(2) << "Succesfully delinearized: " << *this
145 << "\n");
146 }
147
hasSpacialReuse(const IndexedReference & Other,unsigned CLS,AAResults & AA) const148 Optional<bool> IndexedReference::hasSpacialReuse(const IndexedReference &Other,
149 unsigned CLS,
150 AAResults &AA) const {
151 assert(IsValid && "Expecting a valid reference");
152
153 if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
154 LLVM_DEBUG(dbgs().indent(2)
155 << "No spacial reuse: different base pointers\n");
156 return false;
157 }
158
159 unsigned NumSubscripts = getNumSubscripts();
160 if (NumSubscripts != Other.getNumSubscripts()) {
161 LLVM_DEBUG(dbgs().indent(2)
162 << "No spacial reuse: different number of subscripts\n");
163 return false;
164 }
165
166 // all subscripts must be equal, except the leftmost one (the last one).
167 for (auto SubNum : seq<unsigned>(0, NumSubscripts - 1)) {
168 if (getSubscript(SubNum) != Other.getSubscript(SubNum)) {
169 LLVM_DEBUG(dbgs().indent(2) << "No spacial reuse, different subscripts: "
170 << "\n\t" << *getSubscript(SubNum) << "\n\t"
171 << *Other.getSubscript(SubNum) << "\n");
172 return false;
173 }
174 }
175
176 // the difference between the last subscripts must be less than the cache line
177 // size.
178 const SCEV *LastSubscript = getLastSubscript();
179 const SCEV *OtherLastSubscript = Other.getLastSubscript();
180 const SCEVConstant *Diff = dyn_cast<SCEVConstant>(
181 SE.getMinusSCEV(LastSubscript, OtherLastSubscript));
182
183 if (Diff == nullptr) {
184 LLVM_DEBUG(dbgs().indent(2)
185 << "No spacial reuse, difference between subscript:\n\t"
186 << *LastSubscript << "\n\t" << OtherLastSubscript
187 << "\nis not constant.\n");
188 return None;
189 }
190
191 bool InSameCacheLine = (Diff->getValue()->getSExtValue() < CLS);
192
193 LLVM_DEBUG({
194 if (InSameCacheLine)
195 dbgs().indent(2) << "Found spacial reuse.\n";
196 else
197 dbgs().indent(2) << "No spacial reuse.\n";
198 });
199
200 return InSameCacheLine;
201 }
202
hasTemporalReuse(const IndexedReference & Other,unsigned MaxDistance,const Loop & L,DependenceInfo & DI,AAResults & AA) const203 Optional<bool> IndexedReference::hasTemporalReuse(const IndexedReference &Other,
204 unsigned MaxDistance,
205 const Loop &L,
206 DependenceInfo &DI,
207 AAResults &AA) const {
208 assert(IsValid && "Expecting a valid reference");
209
210 if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
211 LLVM_DEBUG(dbgs().indent(2)
212 << "No temporal reuse: different base pointer\n");
213 return false;
214 }
215
216 std::unique_ptr<Dependence> D =
217 DI.depends(&StoreOrLoadInst, &Other.StoreOrLoadInst, true);
218
219 if (D == nullptr) {
220 LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: no dependence\n");
221 return false;
222 }
223
224 if (D->isLoopIndependent()) {
225 LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
226 return true;
227 }
228
229 // Check the dependence distance at every loop level. There is temporal reuse
230 // if the distance at the given loop's depth is small (|d| <= MaxDistance) and
231 // it is zero at every other loop level.
