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/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 
35 using namespace llvm;
36 
37 #define DEBUG_TYPE "loop-cache-cost"
38 
39 static cl::opt<unsigned> DefaultTripCount(
40     "default-trip-count", cl::init(100), cl::Hidden,
41     cl::desc("Use this to specify the default trip count of a loop"));
42 
43 // In this analysis two array references are considered to exhibit temporal
44 // reuse if they access either the same memory location, or a memory location
45 // with distance smaller than a configurable threshold.
46 static cl::opt<unsigned> TemporalReuseThreshold(
47     "temporal-reuse-threshold", cl::init(2), cl::Hidden,
48     cl::desc("Use this to specify the max. distance between array elements "
49              "accessed in a loop so that the elements are classified to have "
50              "temporal reuse"));
51 
52 /// Retrieve the innermost loop in the given loop nest \p Loops. It returns a
53 /// nullptr if any loops in the loop vector supplied has more than one sibling.
54 /// The loop vector is expected to contain loops collected in breadth-first
55 /// order.
56 static Loop *getInnerMostLoop(const LoopVectorTy &Loops) {
57   assert(!Loops.empty() && "Expecting a non-empy loop vector");
58 
59   Loop *LastLoop = Loops.back();
60   Loop *ParentLoop = LastLoop->getParentLoop();
61 
62   if (ParentLoop == nullptr) {
63     assert(Loops.size() == 1 && "Expecting a single loop");
64     return LastLoop;
65   }
66 
67   return (std::is_sorted(Loops.begin(), Loops.end(),
68                          [](const Loop *L1, const Loop *L2) {
69                            return L1->getLoopDepth() < L2->getLoopDepth();
70                          }))
71              ? LastLoop
72              : nullptr;
73 }
74 
75 static bool isOneDimensionalArray(const SCEV &AccessFn, const SCEV &ElemSize,
76                                   const Loop &L, ScalarEvolution &SE) {
77   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&AccessFn);
78   if (!AR || !AR->isAffine())
79     return false;
80 
81   assert(AR->getLoop() && "AR should have a loop");
82 
83   // Check that start and increment are not add recurrences.
84   const SCEV *Start = AR->getStart();
85   const SCEV *Step = AR->getStepRecurrence(SE);
86   if (isa<SCEVAddRecExpr>(Start) || isa<SCEVAddRecExpr>(Step))
87     return false;
88 
89   // Check that start and increment are both invariant in the loop.
90   if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
91     return false;
92 
93   return AR->getStepRecurrence(SE) == &ElemSize;
94 }
95 
96 /// Compute the trip count for the given loop \p L. Return the SCEV expression
97 /// for the trip count or nullptr if it cannot be computed.
98 static const SCEV *computeTripCount(const Loop &L, ScalarEvolution &SE) {
99   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(&L);
100   if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) ||
101       !isa<SCEVConstant>(BackedgeTakenCount))
102     return nullptr;
103 
104   return SE.getAddExpr(BackedgeTakenCount,
105                        SE.getOne(BackedgeTakenCount->getType()));
106 }
107 
108 //===----------------------------------------------------------------------===//
109 // IndexedReference implementation
110 //
111 raw_ostream &llvm::operator<<(raw_ostream &OS, const IndexedReference &R) {
112   if (!R.IsValid) {
113     OS << R.StoreOrLoadInst;
114     OS << ", IsValid=false.";
115     return OS;
116   }
117 
118   OS << *R.BasePointer;
119   for (const SCEV *Subscript : R.Subscripts)
120     OS << "[" << *Subscript << "]";
121 
122   OS << ", Sizes: ";
123   for (const SCEV *Size : R.Sizes)
124     OS << "[" << *Size << "]";
125 
126   return OS;
127 }
128 
129 IndexedReference::IndexedReference(Instruction &StoreOrLoadInst,
130                                    const LoopInfo &LI, ScalarEvolution &SE)
131     : StoreOrLoadInst(StoreOrLoadInst), SE(SE) {
132   assert((isa<StoreInst>(StoreOrLoadInst) || isa<LoadInst>(StoreOrLoadInst)) &&
133          "Expecting a load or store instruction");
134 
135   IsValid = delinearize(LI);
136   if (IsValid)
137     LLVM_DEBUG(dbgs().indent(2) << "Succesfully delinearized: " << *this
138                                 << "\n");
139 }
140 
141 Optional<bool> IndexedReference::hasSpacialReuse(const IndexedReference &Other,
142                                                  unsigned CLS,
143                                                  AliasAnalysis &AA) const {
144   assert(IsValid && "Expecting a valid reference");
145 
146   if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
147     LLVM_DEBUG(dbgs().indent(2)
148                << "No spacial reuse: different base pointers\n");
149     return false;
150   }
151 
152   unsigned NumSubscripts = getNumSubscripts();
153   if (NumSubscripts != Other.getNumSubscripts()) {
154     LLVM_DEBUG(dbgs().indent(2)
155                << "No spacial reuse: different number of subscripts\n");
156     return false;
157   }
158 
159   // all subscripts must be equal, except the leftmost one (the last one).
