| 1 | //===- SLPCompatibilityAnalysis.cpp - SLP same-opcode helpers -------------===// |
| 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 | #include "SLPCompatibilityAnalysis.h" |
| 10 | #include "SLPUtils.h" |
| 11 | |
| 12 | #include "llvm/ADT/APInt.h" |
| 13 | #include "llvm/ADT/ArrayRef.h" |
| 14 | #include "llvm/ADT/STLExtras.h" |
| 15 | #include "llvm/ADT/SetVector.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/ADT/SmallVectorExtras.h" |
| 18 | #include "llvm/ADT/bit.h" |
| 19 | #include "llvm/Analysis/VectorUtils.h" |
| 20 | #include "llvm/IR/Constants.h" |
| 21 | #include "llvm/IR/InstrTypes.h" |
| 22 | #include "llvm/IR/Instruction.h" |
| 23 | #include "llvm/IR/Instructions.h" |
| 24 | #include "llvm/IR/IntrinsicInst.h" |
| 25 | #include "llvm/IR/Intrinsics.h" |
| 26 | #include "llvm/IR/PatternMatch.h" |
| 27 | #include "llvm/IR/Value.h" |
| 28 | #include "llvm/Support/Casting.h" |
| 29 | #include "llvm/Support/ErrorHandling.h" |
| 30 | |
| 31 | #include <algorithm> |
| 32 | #include <array> |
| 33 | #include <cassert> |
| 34 | #include <optional> |
| 35 | #include <utility> |
| 36 | |
| 37 | using namespace llvm; |
| 38 | using namespace llvm::PatternMatch; |
| 39 | |
| 40 | namespace llvm::slpvectorizer { |
| 41 | |
| 42 | bool isValidForAlternation(unsigned Opcode) { |
| 43 | return !Instruction::isIntDivRem(Opcode); |
| 44 | } |
| 45 | |
| 46 | std::pair<Constant *, unsigned> |
| 47 | BinOpSameOpcodeHelper::isBinOpWithConstant(const Instruction *I) { |
| 48 | [[maybe_unused]] unsigned Opcode = I->getOpcode(); |
| 49 | assert(binary_search(SupportedOp, Opcode) && "Unsupported opcode." ); |
| 50 | (void)SupportedOp; |
| 51 | auto *BinOp = cast<BinaryOperator>(Val: I); |
| 52 | auto GetConstant = [](Value *V) -> Constant * { |
| 53 | if (auto *CI = dyn_cast<ConstantInt>(Val: V)) |
| 54 | return CI; |
| 55 | return dyn_cast<ConstantFP>(Val: V); |
| 56 | }; |
| 57 | if (Constant *C = GetConstant(BinOp->getOperand(i_nocapture: 1))) |
| 58 | return {C, 1}; |
| 59 | if (!isCommutative(I)) |
| 60 | return {nullptr, 0}; |
| 61 | if (Constant *C = GetConstant(BinOp->getOperand(i_nocapture: 0))) |
| 62 | return {C, 0}; |
| 63 | return {nullptr, 0}; |
| 64 | } |
| 65 | |
| 66 | bool BinOpSameOpcodeHelper::InterchangeableInfo::trySet( |
| 67 | MaskType OpcodeInMaskForm, MaskType InterchangeableMask) { |
| 68 | if (Mask & InterchangeableMask) { |
| 69 | SeenBefore |= OpcodeInMaskForm; |
| 70 | Mask &= InterchangeableMask; |
| 71 | return true; |
| 72 | } |
| 73 | return false; |
| 74 | } |
| 75 | |
| 76 | unsigned BinOpSameOpcodeHelper::InterchangeableInfo::getOpcode() const { |
| 77 | MaskType Candidate = Mask & SeenBefore; |
| 78 | if (Candidate & MainOpBIT) |
| 79 | return I->getOpcode(); |
| 80 | if (Candidate & ShlBIT) |
| 81 | return Instruction::Shl; |
| 82 | if (Candidate & AShrBIT) |
| 83 | return Instruction::AShr; |
| 84 | if (Candidate & MulBIT) |
| 85 | return Instruction::Mul; |
| 86 | if (Candidate & AddBIT) |
| 87 | return Instruction::Add; |
| 88 | if (Candidate & SubBIT) |
| 89 | return Instruction::Sub; |
| 90 | if (Candidate & FAddBIT) |
| 91 | return Instruction::FAdd; |
| 92 | if (Candidate & FSubBIT) |
| 93 | return Instruction::FSub; |
| 94 | if (Candidate & AndBIT) |
| 95 | return Instruction::And; |
| 96 | if (Candidate & OrBIT) |
| 97 | return Instruction::Or; |
| 98 | if (Candidate & XorBIT) |
| 99 | return Instruction::Xor; |
| 100 | llvm_unreachable("Cannot find interchangeable instruction." ); |
| 101 | } |
| 102 | |
| 103 | bool BinOpSameOpcodeHelper::InterchangeableInfo::hasCandidateOpcode( |
| 104 | unsigned Opcode) const { |
| 105 | MaskType Candidate = Mask & SeenBefore; |
| 106 | switch (Opcode) { |
| 107 | case Instruction::Shl: |
| 108 | return Candidate & ShlBIT; |
| 109 | case Instruction::AShr: |
| 110 | return Candidate & AShrBIT; |
| 111 | case Instruction::Mul: |
| 112 | return Candidate & MulBIT; |
| 113 | case Instruction::Add: |
| 114 | return Candidate & AddBIT; |
| 115 | case Instruction::Sub: |
| 116 | return Candidate & SubBIT; |
| 117 | case Instruction::And: |
| 118 | return Candidate & AndBIT; |
| 119 | case Instruction::Or: |
| 120 | return Candidate & OrBIT; |
| 121 | case Instruction::Xor: |
| 122 | return Candidate & XorBIT; |
| 123 | case Instruction::FAdd: |
| 124 | return Candidate & FAddBIT; |
| 125 | case Instruction::FSub: |
| 126 | return Candidate & FSubBIT; |
| 127 | case Instruction::LShr: |
| 128 | case Instruction::FMul: |
| 129 | case Instruction::SDiv: |
| 130 | case Instruction::UDiv: |
| 131 | case Instruction::FDiv: |
| 132 | case Instruction::SRem: |
| 133 | case Instruction::URem: |
| 134 | case Instruction::FRem: |
| 135 | return false; |
| 136 | default: |
| 137 | break; |
| 138 | } |
| 139 | llvm_unreachable("Cannot find interchangeable instruction." ); |
| 140 | } |
| 141 | |
| 142 | SmallVector<Value *> BinOpSameOpcodeHelper::InterchangeableInfo::getOperand( |
| 143 | const Instruction *To) const { |
| 144 | unsigned ToOpcode = To->getOpcode(); |
| 145 | unsigned FromOpcode = I->getOpcode(); |
| 146 | if (FromOpcode == ToOpcode) |
| 147 | return SmallVector<Value *>(I->operands()); |
| 148 | assert(binary_search(SupportedOp, ToOpcode) && "Unsupported opcode." ); |
| 149 | auto [C, Pos] = isBinOpWithConstant(I); |
| 150 | Type *RHSType = I->getOperand(i: Pos)->getType(); |
| 151 | Constant *RHS; |
| 152 | if (auto *CFP = dyn_cast<ConstantFP>(Val: C)) { |
| 153 | // fsub(x, c) == fadd(x, -c) for every FP constant c, since IEEE 754 |
| 154 | // defines subtraction as addition of the negated operand. |
| 155 | assert(is_contained({Instruction::FAdd, Instruction::FSub}, ToOpcode) && |
| 156 | "Cannot convert the instruction." ); |
| 157 | RHS = ConstantFP::get(Ty: RHSType, V: -CFP->getValueAPF()); |
| 158 | } else { |
| 159 | auto *CI = cast<ConstantInt>(Val: C); |
| 160 | const APInt &FromCIValue = CI->getValue(); |
