| 1 | //===- SLPUtils.h - SLP Vectorizer free utility helpers --------*- 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 | // Internal header used by SLPVectorizer.cpp. It declares free helper |
| 10 | // functions that do not depend on BoUpSLP, InstructionsState, or any other |
| 11 | // SLP-private type. Splitting them out keeps SLPVectorizer.cpp focused on |
| 12 | // the build / legality / cost / codegen pipeline. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
| 16 | #ifndef LLVM_LIB_TRANSFORMS_VECTORIZE_SLPVECTORIZER_SLPUTILS_H |
| 17 | #define LLVM_LIB_TRANSFORMS_VECTORIZE_SLPVECTORIZER_SLPUTILS_H |
| 18 | |
| 19 | #include "llvm/ADT/APInt.h" |
| 20 | #include "llvm/ADT/ArrayRef.h" |
| 21 | #include "llvm/ADT/SmallBitVector.h" |
| 22 | #include "llvm/ADT/SmallVector.h" |
| 23 | #include "llvm/Analysis/MemoryLocation.h" |
| 24 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 25 | #include "llvm/IR/Intrinsics.h" |
| 26 | |
| 27 | #include <optional> |
| 28 | #include <string> |
| 29 | |
| 30 | namespace llvm { |
| 31 | class Constant; |
| 32 | class DataLayout; |
| 33 | class Instruction; |
| 34 | class TargetLibraryInfo; |
| 35 | class Type; |
| 36 | class Value; |
| 37 | } // namespace llvm |
| 38 | |
| 39 | namespace llvm::slpvectorizer { |
| 40 | |
| 41 | /// Limit of the number of uses for potentially transformed instructions/values, |
| 42 | /// used in checks to avoid compile-time explode. |
| 43 | inline constexpr int UsesLimit = 64; |
| 44 | |
| 45 | /// \returns True if the value is a constant (but not globals/constant |
| 46 | /// expressions). |
| 47 | bool isConstant(Value *V); |
| 48 | |
| 49 | /// \returns True if \p V is the integer identity constant for binary \p Opcode |
| 50 | /// (e.g. 0 for add, 1 for mul, all-ones for and). Floating-point identities are |
| 51 | /// excluded: a ConstantInt never matches the ConstantFP getBinOpIdentity() |
| 52 | /// returns for FAdd/FMul, whose identity fast-math may break anyway. |
| 53 | bool isBinOpIdentityConstant(const Value *V, unsigned Opcode); |
| 54 | |
| 55 | /// \returns the opcode of the combines emitted for a reassociated node: |
| 56 | /// subtract chains regroup their positive and negative operand columns with |
| 57 | /// plain adds. |
| 58 | unsigned getReassocCombineOpcode(unsigned Opcode); |
| 59 | |
| 60 | /// \returns True if \p I can be a link of a flattenable binary chain: |
| 61 | /// subtracts flatten as adds of a negated leaf, float subtracts need reassoc |
| 62 | /// to allow the regrouping. |
| 63 | bool isReassocChainLink(const Instruction *I); |
| 64 | |
| 65 | /// Checks if \p V is one of vector-like instructions, i.e. undef, |
| 66 | /// insertelement/extractelement with constant indices for fixed vector type |
| 67 | /// or extractvalue instruction. |
| 68 | bool isVectorLikeInstWithConstOps(Value *V); |
| 69 | |
| 70 | /// \returns the number of elements for Ty. |
| 71 | unsigned getNumElements(Type *Ty); |
| 72 | |
| 73 | /// Returns power-of-2 number of elements in a single register (part), given |
| 74 | /// the total number of elements \p Size and number of registers (parts) \p |
| 75 | /// NumParts. |
| 76 | unsigned getPartNumElems(unsigned Size, unsigned NumParts); |
| 77 | |
| 78 | /// Returns correct remaining number of elements, considering total amount |
| 79 | /// \p Size, (power-of-2 number) of elements in a single register |
| 80 | /// \p PartNumElems and current register (part) \p Part. |
| 81 | unsigned getNumElems(unsigned Size, unsigned PartNumElems, unsigned Part); |
| 82 | |
| 83 | #if !defined(NDEBUG) |
| 84 | /// Print a short descriptor of the instruction bundle suitable for debug |
| 85 | /// output. |
| 86 | std::string shortBundleName(ArrayRef<Value *> VL, int Idx = -1); |
| 87 | #endif |
| 88 | |
| 89 | /// \returns True if all of the instructions in \p VL are in the same block. |
| 90 | bool allSameBlock(ArrayRef<Value *> VL); |
| 91 | |
| 92 | /// \returns True if all of the values in \p VL are constants (but not |