160   for (auto SubNum : seq<unsigned>(0, NumSubscripts - 1)) {
161     if (getSubscript(SubNum) != Other.getSubscript(SubNum)) {
162       LLVM_DEBUG(dbgs().indent(2) << "No spacial reuse, different subscripts: "
163                                   << "\n\t" << *getSubscript(SubNum) << "\n\t"
164                                   << *Other.getSubscript(SubNum) << "\n");
165       return false;
166     }
167   }
168 
169   // the difference between the last subscripts must be less than the cache line
170   // size.
171   const SCEV *LastSubscript = getLastSubscript();
172   const SCEV *OtherLastSubscript = Other.getLastSubscript();
173   const SCEVConstant *Diff = dyn_cast<SCEVConstant>(
174       SE.getMinusSCEV(LastSubscript, OtherLastSubscript));
175 
176   if (Diff == nullptr) {
177     LLVM_DEBUG(dbgs().indent(2)
178                << "No spacial reuse, difference between subscript:\n\t"
179                << *LastSubscript << "\n\t" << OtherLastSubscript
180                << "\nis not constant.\n");
181     return None;
182   }
183 
184   bool InSameCacheLine = (Diff->getValue()->getSExtValue() < CLS);
185 
186   LLVM_DEBUG({
187     if (InSameCacheLine)
188       dbgs().indent(2) << "Found spacial reuse.\n";
189     else
190       dbgs().indent(2) << "No spacial reuse.\n";
191   });
192 
193   return InSameCacheLine;
194 }
195 
196 Optional<bool> IndexedReference::hasTemporalReuse(const IndexedReference &Other,
197                                                   unsigned MaxDistance,
198                                                   const Loop &L,
199                                                   DependenceInfo &DI,
200                                                   AliasAnalysis &AA) const {
201   assert(IsValid && "Expecting a valid reference");
202 
203   if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
204     LLVM_DEBUG(dbgs().indent(2)
205                << "No temporal reuse: different base pointer\n");
206     return false;
207   }
208 
209   std::unique_ptr<Dependence> D =
210       DI.depends(&StoreOrLoadInst, &Other.StoreOrLoadInst, true);
211 
212   if (D == nullptr) {
213     LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: no dependence\n");
214     return false;
215   }
216 
217   if (D->isLoopIndependent()) {
218     LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
219     return true;
220   }
221 
222   // Check the dependence distance at every loop level. There is temporal reuse
223   // if the distance at the given loop's depth is small (|d| <= MaxDistance) and
224   // it is zero at every other loop level.
225   int LoopDepth = L.getLoopDepth();
226   int Levels = D->getLevels();
227   for (int Level = 1; Level <= Levels; ++Level) {
228     const SCEV *Distance = D->getDistance(Level);
229     const SCEVConstant *SCEVConst = dyn_cast_or_null<SCEVConstant>(Distance);
230 
231     if (SCEVConst == nullptr) {
232       LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: distance unknown\n");
233       return None;
234     }
235 
236     const ConstantInt &CI = *SCEVConst->getValue();
237     if (Level != LoopDepth && !CI.isZero()) {
238       LLVM_DEBUG(dbgs().indent(2)
239                  << "No temporal reuse: distance is not zero at depth=" << Level
240                  << "\n");
241       return false;
242     } else if (Level == LoopDepth && CI.getSExtValue() > MaxDistance) {
243       LLVM_DEBUG(
244           dbgs().indent(2)
245           << "No temporal reuse: distance is greater than MaxDistance at depth="
246           << Level << "\n");
247       return false;
248     }
249   }
250 
251   LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
252   return true;
253 }
254 
255 CacheCostTy IndexedReference::computeRefCost(const Loop &L,
256                                              unsigned CLS) const {
257   assert(IsValid && "Expecting a valid reference");
258   LLVM_DEBUG({
259     dbgs().indent(2) << "Computing cache cost for:\n";
260     dbgs().indent(4) << *this << "\n";
261   });
262 
263   // If the indexed reference is loop invariant the cost is one.
