| 1 | //===- Legality.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 | // Legality checks for the Sandbox Vectorizer. |
| 10 | // |
| 11 | |
| 12 | #ifndef LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_LEGALITY_H |
| 13 | #define LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_LEGALITY_H |
| 14 | |
| 15 | #include "llvm/ADT/ArrayRef.h" |
| 16 | #include "llvm/Analysis/ScalarEvolution.h" |
| 17 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 18 | #include "llvm/IR/DataLayout.h" |
| 19 | #include "llvm/Support/Casting.h" |
| 20 | #include "llvm/Support/Compiler.h" |
| 21 | #include "llvm/Support/raw_ostream.h" |
| 22 | #include "llvm/Transforms/Vectorize/SandboxVectorizer/InstrMaps.h" |
| 23 | #include "llvm/Transforms/Vectorize/SandboxVectorizer/Scheduler.h" |
| 24 | |
| 25 | namespace llvm::sandboxir { |
| 26 | |
| 27 | class LegalityAnalysis; |
| 28 | class Value; |
| 29 | class InstrMaps; |
| 30 | |
| 31 | class ShuffleMask { |
| 32 | public: |
| 33 | using IndicesVecT = SmallVector<int, 8>; |
| 34 | |
| 35 | private: |
| 36 | IndicesVecT Indices; |
| 37 | |
| 38 | public: |
| 39 | ShuffleMask(SmallVectorImpl<int> &&Indices) : Indices(std::move(Indices)) {} |
| 40 | ShuffleMask(std::initializer_list<int> Indices) : Indices(Indices) {} |
| 41 | explicit ShuffleMask(ArrayRef<int> Indices) : Indices(Indices) {} |
| 42 | operator ArrayRef<int>() const { return Indices; } |
| 43 | /// Creates and returns an identity shuffle mask of size \p Sz. |
| 44 | /// For example if Sz == 4 the returned mask is {0, 1, 2, 3}. |
| 45 | static ShuffleMask getIdentity(unsigned Sz) { |
| 46 | IndicesVecT Indices; |
| 47 | Indices.reserve(N: Sz); |
| 48 | llvm::append_range(C&: Indices, R: seq<int>(Begin: 0, End: (int)Sz)); |
| 49 | return ShuffleMask(std::move(Indices)); |
| 50 | } |
| 51 | /// \Returns true if the mask is a perfect identity mask with consecutive |
| 52 | /// indices, i.e., performs no lane shuffling, like 0,1,2,3... |
| 53 | bool isIdentity() const { |
| 54 | for (auto [Idx, Elm] : enumerate(First: Indices)) { |
| 55 | if ((int)Idx != Elm) |
| 56 | return false; |
| 57 | } |
| 58 | return true; |
| 59 | } |
| 60 | bool operator==(const ShuffleMask &Other) const { |
| 61 | return Indices == Other.Indices; |
| 62 | } |
| 63 | bool operator!=(const ShuffleMask &Other) const { return !(*this == Other); } |
| 64 | size_t size() const { return Indices.size(); } |
| 65 | int operator[](int Idx) const { return Indices[Idx]; } |
| 66 | using const_iterator = IndicesVecT::const_iterator; |
| 67 | const_iterator begin() const { return Indices.begin(); } |
| 68 | const_iterator end() const { return Indices.end(); } |
| 69 | #ifndef NDEBUG |
| 70 | friend raw_ostream &operator<<(raw_ostream &OS, const ShuffleMask &Mask) { |
| 71 | Mask.print(OS); |
| 72 | return OS; |
| 73 | } |
| 74 | void print(raw_ostream &OS) const { |
| 75 | interleave(Indices, OS, [&OS](auto Elm) { OS << Elm; }, "," ); |
| 76 | } |
| 77 | LLVM_DUMP_METHOD void dump() const; |
| 78 | #endif |
| 79 | }; |
| 80 | |
| 81 | enum class LegalityResultID { |
| 82 | Pack, ///> Collect scalar values. |
| 83 | Widen, ///> Vectorize by combining scalars to a vector. |