| 93 | /// globals/constant expressions). |
| 94 | bool allConstant(ArrayRef<Value *> VL); |
| 95 | |
| 96 | /// \returns True if all of the values in \p VL are identical or some of them |
| 97 | /// are UndefValue. |
| 98 | bool isSplat(ArrayRef<Value *> VL); |
| 99 | |
| 100 | /// Checks if \p LHS and \p RHS are the same intrinsic, or one is llvm.fma |
| 101 | /// and the other is llvm.fmuladd, since both lower to the same fused |
| 102 | /// vector operation. |
| 103 | /// \returns the intrinsic ID to use for the pair (\p RHS if the IDs match, |
| 104 | /// otherwise Intrinsic::fma), or Intrinsic::not_intrinsic if they are not |
| 105 | /// equivalent. |
| 106 | Intrinsic::ID isEquivalentIntrinsicID(Intrinsic::ID LHS, Intrinsic::ID RHS); |
| 107 | |
| 108 | /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator. |
| 109 | /// For BinaryOperator, it also checks if \p ValWithUses is used in specific |
| 110 | /// patterns that make it effectively commutative (like equality comparisons |
| 111 | /// with zero). |
| 112 | /// In most cases, users should not call this function directly (since \p I and |
| 113 | /// \p ValWithUses are the same). However, when analyzing interchangeable |
| 114 | /// instructions, we need to use the converted opcode along with the original |
| 115 | /// uses. |
| 116 | /// \param I The instruction to check for commutativity |
| 117 | /// \param ValWithUses The value whose uses are analyzed for special |
| 118 | /// patterns |
| 119 | bool isCommutative(const Instruction *I, const Value *ValWithUses, |
| 120 | bool IsCopyable = false); |
| 121 | |
| 122 | /// This is a helper function to check whether \p I is commutative. |
| 123 | /// This is a convenience wrapper that calls the two-parameter version of |
| 124 | /// isCommutative with the same instruction for both parameters. This is |
| 125 | /// the common case where the instruction being checked for commutativity |
| 126 | /// is the same as the instruction whose uses are analyzed for special |
| 127 | /// patterns (see the two-parameter version above for details). |
| 128 | /// \param I The instruction to check for commutativity |
| 129 | /// \returns true if the instruction is commutative, false otherwise |
| 130 | bool isCommutative(const Instruction *I); |
| 131 | |
| 132 | /// Checks if the operand is commutative. In commutative operations, not all |
| 133 | /// operands might commutable, e.g. for fmuladd only 2 first operands are |
| 134 | /// commutable. |
| 135 | bool isCommutableOperand(const Instruction *I, Value *ValWithUses, unsigned Op, |
| 136 | bool IsCopyable = false); |
| 137 | |
| 138 | /// \returns number of operands of \p I, considering commutativity. Returns 2 |
| 139 | /// for commutative intrinsics. |
| 140 | /// \param I The instruction to check for commutativity |
| 141 | unsigned getNumberOfPotentiallyCommutativeOps(Instruction *I); |
| 142 | |
| 143 | /// \returns inserting or extracting index of InsertElement, ExtractElement |
| 144 | /// or InsertValue instruction, using \p Offset as base offset for index. |
| 145 | /// \returns std::nullopt if the index is not an immediate. |
| 146 | std::optional<unsigned> getElementIndex(const Value *Inst, unsigned Offset = 0); |
| 147 | |
| 148 | /// \returns True if all of the values in \p VL use the same opcode. |
| 149 | /// For comparison instructions, also checks if predicates match. |
| 150 | /// PoisonValues are considered matching. Interchangeable instructions are |
| 151 | /// not considered. |
| 152 | bool allSameOpcode(ArrayRef<Value *> VL); |
| 153 | |
| 154 | /// \returns Optional element Idx for Extract{Value,Element} instructions. |
| 155 | std::optional<unsigned> (const Instruction *E); |
| 156 | |
| 157 | /// Compute the inverse permutation \p Mask of \p Indices. |
| 158 | void inversePermutation(ArrayRef<unsigned> Indices, SmallVectorImpl<int> &Mask); |
| 159 | |
| 160 | /// Reorders the list of scalars in accordance with the given \p Mask. |
