Olivier Deprez | f4ef2d0 | 2021-04-20 13:36:24 +0200 | [diff] [blame^] | 1 | //===- llvm/Analysis/LoopCacheAnalysis.h ------------------------*- C++ -*-===// |
| 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 | /// \file |
| 10 | /// This file defines the interface for the loop cache analysis. |
| 11 | /// |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef LLVM_ANALYSIS_LOOPCACHEANALYSIS_H |
| 15 | #define LLVM_ANALYSIS_LOOPCACHEANALYSIS_H |
| 16 | |
| 17 | #include "llvm/Analysis/LoopAnalysisManager.h" |
| 18 | #include "llvm/IR/Instructions.h" |
| 19 | #include "llvm/IR/PassManager.h" |
| 20 | #include "llvm/Support/raw_ostream.h" |
| 21 | |
| 22 | namespace llvm { |
| 23 | |
| 24 | class AAResults; |
| 25 | class DependenceInfo; |
| 26 | class LPMUpdater; |
| 27 | class ScalarEvolution; |
| 28 | class SCEV; |
| 29 | class TargetTransformInfo; |
| 30 | |
| 31 | using CacheCostTy = int64_t; |
| 32 | using LoopVectorTy = SmallVector<Loop *, 8>; |
| 33 | |
| 34 | /// Represents a memory reference as a base pointer and a set of indexing |
| 35 | /// operations. For example given the array reference A[i][2j+1][3k+2] in a |
| 36 | /// 3-dim loop nest: |
| 37 | /// for(i=0;i<n;++i) |
| 38 | /// for(j=0;j<m;++j) |
| 39 | /// for(k=0;k<o;++k) |
| 40 | /// ... A[i][2j+1][3k+2] ... |
| 41 | /// We expect: |
| 42 | /// BasePointer -> A |
| 43 | /// Subscripts -> [{0,+,1}<%for.i>][{1,+,2}<%for.j>][{2,+,3}<%for.k>] |
| 44 | /// Sizes -> [m][o][4] |
| 45 | class IndexedReference { |
| 46 | friend raw_ostream &operator<<(raw_ostream &OS, const IndexedReference &R); |
| 47 | |
| 48 | public: |
| 49 | /// Construct an indexed reference given a \p StoreOrLoadInst instruction. |
| 50 | IndexedReference(Instruction &StoreOrLoadInst, const LoopInfo &LI, |
| 51 | ScalarEvolution &SE); |
| 52 | |
| 53 | bool isValid() const { return IsValid; } |
| 54 | const SCEV *getBasePointer() const { return BasePointer; } |
| 55 | size_t getNumSubscripts() const { return Subscripts.size(); } |
| 56 | const SCEV *getSubscript(unsigned SubNum) const { |
| 57 | assert(SubNum < getNumSubscripts() && "Invalid subscript number"); |
| 58 | return Subscripts[SubNum]; |
| 59 | } |
| 60 | const SCEV *getFirstSubscript() const { |
| 61 | assert(!Subscripts.empty() && "Expecting non-empty container"); |
| 62 | return Subscripts.front(); |
| 63 | } |
| 64 | const SCEV *getLastSubscript() const { |
| 65 | assert(!Subscripts.empty() && "Expecting non-empty container"); |
| 66 | return Subscripts.back(); |
| 67 | } |
| 68 | |
| 69 | /// Return true/false if the current object and the indexed reference \p Other |
| 70 | /// are/aren't in the same cache line of size \p CLS. Two references are in |
| 71 | /// the same chace line iff the distance between them in the innermost |
| 72 | /// dimension is less than the cache line size. Return None if unsure. |
| 73 | Optional<bool> hasSpacialReuse(const IndexedReference &Other, unsigned CLS, |
| 74 | AAResults &AA) const; |
| 75 | |
| 76 | /// Return true if the current object and the indexed reference \p Other |
| 77 | /// have distance smaller than \p MaxDistance in the dimension associated with |
| 78 | /// the given loop \p L. Return false if the distance is not smaller than \p |
| 79 | /// MaxDistance and None if unsure. |
| 80 | Optional<bool> hasTemporalReuse(const IndexedReference &Other, |
| 81 | unsigned MaxDistance, const Loop &L, |