| 161 | unsigned FromCIValueBitWidth = FromCIValue.getBitWidth(); |
| 162 | switch (FromOpcode) { |
| 163 | case Instruction::Shl: |
| 164 | if (ToOpcode == Instruction::Add && FromCIValue.isOne()) |
| 165 | return {I->getOperand(i: 0), I->getOperand(i: 0)}; |
| 166 | if (ToOpcode == Instruction::Mul) { |
| 167 | RHS = ConstantInt::get(Ty: RHSType, |
| 168 | V: APInt::getOneBitSet(numBits: FromCIValueBitWidth, |
| 169 | BitNo: FromCIValue.getZExtValue())); |
| 170 | } else { |
| 171 | assert(FromCIValue.isZero() && "Cannot convert the instruction." ); |
| 172 | RHS = ConstantExpr::getBinOpIdentity(Opcode: ToOpcode, Ty: RHSType, |
| 173 | /*AllowRHSConstant=*/true); |
| 174 | } |
| 175 | break; |
| 176 | case Instruction::Mul: |
| 177 | assert(FromCIValue.isPowerOf2() && "Cannot convert the instruction." ); |
| 178 | if (ToOpcode == Instruction::Shl) { |
| 179 | RHS = ConstantInt::get( |
| 180 | Ty: RHSType, V: APInt(FromCIValueBitWidth, FromCIValue.logBase2())); |
| 181 | } else { |
| 182 | assert(FromCIValue.isOne() && "Cannot convert the instruction." ); |
| 183 | RHS = ConstantExpr::getBinOpIdentity(Opcode: ToOpcode, Ty: RHSType, |
| 184 | /*AllowRHSConstant=*/true); |
| 185 | } |
| 186 | break; |
| 187 | case Instruction::Add: |
| 188 | case Instruction::Sub: |
| 189 | if (FromCIValue.isZero()) { |
| 190 | RHS = ConstantExpr::getBinOpIdentity(Opcode: ToOpcode, Ty: RHSType, |
| 191 | /*AllowRHSConstant=*/true); |
| 192 | } else { |
| 193 | assert(is_contained({Instruction::Add, Instruction::Sub}, ToOpcode) && |
| 194 | "Cannot convert the instruction." ); |
| 195 | APInt NegatedVal = APInt(FromCIValue); |
| 196 | NegatedVal.negate(); |
| 197 | RHS = ConstantInt::get(Ty: RHSType, V: NegatedVal); |
| 198 | } |
| 199 | break; |
| 200 | case Instruction::And: |
| 201 | assert(FromCIValue.isAllOnes() && "Cannot convert the instruction." ); |
| 202 | RHS = ConstantExpr::getBinOpIdentity(Opcode: ToOpcode, Ty: RHSType, |
| 203 | /*AllowRHSConstant=*/true); |
| 204 | break; |
| 205 | default: |
| 206 | assert(FromCIValue.isZero() && "Cannot convert the instruction." ); |
| 207 | RHS = ConstantExpr::getBinOpIdentity(Opcode: ToOpcode, Ty: RHSType, |
| 208 | /*AllowRHSConstant=*/true); |
| 209 | break; |
| 210 | } |
| 211 | } |
| 212 | Value *LHS = I->getOperand(i: 1 - Pos); |
| 213 | // If the target opcode is non-commutative (e.g., shl, sub), |
| 214 | // force the variable to the left and the constant to the right. |
| 215 | if (Pos == 1 || !Instruction::isCommutative(Opcode: ToOpcode)) |
| 216 | return SmallVector<Value *>({LHS, RHS}); |
| 217 | |
| 218 | return SmallVector<Value *>({RHS, LHS}); |
| 219 | } |
| 220 | |
| 221 | bool BinOpSameOpcodeHelper::isValidForAlternation(const Instruction *I) const { |
| 222 | return slpvectorizer::isValidForAlternation(Opcode: MainOp.I->getOpcode()) && |
| 223 | slpvectorizer::isValidForAlternation(Opcode: I->getOpcode()); |
| 224 | } |
| 225 | |
| 226 | bool BinOpSameOpcodeHelper::initializeAltOp(const Instruction *I) { |
| 227 | if (AltOp.I) |
| 228 | return true; |
| 229 | if (!isValidForAlternation(I)) |
| 230 | return false; |
| 231 | AltOp.I = I; |
| 232 | return true; |
| 233 | } |
| 234 | |
| 235 | bool BinOpSameOpcodeHelper::add(const Instruction *I) { |
| 236 | assert(isa<BinaryOperator>(I) && |
| 237 | "BinOpSameOpcodeHelper only accepts BinaryOperator." ); |
| 238 | unsigned Opcode = I->getOpcode(); |
| 239 | MaskType OpcodeInMaskForm; |
| 240 | // Prefer Shl, AShr, Mul, Add, Sub, And, Or, Xor, FAdd and FSub over |
| 241 | // MainOp. |
| 242 | switch (Opcode) { |
| 243 | case Instruction::Shl: |
| 244 | OpcodeInMaskForm = ShlBIT; |
| 245 | break; |
| 246 | case Instruction::AShr: |
| 247 | OpcodeInMaskForm = AShrBIT; |
| 248 | break; |
| 249 | case Instruction::Mul: |
| 250 | OpcodeInMaskForm = MulBIT; |
| 251 | break; |
| 252 | case Instruction::Add: |
| 253 | OpcodeInMaskForm = AddBIT; |
| 254 | break; |
| 255 | case Instruction::Sub: |
| 256 | OpcodeInMaskForm = SubBIT; |
| 257 | break; |
| 258 | case Instruction::And: |
| 259 | OpcodeInMaskForm = AndBIT; |
| 260 | break; |
| 261 | case Instruction::Or: |
| 262 | OpcodeInMaskForm = OrBIT; |
| 263 | break; |
| 264 | case Instruction::Xor: |
| 265 | OpcodeInMaskForm = XorBIT; |
| 266 | break; |
| 267 | case Instruction::FAdd: |
| 268 | OpcodeInMaskForm = FAddBIT; |
| 269 | break; |
| 270 | case Instruction::FSub: |
| 271 | OpcodeInMaskForm = FSubBIT; |
| 272 | break; |
| 273 | default: |
| 274 | return MainOp.equal(Opcode) || (initializeAltOp(I) && AltOp.equal(Opcode)); |
| 275 | } |
| 276 | MaskType InterchangeableMask = OpcodeInMaskForm; |
| 277 | auto [C, Pos] = isBinOpWithConstant(I); |
| 278 | if (auto *CI = dyn_cast_or_null<ConstantInt>(Val: C)) { |
| 279 | constexpr MaskType CanBeAll = |
| 280 | XorBIT | OrBIT | AndBIT | SubBIT | AddBIT | MulBIT | AShrBIT | ShlBIT; |
| 281 | const APInt &CIValue = CI->getValue(); |
| 282 | switch (Opcode) { |
| 283 | case Instruction::Shl: |
| 284 | if (CIValue.ult(RHS: CIValue.getBitWidth())) |
| 285 | InterchangeableMask = CIValue.isZero() ? CanBeAll : MulBIT | ShlBIT; |
| 286 | if (CIValue.isOne()) |
| 287 | InterchangeableMask |= AddBIT; |
| 288 | break; |
| 289 | case Instruction::Mul: |
| 290 | if (CIValue.isOne()) { |
| 291 | InterchangeableMask = CanBeAll; |
| 292 | break; |
| 293 | } |
| 294 | if (CIValue.isPowerOf2()) |
| 295 | InterchangeableMask = MulBIT | ShlBIT; |
| 296 | break; |
| 297 | case Instruction::Add: |
| 298 | case Instruction::Sub: |
| 299 | InterchangeableMask = CIValue.isZero() ? CanBeAll : SubBIT | AddBIT; |
| 300 | break; |
| 301 | case Instruction::And: |
| 302 | if (CIValue.isAllOnes()) |
| 303 | InterchangeableMask = CanBeAll; |
| 304 | break; |
| 305 | case Instruction::Xor: |
| 306 | if (CIValue.isZero()) |
| 307 | InterchangeableMask = XorBIT | OrBIT | SubBIT | AddBIT; |
| 308 | break; |
| 309 | default: |
| 310 | if (CIValue.isZero()) |