264   if (isLoopInvariant(L)) {
265     LLVM_DEBUG(dbgs().indent(4) << "Reference is loop invariant: RefCost=1\n");
266     return 1;
267   }
268 
269   const SCEV *TripCount = computeTripCount(L, SE);
270   if (!TripCount) {
271     LLVM_DEBUG(dbgs() << "Trip count of loop " << L.getName()
272                       << " could not be computed, using DefaultTripCount\n");
273     const SCEV *ElemSize = Sizes.back();
274     TripCount = SE.getConstant(ElemSize->getType(), DefaultTripCount);
275   }
276   LLVM_DEBUG(dbgs() << "TripCount=" << *TripCount << "\n");
277 
278   // If the indexed reference is 'consecutive' the cost is
279   // (TripCount*Stride)/CLS, otherwise the cost is TripCount.
280   const SCEV *RefCost = TripCount;
281 
282   if (isConsecutive(L, CLS)) {
283     const SCEV *Coeff = getLastCoefficient();
284     const SCEV *ElemSize = Sizes.back();
285     const SCEV *Stride = SE.getMulExpr(Coeff, ElemSize);
286     const SCEV *CacheLineSize = SE.getConstant(Stride->getType(), CLS);
287     Type *WiderType = SE.getWiderType(Stride->getType(), TripCount->getType());
288     Stride = SE.getNoopOrSignExtend(Stride, WiderType);
289     TripCount = SE.getNoopOrAnyExtend(TripCount, WiderType);
290     const SCEV *Numerator = SE.getMulExpr(Stride, TripCount);
291     RefCost = SE.getUDivExpr(Numerator, CacheLineSize);
292     LLVM_DEBUG(dbgs().indent(4)
293                << "Access is consecutive: RefCost=(TripCount*Stride)/CLS="
294                << *RefCost << "\n");
295   } else
296     LLVM_DEBUG(dbgs().indent(4)
297                << "Access is not consecutive: RefCost=TripCount=" << *RefCost
298                << "\n");
299 
300   // Attempt to fold RefCost into a constant.
301   if (auto ConstantCost = dyn_cast<SCEVConstant>(RefCost))
302     return ConstantCost->getValue()->getSExtValue();
303 
304   LLVM_DEBUG(dbgs().indent(4)
305              << "RefCost is not a constant! Setting to RefCost=InvalidCost "
306                 "(invalid value).\n");
307 
308   return CacheCost::InvalidCost;
309 }
310 
311 bool IndexedReference::delinearize(const LoopInfo &LI) {
312   assert(Subscripts.empty() && "Subscripts should be empty");
313   assert(Sizes.empty() && "Sizes should be empty");
314   assert(!IsValid && "Should be called once from the constructor");
315   LLVM_DEBUG(dbgs() << "Delinearizing: " << StoreOrLoadInst << "\n");
316 
317   const SCEV *ElemSize = SE.getElementSize(&StoreOrLoadInst);
318   const BasicBlock *BB = StoreOrLoadInst.getParent();
319 
320   if (Loop *L = LI.getLoopFor(BB)) {
321     const SCEV *AccessFn =
322         SE.getSCEVAtScope(getPointerOperand(&StoreOrLoadInst), L);
323 
324     BasePointer = dyn_cast<SCEVUnknown>(SE.getPointerBase(AccessFn));
325     if (BasePointer == nullptr) {
326       LLVM_DEBUG(
327           dbgs().indent(2)
328           << "ERROR: failed to delinearize, can't identify base pointer\n");
329       return false;
330     }
331 
332     AccessFn = SE.getMinusSCEV(AccessFn, BasePointer);
333 
334     LLVM_DEBUG(dbgs().indent(2) << "In Loop '" << L->getName()
335                                 << "', AccessFn: " << *AccessFn << "\n");
336 
337     SE.delinearize(AccessFn, Subscripts, Sizes,
338                    SE.getElementSize(&StoreOrLoadInst));
339 
340     if (Subscripts.empty() || Sizes.empty() ||
341         Subscripts.size() != Sizes.size()) {
342       // Attempt to determine whether we have a single dimensional array access.
