xref: /freebsd/contrib/llvm-project/llvm/lib/Analysis/LoopCacheAnalysis.cpp (revision ba7b7f94c239ce43343d7af403734fdc941b7664)
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/Delinearization.h"
34 #include "llvm/Analysis/DependenceAnalysis.h"
35 #include "llvm/Analysis/LoopInfo.h"
36 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
37 #include "llvm/Analysis/TargetTransformInfo.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 
41 using namespace llvm;
42 
43 #define DEBUG_TYPE "loop-cache-cost"
44 
45 static cl::opt<unsigned> DefaultTripCount(
46     "default-trip-count", cl::init(100), cl::Hidden,
47     cl::desc("Use this to specify the default trip count of a loop"));
48 
49 // In this analysis two array references are considered to exhibit temporal
50 // reuse if they access either the same memory location, or a memory location
51 // with distance smaller than a configurable threshold.
52 static cl::opt<unsigned> TemporalReuseThreshold(
53     "temporal-reuse-threshold", cl::init(2), cl::Hidden,
54     cl::desc("Use this to specify the max. distance between array elements "
55              "accessed in a loop so that the elements are classified to have "
56              "temporal reuse"));
57 
58 /// Retrieve the innermost loop in the given loop nest \p Loops. It returns a
59 /// nullptr if any loops in the loop vector supplied has more than one sibling.
60 /// The loop vector is expected to contain loops collected in breadth-first
61 /// order.
62 static Loop *getInnerMostLoop(const LoopVectorTy &Loops) {
63   assert(!Loops.empty() && "Expecting a non-empy loop vector");
64 
65   Loop *LastLoop = Loops.back();
66   Loop *ParentLoop = LastLoop->getParentLoop();
67 
68   if (ParentLoop == nullptr) {
69     assert(Loops.size() == 1 && "Expecting a single loop");
70     return LastLoop;
71   }
72 
73   return (llvm::is_sorted(Loops,
74                           [](const Loop *L1, const Loop *L2) {
75                             return L1->getLoopDepth() < L2->getLoopDepth();
76                           }))
77              ? LastLoop
78              : nullptr;
79 }
80 
81 static bool isOneDimensionalArray(const SCEV &AccessFn, const SCEV &ElemSize,
82                                   const Loop &L, ScalarEvolution &SE) {
83   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&AccessFn);
84   if (!AR || !AR->isAffine())
85     return false;
86 
87   assert(AR->getLoop() && "AR should have a loop");
88 
89   // Check that start and increment are not add recurrences.
90   const SCEV *Start = AR->getStart();
91   const SCEV *Step = AR->getStepRecurrence(SE);
92   if (isa<SCEVAddRecExpr>(Start) || isa<SCEVAddRecExpr>(Step))
93     return false;
94 
95   // Check that start and increment are both invariant in the loop.
96   if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
97     return false;
98 
99   const SCEV *StepRec = AR->getStepRecurrence(SE);
100   if (StepRec && SE.isKnownNegative(StepRec))
101     StepRec = SE.getNegativeSCEV(StepRec);
102 
103   return StepRec == &ElemSize;
104 }
105 
106 /// Compute the trip count for the given loop \p L or assume a default value if
107 /// it is not a compile time constant. Return the SCEV expression for the trip
108 /// count.