| 311 | InterchangeableMask = CanBeAll; |
| 312 | break; |
| 313 | } |
| 314 | } else if (C && Pos == 1) { |
| 315 | // FAdd/FSub with a constant RHS: negating the constant always |
| 316 | // converts one into the other, so no value check is needed. A |
| 317 | // constant LHS (Pos == 0, e.g. "0.0 - x") is excluded: unlike a |
| 318 | // constant RHS, it cannot be moved to the other opcode without also |
| 319 | // swapping the variable operand, which would misalign it against |
| 320 | // lanes that keep their native opcode (their variable operand stays |
| 321 | // on the other side). |
| 322 | InterchangeableMask = FSubBIT | FAddBIT; |
| 323 | } |
| 324 | return MainOp.trySet(OpcodeInMaskForm, InterchangeableMask) || |
| 325 | (initializeAltOp(I) && |
| 326 | AltOp.trySet(OpcodeInMaskForm, InterchangeableMask)); |
| 327 | } |
| 328 | |
| 329 | /// If the comparison (Pred, X, C) is a single-element or single-complement |
| 330 | /// range check, returns its boundary family: false + K for the singleton |
| 331 | /// {K} (eq forms), true + K for the complement of {K} (ne forms). |
| 332 | static std::optional<std::pair<bool, APInt>> |
| 333 | getCmpBoundaryFamily(CmpInst::Predicate Pred, const APInt &C) { |
| 334 | const unsigned BW = C.getBitWidth(); |
| 335 | const bool IsSigned = CmpInst::isSigned(Pred); |
| 336 | const APInt Min = IsSigned ? APInt::getSignedMinValue(numBits: BW) : APInt(BW, 0); |
| 337 | const APInt Max = |
| 338 | IsSigned ? APInt::getSignedMaxValue(numBits: BW) : APInt::getMaxValue(numBits: BW); |
| 339 | switch (Pred) { |
| 340 | case CmpInst::ICMP_EQ: |
| 341 | return std::make_pair(x: false, y: C); |
| 342 | case CmpInst::ICMP_NE: |
| 343 | return std::make_pair(x: true, y: C); |
| 344 | case CmpInst::ICMP_ULT: |
| 345 | case CmpInst::ICMP_SLT: |
| 346 | if (C == Min + 1) |
| 347 | return std::make_pair(x: false, y: Min); |
| 348 | if (C == Max) |
| 349 | return std::make_pair(x: true, y: C); |
| 350 | break; |
| 351 | case CmpInst::ICMP_ULE: |
| 352 | case CmpInst::ICMP_SLE: |
| 353 | if (C == Min) |
| 354 | return std::make_pair(x: false, y: C); |
| 355 | if (C == Max - 1) |
| 356 | return std::make_pair(x: true, y: Max); |
| 357 | break; |
| 358 | case CmpInst::ICMP_UGT: |
| 359 | case CmpInst::ICMP_SGT: |
| 360 | if (C == Min) |
| 361 | return std::make_pair(x: true, y: C); |
| 362 | if (C == Max - 1) |
| 363 | return std::make_pair(x: false, y: Max); |
| 364 | break; |
| 365 | case CmpInst::ICMP_UGE: |
| 366 | case CmpInst::ICMP_SGE: |
| 367 | if (C == Min + 1) |
| 368 | return std::make_pair(x: true, y: Min); |
| 369 | if (C == Max) |
| 370 | return std::make_pair(x: false, y: C); |
| 371 | break; |
| 372 | default: |
| 373 | break; |
| 374 | } |
| 375 | return std::nullopt; |
| 376 | } |
| 377 | |
| 378 | CmpSamePredicateHelper::MaskType |
| 379 | CmpSamePredicateHelper::getFormsMask(CmpInst::Predicate Pred, const APInt &C) { |
| 380 | MaskType M = getBit(P: Pred); |
| 381 | std::optional<std::pair<bool, APInt>> Family = getCmpBoundaryFamily(Pred, C); |
| 382 | if (!Family) |
| 383 | return M; |
| 384 | const auto &[IsComplement, K] = *Family; |
| 385 | // At each type boundary the two range checks covering exactly {K}; the |
| 386 | // complement family uses their inverses, covering everything but {K}. |
| 387 | const MaskType LoU = getBit(P: CmpInst::ICMP_ULT) | getBit(P: CmpInst::ICMP_ULE); |
| 388 | const MaskType HiU = getBit(P: CmpInst::ICMP_UGT) | getBit(P: CmpInst::ICMP_UGE); |
| 389 | const MaskType LoS = getBit(P: CmpInst::ICMP_SLT) | getBit(P: CmpInst::ICMP_SLE); |
| 390 | const MaskType HiS = getBit(P: CmpInst::ICMP_SGT) | getBit(P: CmpInst::ICMP_SGE); |
| 391 | if (K.isZero()) |
| 392 | M |= IsComplement ? HiU : LoU; |
| 393 | if (K.isMaxValue()) |
| 394 | M |= IsComplement ? LoU : HiU; |
| 395 | if (K.isMinSignedValue()) |
| 396 | M |= IsComplement ? HiS : LoS; |
| 397 | if (K.isMaxSignedValue()) |
| 398 | M |= IsComplement ? LoS : HiS; |
| 399 | return M | getBit(P: IsComplement ? CmpInst::ICMP_NE : CmpInst::ICMP_EQ); |
| 400 | } |
| 401 | |
| 402 | APInt CmpSamePredicateHelper::getFamilyConstant(bool IsComplement, |
| 403 | const APInt &K, |
| 404 | CmpInst::Predicate Pred) { |
| 405 | if (Pred == CmpInst::ICMP_EQ || Pred == CmpInst::ICMP_NE) |
| 406 | return K; |
| 407 | // The complement form uses the singleton constant of the inverse |
| 408 | // predicate. |
| 409 | if (IsComplement) |
| 410 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 411 | switch (Pred) { |
| 412 | case CmpInst::ICMP_ULT: |
| 413 | case CmpInst::ICMP_SLT: |
| 414 | return K + 1; |
| 415 | case CmpInst::ICMP_UGT: |
| 416 | case CmpInst::ICMP_SGT: |
| 417 | return K - 1; |
| 418 | default: |
| 419 | return K; |
| 420 | } |
| 421 | } |
| 422 | |
| 423 | bool CmpSamePredicateHelper::add(const ICmpInst *CI) { |
| 424 | MaskType LaneMask = getBit(P: CI->getPredicate()); |
| 425 | if (auto *C = dyn_cast<ConstantInt>(Val: CI->getOperand(i_nocapture: 1))) |
| 426 | LaneMask = getFormsMask(Pred: CI->getPredicate(), C: C->getValue()); |
| 427 | SeenBefore |= getBit(P: CI->getPredicate()); |
| 428 | Mask &= LaneMask; |
| 429 | return Mask != 0; |
| 430 | } |
| 431 | |
| 432 | CmpInst::Predicate |
| 433 | CmpSamePredicateHelper::getPredicate(const ICmpInst *Preferred) const { |
| 434 | MaskType Candidate = Mask & SeenBefore; |
| 435 | if (!Candidate) |
| 436 | return CmpInst::BAD_ICMP_PREDICATE; |
| 437 | if (Candidate & getBit(P: Preferred->getPredicate())) |
| 438 | return Preferred->getPredicate(); |
| 439 | return static_cast<CmpInst::Predicate>(CmpInst::ICMP_EQ + |
| 440 | countr_zero(Val: Candidate)); |
| 441 | } |
| 442 | |
| 443 | CmpInst::Predicate |
| 444 | CmpSamePredicateHelper::getSharedPredicate(ArrayRef<Value *> VL, |
| 445 | const ICmpInst *Preferred) { |
| 446 | CmpSamePredicateHelper Helper; |
| 447 | if (!all_of(Range&: VL, P: [&](Value *V) { |
| 448 | auto *CI = dyn_cast<ICmpInst>(Val: V); |