343       // before giving up.
344       if (!isOneDimensionalArray(*AccessFn, *ElemSize, *L, SE)) {
345         LLVM_DEBUG(dbgs().indent(2)
346                    << "ERROR: failed to delinearize reference\n");
347         Subscripts.clear();
348         Sizes.clear();
349         return false;
350       }
351 
352       const SCEV *Div = SE.getUDivExactExpr(AccessFn, ElemSize);
353       Subscripts.push_back(Div);
354       Sizes.push_back(ElemSize);
355     }
356 
357     return all_of(Subscripts, [&](const SCEV *Subscript) {
358       return isSimpleAddRecurrence(*Subscript, *L);
359     });
360   }
361 
362   return false;
363 }
364 
365 bool IndexedReference::isLoopInvariant(const Loop &L) const {
366   Value *Addr = getPointerOperand(&StoreOrLoadInst);
367   assert(Addr != nullptr && "Expecting either a load or a store instruction");
368   assert(SE.isSCEVable(Addr->getType()) && "Addr should be SCEVable");
369 
370   if (SE.isLoopInvariant(SE.getSCEV(Addr), &L))
371     return true;
372 
373   // The indexed reference is loop invariant if none of the coefficients use
374   // the loop induction variable.
375   bool allCoeffForLoopAreZero = all_of(Subscripts, [&](const SCEV *Subscript) {
376     return isCoeffForLoopZeroOrInvariant(*Subscript, L);
377   });
378 
379   return allCoeffForLoopAreZero;
380 }
381 
382 bool IndexedReference::isConsecutive(const Loop &L, unsigned CLS) const {
383   // The indexed reference is 'consecutive' if the only coefficient that uses
384   // the loop induction variable is the last one...
385   const SCEV *LastSubscript = Subscripts.back();
386   for (const SCEV *Subscript : Subscripts) {
387     if (Subscript == LastSubscript)
388       continue;
389     if (!isCoeffForLoopZeroOrInvariant(*Subscript, L))
390       return false;
391   }
392 
393   // ...and the access stride is less than the cache line size.
394   const SCEV *Coeff = getLastCoefficient();
395   const SCEV *ElemSize = Sizes.back();
396   const SCEV *Stride = SE.getMulExpr(Coeff, ElemSize);
397   const SCEV *CacheLineSize = SE.getConstant(Stride->getType(), CLS);
398 
399   return SE.isKnownPredicate(ICmpInst::ICMP_ULT, Stride, CacheLineSize);
400 }
401 
402 const SCEV *IndexedReference::getLastCoefficient() const {
403   const SCEV *LastSubscript = getLastSubscript();
404   assert(isa<SCEVAddRecExpr>(LastSubscript) &&
405          "Expecting a SCEV add recurrence expression");
406   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LastSubscript);
407   return AR->getStepRecurrence(SE);
408 }
409 
410 bool IndexedReference::isCoeffForLoopZeroOrInvariant(const SCEV &Subscript,
411                                                      const Loop &L) const {
412   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&Subscript);
413   return (AR != nullptr) ? AR->getLoop() != &L
414                          : SE.isLoopInvariant(&Subscript, &L);
415 }
416 
417 bool IndexedReference::isSimpleAddRecurrence(const SCEV &Subscript,
418                                              const Loop &L) const {
419   if (!isa<SCEVAddRecExpr>(Subscript))
420     return false;
421 
422   const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(&Subscript);
423   assert(AR->getLoop() && "AR should have a loop");
424 
425   if (!AR->isAffine())
426     return false;
427 
428   const SCEV *Start = AR->getStart();
429   const SCEV *Step = AR->getStepRecurrence(SE);
430 
431   if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
432     return false;
433 
434   return true;
435 }
436 