109 static const SCEV *computeTripCount(const Loop &L, const SCEV &ElemSize,
110                                     ScalarEvolution &SE) {
111   const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(&L);
112   const SCEV *TripCount = (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) &&
113                            isa<SCEVConstant>(BackedgeTakenCount))
114                               ? SE.getTripCountFromExitCount(BackedgeTakenCount)
115                               : nullptr;
116 
117   if (!TripCount) {
118     LLVM_DEBUG(dbgs() << "Trip count of loop " << L.getName()
119                << " could not be computed, using DefaultTripCount\n");
120     TripCount = SE.getConstant(ElemSize.getType(), DefaultTripCount);
121   }
122 
123   return TripCount;
124 }
125 
126 //===----------------------------------------------------------------------===//
127 // IndexedReference implementation
128 //
129 raw_ostream &llvm::operator<<(raw_ostream &OS, const IndexedReference &R) {
130   if (!R.IsValid) {
131     OS << R.StoreOrLoadInst;
132     OS << ", IsValid=false.";
133     return OS;
134   }
135 
136   OS << *R.BasePointer;
137   for (const SCEV *Subscript : R.Subscripts)
138     OS << "[" << *Subscript << "]";
139 
140   OS << ", Sizes: ";
141   for (const SCEV *Size : R.Sizes)
142     OS << "[" << *Size << "]";
143 
144   return OS;
145 }
146 
147 IndexedReference::IndexedReference(Instruction &StoreOrLoadInst,
148                                    const LoopInfo &LI, ScalarEvolution &SE)
149     : StoreOrLoadInst(StoreOrLoadInst), SE(SE) {
150   assert((isa<StoreInst>(StoreOrLoadInst) || isa<LoadInst>(StoreOrLoadInst)) &&
151          "Expecting a load or store instruction");
152 
153   IsValid = delinearize(LI);
154   if (IsValid)
155     LLVM_DEBUG(dbgs().indent(2) << "Succesfully delinearized: " << *this
156                                 << "\n");
157 }
158 
159 std::optional<bool>
160 IndexedReference::hasSpacialReuse(const IndexedReference &Other, unsigned CLS,
161                                   AAResults &AA) const {
162   assert(IsValid && "Expecting a valid reference");
163 
164   if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
165     LLVM_DEBUG(dbgs().indent(2)
166                << "No spacial reuse: different base pointers\n");
167     return false;
168   }
169 
170   unsigned NumSubscripts = getNumSubscripts();
171   if (NumSubscripts != Other.getNumSubscripts()) {
172     LLVM_DEBUG(dbgs().indent(2)
173                << "No spacial reuse: different number of subscripts\n");
174     return false;
175   }
176 
177   // all subscripts must be equal, except the leftmost one (the last one).
178   for (auto SubNum : seq<unsigned>(0, NumSubscripts - 1)) {
179     if (getSubscript(SubNum) != Other.getSubscript(SubNum)) {
180       LLVM_DEBUG(dbgs().indent(2) << "No spacial reuse, different subscripts: "
181                                   << "\n\t" << *getSubscript(SubNum) << "\n\t"
182                                   << *Other.getSubscript(SubNum) << "\n");
183       return false;
184     }
185   }
186 
187   // the difference between the last subscripts must be less than the cache line
188   // size.
189   const SCEV *LastSubscript = getLastSubscript();
190   const SCEV *OtherLastSubscript = Other.getLastSubscript();
191   const SCEVConstant *Diff = dyn_cast<SCEVConstant>(
192       SE.getMinusSCEV(LastSubscript, OtherLastSubscript));
193 
194   if (Diff == nullptr) {
195     LLVM_DEBUG(dbgs().indent(2)
196                << "No spacial reuse, difference between subscript:\n\t"
197                << *LastSubscript << "\n\t" << OtherLastSubscript
198                << "\nis not constant.\n");
199     return std::nullopt;
200   }
201 
202   bool InSameCacheLine = (Diff->getValue()->getSExtValue() < CLS);
203 
204   LLVM_DEBUG({
205     if (InSameCacheLine)
206       dbgs().indent(2) << "Found spacial reuse.\n";
207     else
208       dbgs().indent(2) << "No spacial reuse.\n";
209   });
210 
211   return InSameCacheLine;
212 }
213 
214 std::optional<bool>
215 IndexedReference::hasTemporalReuse(const IndexedReference &Other,
216                                    unsigned MaxDistance, const Loop &L,
217                                    DependenceInfo &DI, AAResults &AA) const {
218   assert(IsValid && "Expecting a valid reference");
219 
220   if (BasePointer != Other.getBasePointer() && !isAliased(Other, AA)) {
221     LLVM_DEBUG(dbgs().indent(2)
222                << "No temporal reuse: different base pointer\n");
223     return false;
224   }
225 
226   std::unique_ptr<Dependence> D =
227       DI.depends(&StoreOrLoadInst, &Other.StoreOrLoadInst, true);
228 
229   if (D == nullptr) {
230     LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: no dependence\n");
231     return false;
232   }
233 
234   if (D->isLoopIndependent()) {
235     LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
236     return true;
237   }
238 
239   // Check the dependence distance at every loop level. There is temporal reuse
240   // if the distance at the given loop's depth is small (|d| <= MaxDistance) and
241   // it is zero at every other loop level.