| 449 | return isa<PoisonValue>(Val: V) || (CI && Helper.add(CI)); |
| 450 | })) |
| 451 | return CmpInst::BAD_ICMP_PREDICATE; |
| 452 | return Helper.getPredicate(Preferred); |
| 453 | } |
| 454 | |
| 455 | bool CmpSamePredicateHelper::canConvertTo(const CmpInst *CI, |
| 456 | CmpInst::Predicate Pred) { |
| 457 | auto *ICI = dyn_cast<ICmpInst>(Val: CI); |
| 458 | if (!ICI || !CmpInst::isIntPredicate(P: Pred)) |
| 459 | return false; |
| 460 | if (ICI->getPredicate() == Pred) |
| 461 | return true; |
| 462 | auto *C = dyn_cast<ConstantInt>(Val: ICI->getOperand(i_nocapture: 1)); |
| 463 | return C && |
| 464 | (getFormsMask(Pred: ICI->getPredicate(), C: C->getValue()) & getBit(P: Pred)) != 0; |
| 465 | } |
| 466 | |
| 467 | ConstantInt * |
| 468 | CmpSamePredicateHelper::getAdjustedConstant(const CmpInst *CI, |
| 469 | CmpInst::Predicate Pred) { |
| 470 | if (!canConvertTo(CI, Pred)) |
| 471 | return nullptr; |
| 472 | auto *ICI = cast<ICmpInst>(Val: CI); |
| 473 | if (ICI->getPredicate() == Pred) |
| 474 | return nullptr; |
| 475 | auto *C = cast<ConstantInt>(Val: ICI->getOperand(i_nocapture: 1)); |
| 476 | std::optional<std::pair<bool, APInt>> Family = |
| 477 | getCmpBoundaryFamily(Pred: ICI->getPredicate(), C: C->getValue()); |
| 478 | assert(Family && "Expected a boundary family for a convertible compare." ); |
| 479 | const auto &[IsComplement, K] = *Family; |
| 480 | return ConstantInt::get(Context&: CI->getContext(), |
| 481 | V: getFamilyConstant(IsComplement, K, Pred)); |
| 482 | } |
| 483 | |
| 484 | bool InstructionsState::isSameOperation(const Instruction *I, |
| 485 | const Instruction *Op) { |
| 486 | if (I->getOpcode() != Op->getOpcode()) |
| 487 | return false; |
| 488 | const auto *II = dyn_cast<IntrinsicInst>(Val: I); |
| 489 | const auto *IOp = dyn_cast<IntrinsicInst>(Val: Op); |
| 490 | if (II || IOp) |
| 491 | return II && IOp && |
| 492 | isEquivalentIntrinsicID(LHS: II->getIntrinsicID(), |
| 493 | RHS: IOp->getIntrinsicID()) != |
| 494 | Intrinsic::not_intrinsic; |
| 495 | return true; |
| 496 | } |
| 497 | |
| 498 | Instruction *InstructionsState::getMatchingMainOpOrAltOp(Instruction *I) const { |
| 499 | assert(MainOp && "MainOp cannot be nullptr." ); |
| 500 | if (isSameOperation(I, Op: MainOp)) |
| 501 | return MainOp; |
| 502 | if (MainOp->getOpcode() == Instruction::Select && |
| 503 | I->getOpcode() == Instruction::ZExt && !isAltShuffle()) |
| 504 | return MainOp; |
| 505 | // Prefer AltOp instead of interchangeable instruction of MainOp. |
| 506 | assert(AltOp && "AltOp cannot be nullptr." ); |
| 507 | if (isSameOperation(I, Op: AltOp)) |
| 508 | return AltOp; |
| 509 | // BinOpSameOpcodeHelper handles only BinaryOperators; a call cannot match. |
| 510 | if (!I->isBinaryOp() || !MainOp->isBinaryOp()) |
| 511 | return nullptr; |
| 512 | BinOpSameOpcodeHelper Converter(MainOp); |
| 513 | if (!Converter.add(I) || !Converter.add(I: MainOp)) |
| 514 | return nullptr; |
| 515 | if (isAltShuffle() && !Converter.hasCandidateOpcode(Opcode: MainOp->getOpcode())) { |
| 516 | BinOpSameOpcodeHelper AltConverter(AltOp); |
| 517 | if (AltConverter.add(I) && AltConverter.add(I: AltOp) && |
| 518 | AltConverter.hasCandidateOpcode(Opcode: AltOp->getOpcode())) |
| 519 | return AltOp; |
| 520 | } |
| 521 | if (Converter.hasAltOp() && !isAltShuffle()) |
| 522 | return nullptr; |
| 523 | return Converter.hasAltOp() ? AltOp : MainOp; |
| 524 | } |
| 525 | |
| 526 | bool InstructionsState::isMulDivLikeOp() const { |
| 527 | constexpr std::array<unsigned, 8> MulDiv = { |
| 528 | Instruction::Mul, Instruction::FMul, Instruction::SDiv, |
| 529 | Instruction::UDiv, Instruction::FDiv, Instruction::SRem, |
| 530 | Instruction::URem, Instruction::FRem}; |
| 531 | return is_contained(Range: MulDiv, Element: getOpcode()) && |
| 532 | is_contained(Range: MulDiv, Element: getAltOpcode()); |
| 533 | } |
| 534 | |
| 535 | bool InstructionsState::isAddSubLikeOp() const { |
| 536 | constexpr std::array<unsigned, 4> AddSub = { |
| 537 | Instruction::Add, Instruction::Sub, Instruction::FAdd, Instruction::FSub}; |
| 538 | return is_contained(Range: AddSub, Element: getOpcode()) && |
| 539 | is_contained(Range: AddSub, Element: getAltOpcode()); |
| 540 | } |
| 541 | |
| 542 | bool InstructionsState::isCopyableElement(Value *V) const { |
| 543 | assert(valid() && "InstructionsState is invalid." ); |
| 544 | if (!HasCopyables) |
| 545 | return false; |
| 546 | if (isAltShuffle() || getOpcode() == Instruction::GetElementPtr) |
| 547 | return false; |
| 548 | auto *I = dyn_cast<Instruction>(Val: V); |
| 549 | if (!I) |
| 550 | return !isa<PoisonValue>(Val: V); |
| 551 | if (I->getParent() != MainOp->getParent() && |
| 552 | (!isVectorLikeInstWithConstOps(V: I) || |
| 553 | !isVectorLikeInstWithConstOps(V: MainOp))) |
| 554 | return true; |
| 555 | if (isSameOperation(I, Op: MainOp)) |
| 556 | return false; |
| 557 | // BinOpSameOpcodeHelper handles only BinaryOperators; a call is copyable. |
| 558 | if (!I->isBinaryOp() || !MainOp->isBinaryOp()) |
| 559 | return true; |
| 560 | BinOpSameOpcodeHelper Converter(MainOp); |
| 561 | return !Converter.add(I) || !Converter.add(I: MainOp) || Converter.hasAltOp() || |
| 562 | !Converter.hasCandidateOpcode(Opcode: getOpcode()); |
| 563 | } |
| 564 | |
| 565 | bool isAbsorbableFMulOrFAdd(ArrayRef<Value *> VL, Value *V) { |
| 566 | auto *I = dyn_cast<Instruction>(Val: V); |
| 567 | return I && |
| 568 | (I->getOpcode() == Instruction::FMul || |
| 569 | I->getOpcode() == Instruction::FAdd) && |
| 570 | I->hasOneUse() && none_of(Range: I->operands(), P: [&](Value *Op) { |
| 571 | return is_contained(Range&: VL, Element: Op); |
| 572 | }); |
| 573 | } |
| 574 | |
| 575 | bool isAbsorbableCopyableFMulOrFAdd(const InstructionsState &S, Value *V) { |
| 576 | auto *I = dyn_cast<Instruction>(Val: V); |
| 577 | return I && S.isCopyableElement(V: I) && |
| 578 | (I->getOpcode() == Instruction::FMul || |