437 bool IndexedReference::isAliased(const IndexedReference &Other,
438                                  AliasAnalysis &AA) const {
439   const auto &Loc1 = MemoryLocation::get(&StoreOrLoadInst);
440   const auto &Loc2 = MemoryLocation::get(&Other.StoreOrLoadInst);
441   return AA.isMustAlias(Loc1, Loc2);
442 }
443 
444 //===----------------------------------------------------------------------===//
445 // CacheCost implementation
446 //
447 raw_ostream &llvm::operator<<(raw_ostream &OS, const CacheCost &CC) {
448   for (const auto &LC : CC.LoopCosts) {
449     const Loop *L = LC.first;
450     OS << "Loop '" << L->getName() << "' has cost = " << LC.second << "\n";
451   }
452   return OS;
453 }
454 
455 CacheCost::CacheCost(const LoopVectorTy &Loops, const LoopInfo &LI,
456                      ScalarEvolution &SE, TargetTransformInfo &TTI,
457                      AliasAnalysis &AA, DependenceInfo &DI,
458                      Optional<unsigned> TRT)
459     : Loops(Loops), TripCounts(), LoopCosts(),
460       TRT((TRT == None) ? Optional<unsigned>(TemporalReuseThreshold) : TRT),
461       LI(LI), SE(SE), TTI(TTI), AA(AA), DI(DI) {
462   assert(!Loops.empty() && "Expecting a non-empty loop vector.");
463 
464   for (const Loop *L : Loops) {
465     unsigned TripCount = SE.getSmallConstantTripCount(L);
466     TripCount = (TripCount == 0) ? DefaultTripCount : TripCount;
467     TripCounts.push_back({L, TripCount});
468   }
469 
470   calculateCacheFootprint();
471 }
472 
473 std::unique_ptr<CacheCost>
474 CacheCost::getCacheCost(Loop &Root, LoopStandardAnalysisResults &AR,
475                         DependenceInfo &DI, Optional<unsigned> TRT) {
476   if (Root.getParentLoop()) {
477     LLVM_DEBUG(dbgs() << "Expecting the outermost loop in a loop nest\n");
478     return nullptr;
479   }
480 
481   LoopVectorTy Loops;
482   for (Loop *L : breadth_first(&Root))
483     Loops.push_back(L);
484 
485   if (!getInnerMostLoop(Loops)) {
486     LLVM_DEBUG(dbgs() << "Cannot compute cache cost of loop nest with more "
487                          "than one innermost loop\n");
488     return nullptr;
489   }
490 
491   return std::make_unique<CacheCost>(Loops, AR.LI, AR.SE, AR.TTI, AR.AA, DI, TRT);
492 }
493 
494 void CacheCost::calculateCacheFootprint() {
495   LLVM_DEBUG(dbgs() << "POPULATING REFERENCE GROUPS\n");
496   ReferenceGroupsTy RefGroups;
497   if (!populateReferenceGroups(RefGroups))
498     return;
499 
500   LLVM_DEBUG(dbgs() << "COMPUTING LOOP CACHE COSTS\n");
501   for (const Loop *L : Loops) {
502     assert((std::find_if(LoopCosts.begin(), LoopCosts.end(),
503                          [L](const LoopCacheCostTy &LCC) {
504                            return LCC.first == L;
505                          }) == LoopCosts.end()) &&
506            "Should not add duplicate element");
507     CacheCostTy LoopCost = computeLoopCacheCost(*L, RefGroups);
508     LoopCosts.push_back(std::make_pair(L, LoopCost));
509   }
510 
511   sortLoopCosts();
512   RefGroups.clear();
513 }
514 
515 bool CacheCost::populateReferenceGroups(ReferenceGroupsTy &RefGroups) const {
516   assert(RefGroups.empty() && "Reference groups should be empty");
517 
518   unsigned CLS = TTI.getCacheLineSize();
519   Loop *InnerMostLoop = getInnerMostLoop(Loops);
520   assert(InnerMostLoop != nullptr && "Expecting a valid innermost loop");
521 
522   for (BasicBlock *BB : InnerMostLoop->getBlocks()) {
523     for (Instruction &I : *BB) {
524       if (!isa<StoreInst>(I) && !isa<LoadInst>(I))
525         continue;
526 