| 84 | DiamondReuse, ///> Don't generate new code, reuse existing vector. |
| 85 | DiamondReuseWithShuffle, ///> Reuse the existing vector but add a shuffle. |
| 86 | DiamondReuseMultiInput, ///> Reuse more than one vector and/or scalars. |
| 87 | }; |
| 88 | |
| 89 | /// The reason for vectorizing or not vectorizing. |
| 90 | enum class ResultReason { |
| 91 | NotInstructions, |
| 92 | DiffOpcodes, |
| 93 | DiffTypes, |
| 94 | DiffMathFlags, |
| 95 | DiffWrapFlags, |
| 96 | DiffBBs, |
| 97 | RepeatedInstrs, |
| 98 | NotConsecutive, |
| 99 | AlignmentNotSupported, |
| 100 | CantSchedule, |
| 101 | Unimplemented, |
| 102 | Infeasible, |
| 103 | ForcePackForDebugging, |
| 104 | }; |
| 105 | |
| 106 | #ifndef NDEBUG |
| 107 | struct ToStr { |
| 108 | static const char *getLegalityResultID(LegalityResultID ID) { |
| 109 | switch (ID) { |
| 110 | case LegalityResultID::Pack: |
| 111 | return "Pack" ; |
| 112 | case LegalityResultID::Widen: |
| 113 | return "Widen" ; |
| 114 | case LegalityResultID::DiamondReuse: |
| 115 | return "DiamondReuse" ; |
| 116 | case LegalityResultID::DiamondReuseWithShuffle: |
| 117 | return "DiamondReuseWithShuffle" ; |
| 118 | case LegalityResultID::DiamondReuseMultiInput: |
| 119 | return "DiamondReuseMultiInput" ; |
| 120 | } |
| 121 | llvm_unreachable("Unknown LegalityResultID enum" ); |
| 122 | } |
| 123 | |
| 124 | static const char *getVecReason(ResultReason Reason) { |
| 125 | switch (Reason) { |
| 126 | case ResultReason::NotInstructions: |
| 127 | return "NotInstructions" ; |
| 128 | case ResultReason::DiffOpcodes: |
| 129 | return "DiffOpcodes" ; |
| 130 | case ResultReason::DiffTypes: |
| 131 | return "DiffTypes" ; |
| 132 | case ResultReason::DiffMathFlags: |
| 133 | return "DiffMathFlags" ; |
| 134 | case ResultReason::DiffWrapFlags: |
| 135 | return "DiffWrapFlags" ; |
| 136 | case ResultReason::DiffBBs: |
| 137 | return "DiffBBs" ; |
| 138 | case ResultReason::RepeatedInstrs: |
| 139 | return "RepeatedInstrs" ; |
| 140 | case ResultReason::NotConsecutive: |
| 141 | return "NotConsecutive" ; |
| 142 | case ResultReason::AlignmentNotSupported: |
| 143 | return "AlignmentNotSupported" ; |
| 144 | case ResultReason::CantSchedule: |
| 145 | return "CantSchedule" ; |
| 146 | case ResultReason::Unimplemented: |
| 147 | return "Unimplemented" ; |
| 148 | case ResultReason::Infeasible: |
| 149 | return "Infeasible" ; |
| 150 | case ResultReason::ForcePackForDebugging: |
| 151 | return "ForcePackForDebugging" ; |
| 152 | } |
| 153 | llvm_unreachable("Unknown ResultReason enum" ); |
| 154 | } |
| 155 | }; |
| 156 | #endif // NDEBUG |
| 157 | |
| 158 | /// The legality outcome is represented by a class rather than an enum class |
| 159 | /// because in some cases the legality checks are expensive and look for a |
| 160 | /// particular instruction that can be passed along to the vectorizer to avoid |
| 161 | /// repeating the same expensive computation. |
| 162 | class LegalityResult { |
| 163 | protected: |
| 164 | LegalityResultID ID; |
| 165 | /// Only Legality can create LegalityResults. |
| 166 | LegalityResult(LegalityResultID ID) : ID(ID) {} |
| 167 | friend class LegalityAnalysis; |