242   int LoopDepth = L.getLoopDepth();
243   int Levels = D->getLevels();
244   for (int Level = 1; Level <= Levels; ++Level) {
245     const SCEV *Distance = D->getDistance(Level);
246     const SCEVConstant *SCEVConst = dyn_cast_or_null<SCEVConstant>(Distance);
247 
248     if (SCEVConst == nullptr) {
249       LLVM_DEBUG(dbgs().indent(2) << "No temporal reuse: distance unknown\n");
250       return std::nullopt;
251     }
252 
253     const ConstantInt &CI = *SCEVConst->getValue();
254     if (Level != LoopDepth && !CI.isZero()) {
255       LLVM_DEBUG(dbgs().indent(2)
256                  << "No temporal reuse: distance is not zero at depth=" << Level
257                  << "\n");
258       return false;
259     } else if (Level == LoopDepth && CI.getSExtValue() > MaxDistance) {
260       LLVM_DEBUG(
261           dbgs().indent(2)
262           << "No temporal reuse: distance is greater than MaxDistance at depth="
263           << Level << "\n");
264       return false;
265     }
266   }
267 
268   LLVM_DEBUG(dbgs().indent(2) << "Found temporal reuse\n");
269   return true;
270 }
271 
272 CacheCostTy IndexedReference::computeRefCost(const Loop &L,
273                                              unsigned CLS) const {
274   assert(IsValid && "Expecting a valid reference");
275   LLVM_DEBUG({
276     dbgs().indent(2) << "Computing cache cost for:\n";
277     dbgs().indent(4) << *this << "\n";
278   });
279 
280   // If the indexed reference is loop invariant the cost is one.
281   if (isLoopInvariant(L)) {
282     LLVM_DEBUG(dbgs().indent(4) << "Reference is loop invariant: RefCost=1\n");
283     return 1;
284   }
285 
286   const SCEV *TripCount = computeTripCount(L, *Sizes.back(), SE);
287   assert(TripCount && "Expecting valid TripCount");
288   LLVM_DEBUG(dbgs() << "TripCount=" << *TripCount << "\n");
289 
290   const SCEV *RefCost = nullptr;
291   const SCEV *Stride = nullptr;
292   if (isConsecutive(L, Stride, CLS)) {
293     // If the indexed reference is 'consecutive' the cost is
294     // (TripCount*Stride)/CLS.
295     assert(Stride != nullptr &&
296            "Stride should not be null for consecutive access!");
297     Type *WiderType = SE.getWiderType(Stride->getType(), TripCount->getType());
298     const SCEV *CacheLineSize = SE.getConstant(WiderType, CLS);
299     Stride = SE.getNoopOrAnyExtend(Stride, WiderType);
300     TripCount = SE.getNoopOrZeroExtend(TripCount, WiderType);
301     const SCEV *Numerator = SE.getMulExpr(Stride, TripCount);
302     RefCost = SE.getUDivExpr(Numerator, CacheLineSize);
303 
304     LLVM_DEBUG(dbgs().indent(4)
305                << "Access is consecutive: RefCost=(TripCount*Stride)/CLS="
306                << *RefCost << "\n");
307   } else {
308     // If the indexed reference is not 'consecutive' the cost is proportional to
309     // the trip count and the depth of the dimension which the subject loop
310     // subscript is accessing. We try to estimate this by multiplying the cost
311     // by the trip counts of loops corresponding to the inner dimensions. For
312     // example, given the indexed reference 'A[i][j][k]', and assuming the
313     // i-loop is in the innermost position, the cost would be equal to the
314     // iterations of the i-loop multiplied by iterations of the j-loop.