232 int LoopDepth = L.getLoopDepth();
233 int Levels = D->getLevels();
234 for (int Level = 1; Level <= Levels; ++Level) {
235 const SCEV *Distance = D->getDistance(Level);
236 const SCEVConstant *SCEVConst = dyn_cast_or_null<SCEVConstant>(Distance);
237
238 if (SCEVConst == nullptr) {
239 LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: distance unknown\n");
240 return None;
241 }
242
243 const ConstantInt &CI = *SCEVConst->getValue();
244 if (Level != LoopDepth && !CI.isZero()) {
245 LLVM_DEBUG(dbgs().indent(2)
246 << "No temporal reuse: distance is not zero at depth=" << Level
247 << "\n");
248 return false;
249 } else if (Level == LoopDepth && CI.getSExtValue() > MaxDistance) {
250 LLVM_DEBUG(
251 dbgs().indent(2)
252 << "No temporal reuse: distance is greater than MaxDistance at depth="
253 << Level << "\n");
254 return false;
255 }
256 }
257
258 LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
259 return true;
260 }
261
computeRefCost(const Loop & L,unsigned CLS) const262 CacheCostTy IndexedReference::computeRefCost(const Loop &L,
263 unsigned CLS) const {
264 assert(IsValid && "Expecting a valid reference");
265 LLVM_DEBUG({
266 dbgs().indent(2) << "Computing cache cost for:\n";
267 dbgs().indent(4) << *this << "\n";
268 });
269
270 // If the indexed reference is loop invariant the cost is one.
271 if (isLoopInvariant(L)) {
272 LLVM_DEBUG(dbgs().indent(4) << "Reference is loop invariant: RefCost=1\n");
273 return 1;
274 }
275
276 const SCEV *TripCount = computeTripCount(L, SE);
277 if (!TripCount) {
278 LLVM_DEBUG(dbgs() << "Trip count of loop " << L.getName()
279 << " could not be computed, using DefaultTripCount\n");
280 const SCEV *ElemSize = Sizes.back();
281 TripCount = SE.getConstant(ElemSize->getType(), DefaultTripCount);
282 }
283 LLVM_DEBUG(dbgs() << "TripCount=" << *TripCount << "\n");
284
285 // If the indexed reference is 'consecutive' the cost is
286 // (TripCount*Stride)/CLS, otherwise the cost is TripCount.
287 const SCEV *RefCost = TripCount;
288
289 if (isConsecutive(L, CLS)) {
290 const SCEV *Coeff = getLastCoefficient();
291 const SCEV *ElemSize = Sizes.back();
292 const SCEV *Stride = SE.getMulExpr(Coeff, ElemSize);
293 const SCEV *CacheLineSize = SE.getConstant(Stride->getType(), CLS);
294 Type *WiderType = SE.getWiderType(Stride->getType(), TripCount->getType());
295 if (SE.isKnownNegative(Stride))
296 Stride = SE.getNegativeSCEV(Stride);
297 Stride = SE.getNoopOrAnyExtend(Stride, WiderType);
298 TripCount = SE.getNoopOrAnyExtend(TripCount, WiderType);
299 const SCEV *Numerator = SE.getMulExpr(Stride, TripCount);
300 RefCost = SE.getUDivExpr(Numerator, CacheLineSize);
301
302 LLVM_DEBUG(dbgs().indent(4)
303 << "Access is consecutive: RefCost=(TripCount*Stride)/CLS="
304 << *RefCost << "\n");
305 } else
306 LLVM_DEBUG(dbgs().indent(4)
307 << "Access is not consecutive: RefCost=TripCount=" << *RefCost
308 << "\n");
309
310 // Attempt to fold RefCost into a constant.