| 168 | |
| 169 | /// We shouldn't need copies. |
| 170 | LegalityResult(const LegalityResult &) = delete; |
| 171 | LegalityResult &operator=(const LegalityResult &) = delete; |
| 172 | |
| 173 | public: |
| 174 | virtual ~LegalityResult() = default; |
| 175 | LegalityResultID getSubclassID() const { return ID; } |
| 176 | #ifndef NDEBUG |
| 177 | virtual void print(raw_ostream &OS) const { |
| 178 | OS << ToStr::getLegalityResultID(ID); |
| 179 | } |
| 180 | LLVM_DUMP_METHOD void dump() const; |
| 181 | friend raw_ostream &operator<<(raw_ostream &OS, const LegalityResult &LR) { |
| 182 | LR.print(OS); |
| 183 | return OS; |
| 184 | } |
| 185 | #endif // NDEBUG |
| 186 | }; |
| 187 | |
| 188 | /// Base class for results with reason. |
| 189 | class LegalityResultWithReason : public LegalityResult { |
| 190 | ResultReason Reason; |
| 191 | LegalityResultWithReason(LegalityResultID ID, ResultReason Reason) |
| 192 | : LegalityResult(ID), Reason(Reason) {} |
| 193 | friend class Pack; // For constructor. |
| 194 | |
| 195 | public: |
| 196 | ResultReason getReason() const { return Reason; } |
| 197 | #ifndef NDEBUG |
| 198 | void print(raw_ostream &OS) const override { |
| 199 | LegalityResult::print(OS); |
| 200 | OS << " Reason: " << ToStr::getVecReason(Reason); |
| 201 | } |
| 202 | #endif |
| 203 | }; |
| 204 | |
| 205 | class Widen final : public LegalityResult { |
| 206 | friend class LegalityAnalysis; |
| 207 | Widen() : LegalityResult(LegalityResultID::Widen) {} |
| 208 | |
| 209 | public: |
| 210 | static bool classof(const LegalityResult *From) { |
| 211 | return From->getSubclassID() == LegalityResultID::Widen; |
| 212 | } |
| 213 | }; |
| 214 | |
| 215 | class DiamondReuse final : public LegalityResult { |
| 216 | friend class LegalityAnalysis; |
| 217 | Action *Vec; |
| 218 | DiamondReuse(Action *Vec) |
| 219 | : LegalityResult(LegalityResultID::DiamondReuse), Vec(Vec) {} |
| 220 | |
| 221 | public: |
| 222 | static bool classof(const LegalityResult *From) { |
| 223 | return From->getSubclassID() == LegalityResultID::DiamondReuse; |
| 224 | } |
| 225 | Action *getVector() const { return Vec; } |
| 226 | }; |
| 227 | |
| 228 | class DiamondReuseWithShuffle final : public LegalityResult { |
| 229 | friend class LegalityAnalysis; |
| 230 | Action *Vec; |
| 231 | ShuffleMask Mask; |
| 232 | DiamondReuseWithShuffle(Action *Vec, const ShuffleMask &Mask) |
| 233 | : LegalityResult(LegalityResultID::DiamondReuseWithShuffle), Vec(Vec), |
| 234 | Mask(Mask) {} |
| 235 | |
| 236 | public: |
| 237 | static bool classof(const LegalityResult *From) { |
| 238 | return From->getSubclassID() == LegalityResultID::DiamondReuseWithShuffle; |
| 239 | } |
| 240 | Action *getVector() const { return Vec; } |
| 241 | const ShuffleMask &getMask() const { return Mask; } |
| 242 | }; |
| 243 | |
| 244 | class Pack final : public LegalityResultWithReason { |
| 245 | Pack(ResultReason Reason) |
| 246 | : LegalityResultWithReason(LegalityResultID::Pack, Reason) {} |
| 247 | friend class LegalityAnalysis; // For constructor. |
| 248 | |
| 249 | public: |
| 250 | static bool classof(const LegalityResult *From) { |