| 579 | I->getOpcode() == Instruction::FAdd) && |
| 580 | I->hasOneUse(); |
| 581 | } |
| 582 | |
| 583 | bool hasOnlyAbsorbableCopyableFMulOrFAdds(ArrayRef<Value *> VL) { |
| 584 | bool HasFMulOrFAdd = false; |
| 585 | for (Value *V : VL) { |
| 586 | if (isa<PoisonValue>(Val: V)) |
| 587 | continue; |
| 588 | auto *I = dyn_cast<Instruction>(Val: V); |
| 589 | if (I && RecurrenceDescriptor::isFMulAddIntrinsic(I)) |
| 590 | continue; |
| 591 | if (!isAbsorbableFMulOrFAdd(VL, V)) |
| 592 | return false; |
| 593 | HasFMulOrFAdd = true; |
| 594 | } |
| 595 | return HasFMulOrFAdd; |
| 596 | } |
| 597 | |
| 598 | bool InstructionsState::isExpandedBinOp(Value *V) const { |
| 599 | assert(valid() && "InstructionsState is invalid." ); |
| 600 | if (isCopyableElement(V)) |
| 601 | return false; |
| 602 | auto *ExpandingOp = dyn_cast<Instruction>(Val: V); |
| 603 | if (!ExpandingOp) |
| 604 | return false; |
| 605 | auto CheckForTransformedOpcode = [](const Instruction *RefOp, |
| 606 | const Instruction *ExpandingOp) { |
| 607 | switch (RefOp->getOpcode()) { |
| 608 | case Instruction::Add: |
| 609 | switch (ExpandingOp->getOpcode()) { |
| 610 | case Instruction::Shl: |
| 611 | return match(V: ExpandingOp, P: m_Shl(L: m_Value(), R: m_One())); |
| 612 | default: |
| 613 | break; |
| 614 | } |
| 615 | break; |
| 616 | default: |
| 617 | break; |
| 618 | } |
| 619 | return false; |
| 620 | }; |
| 621 | // getMatchingMainOpOrAltOp() may legitimately return nullptr, e.g. for a |
| 622 | // split node, whose Scalars combine two unrelated operations (main/alt |
| 623 | // ops of the split state), so V is not required to match either of them. |
| 624 | Instruction *MainOp = getMatchingMainOpOrAltOp(I: ExpandingOp); |
| 625 | if (!MainOp) |
| 626 | return false; |
| 627 | return CheckForTransformedOpcode(MainOp, ExpandingOp); |
| 628 | } |
| 629 | |
| 630 | bool InstructionsState::isExpandedOperand(Instruction *I, unsigned Idx) const { |
| 631 | assert(isExpandedBinOp(I) && "Expected an expanded binop." ); |
| 632 | switch (I->getOpcode()) { |
| 633 | case Instruction::Shl: |
| 634 | assert(match(I, m_Shl(m_Value(), m_One())) && "Expected shl x, 1 only." ); |
| 635 | return Idx == 1; |
| 636 | default: |
| 637 | llvm_unreachable("Unexpected opcode for an expanded operand." ); |
| 638 | } |
| 639 | } |
| 640 | |
| 641 | bool InstructionsState::isNonSchedulable(Value *V) const { |
| 642 | assert(valid() && "InstructionsState is invalid." ); |
| 643 | auto *I = dyn_cast<Instruction>(Val: V); |
| 644 | if (!HasCopyables) |
| 645 | return !I || isa<PHINode>(Val: I) || isVectorLikeInstWithConstOps(V: I) || |
| 646 | doesNotNeedToBeScheduled(V); |
| 647 | // MainOp for copyables always schedulable to correctly identify |
| 648 | // non-schedulable copyables. |
| 649 | if (getMainOp() == V) |
| 650 | return false; |
| 651 | if (isCopyableElement(V)) { |
| 652 | auto IsNonSchedulableCopyableElement = [this](Value *V) { |
| 653 | auto *I = dyn_cast<Instruction>(Val: V); |
| 654 | return !I || isa<PHINode>(Val: I) || I->getParent() != MainOp->getParent() || |
| 655 | (doesNotNeedToBeScheduled(V: I) && |
| 656 | // If the copyable instructions comes after MainOp |
| 657 | // (non-schedulable, but used in the block) - cannot vectorize |
| 658 | // it, will possibly generate use before def. |
| 659 | !MainOp->comesBefore(Other: I)); |
| 660 | }; |
| 661 | |
| 662 | return IsNonSchedulableCopyableElement(V); |
| 663 | } |
| 664 | return !I || isa<PHINode>(Val: I) || isVectorLikeInstWithConstOps(V: I) || |
| 665 | doesNotNeedToBeScheduled(V); |
| 666 | } |
| 667 | |
| 668 | /// Find an instruction with a specific opcode in VL. |
| 669 | /// \param VL Array of values to search through. Must contain only Instructions |
| 670 | /// and PoisonValues. |
| 671 | /// \param Opcode The instruction opcode to search for |
| 672 | /// \returns |
| 673 | /// - The first instruction found with matching opcode |
| 674 | /// - nullptr if no matching instruction is found |
| 675 | static Instruction *findInstructionWithOpcode(ArrayRef<Value *> VL, |
| 676 | unsigned Opcode) { |
| 677 | for (Value *V : VL) { |
| 678 | if (isa<PoisonValue>(Val: V)) |
| 679 | continue; |
| 680 | assert(isa<Instruction>(V) && "Only accepts PoisonValue and Instruction." ); |
| 681 | auto *Inst = cast<Instruction>(Val: V); |
| 682 | if (Inst->getOpcode() == Opcode) |
| 683 | return Inst; |
| 684 | } |
| 685 | return nullptr; |
| 686 | } |
| 687 | |
| 688 | /// Checks if the provided operands of 2 cmp instructions are compatible, i.e. |
| 689 | /// compatible instructions or constants, or just some other regular values. |
| 690 | static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0, |
| 691 | Value *Op1, const TargetLibraryInfo &TLI) { |
| 692 | return (isConstant(V: BaseOp0) && isConstant(V: Op0)) || |
| 693 | (isConstant(V: BaseOp1) && isConstant(V: Op1)) || |
| 694 | (!isa<Instruction>(Val: BaseOp0) && !isa<Instruction>(Val: Op0) && |
| 695 | !isa<Instruction>(Val: BaseOp1) && !isa<Instruction>(Val: Op1)) || |
| 696 | BaseOp0 == Op0 || BaseOp1 == Op1 || |
| 697 | getSameOpcode(VL: {BaseOp0, Op0}, TLI) || |
| 698 | getSameOpcode(VL: {BaseOp1, Op1}, TLI); |
| 699 | } |
| 700 | |
| 701 | /// \returns true if a compare instruction \p CI has similar "look" and |
| 702 | /// same predicate as \p BaseCI, "as is" or with its operands and predicate |
| 703 | /// swapped, false otherwise. |
| 704 | static bool isCmpSameOrSwapped(const CmpInst *BaseCI, const CmpInst *CI, |
| 705 | const TargetLibraryInfo &TLI) { |
| 706 | assert(BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() && |
| 707 | "Assessing comparisons of different types?" ); |
| 708 | CmpInst::Predicate BasePred = BaseCI->getPredicate(); |
| 709 | CmpInst::Predicate Pred = CI->getPredicate(); |
| 710 | CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(pred: Pred); |
| 711 | |