311 if (auto ConstantCost = dyn_cast<SCEVConstant>(RefCost))
312 return ConstantCost->getValue()->getSExtValue();
313
314 LLVM_DEBUG(dbgs().indent(4)
315 << "RefCost is not a constant! Setting to RefCost=InvalidCost "
316 "(invalid value).\n");
317
318 return CacheCost::InvalidCost;
319 }
320
delinearize(const LoopInfo & LI)321 bool IndexedReference::delinearize(const LoopInfo &LI) {
322 assert(Subscripts.empty() && "Subscripts should be empty");
323 assert(Sizes.empty() && "Sizes should be empty");
324 assert(!IsValid && "Should be called once from the constructor");
325 LLVM_DEBUG(dbgs() << "Delinearizing: " << StoreOrLoadInst << "\n");
326
327 const SCEV *ElemSize = SE.getElementSize(&StoreOrLoadInst);
328 const BasicBlock *BB = StoreOrLoadInst.getParent();
329
330 if (Loop *L = LI.getLoopFor(BB)) {
331 const SCEV *AccessFn =
332 SE.getSCEVAtScope(getPointerOperand(&StoreOrLoadInst), L);
333
334 BasePointer = dyn_cast<SCEVUnknown>(SE.getPointerBase(AccessFn));
335 if (BasePointer == nullptr) {
336 LLVM_DEBUG(
337 dbgs().indent(2)
338 << "ERROR: failed to delinearize, can't identify base pointer\n");
339 return false;
340 }
341
342 AccessFn = SE.getMinusSCEV(AccessFn, BasePointer);
343
344 LLVM_DEBUG(dbgs().indent(2) << "In Loop '" << L->getName()
345 << "', AccessFn: " << *AccessFn << "\n");
346
347 SE.delinearize(AccessFn, Subscripts, Sizes,
348 SE.getElementSize(&StoreOrLoadInst));
349
350 if (Subscripts.empty() || Sizes.empty() ||
351 Subscripts.size() != Sizes.size()) {
352 // Attempt to determine whether we have a single dimensional array access.
353 // before giving up.
354 if (!isOneDimensionalArray(*AccessFn, *ElemSize, *L, SE)) {
355 LLVM_DEBUG(dbgs().indent(2)
356 << "ERROR: failed to delinearize reference\n");
357 Subscripts.clear();
358 Sizes.clear();
359 return false;
360 }
361
362 // The array may be accessed in reverse, for example:
363 // for (i = N; i > 0; i--)
364 // A[i] = 0;
365 // In this case, reconstruct the access function using the absolute value
366 // of the step recurrence.
367 const SCEVAddRecExpr *AccessFnAR = dyn_cast<SCEVAddRecExpr>(AccessFn);
368 const SCEV *StepRec = AccessFnAR ? AccessFnAR->getStepRecurrence(SE) : nullptr;
369
370 if (StepRec && SE.isKnownNegative(StepRec))
371 AccessFn = SE.getAddRecExpr(AccessFnAR->getStart(),
372 SE.getNegativeSCEV(StepRec),
373 AccessFnAR->getLoop(),
374 AccessFnAR->getNoWrapFlags());
375 const SCEV *Div = SE.getUDivExactExpr(AccessFn, ElemSize);
376 Subscripts.push_back(Div);
377 Sizes.push_back(ElemSize);
378 }
379
380 return all_of(Subscripts, [&](const SCEV *Subscript) {
381 return isSimpleAddRecurrence(*Subscript, *L);
382 });
383 }
384
385 return false;
386 }
387
isLoopInvariant(const Loop & L) const388 bool IndexedReference::isLoopInvariant(const Loop &L) const {
389 Value *Addr = getPointerOperand(&StoreOrLoadInst);
390 assert(Addr != nullptr && "Expecting either a load or a store instruction");
391 assert(SE.isSCEVable(Addr->getType()) && "Addr should be SCEVable");
392
393 if (SE.isLoopInvariant(SE.getSCEV(Addr), &L))
394 return true;
395
396 // The indexed reference is loop invariant if none of the coefficients use
397 // the loop induction variable.
398 bool allCoeffForLoopAreZero = all_of(Subscripts, [&](const SCEV *Subscript) {
399 return isCoeffForLoopZeroOrInvariant(*Subscript, L);
400 });
401
402 return allCoeffForLoopAreZero;
403 }
404
isConsecutive(const Loop & L,unsigned CLS) const405 bool IndexedReference::isConsecutive(const Loop &L, unsigned CLS) const {
406 // The indexed reference is 'consecutive' if the only coefficient that uses
407 // the loop induction variable is the last one...