315     RefCost = TripCount;
316 
317     int Index = getSubscriptIndex(L);
318     assert(Index >= 0 && "Cound not locate a valid Index");
319 
320     for (unsigned I = Index + 1; I < getNumSubscripts() - 1; ++I) {
321       const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(getSubscript(I));
322       assert(AR && AR->getLoop() && "Expecting valid loop");
323       const SCEV *TripCount =
324           computeTripCount(*AR->getLoop(), *Sizes.back(), SE);
325       Type *WiderType = SE.getWiderType(RefCost->getType(), TripCount->getType());
326       RefCost = SE.getMulExpr(SE.getNoopOrZeroExtend(RefCost, WiderType),
327                               SE.getNoopOrZeroExtend(TripCount, WiderType));
328     }
329 
330     LLVM_DEBUG(dbgs().indent(4)
331                << "Access is not consecutive: RefCost=" << *RefCost << "\n");
332   }
333   assert(RefCost && "Expecting a valid RefCost");
334 
335   // Attempt to fold RefCost into a constant.
336   if (auto ConstantCost = dyn_cast<SCEVConstant>(RefCost))
337     return ConstantCost->getValue()->getZExtValue();
338 
339   LLVM_DEBUG(dbgs().indent(4)
340              << "RefCost is not a constant! Setting to RefCost=InvalidCost "
341                 "(invalid value).\n");
342 
343   return CacheCost::InvalidCost;
344 }
345 
346 bool IndexedReference::tryDelinearizeFixedSize(
347     const SCEV *AccessFn, SmallVectorImpl<const SCEV *> &Subscripts) {
348   SmallVector<int, 4> ArraySizes;
349   if (!tryDelinearizeFixedSizeImpl(&SE, &StoreOrLoadInst, AccessFn, Subscripts,
350                                    ArraySizes))
351     return false;
352 
353   // Populate Sizes with scev expressions to be used in calculations later.
354   for (auto Idx : seq<unsigned>(1, Subscripts.size()))
355     Sizes.push_back(
356         SE.getConstant(Subscripts[Idx]->getType(), ArraySizes[Idx - 1]));
357 
358   LLVM_DEBUG({
359     dbgs() << "Delinearized subscripts of fixed-size array\n"
360            << "GEP:" << *getLoadStorePointerOperand(&StoreOrLoadInst)
361            << "\n";
362   });
363   return true;
364 }
365 
366 bool IndexedReference::delinearize(const LoopInfo &LI) {
367   assert(Subscripts.empty() && "Subscripts should be empty");
368   assert(Sizes.empty() && "Sizes should be empty");
369   assert(!IsValid && "Should be called once from the constructor");
370   LLVM_DEBUG(dbgs() << "Delinearizing: " << StoreOrLoadInst << "\n");
371 
372   const SCEV *ElemSize = SE.getElementSize(&StoreOrLoadInst);
373   const BasicBlock *BB = StoreOrLoadInst.getParent();
374 
375   if (Loop *L = LI.getLoopFor(BB)) {
376     const SCEV *AccessFn =
377         SE.getSCEVAtScope(getPointerOperand(&StoreOrLoadInst), L);
378 
379     BasePointer = dyn_cast<SCEVUnknown>(SE.getPointerBase(AccessFn));
380     if (BasePointer == nullptr) {
381       LLVM_DEBUG(
382           dbgs().indent(2)
383           << "ERROR: failed to delinearize, can't identify base pointer\n");
384       return false;
385     }
386 
387     bool IsFixedSize = false;
388     // Try to delinearize fixed-size arrays.