| 712 | Value *BaseOp0 = BaseCI->getOperand(i_nocapture: 0); |
| 713 | Value *BaseOp1 = BaseCI->getOperand(i_nocapture: 1); |
| 714 | Value *Op0 = CI->getOperand(i_nocapture: 0); |
| 715 | Value *Op1 = CI->getOperand(i_nocapture: 1); |
| 716 | |
| 717 | return (BasePred == Pred && |
| 718 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1, TLI)) || |
| 719 | (BasePred == SwappedPred && |
| 720 | areCompatibleCmpOps(BaseOp0, BaseOp1, Op0: Op1, Op1: Op0, TLI)); |
| 721 | } |
| 722 | |
| 723 | InstructionsState getSameOpcode(ArrayRef<Value *> VL, |
| 724 | const TargetLibraryInfo &TLI) { |
| 725 | // Make sure these are all Instructions. |
| 726 | if (!all_of(Range&: VL, P: IsaPred<Instruction, PoisonValue>)) |
| 727 | return InstructionsState::invalid(); |
| 728 | |
| 729 | auto *It = find_if(Range&: VL, P: IsaPred<Instruction>); |
| 730 | if (It == VL.end()) |
| 731 | return InstructionsState::invalid(); |
| 732 | |
| 733 | Instruction *MainOp = cast<Instruction>(Val: *It); |
| 734 | unsigned InstCnt = std::count_if(first: It, last: VL.end(), pred: IsaPred<Instruction>); |
| 735 | if ((VL.size() > 2 && !isa<PHINode>(Val: MainOp) && InstCnt < VL.size() / 2) || |
| 736 | (VL.size() == 2 && InstCnt < 2)) |
| 737 | return InstructionsState::invalid(); |
| 738 | |
| 739 | bool IsCastOp = isa<CastInst>(Val: MainOp); |
| 740 | bool IsBinOp = isa<BinaryOperator>(Val: MainOp); |
| 741 | bool IsCmpOp = isa<CmpInst>(Val: MainOp); |
| 742 | CmpInst::Predicate BasePred = IsCmpOp ? cast<CmpInst>(Val: MainOp)->getPredicate() |
| 743 | : CmpInst::BAD_ICMP_PREDICATE; |
| 744 | Instruction *AltOp = MainOp; |
| 745 | unsigned Opcode = MainOp->getOpcode(); |
| 746 | unsigned AltOpcode = Opcode; |
| 747 | |
| 748 | BinOpSameOpcodeHelper BinOpHelper(MainOp); |
| 749 | bool SwappedPredsCompatible = IsCmpOp && [&]() { |
| 750 | SetVector<unsigned> UniquePreds, UniqueNonSwappedPreds; |
| 751 | UniquePreds.insert(X: BasePred); |
| 752 | UniqueNonSwappedPreds.insert(X: BasePred); |
| 753 | for (Value *V : VL) { |
| 754 | auto *I = dyn_cast<CmpInst>(Val: V); |
| 755 | if (!I) |
| 756 | return false; |
| 757 | CmpInst::Predicate CurrentPred = I->getPredicate(); |
| 758 | CmpInst::Predicate SwappedCurrentPred = |
| 759 | CmpInst::getSwappedPredicate(pred: CurrentPred); |
| 760 | UniqueNonSwappedPreds.insert(X: CurrentPred); |
| 761 | if (!UniquePreds.contains(key: CurrentPred) && |
| 762 | !UniquePreds.contains(key: SwappedCurrentPred)) |
| 763 | UniquePreds.insert(X: CurrentPred); |
| 764 | } |
| 765 | // Total number of predicates > 2, but if consider swapped predicates |
| 766 | // compatible only 2, consider swappable predicates as compatible opcodes, |
| 767 | // not alternate. |
| 768 | return UniqueNonSwappedPreds.size() > 2 && UniquePreds.size() == 2; |
| 769 | }(); |
| 770 | // Find the predicate the whole bundle can share, if any, treating |
| 771 | // boundary comparisons canonicalized to eq/ne as interchangeable. |
| 772 | CmpInst::Predicate InterchangeablePred = CmpInst::BAD_ICMP_PREDICATE; |
| 773 | if (IsCmpOp && isa<ICmpInst>(Val: MainOp)) |
| 774 | InterchangeablePred = |
| 775 | CmpSamePredicateHelper::getSharedPredicate(VL, Preferred: cast<ICmpInst>(Val: MainOp)); |
| 776 | // Check for one alternate opcode from another BinaryOperator. |
| 777 | // TODO - generalize to support all operators (types, calls etc.). |
| 778 | Intrinsic::ID BaseID = 0; |
| 779 | SmallVector<VFInfo, 4> BaseMappings; |
| 780 | if (auto *CallBase = dyn_cast<CallInst>(Val: MainOp)) { |
| 781 | BaseID = getVectorIntrinsicIDForCall(CI: CallBase, TLI: &TLI); |
| 782 | BaseMappings = VFDatabase(*CallBase).getMappings(CI: *CallBase); |
| 783 | if (!isTriviallyVectorizable(ID: BaseID) && BaseMappings.empty()) |
| 784 | return InstructionsState::invalid(); |
| 785 | } |
| 786 | bool AnyPoison = InstCnt != VL.size(); |
| 787 | // Check MainOp too to be sure that it matches the requirements for the |
| 788 | // instructions. |
| 789 | for (Value *V : iterator_range(It, VL.end())) { |
| 790 | auto *I = dyn_cast<Instruction>(Val: V); |
| 791 | if (!I) |
| 792 | continue; |
| 793 | |
| 794 | // Cannot combine poison and divisions. |
| 795 | // TODO: do some smart analysis of the CallInsts to exclude divide-like |
| 796 | // intrinsics/functions only. |
| 797 | if (AnyPoison && (I->isIntDivRem() || I->isFPDivRem() || isa<CallInst>(Val: I))) |
| 798 | return InstructionsState::invalid(); |
| 799 | unsigned InstOpcode = I->getOpcode(); |
| 800 | if (IsBinOp && isa<BinaryOperator>(Val: I)) { |
| 801 | if (BinOpHelper.add(I)) |
| 802 | continue; |
| 803 | } else if (IsCastOp && isa<CastInst>(Val: I)) { |
| 804 | Value *Op0 = MainOp->getOperand(i: 0); |
| 805 | Type *Ty0 = Op0->getType(); |
| 806 | Value *Op1 = I->getOperand(i: 0); |
| 807 | Type *Ty1 = Op1->getType(); |
| 808 | if (Ty0 == Ty1) { |
| 809 | if (InstOpcode == Opcode || InstOpcode == AltOpcode) |
| 810 | continue; |
| 811 | if (Opcode == AltOpcode) { |
| 812 | assert(isValidForAlternation(Opcode) && |
| 813 | isValidForAlternation(InstOpcode) && |
| 814 | "Cast isn't safe for alternation, logic needs to be updated!" ); |
| 815 | AltOpcode = InstOpcode; |
| 816 | AltOp = I; |
| 817 | continue; |
| 818 | } |
| 819 | } |
| 820 | } else if (auto *Inst = dyn_cast<CmpInst>(Val: I); Inst && IsCmpOp) { |
| 821 | auto *BaseInst = cast<CmpInst>(Val: MainOp); |
| 822 | Type *Ty0 = BaseInst->getOperand(i_nocapture: 0)->getType(); |
| 823 | Type *Ty1 = Inst->getOperand(i_nocapture: 0)->getType(); |
| 824 | if (Ty0 == Ty1) { |
| 825 | assert(InstOpcode == Opcode && "Expected same CmpInst opcode." ); |
| 826 | assert(InstOpcode == AltOpcode && |
| 827 | "Alternate instructions are only supported by BinaryOperator " |
| 828 | "and CastInst." ); |
| 829 | // Check for compatible operands. If the corresponding operands are not |
| 830 | // compatible - need to perform alternate vectorization. |
| 831 | CmpInst::Predicate CurrentPred = Inst->getPredicate(); |
