1 //===- InstCombineSelect.cpp ----------------------------------------------===// 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 // This file implements the visitSelect function. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "InstCombineInternal.h" 14 #include "llvm/ADT/APInt.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallVector.h" 17 #include "llvm/Analysis/AssumptionCache.h" 18 #include "llvm/Analysis/CmpInstAnalysis.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/OverflowInstAnalysis.h" 21 #include "llvm/Analysis/ValueTracking.h" 22 #include "llvm/Analysis/VectorUtils.h" 23 #include "llvm/IR/BasicBlock.h" 24 #include "llvm/IR/Constant.h" 25 #include "llvm/IR/ConstantRange.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DerivedTypes.h" 28 #include "llvm/IR/IRBuilder.h" 29 #include "llvm/IR/InstrTypes.h" 30 #include "llvm/IR/Instruction.h" 31 #include "llvm/IR/Instructions.h" 32 #include "llvm/IR/IntrinsicInst.h" 33 #include "llvm/IR/Intrinsics.h" 34 #include "llvm/IR/Operator.h" 35 #include "llvm/IR/PatternMatch.h" 36 #include "llvm/IR/Type.h" 37 #include "llvm/IR/User.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/Support/Casting.h" 40 #include "llvm/Support/ErrorHandling.h" 41 #include "llvm/Support/KnownBits.h" 42 #include "llvm/Transforms/InstCombine/InstCombiner.h" 43 #include <cassert> 44 #include <utility> 45 46 #define DEBUG_TYPE "instcombine" 47 #include "llvm/Transforms/Utils/InstructionWorklist.h" 48 49 using namespace llvm; 50 using namespace PatternMatch; 51 52 53 /// Replace a select operand based on an equality comparison with the identity 54 /// constant of a binop. 55 static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, 56 const TargetLibraryInfo &TLI, 57 InstCombinerImpl &IC) { 58 // The select condition must be an equality compare with a constant operand. 59 Value *X; 60 Constant *C; 61 CmpInst::Predicate Pred; 62 if (!match(Sel.getCondition(), m_Cmp(Pred, m_Value(X), m_Constant(C)))) 63 return nullptr; 64 65 bool IsEq; 66 if (ICmpInst::isEquality(Pred)) 67 IsEq = Pred == ICmpInst::ICMP_EQ; 68 else if (Pred == FCmpInst::FCMP_OEQ) 69 IsEq = true; 70 else if (Pred == FCmpInst::FCMP_UNE) 71 IsEq = false; 72 else 73 return nullptr; 74 75 // A select operand must be a binop. 76 BinaryOperator *BO; 77 if (!match(Sel.getOperand(IsEq ? 1 : 2), m_BinOp(BO))) 78 return nullptr; 79 80 // The compare constant must be the identity constant for that binop. 81 // If this a floating-point compare with 0.0, any zero constant will do. 82 Type *Ty = BO->getType(); 83 Constant *IdC = ConstantExpr::getBinOpIdentity(BO->getOpcode(), Ty, true); 84 if (IdC != C) { 85 if (!IdC || !CmpInst::isFPPredicate(Pred)) 86 return nullptr; 87 if (!match(IdC, m_AnyZeroFP()) || !match(C, m_AnyZeroFP())) 88 return nullptr; 89 } 90 91 // Last, match the compare variable operand with a binop operand. 92 Value *Y; 93 if (!BO->isCommutative() && !match(BO, m_BinOp(m_Value(Y), m_Specific(X)))) 94 return nullptr; 95 if (!match(BO, m_c_BinOp(m_Value(Y), m_Specific(X)))) 96 return nullptr; 97 98 // +0.0 compares equal to -0.0, and so it does not behave as required for this 99 // transform. Bail out if we can not exclude that possibility. 100 if (isa<FPMathOperator>(BO)) 101 if (!BO->hasNoSignedZeros() && !CannotBeNegativeZero(Y, &TLI)) 102 return nullptr; 103 104 // BO = binop Y, X 105 // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } 106 // => 107 // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } 108 return IC.replaceOperand(Sel, IsEq ? 1 : 2, Y); 109 } 110 111 /// This folds: 112 /// select (icmp eq (and X, C1)), TC, FC 113 /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. 114 /// To something like: 115 /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC 116 /// Or: 117 /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC 118 /// With some variations depending if FC is larger than TC, or the shift 119 /// isn't needed, or the bit widths don't match. 120 static Value *foldSelectICmpAnd(SelectInst &Sel, ICmpInst *Cmp, 121 InstCombiner::BuilderTy &Builder) { 122 const APInt *SelTC, *SelFC; 123 if (!match(Sel.getTrueValue(), m_APInt(SelTC)) || 124 !match(Sel.getFalseValue(), m_APInt(SelFC))) 125 return nullptr; 126 127 // If this is a vector select, we need a vector compare. 128 Type *SelType = Sel.getType(); 129 if (SelType->isVectorTy() != Cmp->getType()->isVectorTy()) 130 return nullptr; 131 132 Value *V; 133 APInt AndMask; 134 bool CreateAnd = false; 135 ICmpInst::Predicate Pred = Cmp->getPredicate(); 136 if (ICmpInst::isEquality(Pred)) { 137 if (!match(Cmp->getOperand(1), m_Zero())) 138 return nullptr; 139 140 V = Cmp->getOperand(0); 141 const APInt *AndRHS; 142 if (!match(V, m_And(m_Value(), m_Power2(AndRHS)))) 143 return nullptr; 144 145 AndMask = *AndRHS; 146 } else if (decomposeBitTestICmp(Cmp->getOperand(0), Cmp->getOperand(1), 147 Pred, V, AndMask)) { 148 assert(ICmpInst::isEquality(Pred) && "Not equality test?"); 149 if (!AndMask.isPowerOf2()) 150 return nullptr; 151 152 CreateAnd = true; 153 } else { 154 return nullptr; 155 } 156 157 // In general, when both constants are non-zero, we would need an offset to 158 // replace the select. This would require more instructions than we started 159 // with. But there's one special-case that we handle here because it can 160 // simplify/reduce the instructions. 161 APInt TC = *SelTC; 162 APInt FC = *SelFC; 163 if (!TC.isZero() && !FC.isZero()) { 164 // If the select constants differ by exactly one bit and that's the same 165 // bit that is masked and checked by the select condition, the select can 166 // be replaced by bitwise logic to set/clear one bit of the constant result. 