389     if (tryDelinearizeFixedSize(AccessFn, Subscripts)) {
390       IsFixedSize = true;
391       // The last element of Sizes is the element size.
392       Sizes.push_back(ElemSize);
393       LLVM_DEBUG(dbgs().indent(2) << "In Loop '" << L->getName()
394                                   << "', AccessFn: " << *AccessFn << "\n");
395     }
396 
397     AccessFn = SE.getMinusSCEV(AccessFn, BasePointer);
398 
399     // Try to delinearize parametric-size arrays.
400     if (!IsFixedSize) {
401       LLVM_DEBUG(dbgs().indent(2) << "In Loop '" << L->getName()
402                                   << "', AccessFn: " << *AccessFn << "\n");
403       llvm::delinearize(SE, AccessFn, Subscripts, Sizes,
404                         SE.getElementSize(&StoreOrLoadInst));
405     }
406 
407     if (Subscripts.empty() || Sizes.empty() ||
408         Subscripts.size() != Sizes.size()) {
409       // Attempt to determine whether we have a single dimensional array access.
410       // before giving up.
411       if (!isOneDimensionalArray(*AccessFn, *ElemSize, *L, SE)) {
412         LLVM_DEBUG(dbgs().indent(2)
413                    << "ERROR: failed to delinearize reference\n");
414         Subscripts.clear();
415         Sizes.clear();
416         return false;
417       }
418 
419       // The array may be accessed in reverse, for example:
420       //   for (i = N; i > 0; i--)
421       //     A[i] = 0;
422       // In this case, reconstruct the access function using the absolute value
423       // of the step recurrence.
424       const SCEVAddRecExpr *AccessFnAR = dyn_cast<SCEVAddRecExpr>(AccessFn);
425       const SCEV *StepRec = AccessFnAR ? AccessFnAR->getStepRecurrence(SE) : nullptr;
426 
427       if (StepRec && SE.isKnownNegative(StepRec))
428         AccessFn = SE.getAddRecExpr(AccessFnAR->getStart(),
429                                     SE.getNegativeSCEV(StepRec),
430                                     AccessFnAR->getLoop(),
431                                     AccessFnAR->getNoWrapFlags());
432       const SCEV *Div = SE.getUDivExactExpr(AccessFn, ElemSize);
433       Subscripts.push_back(Div);
434       Sizes.push_back(ElemSize);
435     }
436 
437     return all_of(Subscripts, [&](const SCEV *Subscript) {
438       return isSimpleAddRecurrence(*Subscript, *L);
439     });
440   }
441 
442   return false;
443 }
444 
445 bool IndexedReference::isLoopInvariant(const Loop &L) const {
446   Value *Addr = getPointerOperand(&StoreOrLoadInst);
447   assert(Addr != nullptr && "Expecting either a load or a store instruction");
448   assert(SE.isSCEVable(Addr->getType()) && "Addr should be SCEVable");
449 
450   if (SE.isLoopInvariant(SE.getSCEV(Addr), &L))
451     return true;
452 
453   // The indexed reference is loop invariant if none of the coefficients use
454   // the loop induction variable.
455   bool allCoeffForLoopAreZero = all_of(Subscripts, [&](const SCEV *Subscript) {
456     return isCoeffForLoopZeroOrInvariant(*Subscript, L);
457   });
458 
459   return allCoeffForLoopAreZero;
460 }
461 
462 bool IndexedReference::isConsecutive(const Loop &L, const SCEV *&Stride,
463                                      unsigned CLS) const {
464   // The indexed reference is 'consecutive' if the only coefficient that uses
465   // the loop induction variable is the last one...