| 161 | void reorderScalars(SmallVectorImpl<Value *> &Scalars, ArrayRef<int> Mask); |
| 162 | |
| 163 | /// \returns True iff every value in \p VL has the same Type as the first. |
| 164 | bool allSameType(ArrayRef<Value *> VL); |
| 165 | |
| 166 | /// Checks if the provided value does not require scheduling. It does not |
| 167 | /// require scheduling if this is not an instruction or it is an instruction |
| 168 | /// that does not read/write memory and all operands are either not |
| 169 | /// instructions or phi nodes or instructions from different blocks. |
| 170 | bool areAllOperandsNonInsts(Value *V); |
| 171 | |
| 172 | /// Checks if the provided value does not require scheduling. It does not |
| 173 | /// require scheduling if this is not an instruction or it is an instruction |
| 174 | /// that does not read/write memory and all users are phi nodes or |
| 175 | /// instructions from different blocks. |
| 176 | bool isUsedOutsideBlock(Value *V); |
| 177 | |
| 178 | /// Checks if the specified value does not require scheduling. It does not |
| 179 | /// require scheduling if all operands and all users do not need to be |
| 180 | /// scheduled in the current basic block. |
| 181 | bool doesNotNeedToBeScheduled(Value *V); |
| 182 | |
| 183 | /// Checks if the specified array of instructions does not require scheduling. |
| 184 | /// It is so if all either instructions have operands that do not require |
| 185 | /// scheduling or their users do not require scheduling since they are phis or |
| 186 | /// in other basic blocks. |
| 187 | bool doesNotNeedToSchedule(ArrayRef<Value *> VL); |
| 188 | |
| 189 | /// \returns inserting or extracting index of InsertElement / ExtractElement |
| 190 | /// instruction, using \p Offset as base offset for index. Only instantiated |
| 191 | /// for InsertElementInst and ExtractElementInst (see SLPUtils.cpp). |
| 192 | template <typename T> |
| 193 | std::optional<unsigned> (const Value *Inst, |
| 194 | unsigned Offset); |
| 195 | |
| 196 | void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements, |
| 197 | SmallVectorImpl<int> &Mask); |
| 198 | |
| 199 | /// \returns the number of groups of shufflevector |
| 200 | /// A group has the following features |
| 201 | /// 1. All of value in a group are shufflevector. |
| 202 | /// 2. The mask of all shufflevector is isExtractSubvectorMask. |
| 203 | /// 3. The mask of all shufflevector uses all of the elements of the source. |
| 204 | /// e.g., it is 1 group (%0) |
| 205 | /// %1 = shufflevector <16 x i8> %0, <16 x i8> poison, |
| 206 | /// <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7> |
| 207 | /// %2 = shufflevector <16 x i8> %0, <16 x i8> poison, |
| 208 | /// <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15> |
| 209 | /// it is 2 groups (%3 and %4) |
| 210 | /// %5 = shufflevector <8 x i16> %3, <8 x i16> poison, |
| 211 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 212 | /// %6 = shufflevector <8 x i16> %3, <8 x i16> poison, |
| 213 | /// <4 x i32> <i32 4, i32 5, i32 6, i32 7> |
| 214 | /// %7 = shufflevector <8 x i16> %4, <8 x i16> poison, |
| 215 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 216 | /// %8 = shufflevector <8 x i16> %4, <8 x i16> poison, |
| 217 | /// <4 x i32> <i32 4, i32 5, i32 6, i32 7> |
| 218 | /// it is 0 group |
| 219 | /// %12 = shufflevector <8 x i16> %10, <8 x i16> poison, |
| 220 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 221 | /// %13 = shufflevector <8 x i16> %11, <8 x i16> poison, |
| 222 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 223 | unsigned getShufflevectorNumGroups(ArrayRef<Value *> VL); |
| 224 | |
| 225 | /// \returns a shufflevector mask which is used to vectorize shufflevectors |
| 226 | /// e.g., |
| 227 | /// %5 = shufflevector <8 x i16> %3, <8 x i16> poison, |
| 228 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 229 | /// %6 = shufflevector <8 x i16> %3, <8 x i16> poison, |
| 230 | /// <4 x i32> <i32 4, i32 5, i32 6, i32 7> |
| 231 | /// %7 = shufflevector <8 x i16> %4, <8 x i16> poison, |