167 if (TC.getBitWidth() != AndMask.getBitWidth() || (TC ^ FC) != AndMask) 168 return nullptr; 169 if (CreateAnd) { 170 // If we have to create an 'and', then we must kill the cmp to not 171 // increase the instruction count. 172 if (!Cmp->hasOneUse()) 173 return nullptr; 174 V = Builder.CreateAnd(V, ConstantInt::get(SelType, AndMask)); 175 } 176 bool ExtraBitInTC = TC.ugt(FC); 177 if (Pred == ICmpInst::ICMP_EQ) { 178 // If the masked bit in V is clear, clear or set the bit in the result: 179 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) ^ TC 180 // (V & AndMaskC) == 0 ? TC : FC --> (V & AndMaskC) | TC 181 Constant *C = ConstantInt::get(SelType, TC); 182 return ExtraBitInTC ? Builder.CreateXor(V, C) : Builder.CreateOr(V, C); 183 } 184 if (Pred == ICmpInst::ICMP_NE) { 185 // If the masked bit in V is set, set or clear the bit in the result: 186 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) | FC 187 // (V & AndMaskC) != 0 ? TC : FC --> (V & AndMaskC) ^ FC 188 Constant *C = ConstantInt::get(SelType, FC); 189 return ExtraBitInTC ? Builder.CreateOr(V, C) : Builder.CreateXor(V, C); 190 } 191 llvm_unreachable("Only expecting equality predicates"); 192 } 193 194 // Make sure one of the select arms is a power-of-2. 195 if (!TC.isPowerOf2() && !FC.isPowerOf2()) 196 return nullptr; 197 198 // Determine which shift is needed to transform result of the 'and' into the 199 // desired result. 200 const APInt &ValC = !TC.isZero() ? TC : FC; 201 unsigned ValZeros = ValC.logBase2(); 202 unsigned AndZeros = AndMask.logBase2(); 203 204 // Insert the 'and' instruction on the input to the truncate. 205 if (CreateAnd) 206 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), AndMask)); 207 208 // If types don't match, we can still convert the select by introducing a zext 209 // or a trunc of the 'and'. 210 if (ValZeros > AndZeros) { 211 V = Builder.CreateZExtOrTrunc(V, SelType); 212 V = Builder.CreateShl(V, ValZeros - AndZeros); 213 } else if (ValZeros < AndZeros) { 214 V = Builder.CreateLShr(V, AndZeros - ValZeros); 215 V = Builder.CreateZExtOrTrunc(V, SelType); 216 } else { 217 V = Builder.CreateZExtOrTrunc(V, SelType); 218 } 219 220 // Okay, now we know that everything is set up, we just don't know whether we 221 // have a icmp_ne or icmp_eq and whether the true or false val is the zero. 222 bool ShouldNotVal = !TC.isZero(); 223 ShouldNotVal ^= Pred == ICmpInst::ICMP_NE; 224 if (ShouldNotVal) 225 V = Builder.CreateXor(V, ValC); 226 227 return V; 228 } 229 230 /// We want to turn code that looks like this: 231 /// %C = or %A, %B 232 /// %D = select %cond, %C, %A 233 /// into: 234 /// %C = select %cond, %B, 0 235 /// %D = or %A, %C 236 /// 237 /// Assuming that the specified instruction is an operand to the select, return 238 /// a bitmask indicating which operands of this instruction are foldable if they 239 /// equal the other incoming value of the select. 240 static unsigned getSelectFoldableOperands(BinaryOperator *I) { 241 switch (I->getOpcode()) { 242 case Instruction::Add: 243 case Instruction::FAdd: 244 case Instruction::Mul: 245 case Instruction::FMul: 246 case Instruction::And: 247 case Instruction::Or: 248 case Instruction::Xor: 249 return 3; // Can fold through either operand. 250 case Instruction::Sub: // Can only fold on the amount subtracted. 251 case Instruction::FSub: 252 case Instruction::FDiv: // Can only fold on the divisor amount. 253 case Instruction::Shl: // Can only fold on the shift amount. 254 case Instruction::LShr: 255 case Instruction::AShr: 256 return 1; 257 default: 258 return 0; // Cannot fold 259 } 260 } 261 262 /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. 263 Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, 264 Instruction *FI) { 265 // Don't break up min/max patterns. The hasOneUse checks below prevent that 266 // for most cases, but vector min/max with bitcasts can be transformed. If the 267 // one-use restrictions are eased for other patterns, we still don't want to 268 // obfuscate min/max. 269 if ((match(&SI, m_SMin(m_Value(), m_Value())) || 270 match(&SI, m_SMax(m_Value(), m_Value())) || 271 match(&SI, m_UMin(m_Value(), m_Value())) || 272 match(&SI, m_UMax(m_Value(), m_Value())))) 273 return nullptr; 274 275 // If this is a cast from the same type, merge. 276 Value *Cond = SI.getCondition(); 277 Type *CondTy = Cond->getType(); 278 if (TI->getNumOperands() == 1 && TI->isCast()) { 279 Type *FIOpndTy = FI->getOperand(0)->getType(); 280 if (TI->getOperand(0)->getType() != FIOpndTy) 281 return nullptr; 282 283 // The select condition may be a vector. We may only change the operand 284 // type if the vector width remains the same (and matches the condition). 285 if (auto *CondVTy = dyn_cast<VectorType>(CondTy)) { 286 if (!FIOpndTy->isVectorTy() || 287 CondVTy->getElementCount() != 288 cast<VectorType>(FIOpndTy)->getElementCount()) 289 return nullptr; 290 291 // TODO: If the backend knew how to deal with casts better, we could 292 // remove this limitation. For now, there's too much potential to create 293 // worse codegen by promoting the select ahead of size-altering casts 294 // (PR28160). 295 // 296 // Note that ValueTracking's matchSelectPattern() looks through casts 297 // without checking 'hasOneUse' when it matches min/max patterns, so this 298 // transform may end up happening anyway. 299 if (TI->getOpcode() != Instruction::BitCast && 300 (!TI->hasOneUse() || !FI->hasOneUse())) 301 return nullptr; 302 } else if (!TI->hasOneUse() || !FI->hasOneUse()) { 303 // TODO: The one-use restrictions for a scalar select could be eased if 304 // the fold of a select in visitLoadInst() was enhanced to match a pattern 305 // that includes a cast. 306 return nullptr; 307 } 308 309 // Fold this by inserting a select from the input values. 