466   const SCEV *LastSubscript = Subscripts.back();
467   for (const SCEV *Subscript : Subscripts) {
468     if (Subscript == LastSubscript)
469       continue;
470     if (!isCoeffForLoopZeroOrInvariant(*Subscript, L))
471       return false;
472   }
473 
474   // ...and the access stride is less than the cache line size.
475   const SCEV *Coeff = getLastCoefficient();
476   const SCEV *ElemSize = Sizes.back();
477   Type *WiderType = SE.getWiderType(Coeff->getType(), ElemSize->getType());
478   // FIXME: This assumes that all values are signed integers which may
479   // be incorrect in unusual codes and incorrectly use sext instead of zext.
480   // for (uint32_t i = 0; i < 512; ++i) {
481   //   uint8_t trunc = i;
482   //   A[trunc] = 42;
483   // }
484   // This consecutively iterates twice over A. If `trunc` is sign-extended,
485   // we would conclude that this may iterate backwards over the array.
486   // However, LoopCacheAnalysis is heuristic anyway and transformations must
487   // not result in wrong optimizations if the heuristic was incorrect.
488   Stride = SE.getMulExpr(SE.getNoopOrSignExtend(Coeff, WiderType),
489                          SE.getNoopOrSignExtend(ElemSize, WiderType));
490   const SCEV *CacheLineSize = SE.getConstant(Stride->getType(), CLS);
491 
492   Stride = SE.isKnownNegative(Stride) ? SE.getNegativeSCEV(Stride) : Stride;
493   return SE.isKnownPredicate(ICmpInst::ICMP_ULT, Stride, CacheLineSize);
494 }
495 
496 int IndexedReference::getSubscriptIndex(const Loop &L) const {
497   for (auto Idx : seq<int>(0, getNumSubscripts())) {
498     const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(getSubscript(Idx));
499     if (AR && AR->getLoop() == &L) {
500       return Idx;
501     }
502   }
503   return -1;
504 }
505 
506 const SCEV *IndexedReference::getLastCoefficient() const {
507   const SCEV *LastSubscript = getLastSubscript();
508   auto *AR = cast<SCEVAddRecExpr>(LastSubscript);
509   return AR->getStepRecurrence(SE);
510 }
511 
512 bool IndexedReference::isCoeffForLoopZeroOrInvariant(const SCEV &Subscript,
513                                                      const Loop &L) const {
514   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(&Subscript);
515   return (AR != nullptr) ? AR->getLoop() != &L
516                          : SE.isLoopInvariant(&Subscript, &L);
517 }
518 
519 bool IndexedReference::isSimpleAddRecurrence(const SCEV &Subscript,
520                                              const Loop &L) const {
521   if (!isa<SCEVAddRecExpr>(Subscript))
522     return false;
523 
524   const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(&Subscript);
525   assert(AR->getLoop() && "AR should have a loop");
526 
527   if (!AR->isAffine())
528     return false;
529 
530   const SCEV *Start = AR->getStart();
531   const SCEV *Step = AR->getStepRecurrence(SE);
532 
533   if (!SE.isLoopInvariant(Start, &L) || !SE.isLoopInvariant(Step, &L))
534     return false;
535 
536   return true;
537 }
538 
539 bool IndexedReference::isAliased(const IndexedReference &Other,
540                                  AAResults &AA) const {
541   const auto &Loc1 = MemoryLocation::get(&StoreOrLoadInst);
542   const auto &Loc2 = MemoryLocation::get(&Other.StoreOrLoadInst);
543   return AA.isMustAlias(Loc1, Loc2);
544 }
545 
546 //===----------------------------------------------------------------------===//
547 // CacheCost implementation
548 //
549 raw_ostream &llvm::operator<<(raw_ostream &OS, const CacheCost &CC) {