| 251 | return From->getSubclassID() == LegalityResultID::Pack; |
| 252 | } |
| 253 | }; |
| 254 | |
| 255 | /// Describes how to collect the values needed by each lane. |
| 256 | class CollectDescr { |
| 257 | public: |
| 258 | /// Describes how to get a value element. If the value is a vector then it |
| 259 | /// also provides the index to extract it from. |
| 260 | class ExtractElementDescr { |
| 261 | PointerUnion<Action *, Value *> V = nullptr; |
| 262 | /// The index in `V` that the value can be extracted from. |
| 263 | int ExtractIdx = 0; |
| 264 | |
| 265 | public: |
| 266 | ExtractElementDescr(Action *V, int ) |
| 267 | : V(V), ExtractIdx(ExtractIdx) {} |
| 268 | ExtractElementDescr(Value *V) : V(V) {} |
| 269 | Action *getValue() const { return cast<Action *>(Val: V); } |
| 270 | Value *getScalar() const { return cast<Value *>(Val: V); } |
| 271 | bool needsExtract() const { return isa<Action *>(Val: V); } |
| 272 | int getExtractIdx() const { return ExtractIdx; } |
| 273 | }; |
| 274 | |
| 275 | using DescrVecT = SmallVector<ExtractElementDescr, 4>; |
| 276 | DescrVecT Descrs; |
| 277 | |
| 278 | public: |
| 279 | CollectDescr(SmallVectorImpl<ExtractElementDescr> &&Descrs) |
| 280 | : Descrs(std::move(Descrs)) {} |
| 281 | /// If all elements come from a single vector input, then return that vector |
| 282 | /// and also the shuffle mask required to get them in order. |
| 283 | std::optional<std::pair<Action *, ShuffleMask>> getSingleInput() const { |
| 284 | const auto &Descr0 = *Descrs.begin(); |
| 285 | if (!Descr0.needsExtract()) |
| 286 | return std::nullopt; |
| 287 | auto *V0 = Descr0.getValue(); |
| 288 | ShuffleMask::IndicesVecT MaskIndices; |
| 289 | MaskIndices.push_back(Elt: Descr0.getExtractIdx()); |
| 290 | for (const auto &Descr : drop_begin(RangeOrContainer: Descrs)) { |
| 291 | if (!Descr.needsExtract()) |
| 292 | return std::nullopt; |
| 293 | if (Descr.getValue() != V0) |
| 294 | return std::nullopt; |
| 295 | MaskIndices.push_back(Elt: Descr.getExtractIdx()); |
| 296 | } |
| 297 | return std::make_pair(x&: V0, y: ShuffleMask(std::move(MaskIndices))); |
| 298 | } |
| 299 | bool hasVectorInputs() const { |
| 300 | return any_of(Range: Descrs, P: [](const auto &D) { return D.needsExtract(); }); |
| 301 | } |
| 302 | const SmallVector<ExtractElementDescr, 4> &getDescrs() const { |
| 303 | return Descrs; |
| 304 | } |
| 305 | }; |
| 306 | |
| 307 | class DiamondReuseMultiInput final : public LegalityResult { |
| 308 | friend class LegalityAnalysis; |
| 309 | CollectDescr Descr; |
| 310 | DiamondReuseMultiInput(CollectDescr &&Descr) |
| 311 | : LegalityResult(LegalityResultID::DiamondReuseMultiInput), |
| 312 | Descr(std::move(Descr)) {} |
| 313 | |
| 314 | public: |
| 315 | static bool classof(const LegalityResult *From) { |
| 316 | return From->getSubclassID() == LegalityResultID::DiamondReuseMultiInput; |
| 317 | } |
| 318 | const CollectDescr &getCollectDescr() const { return Descr; } |
| 319 | }; |
| 320 | |
| 321 | /// Performs the legality analysis and returns a LegalityResult object. |
| 322 | class LegalityAnalysis { |
| 323 | Scheduler Sched; |
| 324 | /// Owns the legality result objects created by createLegalityResult(). |