310 Value *NewSI = 311 Builder.CreateSelect(Cond, TI->getOperand(0), FI->getOperand(0), 312 SI.getName() + ".v", &SI); 313 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI, 314 TI->getType()); 315 } 316 317 Value *OtherOpT, *OtherOpF; 318 bool MatchIsOpZero; 319 auto getCommonOp = [&](Instruction *TI, Instruction *FI, bool Commute, 320 bool Swapped = false) -> Value * { 321 assert(!(Commute && Swapped) && 322 "Commute and Swapped can't set at the same time"); 323 if (!Swapped) { 324 if (TI->getOperand(0) == FI->getOperand(0)) { 325 OtherOpT = TI->getOperand(1); 326 OtherOpF = FI->getOperand(1); 327 MatchIsOpZero = true; 328 return TI->getOperand(0); 329 } else if (TI->getOperand(1) == FI->getOperand(1)) { 330 OtherOpT = TI->getOperand(0); 331 OtherOpF = FI->getOperand(0); 332 MatchIsOpZero = false; 333 return TI->getOperand(1); 334 } 335 } 336 337 if (!Commute && !Swapped) 338 return nullptr; 339 340 // If we are allowing commute or swap of operands, then 341 // allow a cross-operand match. In that case, MatchIsOpZero 342 // means that TI's operand 0 (FI's operand 1) is the common op. 343 if (TI->getOperand(0) == FI->getOperand(1)) { 344 OtherOpT = TI->getOperand(1); 345 OtherOpF = FI->getOperand(0); 346 MatchIsOpZero = true; 347 return TI->getOperand(0); 348 } else if (TI->getOperand(1) == FI->getOperand(0)) { 349 OtherOpT = TI->getOperand(0); 350 OtherOpF = FI->getOperand(1); 351 MatchIsOpZero = false; 352 return TI->getOperand(1); 353 } 354 return nullptr; 355 }; 356 357 if (TI->hasOneUse() || FI->hasOneUse()) { 358 // Cond ? -X : -Y --> -(Cond ? X : Y) 359 Value *X, *Y; 360 if (match(TI, m_FNeg(m_Value(X))) && match(FI, m_FNeg(m_Value(Y)))) { 361 // Intersect FMF from the fneg instructions and union those with the 362 // select. 363 FastMathFlags FMF = TI->getFastMathFlags(); 364 FMF &= FI->getFastMathFlags(); 365 FMF |= SI.getFastMathFlags(); 366 Value *NewSel = 367 Builder.CreateSelect(Cond, X, Y, SI.getName() + ".v", &SI); 368 if (auto *NewSelI = dyn_cast<Instruction>(NewSel)) 369 NewSelI->setFastMathFlags(FMF); 370 Instruction *NewFNeg = UnaryOperator::CreateFNeg(NewSel); 371 NewFNeg->setFastMathFlags(FMF); 372 return NewFNeg; 373 } 374 375 // Min/max intrinsic with a common operand can have the common operand 376 // pulled after the select. This is the same transform as below for binops, 377 // but specialized for intrinsic matching and without the restrictive uses 378 // clause. 379 auto *TII = dyn_cast<IntrinsicInst>(TI); 380 auto *FII = dyn_cast<IntrinsicInst>(FI); 381 if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID()) { 382 if (match(TII, m_MaxOrMin(m_Value(), m_Value()))) { 383 if (Value *MatchOp = getCommonOp(TI, FI, true)) { 384 Value *NewSel = 385 Builder.CreateSelect(Cond, OtherOpT, OtherOpF, "minmaxop", &SI); 386 return CallInst::Create(TII->getCalledFunction(), {NewSel, MatchOp}); 387 } 388 } 389 } 390 391 // icmp with a common operand also can have the common operand 392 // pulled after the select. 393 ICmpInst::Predicate TPred, FPred; 394 if (match(TI, m_ICmp(TPred, m_Value(), m_Value())) && 395 match(FI, m_ICmp(FPred, m_Value(), m_Value()))) { 396 if (TPred == FPred || TPred == CmpInst::getSwappedPredicate(FPred)) { 397 bool Swapped = TPred != FPred; 398 if (Value *MatchOp = 399 getCommonOp(TI, FI, ICmpInst::isEquality(TPred), Swapped)) { 400 Value *NewSel = Builder.CreateSelect(Cond, OtherOpT, OtherOpF, 401 SI.getName() + ".v", &SI); 402 return new ICmpInst( 403 MatchIsOpZero ? TPred : CmpInst::getSwappedPredicate(TPred), 404 MatchOp, NewSel); 405 } 406 } 407 } 408 } 409 410 // Only handle binary operators (including two-operand getelementptr) with 411 // one-use here. As with the cast case above, it may be possible to relax the 412 // one-use constraint, but that needs be examined carefully since it may not 413 // reduce the total number of instructions. 414 if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || 415 !TI->isSameOperationAs(FI) || 416 (!isa<BinaryOperator>(TI) && !isa<GetElementPtrInst>(TI)) || 417 !TI->hasOneUse() || !FI->hasOneUse()) 418 return nullptr; 419 420 // Figure out if the operations have any operands in common. 421 Value *MatchOp = getCommonOp(TI, FI, TI->isCommutative()); 422 if (!MatchOp) 423 return nullptr; 424 425 // If the select condition is a vector, the operands of the original select's 426 // operands also must be vectors. This may not be the case for getelementptr 427 // for example. 428 if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || 429 !OtherOpF->getType()->isVectorTy())) 430 return nullptr; 431 432 // If we reach here, they do have operations in common. 433 Value *NewSI = Builder.CreateSelect(Cond, OtherOpT, OtherOpF, 434 SI.getName() + ".v", &SI); 435 Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; 436 Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; 437 if (auto *BO = dyn_cast<BinaryOperator>(TI)) { 438 BinaryOperator *NewBO = BinaryOperator::Create(BO->getOpcode(), Op0, Op1); 439 NewBO->copyIRFlags(TI); 440 NewBO->andIRFlags(FI); 441 return NewBO; 442 } 443 if (auto *TGEP = dyn_cast<GetElementPtrInst>(TI)) { 444 auto *FGEP = cast<GetElementPtrInst>(FI); 445 Type *ElementType = TGEP->getResultElementType(); 446 return TGEP->isInBounds() && FGEP->isInBounds() 447 ? GetElementPtrInst::CreateInBounds(ElementType, Op0, {Op1}) 448 : GetElementPtrInst::Create(ElementType, Op0, {Op1}); 449 } 450 llvm_unreachable("Expected BinaryOperator or GEP"); 451 return nullptr; 452 } 453 454 static bool isSelect01(const APInt &C1I, const APInt &C2I) { 455 if (!C1I.isZero() && !C2I.isZero()) // One side must be zero. 456 return false; 457 return C1I.isOne() || C1I.isAllOnes() || C2I.isOne() || C2I.isAllOnes(); 458 } 459 460 /// Try to fold the select into one of the operands to allow further 461 /// optimization. 