550   for (const auto &LC : CC.LoopCosts) {
551     const Loop *L = LC.first;
552     OS << "Loop '" << L->getName() << "' has cost = " << LC.second << "\n";
553   }
554   return OS;
555 }
556 
557 CacheCost::CacheCost(const LoopVectorTy &Loops, const LoopInfo &LI,
558                      ScalarEvolution &SE, TargetTransformInfo &TTI,
559                      AAResults &AA, DependenceInfo &DI,
560                      std::optional<unsigned> TRT)
561     : Loops(Loops), TRT(TRT.value_or(TemporalReuseThreshold)), LI(LI), SE(SE),
562       TTI(TTI), AA(AA), DI(DI) {
563   assert(!Loops.empty() && "Expecting a non-empty loop vector.");
564 
565   for (const Loop *L : Loops) {
566     unsigned TripCount = SE.getSmallConstantTripCount(L);
567     TripCount = (TripCount == 0) ? DefaultTripCount : TripCount;
568     TripCounts.push_back({L, TripCount});
569   }
570 
571   calculateCacheFootprint();
572 }
573 
574 std::unique_ptr<CacheCost>
575 CacheCost::getCacheCost(Loop &Root, LoopStandardAnalysisResults &AR,
576                         DependenceInfo &DI, std::optional<unsigned> TRT) {
577   if (!Root.isOutermost()) {
578     LLVM_DEBUG(dbgs() << "Expecting the outermost loop in a loop nest\n");
579     return nullptr;
580   }
581 
582   LoopVectorTy Loops;
583   append_range(Loops, breadth_first(&Root));
584 
585   if (!getInnerMostLoop(Loops)) {
586     LLVM_DEBUG(dbgs() << "Cannot compute cache cost of loop nest with more "
587                          "than one innermost loop\n");
588     return nullptr;
589   }
590 
591   return std::make_unique<CacheCost>(Loops, AR.LI, AR.SE, AR.TTI, AR.AA, DI, TRT);
592 }
593 
594 void CacheCost::calculateCacheFootprint() {
595   LLVM_DEBUG(dbgs() << "POPULATING REFERENCE GROUPS\n");
596   ReferenceGroupsTy RefGroups;
597   if (!populateReferenceGroups(RefGroups))
598     return;
599 
600   LLVM_DEBUG(dbgs() << "COMPUTING LOOP CACHE COSTS\n");
601   for (const Loop *L : Loops) {
602     assert(llvm::none_of(
603                LoopCosts,
604                [L](const LoopCacheCostTy &LCC) { return LCC.first == L; }) &&
605            "Should not add duplicate element");
606     CacheCostTy LoopCost = computeLoopCacheCost(*L, RefGroups);
607     LoopCosts.push_back(std::make_pair(L, LoopCost));
608   }
609 
610   sortLoopCosts();
611   RefGroups.clear();
612 }
613 
614 bool CacheCost::populateReferenceGroups(ReferenceGroupsTy &RefGroups) const {
615   assert(RefGroups.empty() && "Reference groups should be empty");
616 
617   unsigned CLS = TTI.getCacheLineSize();
618   Loop *InnerMostLoop = getInnerMostLoop(Loops);
619   assert(InnerMostLoop != nullptr && "Expecting a valid innermost loop");
620 
621   for (BasicBlock *BB : InnerMostLoop->getBlocks()) {
622     for (Instruction &I : *BB) {
623       if (!isa<StoreInst>(I) && !isa<LoadInst>(I))
624         continue;
625 
626       std::unique_ptr<IndexedReference> R(new IndexedReference(I, LI, SE));
627       if (!R->isValid())
628         continue;
629 
630       bool Added = false;
631       for (ReferenceGroupTy &RefGroup : RefGroups) {
632         const IndexedReference &Representative = *RefGroup.front();
633         LLVM_DEBUG({
634           dbgs() << "References:\n";
635           dbgs().indent(2) << *R << "\n";
636           dbgs().indent(2) << Representative << "\n";
637         });
638 
639 
640        // FIXME: Both positive and negative access functions will be placed
641        // into the same reference group, resulting in a bi-directional array
642        // access such as:
643        //   for (i = N; i > 0; i--)
644        //     A[i] = A[N - i];
645        // having the same cost calculation as a single dimention access pattern
646        //   for (i = 0; i < N; i++)
647        //     A[i] = A[i];
648        // when in actuality, depending on the array size, the first example
649        // should have a cost closer to 2x the second due to the two cache
650        // access per iteration from opposite ends of the array
651         std::optional<bool> HasTemporalReuse =
652             R->hasTemporalReuse(Representative, *TRT, *InnerMostLoop, DI, AA);
653         std::optional<bool> HasSpacialReuse =
654             R->hasSpacialReuse(Representative, CLS, AA);
655 
656         if ((HasTemporalReuse && *HasTemporalReuse) ||
657             (HasSpacialReuse && *HasSpacialReuse)) {
658           RefGroup.push_back(std::move(R));
659           Added = true;
660           break;
661         }
662       }
663 
664       if (!Added) {
665         ReferenceGroupTy RG;
666         RG.push_back(std::move(R));
667         RefGroups.push_back(std::move(RG));
668       }
669     }
670   }
671 
672   if (RefGroups.empty())
673     return false;
674 
675   LLVM_DEBUG({
676     dbgs() << "\nIDENTIFIED REFERENCE GROUPS:\n";
677     int n = 1;
678     for (const ReferenceGroupTy &RG : RefGroups) {
679       dbgs().indent(2) << "RefGroup " << n << ":\n";
680       for (const auto &IR : RG)
681         dbgs().indent(4) << *IR << "\n";
682       n++;
683     }
684     dbgs() << "\n";
685   });
686 
687   return true;
688 }
689 
690 CacheCostTy
691 CacheCost::computeLoopCacheCost(const Loop &L,
692                                 const ReferenceGroupsTy &RefGroups) const {
693   if (!L.isLoopSimplifyForm())
694     return InvalidCost;
695 
696   LLVM_DEBUG(dbgs() << "Considering loop '" << L.getName()
697                     << "' as innermost loop.\n");
698 
699   // Compute the product of the trip counts of each other loop in the nest.
700   CacheCostTy TripCountsProduct = 1;
701   for (const auto &TC : TripCounts) {
702     if (TC.first == &L)
703       continue;
704     TripCountsProduct *= TC.second;
705   }
706 
707   CacheCostTy LoopCost = 0;
708   for (const ReferenceGroupTy &RG : RefGroups) {
709     CacheCostTy RefGroupCost = computeRefGroupCacheCost(RG, L);
710     LoopCost += RefGroupCost * TripCountsProduct;
711   }
712 
713   LLVM_DEBUG(dbgs().indent(2) << "Loop '" << L.getName()
714                               << "' has cost=" << LoopCost << "\n");
715 
716   return LoopCost;
717 }
718 
719 CacheCostTy CacheCost::computeRefGroupCacheCost(const ReferenceGroupTy &RG,
720                                                 const Loop &L) const {
721   assert(!RG.empty() && "Reference group should have at least one member.");
722 
723   const IndexedReference *Representative = RG.front().get();
724   return Representative->computeRefCost(L, TTI.getCacheLineSize());
725 }
726 
727 //===----------------------------------------------------------------------===//
728 // LoopCachePrinterPass implementation
729 //
730 PreservedAnalyses LoopCachePrinterPass::run(Loop &L, LoopAnalysisManager &AM,
731                                             LoopStandardAnalysisResults &AR,
732                                             LPMUpdater &U) {
733   Function *F = L.getHeader()->getParent();
734   DependenceInfo DI(F, &AR.AA, &AR.SE, &AR.LI);
735 
736   if (auto CC = CacheCost::getCacheCost(L, AR, DI))
737     OS << *CC;
738 
739   return PreservedAnalyses::all();
740 }
741