| 832 | CmpInst::Predicate SwappedCurrentPred = |
| 833 | CmpInst::getSwappedPredicate(pred: CurrentPred); |
| 834 | |
| 835 | if ((VL.size() == 2 || SwappedPredsCompatible) && |
| 836 | (BasePred == CurrentPred || BasePred == SwappedCurrentPred)) |
| 837 | continue; |
| 838 | |
| 839 | if (isCmpSameOrSwapped(BaseCI: BaseInst, CI: Inst, TLI)) |
| 840 | continue; |
| 841 | if (CmpSamePredicateHelper::canConvertTo(CI: Inst, Pred: InterchangeablePred)) |
| 842 | continue; |
| 843 | auto *AltInst = cast<CmpInst>(Val: AltOp); |
| 844 | if (MainOp != AltOp) { |
| 845 | if (isCmpSameOrSwapped(BaseCI: AltInst, CI: Inst, TLI)) |
| 846 | continue; |
| 847 | } else if (BasePred != CurrentPred) { |
| 848 | assert( |
| 849 | isValidForAlternation(InstOpcode) && |
| 850 | "CmpInst isn't safe for alternation, logic needs to be updated!" ); |
| 851 | AltOp = I; |
| 852 | continue; |
| 853 | } |
| 854 | CmpInst::Predicate AltPred = AltInst->getPredicate(); |
| 855 | if (BasePred == CurrentPred || BasePred == SwappedCurrentPred || |
| 856 | AltPred == CurrentPred || AltPred == SwappedCurrentPred) |
| 857 | continue; |
| 858 | } |
| 859 | } else if (InstOpcode == Opcode) { |
| 860 | assert(InstOpcode == AltOpcode && |
| 861 | "Alternate instructions are only supported by BinaryOperator and " |
| 862 | "CastInst." ); |
| 863 | if (auto *Gep = dyn_cast<GetElementPtrInst>(Val: I)) { |
| 864 | if (Gep->getNumOperands() != 2 || |
| 865 | Gep->getOperand(i_nocapture: 0)->getType() != MainOp->getOperand(i: 0)->getType()) |
| 866 | return InstructionsState::invalid(); |
| 867 | } else if (auto *EI = dyn_cast<ExtractElementInst>(Val: I)) { |
| 868 | if (!isVectorLikeInstWithConstOps(V: EI)) |
| 869 | return InstructionsState::invalid(); |
| 870 | } else if (auto *LI = dyn_cast<LoadInst>(Val: I)) { |
| 871 | auto *BaseLI = cast<LoadInst>(Val: MainOp); |
| 872 | if (!LI->isSimple() || !BaseLI->isSimple()) |
| 873 | return InstructionsState::invalid(); |
| 874 | } else if (auto *Call = dyn_cast<CallInst>(Val: I)) { |
| 875 | auto *CallBase = cast<CallInst>(Val: MainOp); |
| 876 | Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI: Call, TLI: &TLI); |
| 877 | Intrinsic::ID Equivalent = isEquivalentIntrinsicID(LHS: ID, RHS: BaseID); |
| 878 | if (Call->getCalledFunction() != CallBase->getCalledFunction() && |
| 879 | isEquivalentIntrinsicID(LHS: Equivalent, RHS: Intrinsic::fmuladd) == |
| 880 | Intrinsic::not_intrinsic) |
| 881 | return InstructionsState::invalid(); |
| 882 | if (Call->hasOperandBundles() && |
| 883 | (!CallBase->hasOperandBundles() || |
| 884 | !std::equal(first1: Call->op_begin() + Call->getBundleOperandsStartIndex(), |
| 885 | last1: Call->op_begin() + Call->getBundleOperandsEndIndex(), |
| 886 | first2: CallBase->op_begin() + |
| 887 | CallBase->getBundleOperandsStartIndex()))) |
| 888 | return InstructionsState::invalid(); |
| 889 | if (ID != BaseID && Equivalent == Intrinsic::not_intrinsic) |
| 890 | return InstructionsState::invalid(); |
| 891 | if (!ID) { |
| 892 | SmallVector<VFInfo, 4> Mappings = |
| 893 | VFDatabase(*Call).getMappings(CI: *Call); |
| 894 | if (Mappings.size() != BaseMappings.size() || |
| 895 | Mappings.front().ISA != BaseMappings.front().ISA || |
| 896 | Mappings.front().ScalarName != BaseMappings.front().ScalarName || |
| 897 | Mappings.front().VectorName != BaseMappings.front().VectorName || |
| 898 | Mappings.front().Shape.VF != BaseMappings.front().Shape.VF || |
| 899 | Mappings.front().Shape.Parameters != |
| 900 | BaseMappings.front().Shape.Parameters) |
| 901 | return InstructionsState::invalid(); |
| 902 | } |
| 903 | } |
| 904 | continue; |
| 905 | } |
| 906 | return InstructionsState::invalid(); |
| 907 | } |
| 908 | |
| 909 | if (IsBinOp) { |
| 910 | if (!BinOpHelper.hasDefinedMainOpcode() || |
| 911 | !BinOpHelper.hasDefinedAltOpcode()) |
| 912 | return InstructionsState::invalid(); |
| 913 | MainOp = findInstructionWithOpcode(VL, Opcode: BinOpHelper.getMainOpcode()); |
| 914 | assert(MainOp && "Cannot find MainOp with Opcode from BinOpHelper." ); |
| 915 | AltOp = findInstructionWithOpcode(VL, Opcode: BinOpHelper.getAltOpcode()); |
| 916 | assert(AltOp && "Cannot find AltOp with Opcode from BinOpHelper." ); |
| 917 | } else if (auto *CB = dyn_cast<CallInst>(Val: MainOp); |
| 918 | CB && |
| 919 | getVectorIntrinsicIDForCall(CI: CB, TLI: &TLI) == Intrinsic::fmuladd) { |
| 920 | // fma and fmuladd share a single vector fma node; use the fma as the |
| 921 | // representative so the fused form is not weakened to fmuladd. |
| 922 | auto *It = find_if(Range&: VL, P: [&](Value *V) { |
| 923 | auto *CI = dyn_cast<CallInst>(Val: V); |
| 924 | return CI && getVectorIntrinsicIDForCall(CI, TLI: &TLI) == Intrinsic::fma; |
| 925 | }); |
| 926 | if (It != VL.end()) |
| 927 | MainOp = AltOp = cast<Instruction>(Val: *It); |
| 928 | } |
| 929 | if (IsCmpOp && InterchangeablePred != CmpInst::BAD_ICMP_PREDICATE && |
| 930 | InterchangeablePred != BasePred) { |
| 931 | // Every lane is convertible to the shared predicate, so the alternate |
| 932 | // operation is never set for such bundles. |
| 933 | auto *SharedIt = find_if(Range&: VL, P: [&](Value *V) { |
| 934 | auto *CI = dyn_cast<ICmpInst>(Val: V); |
| 935 | return CI && CI->getPredicate() == InterchangeablePred; |
| 936 | }); |
| 937 | assert(SharedIt != VL.end() && |
| 938 | "Expected an instruction with the shared predicate." ); |
| 939 | MainOp = AltOp = cast<Instruction>(Val: *SharedIt); |
| 940 | } |
| 941 | assert((MainOp == AltOp || !allSameOpcode(VL)) && |
| 942 | "Incorrect implementation of allSameOpcode." ); |
| 943 | InstructionsState S(MainOp, AltOp); |
| 944 | assert(all_of(VL, |
| 945 | [&](Value *V) { |
| 946 | return isa<PoisonValue>(V) || |
| 947 | S.getMatchingMainOpOrAltOp(cast<Instruction>(V)); |
| 948 | }) && |
| 949 | "Invalid InstructionsState." ); |
| 950 | return S; |
| 951 | } |
| 952 | |