462 Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, 463 Value *FalseVal) { 464 // See the comment above GetSelectFoldableOperands for a description of the 465 // transformation we are doing here. 466 auto TryFoldSelectIntoOp = [&](SelectInst &SI, Value *TrueVal, 467 Value *FalseVal, 468 bool Swapped) -> Instruction * { 469 auto *TVI = dyn_cast<BinaryOperator>(TrueVal); 470 if (!TVI || !TVI->hasOneUse() || isa<Constant>(FalseVal)) 471 return nullptr; 472 473 unsigned SFO = getSelectFoldableOperands(TVI); 474 unsigned OpToFold = 0; 475 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) 476 OpToFold = 1; 477 else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) 478 OpToFold = 2; 479 480 if (!OpToFold) 481 return nullptr; 482 483 // TODO: We probably ought to revisit cases where the select and FP 484 // instructions have different flags and add tests to ensure the 485 // behaviour is correct. 486 FastMathFlags FMF; 487 if (isa<FPMathOperator>(&SI)) 488 FMF = SI.getFastMathFlags(); 489 Constant *C = ConstantExpr::getBinOpIdentity( 490 TVI->getOpcode(), TVI->getType(), true, FMF.noSignedZeros()); 491 Value *OOp = TVI->getOperand(2 - OpToFold); 492 // Avoid creating select between 2 constants unless it's selecting 493 // between 0, 1 and -1. 494 const APInt *OOpC; 495 bool OOpIsAPInt = match(OOp, m_APInt(OOpC)); 496 if (!isa<Constant>(OOp) || 497 (OOpIsAPInt && isSelect01(C->getUniqueInteger(), *OOpC))) { 498 Value *NewSel = Builder.CreateSelect(SI.getCondition(), Swapped ? C : OOp, 499 Swapped ? OOp : C); 500 if (isa<FPMathOperator>(&SI)) 501 cast<Instruction>(NewSel)->setFastMathFlags(FMF); 502 NewSel->takeName(TVI); 503 BinaryOperator *BO = 504 BinaryOperator::Create(TVI->getOpcode(), FalseVal, NewSel); 505 BO->copyIRFlags(TVI); 506 return BO; 507 } 508 return nullptr; 509 }; 510 511 if (Instruction *R = TryFoldSelectIntoOp(SI, TrueVal, FalseVal, false)) 512 return R; 513 514 if (Instruction *R = TryFoldSelectIntoOp(SI, FalseVal, TrueVal, true)) 515 return R; 516 517 return nullptr; 518 } 519 520 /// We want to turn: 521 /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) 522 /// into: 523 /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) 524 /// Note: 525 /// Z may be 0 if lshr is missing. 526 /// Worst-case scenario is that we will replace 5 instructions with 5 different 527 /// instructions, but we got rid of select. 528 static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, 529 Value *TVal, Value *FVal, 530 InstCombiner::BuilderTy &Builder) { 531 if (!(Cmp->hasOneUse() && Cmp->getOperand(0)->hasOneUse() && 532 Cmp->getPredicate() == ICmpInst::ICMP_EQ && 533 match(Cmp->getOperand(1), m_Zero()) && match(FVal, m_One()))) 534 return nullptr; 535 536 // The TrueVal has general form of: and %B, 1 537 Value *B; 538 if (!match(TVal, m_OneUse(m_And(m_Value(B), m_One())))) 539 return nullptr; 540 541 // Where %B may be optionally shifted: lshr %X, %Z. 542 Value *X, *Z; 543 const bool HasShift = match(B, m_OneUse(m_LShr(m_Value(X), m_Value(Z)))); 544 545 // The shift must be valid. 546 // TODO: This restricts the fold to constant shift amounts. Is there a way to 547 // handle variable shifts safely? PR47012 548 if (HasShift && 549 !match(Z, m_SpecificInt_ICMP(CmpInst::ICMP_ULT, 550 APInt(SelType->getScalarSizeInBits(), 551 SelType->getScalarSizeInBits())))) 552 return nullptr; 553 554 if (!HasShift) 555 X = B; 556 557 Value *Y; 558 if (!match(Cmp->getOperand(0), m_c_And(m_Specific(X), m_Value(Y)))) 559 return nullptr; 560 561 // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 562 // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 563 Constant *One = ConstantInt::get(SelType, 1); 564 Value *MaskB = HasShift ? Builder.CreateShl(One, Z) : One; 565 Value *FullMask = Builder.CreateOr(Y, MaskB); 566 Value *MaskedX = Builder.CreateAnd(X, FullMask); 567 Value *ICmpNeZero = Builder.CreateIsNotNull(MaskedX); 568 return new ZExtInst(ICmpNeZero, SelType); 569 } 570 571 /// We want to turn: 572 /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 573 /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 574 /// into: 575 /// ashr (X, Y) 576 static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, 577 Value *FalseVal, 578 InstCombiner::BuilderTy &Builder) { 579 ICmpInst::Predicate Pred = IC->getPredicate(); 580 Value *CmpLHS = IC->getOperand(0); 581 Value *CmpRHS = IC->getOperand(1); 582 if (!CmpRHS->getType()->isIntOrIntVectorTy()) 583 return nullptr; 584 585 Value *X, *Y; 586 unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); 587 if ((Pred != ICmpInst::ICMP_SGT || 588 !match(CmpRHS, 589 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, -1)))) && 590 (Pred != ICmpInst::ICMP_SLT || 591 !match(CmpRHS, 592 m_SpecificInt_ICMP(ICmpInst::ICMP_SGE, APInt(Bitwidth, 0))))) 593 return nullptr; 594 595 // Canonicalize so that ashr is in FalseVal. 596 if (Pred == ICmpInst::ICMP_SLT) 597 std::swap(TrueVal, FalseVal); 598 599 if (match(TrueVal, m_LShr(m_Value(X), m_Value(Y))) && 600 match(FalseVal, m_AShr(m_Specific(X), m_Specific(Y))) && 601 match(CmpLHS, m_Specific(X))) { 602 const auto *Ashr = cast<Instruction>(FalseVal); 603 // if lshr is not exact and ashr is, this new ashr must not be exact. 