408 const SCEV *LastSubscript = Subscripts.back();
409 for (const SCEV *Subscript : Subscripts) {
410 if (Subscript == LastSubscript)
411 continue;
412 if (!isCoeffForLoopZeroOrInvariant(*Subscript, L))
413 return false;
414 }
415
416 // ...and the access stride is less than the cache line size.
417 const SCEV *Coeff = getLastCoefficient();
418 const SCEV *ElemSize = Sizes.back();
419 const SCEV *Stride = SE.getMulExpr(Coeff, ElemSize);
420 const SCEV *CacheLineSize = SE.getConstant(Stride->getType(), CLS);
421
422 Stride = SE.isKnownNegative(Stride) ? SE.getNegativeSCEV(Stride) : Stride;
423 return SE.isKnownPredicate(ICmpInst::ICMP_ULT, Stride, CacheLineSize);
424 }
425
getLastCoefficient() const426 const SCEV *IndexedReference::getLastCoefficient() const {
427 const SCEV *LastSubscript = getLastSubscript();
428 assert(isa<SCEVAddRecExpr>(LastSubscript) &&
429 "Expecting a SCEV add recurrence expression");
430 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LastSubscript);
431 return AR->getStepRecurrence(SE);
432 }
433
isCoeffForLoopZeroOrInvariant(const SCEV & Subscript,const Loop & L) const434 bool IndexedReference::isCoeffForLoopZeroOrInvariant(const SCEV &Subscript,
435 const Loop &L) const {
436 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&Subscript);
437 return (AR != nullptr) ? AR->getLoop() != &L
438 : SE.isLoopInvariant(&Subscript, &L);
439 }
440
isSimpleAddRecurrence(const SCEV & Subscript,const Loop & L) const441 bool IndexedReference::isSimpleAddRecurrence(const SCEV &Subscript,
442 const Loop &L) const {
443 if (!isa<SCEVAddRecExpr>(Subscript))
444 return false;
445
446 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(&Subscript);
447 assert(AR->getLoop() && "AR should have a loop");
448
449 if (!AR->isAffine())
450 return false;
451
452 const SCEV *Start = AR->getStart();
453 const SCEV *Step = AR->getStepRecurrence(SE);
454
455 if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
456 return false;
457
458 return true;
459 }
460
isAliased(const IndexedReference & Other,AAResults & AA) const461 bool IndexedReference::isAliased(const IndexedReference &Other,
462 AAResults &AA) const {
463 const auto &Loc1 = MemoryLocation::get(&StoreOrLoadInst);
464 const auto &Loc2 = MemoryLocation::get(&Other.StoreOrLoadInst);
465 return AA.isMustAlias(Loc1, Loc2);
466 }
467
468 //===----------------------------------------------------------------------===//
469 // CacheCost implementation
470 //
operator <<(raw_ostream & OS,const CacheCost & CC)471 raw_ostream &llvm::operator<<(raw_ostream &OS, const CacheCost &CC) {
472 for (const auto &LC : CC.LoopCosts) {
473 const Loop *L = LC.first;
474 OS << "Loop '" << L->getName() << "' has cost = " << LC.second << "\n";
475 }
476 return OS;
477 }
478
CacheCost(const LoopVectorTy & Loops,const LoopInfo & LI,ScalarEvolution & SE,TargetTransformInfo & TTI,AAResults & AA,DependenceInfo & DI,Optional<unsigned> TRT)479 CacheCost::CacheCost(const LoopVectorTy &Loops, const LoopInfo &LI,
480 ScalarEvolution &SE, TargetTransformInfo &TTI,
481 AAResults &AA, DependenceInfo &DI,
482 Optional<unsigned> TRT)