| 82 | DependenceInfo &DI, AAResults &AA) const; |
| 83 | |
| 84 | /// Compute the cost of the reference w.r.t. the given loop \p L when it is |
| 85 | /// considered in the innermost position in the loop nest. |
| 86 | /// The cost is defined as: |
| 87 | /// - equal to one if the reference is loop invariant, or |
| 88 | /// - equal to '(TripCount * stride) / cache_line_size' if: |
| 89 | /// + the reference stride is less than the cache line size, and |
| 90 | /// + the coefficient of this loop's index variable used in all other |
| 91 | /// subscripts is zero |
| 92 | /// - or otherwise equal to 'TripCount'. |
| 93 | CacheCostTy computeRefCost(const Loop &L, unsigned CLS) const; |
| 94 | |
| 95 | private: |
| 96 | /// Attempt to delinearize the indexed reference. |
| 97 | bool delinearize(const LoopInfo &LI); |
| 98 | |
| 99 | /// Return true if the index reference is invariant with respect to loop \p L. |
| 100 | bool isLoopInvariant(const Loop &L) const; |
| 101 | |
| 102 | /// Return true if the indexed reference is 'consecutive' in loop \p L. |
| 103 | /// An indexed reference is 'consecutive' if the only coefficient that uses |
| 104 | /// the loop induction variable is the rightmost one, and the access stride is |
| 105 | /// smaller than the cache line size \p CLS. |
| 106 | bool isConsecutive(const Loop &L, unsigned CLS) const; |
| 107 | |
| 108 | /// Return the coefficient used in the rightmost dimension. |
| 109 | const SCEV *getLastCoefficient() const; |
| 110 | |
| 111 | /// Return true if the coefficient corresponding to induction variable of |
| 112 | /// loop \p L in the given \p Subscript is zero or is loop invariant in \p L. |
| 113 | bool isCoeffForLoopZeroOrInvariant(const SCEV &Subscript, |
| 114 | const Loop &L) const; |
| 115 | |
| 116 | /// Verify that the given \p Subscript is 'well formed' (must be a simple add |
| 117 | /// recurrence). |
| 118 | bool isSimpleAddRecurrence(const SCEV &Subscript, const Loop &L) const; |
| 119 | |
| 120 | /// Return true if the given reference \p Other is definetely aliased with |
| 121 | /// the indexed reference represented by this class. |
| 122 | bool isAliased(const IndexedReference &Other, AAResults &AA) const; |
| 123 | |
| 124 | private: |
| 125 | /// True if the reference can be delinearized, false otherwise. |
| 126 | bool IsValid = false; |
| 127 | |
| 128 | /// Represent the memory reference instruction. |
| 129 | Instruction &StoreOrLoadInst; |
| 130 | |
| 131 | /// The base pointer of the memory reference. |
| 132 | const SCEV *BasePointer = nullptr; |
| 133 | |
| 134 | /// The subscript (indexes) of the memory reference. |
| 135 | SmallVector<const SCEV *, 3> Subscripts; |
| 136 | |
| 137 | /// The dimensions of the memory reference. |
| 138 | SmallVector<const SCEV *, 3> Sizes; |
| 139 | |
| 140 | ScalarEvolution &SE; |
| 141 | }; |
| 142 | |
| 143 | /// A reference group represents a set of memory references that exhibit |
| 144 | /// temporal or spacial reuse. Two references belong to the same |
| 145 | /// reference group with respect to a inner loop L iff: |
| 146 | /// 1. they have a loop independent dependency, or |
| 147 | /// 2. they have a loop carried dependence with a small dependence distance |
| 148 | /// (e.g. less than 2) carried by the inner loop, or |
| 149 | /// 3. they refer to the same array, and the subscript in their innermost |
| 150 | /// dimension is less than or equal to 'd' (where 'd' is less than the cache |