| 232 | /// <4 x i32> <i32 0, i32 1, i32 2, i32 3> |
| 233 | /// %8 = shufflevector <8 x i16> %4, <8 x i16> poison, |
| 234 | /// <4 x i32> <i32 4, i32 5, i32 6, i32 7> |
| 235 | /// the result is |
| 236 | /// <0, 1, 2, 3, 12, 13, 14, 15, 16, 17, 18, 19, 28, 29, 30, 31> |
| 237 | SmallVector<int> calculateShufflevectorMask(ArrayRef<Value *> VL); |
| 238 | |
| 239 | /// Specifies the way the mask should be analyzed for undefs/poisonous elements |
| 240 | /// in the shuffle mask. |
| 241 | enum class UseMask { |
| 242 | FirstArg, ///< The mask is expected to be for permutation of 1-2 vectors, |
| 243 | ///< check for the mask elements for the first argument (mask |
| 244 | ///< indices are in range [0:VF)). |
| 245 | SecondArg, ///< The mask is expected to be for permutation of 2 vectors, check |
| 246 | ///< for the mask elements for the second argument (mask indices |
| 247 | ///< are in range [VF:2*VF)) |
| 248 | UndefsAsMask ///< Consider undef mask elements (-1) as placeholders for |
| 249 | ///< future shuffle elements and mark them as ones as being used |
| 250 | ///< in future. Non-undef elements are considered as unused since |
| 251 | ///< they're already marked as used in the mask. |
| 252 | }; |
| 253 | |
| 254 | /// Prepares a use bitset for the given mask either for the first argument or |
| 255 | /// for the second. |
| 256 | SmallBitVector buildUseMask(int VF, ArrayRef<int> Mask, UseMask MaskArg); |
| 257 | |
| 258 | /// Checks if the given value is actually an undefined constant vector. |
| 259 | /// Also, if the \p UseMask is not empty, tries to check if the non-masked |
| 260 | /// elements actually mask the insertelement buildvector, if any. |
| 261 | template <bool IsPoisonOnly = false> |
| 262 | SmallBitVector isUndefVector(const Value *V, |
| 263 | const SmallBitVector &UseMask = {}); |
| 264 | |
| 265 | /// \returns True if in-tree use also needs extract. This refers to |
| 266 | /// possible scalar operand in vectorized instruction. |
| 267 | bool (Value *Scalar, Instruction *UserInst, |
| 268 | TargetLibraryInfo *TLI, |
| 269 | const TargetTransformInfo *TTI); |
| 270 | |
| 271 | /// \returns the AA location that is being access by the instruction. |
| 272 | MemoryLocation getLocation(Instruction *I); |
| 273 | |
| 274 | /// \returns True if the instruction is not a volatile or atomic load/store. |
| 275 | bool isSimple(Instruction *I); |
| 276 | |
| 277 | /// Checks if the loads with scalar type \p ScalarTy and pointer operands |
| 278 | /// \p PointerOps are each (optionally via a constant-offset GEP) a |
| 279 | /// `select Cond, A, B` picking between the same two base pointers A/B on |
| 280 | /// every lane - the shape a fully unrolled `x = cond ? A[i] : B[i]` takes. On |
| 281 | /// success \p TrueBase / \p FalseBase are the candidate bases and |
| 282 | /// \p Conditions holds each lane's `select` condition, used to build the |
| 283 | /// blend mask. Lane \p Idx must be at `Base + Idx * sizeof(ScalarTy)`; only |
| 284 | /// dense, natural lane order starting at the base is recognized (reordered or |
| 285 | /// partial groups fall back to Gather/Scatter). |
| 286 | bool isSelectedBaseLoad(Type *ScalarTy, ArrayRef<Value *> PointerOps, |
| 287 | const DataLayout &DL, Value *&TrueBase, |
| 288 | Value *&FalseBase, |
| 289 | SmallVectorImpl<Value *> &Conditions); |
| 290 | |
| 291 | /// Shuffles \p Mask in accordance with the given \p SubMask. |
| 292 | /// \param ExtendingManyInputs Supports reshuffling of the mask with not only |
| 293 | /// one but two input vectors. |
| 294 | void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask, |
| 295 | bool ExtendingManyInputs = false); |
| 296 | |
| 297 | /// Order may have elements assigned special value (size) which is out of |
| 298 | /// bounds. Such indices only appear on places which correspond to undef values |
| 299 | /// (see canReuseExtract for details) and used in order to avoid undef values |
| 300 | /// have effect on operands ordering. |