483 : Loops(Loops), TripCounts(), LoopCosts(),
484 TRT((TRT == None) ? Optional<unsigned>(TemporalReuseThreshold) : TRT),
485 LI(LI), SE(SE), TTI(TTI), AA(AA), DI(DI) {
486 assert(!Loops.empty() && "Expecting a non-empty loop vector.");
487
488 for (const Loop *L : Loops) {
489 unsigned TripCount = SE.getSmallConstantTripCount(L);
490 TripCount = (TripCount == 0) ? DefaultTripCount : TripCount;
491 TripCounts.push_back({L, TripCount});
492 }
493
494 calculateCacheFootprint();
495 }
496
497 std::unique_ptr<CacheCost>
getCacheCost(Loop & Root,LoopStandardAnalysisResults & AR,DependenceInfo & DI,Optional<unsigned> TRT)498 CacheCost::getCacheCost(Loop &Root, LoopStandardAnalysisResults &AR,
499 DependenceInfo &DI, Optional<unsigned> TRT) {
500 if (!Root.isOutermost()) {
501 LLVM_DEBUG(dbgs() << "Expecting the outermost loop in a loop nest\n");
502 return nullptr;
503 }
504
505 LoopVectorTy Loops;
506 append_range(Loops, breadth_first(&Root));
507
508 if (!getInnerMostLoop(Loops)) {
509 LLVM_DEBUG(dbgs() << "Cannot compute cache cost of loop nest with more "
510 "than one innermost loop\n");
511 return nullptr;
512 }
513
514 return std::make_unique<CacheCost>(Loops, AR.LI, AR.SE, AR.TTI, AR.AA, DI, TRT);
515 }
516
calculateCacheFootprint()517 void CacheCost::calculateCacheFootprint() {
518 LLVM_DEBUG(dbgs() << "POPULATING REFERENCE GROUPS\n");
519 ReferenceGroupsTy RefGroups;
520 if (!populateReferenceGroups(RefGroups))
521 return;
522
523 LLVM_DEBUG(dbgs() << "COMPUTING LOOP CACHE COSTS\n");
524 for (const Loop *L : Loops) {
525 assert((std::find_if(LoopCosts.begin(), LoopCosts.end(),
526 [L](const LoopCacheCostTy &LCC) {
527 return LCC.first == L;
528 }) == LoopCosts.end()) &&
529 "Should not add duplicate element");
530 CacheCostTy LoopCost = computeLoopCacheCost(*L, RefGroups);
531 LoopCosts.push_back(std::make_pair(L, LoopCost));
532 }
533
534 sortLoopCosts();
535 RefGroups.clear();
536 }
537
populateReferenceGroups(ReferenceGroupsTy & RefGroups) const538 bool CacheCost::populateReferenceGroups(ReferenceGroupsTy &RefGroups) const {
539 assert(RefGroups.empty() && "Reference groups should be empty");
540
541 unsigned CLS = TTI.getCacheLineSize();
542 Loop *InnerMostLoop = getInnerMostLoop(Loops);
543 assert(InnerMostLoop != nullptr && "Expecting a valid innermost loop");
544
545 for (BasicBlock *BB : InnerMostLoop->getBlocks()) {
546 for (Instruction &I : *BB) {
547 if (!isa<StoreInst>(I) && !isa<LoadInst>(I))
548 continue;
549
550 std::unique_ptr<IndexedReference> R(new IndexedReference(I, LI, SE));
551 if (!R->isValid())
552 continue;
553
554 bool Added = false;
555 for (ReferenceGroupTy &RefGroup : RefGroups) {
556 const IndexedReference &Representative = *RefGroup.front().get();
557 LLVM_DEBUG({
558 dbgs() << "References:\n";
559 dbgs().indent(2) << *R << "\n";
560 dbgs().indent(2) << Representative << "\n";
561 });
562
563
564 // FIXME: Both positive and negative access functions will be placed
565 // into the same reference group, resulting in a bi-directional array
566 // access such as:
567 // for (i = N; i > 0; i--)
568 // A[i] = A[N - i];