| 151 | /// line size) |
| 152 | /// |
| 153 | /// Intuitively a reference group represents memory references that access |
| 154 | /// the same cache line. Conditions 1,2 above account for temporal reuse, while |
| 155 | /// contition 3 accounts for spacial reuse. |
| 156 | using ReferenceGroupTy = SmallVector<std::unique_ptr<IndexedReference>, 8>; |
| 157 | using ReferenceGroupsTy = SmallVector<ReferenceGroupTy, 8>; |
| 158 | |
| 159 | /// \c CacheCost represents the estimated cost of a inner loop as the number of |
| 160 | /// cache lines used by the memory references it contains. |
| 161 | /// The 'cache cost' of a loop 'L' in a loop nest 'LN' is computed as the sum of |
| 162 | /// the cache costs of all of its reference groups when the loop is considered |
| 163 | /// to be in the innermost position in the nest. |
| 164 | /// A reference group represents memory references that fall into the same cache |
| 165 | /// line. Each reference group is analysed with respect to the innermost loop in |
| 166 | /// a loop nest. The cost of a reference is defined as follow: |
| 167 | /// - one if it is loop invariant w.r.t the innermost loop, |
| 168 | /// - equal to the loop trip count divided by the cache line times the |
| 169 | /// reference stride if the reference stride is less than the cache line |
| 170 | /// size (CLS), and the coefficient of this loop's index variable used in all |
| 171 | /// other subscripts is zero (e.g. RefCost = TripCount/(CLS/RefStride)) |
| 172 | /// - equal to the innermost loop trip count if the reference stride is greater |
| 173 | /// or equal to the cache line size CLS. |
| 174 | class CacheCost { |
| 175 | friend raw_ostream &operator<<(raw_ostream &OS, const CacheCost &CC); |
| 176 | using LoopTripCountTy = std::pair<const Loop *, unsigned>; |
| 177 | using LoopCacheCostTy = std::pair<const Loop *, CacheCostTy>; |
| 178 | |
| 179 | public: |
| 180 | static CacheCostTy constexpr InvalidCost = -1; |
| 181 | |
| 182 | /// Construct a CacheCost object for the loop nest described by \p Loops. |
| 183 | /// The optional parameter \p TRT can be used to specify the max. distance |
| 184 | /// between array elements accessed in a loop so that the elements are |
| 185 | /// classified to have temporal reuse. |
| 186 | CacheCost(const LoopVectorTy &Loops, const LoopInfo &LI, ScalarEvolution &SE, |
| 187 | TargetTransformInfo &TTI, AAResults &AA, DependenceInfo &DI, |
| 188 | Optional<unsigned> TRT = None); |
| 189 | |
| 190 | /// Create a CacheCost for the loop nest rooted by \p Root. |
| 191 | /// The optional parameter \p TRT can be used to specify the max. distance |
| 192 | /// between array elements accessed in a loop so that the elements are |
| 193 | /// classified to have temporal reuse. |
| 194 | static std::unique_ptr<CacheCost> |
| 195 | getCacheCost(Loop &Root, LoopStandardAnalysisResults &AR, DependenceInfo &DI, |
| 196 | Optional<unsigned> TRT = None); |
| 197 | |
| 198 | /// Return the estimated cost of loop \p L if the given loop is part of the |
| 199 | /// loop nest associated with this object. Return -1 otherwise. |
| 200 | CacheCostTy getLoopCost(const Loop &L) const { |
| 201 | auto IT = llvm::find_if(LoopCosts, [&L](const LoopCacheCostTy &LCC) { |
| 202 | return LCC.first == &L; |
| 203 | }); |
| 204 | return (IT != LoopCosts.end()) ? (*IT).second : -1; |
| 205 | } |
| 206 | |
| 207 | /// Return the estimated ordered loop costs. |