| 953 | std::pair<Instruction *, SmallVector<Value *>> |
| 954 | convertTo(Instruction *I, const InstructionsState &S) { |
| 955 | Instruction *SelectedOp = S.getMatchingMainOpOrAltOp(I); |
| 956 | assert(SelectedOp && "Cannot convert the instruction." ); |
| 957 | if (I->isBinaryOp()) { |
| 958 | BinOpSameOpcodeHelper Converter(I); |
| 959 | return std::make_pair(x&: SelectedOp, y: Converter.getOperand(I: SelectedOp)); |
| 960 | } |
| 961 | // Use args() to skip the trailing callee operand in CallInst::operands(). |
| 962 | if (auto *CI = dyn_cast<CallInst>(Val: I)) |
| 963 | return std::make_pair(x&: SelectedOp, y: SmallVector<Value *>(CI->args())); |
| 964 | // A comparison lane interchangeable with the main operation (e.g. x == 0 |
| 965 | // in an x <u C bundle) is emitted with the main predicate and the |
| 966 | // adjusted constant. |
| 967 | if (auto *MainCI = dyn_cast<ICmpInst>(Val: SelectedOp); |
| 968 | MainCI && !S.isAltShuffle()) |
| 969 | if (ConstantInt *C = CmpSamePredicateHelper::getAdjustedConstant( |
| 970 | CI: cast<ICmpInst>(Val: I), Pred: MainCI->getPredicate())) |
| 971 | return std::make_pair(x&: SelectedOp, |
| 972 | y: SmallVector<Value *>{I->getOperand(i: 0), C}); |
| 973 | return std::make_pair(x&: SelectedOp, y: SmallVector<Value *>(I->operands())); |
| 974 | } |
| 975 | |
| 976 | bool isAlternateInstruction(Instruction *I, Instruction *MainOp, |
| 977 | Instruction *AltOp, const TargetLibraryInfo &TLI) { |
| 978 | if (auto *MainCI = dyn_cast<CmpInst>(Val: MainOp)) { |
| 979 | auto *AltCI = cast<CmpInst>(Val: AltOp); |
| 980 | CmpInst::Predicate MainP = MainCI->getPredicate(); |
| 981 | [[maybe_unused]] CmpInst::Predicate AltP = AltCI->getPredicate(); |
| 982 | assert(MainP != AltP && "Expected different main/alternate predicates." ); |
| 983 | auto *CI = cast<CmpInst>(Val: I); |
| 984 | if (isCmpSameOrSwapped(BaseCI: MainCI, CI, TLI)) |
| 985 | return false; |
| 986 | if (isCmpSameOrSwapped(BaseCI: AltCI, CI, TLI)) |
| 987 | return true; |
| 988 | CmpInst::Predicate P = CI->getPredicate(); |
| 989 | CmpInst::Predicate SwappedP = CmpInst::getSwappedPredicate(pred: P); |
| 990 | |
| 991 | assert((MainP == P || AltP == P || MainP == SwappedP || AltP == SwappedP) && |
| 992 | "CmpInst expected to match either main or alternate predicate or " |
| 993 | "their swap." ); |
| 994 | return MainP != P && MainP != SwappedP; |
| 995 | } |
| 996 | return InstructionsState(MainOp, AltOp).getMatchingMainOpOrAltOp(I) == AltOp; |
| 997 | } |
| 998 | |
| 999 | SmallVector<SmallVector<Value *>> scanAltAssociativeOperands( |
| 1000 | const InstructionsState &S, const TargetLibraryInfo &TLI, |
| 1001 | ArrayRef<Value *> VL, ArrayRef<Value *> Op0, ArrayRef<Value *> Op1, |
| 1002 | SmallVectorImpl<Value *> &ReassocScalars, SmallBitVector &SubLanes) { |
| 1003 | assert(S.isAltShuffle() && "Expected an alternate node." ); |
| 1004 | const unsigned NumLanes = VL.size(); |
| 1005 | SmallVector<unsigned> LaneOpcodes = |
| 1006 | map_to_vector(C: seq<unsigned>(Size: NumLanes), F: [&](unsigned Lane) { |
| 1007 | return isAlternateInstruction(I: cast<Instruction>(Val: VL[Lane]), |
| 1008 | MainOp: S.getMainOp(), AltOp: S.getAltOp(), TLI) |
| 1009 | ? S.getAltOpcode() |
| 1010 | : S.getOpcode(); |
| 1011 | }); |
| 1012 | // A lane value peels only as a single-use chain link with the lane's own |
| 1013 | // opcode, keeping every combine level on the same main/alt pattern. |
| 1014 | auto GetChainLink = [&](unsigned Lane, Value *V) -> Instruction * { |
| 1015 | auto *I = dyn_cast<Instruction>(Val: V); |
| 1016 | if (!I || !I->hasOneUse() || I->getOpcode() != LaneOpcodes[Lane] || |
| 1017 | !isReassocChainLink(I)) |
| 1018 | return nullptr; |
| 1019 | return I; |
| 1020 | }; |
| 1021 | SmallVector<SmallVector<Value *>> Columns; |
| 1022 | Columns.emplace_back(Args: Op0.begin(), Args: Op0.end()); |
| 1023 | Columns.emplace_back(Args: Op1.begin(), Args: Op1.end()); |
| 1024 | // The chain link of a commutative lane may sit in the second column; |
| 1025 | // normalize so every lane's link leads. |
| 1026 | for (unsigned Lane : seq<unsigned>(Size: NumLanes)) { |
| 1027 | if (GetChainLink(Lane, Columns[0][Lane])) |
| 1028 | continue; |
| 1029 | Instruction *Link = GetChainLink(Lane, Columns[1][Lane]); |
| 1030 | if (!Link || !Link->isCommutative()) |
| 1031 | return {}; |
| 1032 | std::swap(a&: Columns[0][Lane], b&: Columns[1][Lane]); |
| 1033 | } |
| 1034 | // Peel the leading column while every lane stays a matching chain link. |
| 1035 | while (all_of(Range: seq<unsigned>(Size: NumLanes), P: [&](unsigned Lane) { |
| 1036 | return GetChainLink(Lane, Columns[0][Lane]) != nullptr; |
| 1037 | })) { |
| 1038 | SmallVector<Value *> NewColumn(NumLanes); |
| 1039 | for (unsigned Lane : seq<unsigned>(Size: NumLanes)) { |
| 1040 | Instruction *Link = GetChainLink(Lane, Columns[0][Lane]); |
| 1041 | ReassocScalars.push_back(Elt: Link); |
| 1042 | // The chain of a commutative lane may continue in the second operand; |
| 1043 | // keep the chain link as the running value. |
| 1044 | unsigned RunningOp = Link->isCommutative() && |
| 1045 | !GetChainLink(Lane, Link->getOperand(i: 0)) && |
| 1046 | GetChainLink(Lane, Link->getOperand(i: 1)) |
| 1047 | ? 1 |
| 1048 | : 0; |
| 1049 | NewColumn[Lane] = Link->getOperand(i: 1 - RunningOp); |
| 1050 | Columns[0][Lane] = Link->getOperand(i: RunningOp); |
| 1051 | } |
| 1052 | Columns.insert(I: std::next(x: Columns.begin()), Elt: std::move(NewColumn)); |
| 1053 | } |
| 1054 | assert(!ReassocScalars.empty() && |
| 1055 | "Normalization guarantees at least one peeled level." ); |
| 1056 | SubLanes.resize(N: NumLanes); |
| 1057 | for (unsigned Lane : seq<unsigned>(Size: NumLanes)) |
| 1058 | if (LaneOpcodes[Lane] == Instruction::Sub || |
| 1059 | LaneOpcodes[Lane] == Instruction::FSub) |
| 1060 | SubLanes.set(Lane); |
| 1061 | return Columns; |
| 1062 | } |
| 1063 | } // namespace llvm::slpvectorizer |
| 1064 | |