604 bool IsExact = Ashr->isExact() && cast<Instruction>(TrueVal)->isExact(); 605 return Builder.CreateAShr(X, Y, IC->getName(), IsExact); 606 } 607 608 return nullptr; 609 } 610 611 /// We want to turn: 612 /// (select (icmp eq (and X, C1), 0), Y, (or Y, C2)) 613 /// into: 614 /// (or (shl (and X, C1), C3), Y) 615 /// iff: 616 /// C1 and C2 are both powers of 2 617 /// where: 618 /// C3 = Log(C2) - Log(C1) 619 /// 620 /// This transform handles cases where: 621 /// 1. The icmp predicate is inverted 622 /// 2. The select operands are reversed 623 /// 3. The magnitude of C2 and C1 are flipped 624 static Value *foldSelectICmpAndOr(const ICmpInst *IC, Value *TrueVal, 625 Value *FalseVal, 626 InstCombiner::BuilderTy &Builder) { 627 // Only handle integer compares. Also, if this is a vector select, we need a 628 // vector compare. 629 if (!TrueVal->getType()->isIntOrIntVectorTy() || 630 TrueVal->getType()->isVectorTy() != IC->getType()->isVectorTy()) 631 return nullptr; 632 633 Value *CmpLHS = IC->getOperand(0); 634 Value *CmpRHS = IC->getOperand(1); 635 636 Value *V; 637 unsigned C1Log; 638 bool IsEqualZero; 639 bool NeedAnd = false; 640 if (IC->isEquality()) { 641 if (!match(CmpRHS, m_Zero())) 642 return nullptr; 643 644 const APInt *C1; 645 if (!match(CmpLHS, m_And(m_Value(), m_Power2(C1)))) 646 return nullptr; 647 648 V = CmpLHS; 649 C1Log = C1->logBase2(); 650 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_EQ; 651 } else if (IC->getPredicate() == ICmpInst::ICMP_SLT || 652 IC->getPredicate() == ICmpInst::ICMP_SGT) { 653 // We also need to recognize (icmp slt (trunc (X)), 0) and 654 // (icmp sgt (trunc (X)), -1). 655 IsEqualZero = IC->getPredicate() == ICmpInst::ICMP_SGT; 656 if ((IsEqualZero && !match(CmpRHS, m_AllOnes())) || 657 (!IsEqualZero && !match(CmpRHS, m_Zero()))) 658 return nullptr; 659 660 if (!match(CmpLHS, m_OneUse(m_Trunc(m_Value(V))))) 661 return nullptr; 662 663 C1Log = CmpLHS->getType()->getScalarSizeInBits() - 1; 664 NeedAnd = true; 665 } else { 666 return nullptr; 667 } 668 669 const APInt *C2; 670 bool OrOnTrueVal = false; 671 bool OrOnFalseVal = match(FalseVal, m_Or(m_Specific(TrueVal), m_Power2(C2))); 672 if (!OrOnFalseVal) 673 OrOnTrueVal = match(TrueVal, m_Or(m_Specific(FalseVal), m_Power2(C2))); 674 675 if (!OrOnFalseVal && !OrOnTrueVal) 676 return nullptr; 677 678 Value *Y = OrOnFalseVal ? TrueVal : FalseVal; 679 680 unsigned C2Log = C2->logBase2(); 681 682 bool NeedXor = (!IsEqualZero && OrOnFalseVal) || (IsEqualZero && OrOnTrueVal); 683 bool NeedShift = C1Log != C2Log; 684 bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != 685 V->getType()->getScalarSizeInBits(); 686 687 // Make sure we don't create more instructions than we save. 688 Value *Or = OrOnFalseVal ? FalseVal : TrueVal; 689 if ((NeedShift + NeedXor + NeedZExtTrunc) > 690 (IC->hasOneUse() + Or->hasOneUse())) 691 return nullptr; 692 693 if (NeedAnd) { 694 // Insert the AND instruction on the input to the truncate. 695 APInt C1 = APInt::getOneBitSet(V->getType()->getScalarSizeInBits(), C1Log); 696 V = Builder.CreateAnd(V, ConstantInt::get(V->getType(), C1)); 697 } 698 699 if (C2Log > C1Log) { 700 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 701 V = Builder.CreateShl(V, C2Log - C1Log); 702 } else if (C1Log > C2Log) { 703 V = Builder.CreateLShr(V, C1Log - C2Log); 704 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 705 } else 706 V = Builder.CreateZExtOrTrunc(V, Y->getType()); 707 708 if (NeedXor) 709 V = Builder.CreateXor(V, *C2); 710 711 return Builder.CreateOr(V, Y); 712 } 713 714 /// Canonicalize a set or clear of a masked set of constant bits to 715 /// select-of-constants form. 716 static Instruction *foldSetClearBits(SelectInst &Sel, 717 InstCombiner::BuilderTy &Builder) { 718 Value *Cond = Sel.getCondition(); 719 Value *T = Sel.getTrueValue(); 720 Value *F = Sel.getFalseValue(); 721 Type *Ty = Sel.getType(); 722 Value *X; 723 const APInt *NotC, *C; 724 725 // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) 726 if (match(T, m_And(m_Value(X), m_APInt(NotC))) && 727 match(F, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 728 Constant *Zero = ConstantInt::getNullValue(Ty); 729 Constant *OrC = ConstantInt::get(Ty, *C); 730 Value *NewSel = Builder.CreateSelect(Cond, Zero, OrC, "masksel", &Sel); 731 return BinaryOperator::CreateOr(T, NewSel); 732 } 733 734 // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) 735 if (match(F, m_And(m_Value(X), m_APInt(NotC))) && 736 match(T, m_OneUse(m_Or(m_Specific(X), m_APInt(C)))) && *NotC == ~(*C)) { 737 Constant *Zero = ConstantInt::getNullValue(Ty); 738 Constant *OrC = ConstantInt::get(Ty, *C); 739 Value *NewSel = Builder.CreateSelect(Cond, OrC, Zero, "masksel", &Sel); 740 return BinaryOperator::CreateOr(F, NewSel); 741 } 742 743 return nullptr; 744 } 745 746 // select (x == 0), 0, x * y --> freeze(y) * x 747 // select (y == 0), 0, x * y --> freeze(x) * y 748 // select (x == 0), undef, x * y --> freeze(y) * x 749 // select (x == undef), 0, x * y --> freeze(y) * x 750 // Usage of mul instead of 0 will make the result more poisonous, 751 // so the operand that was not checked in the condition should be frozen. 752 // The latter folding is applied only when a constant compared with x is 753 // is a vector consisting of 0 and undefs. If a constant compared with x 754 // is a scalar undefined value or undefined vector then an expression 755 // should be already folded into a constant. 