| 301 | /// The first loop below simply finds all unused indices and then the next loop |
| 302 | /// nest assigns these indices for undef values positions. |
| 303 | /// As an example below Order has two undef positions and they have assigned |
| 304 | /// values 3 and 7 respectively: |
| 305 | /// before: 6 9 5 4 9 2 1 0 |
| 306 | /// after: 6 3 5 4 7 2 1 0 |
| 307 | void fixupOrderingIndices(MutableArrayRef<unsigned> Order); |
| 308 | |
| 309 | /// \returns a bitset for selecting opcodes. false for Opcode0 and true for |
| 310 | /// Opcode1. |
| 311 | SmallBitVector getAltInstrMask(ArrayRef<Value *> VL, Type *ScalarTy, |
| 312 | unsigned Opcode0, unsigned Opcode1); |
| 313 | |
| 314 | /// Replicates the given \p Val \p VF times. |
| 315 | SmallVector<Constant *> replicateMask(ArrayRef<Constant *> Val, unsigned VF); |
| 316 | |
| 317 | /// \returns the masked division/remainder intrinsic corresponding to \p |
| 318 | /// Opcode. Disabled lanes of these intrinsics are poison rather than UB, |
| 319 | /// unlike the plain opcode. |
| 320 | Intrinsic::ID getMaskedDivRemIntrinsic(unsigned Opcode); |
| 321 | |
| 322 | /// Returns true if \p I forms a vectorizable bundle on its own and its single |
| 323 | /// user does not tear the vector apart. Loads and addresses are excluded: the |
| 324 | /// tree is built without the users, so it does not pay off the extracts. A |
| 325 | /// cast, feeding a multi-used cast, is excluded for the same reason, such a |
| 326 | /// user stays scalar. The fp-to-int conversions move the result to the other |
| 327 | /// register domain, so the extracts are paid on top of the repacking. The |
| 328 | /// values, feeding the inserts, are vectorized together with them by the |
| 329 | /// dedicated attempt. |
| 330 | bool isOnceUsedSeed(const Instruction *I); |
| 331 | |
| 332 | /// If \p V is a single-use fpext of a single-use fptrunc forming a round-trip |
| 333 | /// back to the type of \p V, returns the fptrunc; the round-trip source is its |
| 334 | /// operand, always an instruction of the same type as \p V. If |
| 335 | /// \p MustBeElidable, matches only when the intermediate rounding may be |
| 336 | /// removed: both casts must allow contraction and the widening cast cannot |
| 337 | /// produce nan/inf. |
| 338 | Instruction *lookThroughCastRoundTrip(Value *V, bool MustBeElidable); |
| 339 | |
| 340 | /// Narrow reduction leaf: the value, the shift applied after widening and |
| 341 | /// the mask applied in the narrow type before widening, clearing the bits |
| 342 | /// the absorbed narrow shls shift out and applying the absorbed narrow |
| 343 | /// and-masks. Lossless narrow shls contribute their known-zero bits to the |
| 344 | /// mask so matching lanes can form a splat. All-ones mask means nothing |
| 345 | /// was absorbed and no 'and' is needed. |
| 346 | struct NarrowedLeafInfo { |
| 347 | NarrowedLeafInfo(Value *V, unsigned Shift, APInt Mask) |
| 348 | : V(V), Shift(Shift), Mask(std::move(Mask)) {} |
| 349 | |
| 350 | Value *V; |
| 351 | unsigned Shift; |
| 352 | APInt Mask; |
| 353 | }; |
| 354 | |
| 355 | /// Recursively collects the narrow leaves of the widened reduction value |
| 356 | /// \p V. zext is looked through directly, same-kind binops per operand, |
| 357 | /// shl of a zext - only if no bits are shifted out in the current type, |
| 358 | /// shls in narrower types fold into the shift and ands with a constant into |
| 359 | /// the mask applied in the narrow type. Also collects the looked-through |
| 360 | /// instructions into \p ChainInsts. |
| 361 | void collectNarrowedLeaves(Value *V, unsigned RdxOpcode, unsigned WideBW, |
| 362 | unsigned MaxDepth, |
| 363 | SmallVectorImpl<NarrowedLeafInfo> &Leaves, |
| 364 | SmallVectorImpl<Instruction *> &ChainInsts); |
| 365 | |
| 366 | TargetTransformInfo::TargetCostKind getSLPCostKind(const Function *F); |
| 367 | |
| 368 | } // namespace llvm::slpvectorizer |
| 369 | |
| 370 | #endif // LLVM_LIB_TRANSFORMS_VECTORIZE_SLPVECTORIZER_SLPUTILS_H |
| 371 | |