569 // having the same cost calculation as a single dimention access pattern
570 // for (i = 0; i < N; i++)
571 // A[i] = A[i];
572 // when in actuality, depending on the array size, the first example
573 // should have a cost closer to 2x the second due to the two cache
574 // access per iteration from opposite ends of the array
575 Optional<bool> HasTemporalReuse =
576 R->hasTemporalReuse(Representative, *TRT, *InnerMostLoop, DI, AA);
577 Optional<bool> HasSpacialReuse =
578 R->hasSpacialReuse(Representative, CLS, AA);
579
580 if ((HasTemporalReuse.hasValue() && *HasTemporalReuse) ||
581 (HasSpacialReuse.hasValue() && *HasSpacialReuse)) {
582 RefGroup.push_back(std::move(R));
583 Added = true;
584 break;
585 }
586 }
587
588 if (!Added) {
589 ReferenceGroupTy RG;
590 RG.push_back(std::move(R));
591 RefGroups.push_back(std::move(RG));
592 }
593 }
594 }
595
596 if (RefGroups.empty())
597 return false;
598
599 LLVM_DEBUG({
600 dbgs() << "\nIDENTIFIED REFERENCE GROUPS:\n";
601 int n = 1;
602 for (const ReferenceGroupTy &RG : RefGroups) {
603 dbgs().indent(2) << "RefGroup " << n << ":\n";
604 for (const auto &IR : RG)
605 dbgs().indent(4) << *IR << "\n";
606 n++;
607 }
608 dbgs() << "\n";
609 });
610
611 return true;
612 }
613
614 CacheCostTy
computeLoopCacheCost(const Loop & L,const ReferenceGroupsTy & RefGroups) const615 CacheCost::computeLoopCacheCost(const Loop &L,
616 const ReferenceGroupsTy &RefGroups) const {
617 if (!L.isLoopSimplifyForm())
618 return InvalidCost;
619
620 LLVM_DEBUG(dbgs() << "Considering loop '" << L.getName()
621 << "' as innermost loop.\n");
622
623 // Compute the product of the trip counts of each other loop in the nest.
624 CacheCostTy TripCountsProduct = 1;
625 for (const auto &TC : TripCounts) {
626 if (TC.first == &L)
627 continue;
628 TripCountsProduct *= TC.second;
629 }
630
631 CacheCostTy LoopCost = 0;
632 for (const ReferenceGroupTy &RG : RefGroups) {
633 CacheCostTy RefGroupCost = computeRefGroupCacheCost(RG, L);
634 LoopCost += RefGroupCost * TripCountsProduct;
635 }
636
637 LLVM_DEBUG(dbgs().indent(2) << "Loop '" << L.getName()
638 << "' has cost=" << LoopCost << "\n");
639
640 return LoopCost;
641 }
642
computeRefGroupCacheCost(const ReferenceGroupTy & RG,const Loop & L) const643 CacheCostTy CacheCost::computeRefGroupCacheCost(const ReferenceGroupTy &RG,
644 const Loop &L) const {
645 assert(!RG.empty() && "Reference group should have at least one member.");
646
647 const IndexedReference *Representative = RG.front().get();
648 return Representative->computeRefCost(L, TTI.getCacheLineSize());
649 }
650
651 //===----------------------------------------------------------------------===//
652 // LoopCachePrinterPass implementation
653 //
run(Loop & L,LoopAnalysisManager & AM,LoopStandardAnalysisResults & AR,LPMUpdater & U)654 PreservedAnalyses LoopCachePrinterPass::run(Loop &L, LoopAnalysisManager &AM,
655 LoopStandardAnalysisResults &AR,
656 LPMUpdater &U) {
657 Function *F = L.getHeader()->getParent();
658 DependenceInfo DI(F, &AR.AA, &AR.SE, &AR.LI);
659
660 if (auto CC = CacheCost::getCacheCost(L, AR, DI))
661 OS << *CC;
662
663 return PreservedAnalyses::all();
664 }
665