756 static Instruction *foldSelectZeroOrMul(SelectInst &SI, InstCombinerImpl &IC) { 757 auto *CondVal = SI.getCondition(); 758 auto *TrueVal = SI.getTrueValue(); 759 auto *FalseVal = SI.getFalseValue(); 760 Value *X, *Y; 761 ICmpInst::Predicate Predicate; 762 763 // Assuming that constant compared with zero is not undef (but it may be 764 // a vector with some undef elements). Otherwise (when a constant is undef) 765 // the select expression should be already simplified. 766 if (!match(CondVal, m_ICmp(Predicate, m_Value(X), m_Zero())) || 767 !ICmpInst::isEquality(Predicate)) 768 return nullptr; 769 770 if (Predicate == ICmpInst::ICMP_NE) 771 std::swap(TrueVal, FalseVal); 772 773 // Check that TrueVal is a constant instead of matching it with m_Zero() 774 // to handle the case when it is a scalar undef value or a vector containing 775 // non-zero elements that are masked by undef elements in the compare 776 // constant. 777 auto *TrueValC = dyn_cast<Constant>(TrueVal); 778 if (TrueValC == nullptr || 779 !match(FalseVal, m_c_Mul(m_Specific(X), m_Value(Y))) || 780 !isa<Instruction>(FalseVal)) 781 return nullptr; 782 783 auto *ZeroC = cast<Constant>(cast<Instruction>(CondVal)->getOperand(1)); 784 auto *MergedC = Constant::mergeUndefsWith(TrueValC, ZeroC); 785 // If X is compared with 0 then TrueVal could be either zero or undef. 786 // m_Zero match vectors containing some undef elements, but for scalars 787 // m_Undef should be used explicitly. 788 if (!match(MergedC, m_Zero()) && !match(MergedC, m_Undef())) 789 return nullptr; 790 791 auto *FalseValI = cast<Instruction>(FalseVal); 792 auto *FrY = IC.InsertNewInstBefore(new FreezeInst(Y, Y->getName() + ".fr"), 793 *FalseValI); 794 IC.replaceOperand(*FalseValI, FalseValI->getOperand(0) == Y ? 0 : 1, FrY); 795 return IC.replaceInstUsesWith(SI, FalseValI); 796 } 797 798 /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). 799 /// There are 8 commuted/swapped variants of this pattern. 800 /// TODO: Also support a - UMIN(a,b) patterns. 801 static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, 802 const Value *TrueVal, 803 const Value *FalseVal, 804 InstCombiner::BuilderTy &Builder) { 805 ICmpInst::Predicate Pred = ICI->getPredicate(); 806 Value *A = ICI->getOperand(0); 807 Value *B = ICI->getOperand(1); 808 809 // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 810 // (a == 0) ? 0 : a - 1 -> (a != 0) ? a - 1 : 0 811 if (match(TrueVal, m_Zero())) { 812 Pred = ICmpInst::getInversePredicate(Pred); 813 std::swap(TrueVal, FalseVal); 814 } 815 816 if (!match(FalseVal, m_Zero())) 817 return nullptr; 818 819 // ugt 0 is canonicalized to ne 0 and requires special handling 820 // (a != 0) ? a + -1 : 0 -> usub.sat(a, 1) 821 if (Pred == ICmpInst::ICMP_NE) { 822 if (match(B, m_Zero()) && match(TrueVal, m_Add(m_Specific(A), m_AllOnes()))) 823 return Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, 824 ConstantInt::get(A->getType(), 1)); 825 return nullptr; 826 } 827 828 if (!ICmpInst::isUnsigned(Pred)) 829 return nullptr; 830 831 if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { 832 // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 833 std::swap(A, B); 834 Pred = ICmpInst::getSwappedPredicate(Pred); 835 } 836 837 assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && 838 "Unexpected isUnsigned predicate!"); 839 840 // Ensure the sub is of the form: 841 // (a > b) ? a - b : 0 -> usub.sat(a, b) 842 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 843 // Checking for both a-b and a+(-b) as a constant. 844 bool IsNegative = false; 845 const APInt *C; 846 if (match(TrueVal, m_Sub(m_Specific(B), m_Specific(A))) || 847 (match(A, m_APInt(C)) && 848 match(TrueVal, m_Add(m_Specific(B), m_SpecificInt(-*C))))) 849 IsNegative = true; 850 else if (!match(TrueVal, m_Sub(m_Specific(A), m_Specific(B))) && 851 !(match(B, m_APInt(C)) && 852 match(TrueVal, m_Add(m_Specific(A), m_SpecificInt(-*C))))) 853 return nullptr; 854 855 // If we are adding a negate and the sub and icmp are used anywhere else, we 856 // would end up with more instructions. 857 if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) 858 return nullptr; 859 860 // (a > b) ? a - b : 0 -> usub.sat(a, b) 861 // (a > b) ? b - a : 0 -> -usub.sat(a, b) 862 Value *Result = Builder.CreateBinaryIntrinsic(Intrinsic::usub_sat, A, B); 863 if (IsNegative) 864 Result = Builder.CreateNeg(Result); 865 return Result; 866 } 867 868 static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, 869 InstCombiner::BuilderTy &Builder) { 870 if (!Cmp->hasOneUse()) 871 return nullptr; 872 873 // Match unsigned saturated add with constant. 874 Value *Cmp0 = Cmp->getOperand(0); 875 Value *Cmp1 = Cmp->getOperand(1); 876 ICmpInst::Predicate Pred = Cmp->getPredicate(); 877 Value *X; 878 const APInt *C, *CmpC; 879 if (Pred == ICmpInst::ICMP_ULT && 880 match(TVal, m_Add(m_Value(X), m_APInt(C))) && X == Cmp0 && 881 match(FVal, m_AllOnes()) && match(Cmp1, m_APInt(CmpC)) && *CmpC == ~*C) { 882 // (X u< ~C) ? (X + C) : -1 --> uadd.sat(X, C) 883 return Builder.CreateBinaryIntrinsic( 884 Intrinsic::uadd_sat, X, ConstantInt::get(X->getType(), *C)); 885 } 886 887 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 888 // There are 8 commuted variants. 889 // Canonicalize -1 (saturated result) to true value of the select. 890 if (match(FVal, m_AllOnes())) { 891 std::swap(TVal, FVal); 892 Pred = CmpInst::getInversePredicate(Pred); 893 } 894 if (!match(TVal, m_AllOnes())) 895 return nullptr; 896 897 // Canonicalize predicate to less-than or less-or-equal-than. 898 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { 899 std::swap(Cmp0, Cmp1); 900 Pred = CmpInst::getSwappedPredicate(Pred); 901 } 902 if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) 903 return nullptr; 904 905 // Match unsigned saturated add of 2 variables with an unnecessary 'not'. 