| 325 | SmallVector<std::unique_ptr<LegalityResult>> ResultPool; |
| 326 | /// Checks opcodes, types and other IR-specifics and returns a ResultReason |
| 327 | /// object if not vectorizable, or nullptr otherwise. |
| 328 | std::optional<ResultReason> |
| 329 | notVectorizableBasedOnOpcodesAndTypes(BndlRef<Value *> Bndl); |
| 330 | |
| 331 | ScalarEvolution &SE; |
| 332 | const DataLayout &DL; |
| 333 | TargetTransformInfo &TTI; |
| 334 | InstrMaps &IMaps; |
| 335 | |
| 336 | /// Finds how we can collect the values in \p Bndl from the vectorized or |
| 337 | /// non-vectorized code. It returns a map of the value we should extract from |
| 338 | /// and the corresponding shuffle mask we need to use. |
| 339 | CollectDescr getHowToCollectValues(BndlRef<Value *> Bndl) const; |
| 340 | |
| 341 | public: |
| 342 | LegalityAnalysis(AAResults &AA, ScalarEvolution &SE, const DataLayout &DL, |
| 343 | TargetTransformInfo &TTI, Context &Ctx, InstrMaps &IMaps, |
| 344 | SchedDirection Dir) |
| 345 | : Sched(AA, Ctx, Dir), SE(SE), DL(DL), TTI(TTI), IMaps(IMaps) {} |
| 346 | /// A LegalityResult factory. |
| 347 | template <typename ResultT, typename... ArgsT> |
| 348 | ResultT &createLegalityResult(ArgsT &&...Args) { |
| 349 | ResultPool.push_back( |
| 350 | std::unique_ptr<ResultT>(new ResultT(std::move(Args)...))); |
| 351 | return cast<ResultT>(*ResultPool.back()); |
| 352 | } |
| 353 | |
| 354 | /// \returns true if \p Instrs are in different blocks. |
| 355 | template <typename ValueT> |
| 356 | static bool differentBlock(BndlRef<ValueT *> Instrs) { |
| 357 | auto *BB0 = cast<Instruction>(Instrs[0])->getParent(); |
| 358 | return any_of(drop_begin(Instrs), [BB0](auto *V) { |
| 359 | return cast<Instruction>(V)->getParent() != BB0; |
| 360 | }); |
| 361 | } |
| 362 | |
| 363 | /// \returns true if all values in \p Values are unique. |
| 364 | template <typename ValueT> static bool areUnique(BndlRef<ValueT *> Values) { |
| 365 | SmallPtrSet<Value *, 8> Unique(llvm::from_range, Values); |
| 366 | return Unique.size() == Values.size(); |
| 367 | } |
| 368 | |
| 369 | /// \returns true if the alignment of the vector composed of \p Values has |
| 370 | /// alignment that is supported by the target. |
| 371 | LLVM_ABI bool isAlignmentSupported(ArrayRef<Value *> Values) const; |
| 372 | |
| 373 | /// Checks if it's legal to vectorize the instructions in \p Bndl. |
| 374 | /// \Returns a LegalityResult object owned by LegalityAnalysis. |
| 375 | /// \p SkipScheduling skips the scheduler check and is only meant for testing. |
| 376 | // TODO: Try to remove the SkipScheduling argument by refactoring the tests. |
| 377 | LLVM_ABI const LegalityResult &canVectorize(BndlRef<Value *> Bndl, |
| 378 | bool SkipScheduling = false); |
| 379 | /// \Returns a Pack with reason 'ForcePackForDebugging'. |
| 380 | const LegalityResult &getForcedPackForDebugging() { |
| 381 | return createLegalityResult<Pack>(Args: ResultReason::ForcePackForDebugging); |
| 382 | } |
| 383 | LLVM_ABI void clear(); |
| 384 | }; |
| 385 | |
| 386 | } // namespace llvm::sandboxir |
| 387 | |
| 388 | #endif // LLVM_TRANSFORMS_VECTORIZE_SANDBOXVECTORIZER_LEGALITY_H |
| 389 | |