527       std::unique_ptr<IndexedReference> R(new IndexedReference(I, LI, SE));
528       if (!R->isValid())
529         continue;
530 
531       bool Added = false;
532       for (ReferenceGroupTy &RefGroup : RefGroups) {
533         const IndexedReference &Representative = *RefGroup.front().get();
534         LLVM_DEBUG({
535           dbgs() << "References:\n";
536           dbgs().indent(2) << *R << "\n";
537           dbgs().indent(2) << Representative << "\n";
538         });
539 
540         Optional<bool> HasTemporalReuse =
541             R->hasTemporalReuse(Representative, *TRT, *InnerMostLoop, DI, AA);
542         Optional<bool> HasSpacialReuse =
543             R->hasSpacialReuse(Representative, CLS, AA);
544 
545         if ((HasTemporalReuse.hasValue() && *HasTemporalReuse) ||
546             (HasSpacialReuse.hasValue() && *HasSpacialReuse)) {
547           RefGroup.push_back(std::move(R));
548           Added = true;
549           break;
550         }
551       }
552 
553       if (!Added) {
554         ReferenceGroupTy RG;
555         RG.push_back(std::move(R));
556         RefGroups.push_back(std::move(RG));
557       }
558     }
559   }
560 
561   if (RefGroups.empty())
562     return false;
563 
564   LLVM_DEBUG({
565     dbgs() << "\nIDENTIFIED REFERENCE GROUPS:\n";
566     int n = 1;
567     for (const ReferenceGroupTy &RG : RefGroups) {
568       dbgs().indent(2) << "RefGroup " << n << ":\n";
569       for (const auto &IR : RG)
570         dbgs().indent(4) << *IR << "\n";
571       n++;
572     }
573     dbgs() << "\n";
574   });
575 
576   return true;
577 }
578 
579 CacheCostTy
580 CacheCost::computeLoopCacheCost(const Loop &L,
581                                 const ReferenceGroupsTy &RefGroups) const {
582   if (!L.isLoopSimplifyForm())
583     return InvalidCost;
584 
585   LLVM_DEBUG(dbgs() << "Considering loop '" << L.getName()
586                     << "' as innermost loop.\n");
587 
588   // Compute the product of the trip counts of each other loop in the nest.
589   CacheCostTy TripCountsProduct = 1;
590   for (const auto &TC : TripCounts) {
591     if (TC.first == &L)
592       continue;
593     TripCountsProduct *= TC.second;
594   }
595 
596   CacheCostTy LoopCost = 0;
597   for (const ReferenceGroupTy &RG : RefGroups) {
598     CacheCostTy RefGroupCost = computeRefGroupCacheCost(RG, L);
599     LoopCost += RefGroupCost * TripCountsProduct;
600   }
601 
602   LLVM_DEBUG(dbgs().indent(2) << "Loop '" << L.getName()
603                               << "' has cost=" << LoopCost << "\n");
604 
605   return LoopCost;
606 }
607 
608 CacheCostTy CacheCost::computeRefGroupCacheCost(const ReferenceGroupTy &RG,
609                                                 const Loop &L) const {
610   assert(!RG.empty() && "Reference group should have at least one member.");
611 
612   const IndexedReference *Representative = RG.front().get();
613   return Representative->computeRefCost(L, TTI.getCacheLineSize());
614 }
615 
616 //===----------------------------------------------------------------------===//
617 // LoopCachePrinterPass implementation
618 //
619 PreservedAnalyses LoopCachePrinterPass::run(Loop &L, LoopAnalysisManager &AM,
620                                             LoopStandardAnalysisResults &AR,
621                                             LPMUpdater &U) {
622   Function *F = L.getHeader()->getParent();
623   DependenceInfo DI(F, &AR.AA, &AR.SE, &AR.LI);
624 
625   if (auto CC = CacheCost::getCacheCost(L, AR, DI))
626     OS << *CC;
627 
628   return PreservedAnalyses::all();
629 }
630