906 // Strictness of the comparison is irrelevant. 907 Value *Y; 908 if (match(Cmp0, m_Not(m_Value(X))) && 909 match(FVal, m_c_Add(m_Specific(X), m_Value(Y))) && Y == Cmp1) { 910 // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 911 // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) 912 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, X, Y); 913 } 914 // The 'not' op may be included in the sum but not the compare. 915 // Strictness of the comparison is irrelevant. 916 X = Cmp0; 917 Y = Cmp1; 918 if (match(FVal, m_c_Add(m_Not(m_Specific(X)), m_Specific(Y)))) { 919 // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) 920 // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) 921 BinaryOperator *BO = cast<BinaryOperator>(FVal); 922 return Builder.CreateBinaryIntrinsic( 923 Intrinsic::uadd_sat, BO->getOperand(0), BO->getOperand(1)); 924 } 925 // The overflow may be detected via the add wrapping round. 926 // This is only valid for strict comparison! 927 if (Pred == ICmpInst::ICMP_ULT && 928 match(Cmp0, m_c_Add(m_Specific(Cmp1), m_Value(Y))) && 929 match(FVal, m_c_Add(m_Specific(Cmp1), m_Specific(Y)))) { 930 // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) 931 // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) 932 return Builder.CreateBinaryIntrinsic(Intrinsic::uadd_sat, Cmp1, Y); 933 } 934 935 return nullptr; 936 } 937 938 /// Fold the following code sequence: 939 /// \code 940 /// int a = ctlz(x & -x); 941 // x ? 31 - a : a; 942 /// \code 943 /// 944 /// into: 945 /// cttz(x) 946 static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, 947 Value *FalseVal, 948 InstCombiner::BuilderTy &Builder) { 949 unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); 950 if (!ICI->isEquality() || !match(ICI->getOperand(1), m_Zero())) 951 return nullptr; 952 953 if (ICI->getPredicate() == ICmpInst::ICMP_NE) 954 std::swap(TrueVal, FalseVal); 955 956 if (!match(FalseVal, 957 m_Xor(m_Deferred(TrueVal), m_SpecificInt(BitWidth - 1)))) 958 return nullptr; 959 960 if (!match(TrueVal, m_Intrinsic<Intrinsic::ctlz>())) 961 return nullptr; 962 963 Value *X = ICI->getOperand(0); 964 auto *II = cast<IntrinsicInst>(TrueVal); 965 if (!match(II->getOperand(0), m_c_And(m_Specific(X), m_Neg(m_Specific(X))))) 966 return nullptr; 967 968 Function *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::cttz, 969 II->getType()); 970 return CallInst::Create(F, {X, II->getArgOperand(1)}); 971 } 972 973 /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single 974 /// call to cttz/ctlz with flag 'is_zero_poison' cleared. 975 /// 976 /// For example, we can fold the following code sequence: 977 /// \code 978 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) 979 /// %1 = icmp ne i32 %x, 0 980 /// %2 = select i1 %1, i32 %0, i32 32 981 /// \code 982 /// 983 /// into: 984 /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) 985 static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, 986 InstCombiner::BuilderTy &Builder) { 987 ICmpInst::Predicate Pred = ICI->getPredicate(); 988 Value *CmpLHS = ICI->getOperand(0); 989 Value *CmpRHS = ICI->getOperand(1); 990 991 // Check if the select condition compares a value for equality. 992 if (!ICI->isEquality()) 993 return nullptr; 994 995 Value *SelectArg = FalseVal; 996 Value *ValueOnZero = TrueVal; 997 if (Pred == ICmpInst::ICMP_NE) 998 std::swap(SelectArg, ValueOnZero); 999 1000 // Skip zero extend/truncate. 1001 Value *Count = nullptr; 1002 if (!match(SelectArg, m_ZExt(m_Value(Count))) && 1003 !match(SelectArg, m_Trunc(m_Value(Count)))) 1004 Count = SelectArg; 1005 1006 // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the 1007 // input to the cttz/ctlz is used as LHS for the compare instruction. 1008 Value *X; 1009 if (!match(Count, m_Intrinsic<Intrinsic::cttz>(m_Value(X))) && 1010 !match(Count, m_Intrinsic<Intrinsic::ctlz>(m_Value(X)))) 1011 return nullptr; 1012 1013 // (X == 0) ? BitWidth : ctz(X) 1014 // (X == -1) ? BitWidth : ctz(~X) 1015 if ((X != CmpLHS || !match(CmpRHS, m_Zero())) && 1016 (!match(X, m_Not(m_Specific(CmpLHS))) || !match(CmpRHS, m_AllOnes()))) 1017 return nullptr; 1018 1019 IntrinsicInst *II = cast<IntrinsicInst>(Count); 1020 1021 // Check if the value propagated on zero is a constant number equal to the 1022 // sizeof in bits of 'Count'. 1023 unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); 1024 if (match(ValueOnZero, m_SpecificInt(SizeOfInBits))) { 1025 // Explicitly clear the 'is_zero_poison' flag. It's always valid to go from 1026 // true to false on this flag, so we can replace it for all users. 1027 II->setArgOperand(1, ConstantInt::getFalse(II->getContext())); 1028 return SelectArg; 1029 } 1030 1031 // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional 1032 // zext/trunc) have one use (ending at the select), the cttz/ctlz result will 1033 // not be used if the input is zero. Relax to 'zero is poison' for that case. 1034 if (II->hasOneUse() && SelectArg->hasOneUse() && 1035 !match(II->getArgOperand(1), m_One())) 1036 II->setArgOperand(1, ConstantInt::getTrue(II->getContext())); 1037 1038 return nullptr; 1039 } 1040 1041 /// Return true if we find and adjust an icmp+select pattern where the compare 1042 /// is with a constant that can be incremented or decremented to match the 1043 /// minimum or maximum idiom. 