| 208 | const ArrayRef<LoopCacheCostTy> getLoopCosts() const { return LoopCosts; } |
| 209 | |
| 210 | private: |
| 211 | /// Calculate the cache footprint of each loop in the nest (when it is |
| 212 | /// considered to be in the innermost position). |
| 213 | void calculateCacheFootprint(); |
| 214 | |
| 215 | /// Partition store/load instructions in the loop nest into reference groups. |
| 216 | /// Two or more memory accesses belong in the same reference group if they |
| 217 | /// share the same cache line. |
| 218 | bool populateReferenceGroups(ReferenceGroupsTy &RefGroups) const; |
| 219 | |
| 220 | /// Calculate the cost of the given loop \p L assuming it is the innermost |
| 221 | /// loop in nest. |
| 222 | CacheCostTy computeLoopCacheCost(const Loop &L, |
| 223 | const ReferenceGroupsTy &RefGroups) const; |
| 224 | |
| 225 | /// Compute the cost of a representative reference in reference group \p RG |
| 226 | /// when the given loop \p L is considered as the innermost loop in the nest. |
| 227 | /// The computed cost is an estimate for the number of cache lines used by the |
| 228 | /// reference group. The representative reference cost is defined as: |
| 229 | /// - equal to one if the reference is loop invariant, or |
| 230 | /// - equal to '(TripCount * stride) / cache_line_size' if (a) loop \p L's |
| 231 | /// induction variable is used only in the reference subscript associated |
| 232 | /// with loop \p L, and (b) the reference stride is less than the cache |
| 233 | /// line size, or |
| 234 | /// - TripCount otherwise |
| 235 | CacheCostTy computeRefGroupCacheCost(const ReferenceGroupTy &RG, |
| 236 | const Loop &L) const; |
| 237 | |
| 238 | /// Sort the LoopCosts vector by decreasing cache cost. |
| 239 | void sortLoopCosts() { |
| 240 | sort(LoopCosts, [](const LoopCacheCostTy &A, const LoopCacheCostTy &B) { |
| 241 | return A.second > B.second; |
| 242 | }); |
| 243 | } |
| 244 | |
| 245 | private: |
| 246 | /// Loops in the loop nest associated with this object. |
| 247 | LoopVectorTy Loops; |
| 248 | |
| 249 | /// Trip counts for the loops in the loop nest associated with this object. |
| 250 | SmallVector<LoopTripCountTy, 3> TripCounts; |
| 251 | |
| 252 | /// Cache costs for the loops in the loop nest associated with this object. |
| 253 | SmallVector<LoopCacheCostTy, 3> LoopCosts; |
| 254 | |
| 255 | /// The max. distance between array elements accessed in a loop so that the |
| 256 | /// elements are classified to have temporal reuse. |
| 257 | Optional<unsigned> TRT; |
| 258 | |
| 259 | const LoopInfo &LI; |
| 260 | ScalarEvolution &SE; |
| 261 | TargetTransformInfo &TTI; |
| 262 | AAResults &AA; |
| 263 | DependenceInfo &DI; |
| 264 | }; |
| 265 | |
| 266 | raw_ostream &operator<<(raw_ostream &OS, const IndexedReference &R); |
| 267 | raw_ostream &operator<<(raw_ostream &OS, const CacheCost &CC); |
| 268 | |
| 269 | /// Printer pass for the \c CacheCost results. |
| 270 | class LoopCachePrinterPass : public PassInfoMixin<LoopCachePrinterPass> { |
| 271 | raw_ostream &OS; |
| 272 | |
| 273 | public: |
| 274 | explicit LoopCachePrinterPass(raw_ostream &OS) : OS(OS) {} |
| 275 | |
| 276 | PreservedAnalyses run(Loop &L, LoopAnalysisManager &AM, |
| 277 | LoopStandardAnalysisResults &AR, LPMUpdater &U); |
| 278 | }; |
| 279 | |
| 280 | } // namespace llvm |
| 281 | |
| 282 | #endif // LLVM_ANALYSIS_LOOPCACHEANALYSIS_H |