1044 static bool adjustMinMax(SelectInst &Sel, ICmpInst &Cmp) { 1045 ICmpInst::Predicate Pred = Cmp.getPredicate(); 1046 Value *CmpLHS = Cmp.getOperand(0); 1047 Value *CmpRHS = Cmp.getOperand(1); 1048 Value *TrueVal = Sel.getTrueValue(); 1049 Value *FalseVal = Sel.getFalseValue(); 1050 1051 // We may move or edit the compare, so make sure the select is the only user. 1052 const APInt *CmpC; 1053 if (!Cmp.hasOneUse() || !match(CmpRHS, m_APInt(CmpC))) 1054 return false; 1055 1056 // These transforms only work for selects of integers or vector selects of 1057 // integer vectors. 1058 Type *SelTy = Sel.getType(); 1059 auto *SelEltTy = dyn_cast<IntegerType>(SelTy->getScalarType()); 1060 if (!SelEltTy || SelTy->isVectorTy() != Cmp.getType()->isVectorTy()) 1061 return false; 1062 1063 Constant *AdjustedRHS; 1064 if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_SGT) 1065 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC + 1); 1066 else if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_SLT) 1067 AdjustedRHS = ConstantInt::get(CmpRHS->getType(), *CmpC - 1); 1068 else 1069 return false; 1070 1071 // X > C ? X : C+1 --> X < C+1 ? C+1 : X 1072 // X < C ? X : C-1 --> X > C-1 ? C-1 : X 1073 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) || 1074 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) { 1075 ; // Nothing to do here. Values match without any sign/zero extension. 1076 } 1077 // Types do not match. Instead of calculating this with mixed types, promote 1078 // all to the larger type. This enables scalar evolution to analyze this 1079 // expression. 1080 else if (CmpRHS->getType()->getScalarSizeInBits() < SelEltTy->getBitWidth()) { 1081 Constant *SextRHS = ConstantExpr::getSExt(AdjustedRHS, SelTy); 1082 1083 // X = sext x; x >s c ? X : C+1 --> X = sext x; X <s C+1 ? C+1 : X 1084 // X = sext x; x <s c ? X : C-1 --> X = sext x; X >s C-1 ? C-1 : X 1085 // X = sext x; x >u c ? X : C+1 --> X = sext x; X <u C+1 ? C+1 : X 1086 // X = sext x; x <u c ? X : C-1 --> X = sext x; X >u C-1 ? C-1 : X 1087 if (match(TrueVal, m_SExt(m_Specific(CmpLHS))) && SextRHS == FalseVal) { 1088 CmpLHS = TrueVal; 1089 AdjustedRHS = SextRHS; 1090 } else if (match(FalseVal, m_SExt(m_Specific(CmpLHS))) && 1091 SextRHS == TrueVal) { 1092 CmpLHS = FalseVal; 1093 AdjustedRHS = SextRHS; 1094 } else if (Cmp.isUnsigned()) { 1095 Constant *ZextRHS = ConstantExpr::getZExt(AdjustedRHS, SelTy); 1096 // X = zext x; x >u c ? X : C+1 --> X = zext x; X <u C+1 ? C+1 : X 1097 // X = zext x; x <u c ? X : C-1 --> X = zext x; X >u C-1 ? C-1 : X 1098 // zext + signed compare cannot be changed: 1099 // 0xff <s 0x00, but 0x00ff >s 0x0000 1100 if (match(TrueVal, m_ZExt(m_Specific(CmpLHS))) && ZextRHS == FalseVal) { 1101 CmpLHS = TrueVal; 1102 AdjustedRHS = ZextRHS; 1103 } else if (match(FalseVal, m_ZExt(m_Specific(CmpLHS))) && 1104 ZextRHS == TrueVal) { 1105 CmpLHS = FalseVal; 1106 AdjustedRHS = ZextRHS; 1107 } else { 1108 return false; 1109 } 1110 } else { 1111 return false; 1112 } 1113 } else { 1114 return false; 1115 } 1116 1117 Pred = ICmpInst::getSwappedPredicate(Pred); 1118 CmpRHS = AdjustedRHS; 1119 std::swap(FalseVal, TrueVal); 1120 Cmp.setPredicate(Pred); 1121 Cmp.setOperand(0, CmpLHS); 1122 Cmp.setOperand(1, CmpRHS); 1123 Sel.setOperand(1, TrueVal); 1124 Sel.setOperand(2, FalseVal); 1125 Sel.swapProfMetadata(); 1126 1127 // Move the compare instruction right before the select instruction. Otherwise 1128 // the sext/zext value may be defined after the compare instruction uses it. 1129 Cmp.moveBefore(&Sel); 1130 1131 return true; 1132 } 1133 1134 static Instruction *canonicalizeSPF(SelectInst &Sel, ICmpInst &Cmp, 1135 InstCombinerImpl &IC) { 1136 Value *LHS, *RHS; 1137 // TODO: What to do with pointer min/max patterns? 1138 if (!Sel.getType()->isIntOrIntVectorTy()) 1139 return nullptr; 1140 1141 SelectPatternFlavor SPF = matchSelectPattern(&Sel, LHS, RHS).Flavor; 1142 if (SPF == SelectPatternFlavor::SPF_ABS || 1143 SPF == SelectPatternFlavor::SPF_NABS) { 1144 if (!Cmp.hasOneUse() && !RHS->hasOneUse()) 1145 return nullptr; // TODO: Relax this restriction. 1146 1147 // Note that NSW flag can only be propagated for normal, non-negated abs! 1148 bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && 1149 match(RHS, m_NSWNeg(m_Specific(LHS))); 1150 Constant *IntMinIsPoisonC = 1151 ConstantInt::get(Type::getInt1Ty(Sel.getContext()), IntMinIsPoison); 1152 Instruction *Abs = 1153 IC.Builder.CreateBinaryIntrinsic(Intrinsic::abs, LHS, IntMinIsPoisonC); 1154 1155 if (SPF == SelectPatternFlavor::SPF_NABS) 1156 return BinaryOperator::CreateNeg(Abs); // Always without NSW flag! 1157 return IC.replaceInstUsesWith(Sel, Abs); 1158 } 1159 1160 if (SelectPatternResult::isMinOrMax(SPF)) { 1161 Intrinsic::ID IntrinsicID; 1162 switch (SPF) { 1163 case SelectPatternFlavor::SPF_UMIN: 1164 IntrinsicID = Intrinsic::umin; 1165 break; 1166 case SelectPatternFlavor::SPF_UMAX: 1167 IntrinsicID = Intrinsic::umax; 1168 break; 1169 case SelectPatternFlavor::SPF_SMIN: 1170 IntrinsicID = Intrinsic::smin; 1171 break; 1172 case SelectPatternFlavor::SPF_SMAX: 1173 IntrinsicID = Intrinsic::smax; 1174 break; 1175 default: 1176 llvm_unreachable("Unexpected SPF"); 1177 } 1178 return IC.replaceInstUsesWith( 1179 Sel, IC.Builder.CreateBinaryIntrinsic(IntrinsicID, LHS, RHS)); 1180 } 1181 1182 return nullptr; 1183 } 1184 1185 static bool replaceInInstruction(Value *V, Value *Old, Value *New, 1186 InstCombiner &IC, unsigned Depth = 0) { 1187 // Conservatively limit replacement to two instructions upwards. 1188 if (Depth == 2) 1189 return false; 1190 1191 auto *I = dyn_cast<Instruction>(V); 1192 if (!I || !I->hasOneUse() || !isSafeToSpeculativelyExecute(I)) 1193 return false; 1194 1195 bool Changed = false; 1196 for (Use &U : I->operands()) { 1197 if (U == Old) { 1198 IC.replaceUse(U, New); 1199 Changed = true; 1200 } else { 1201 Changed |= replaceInInstruction(U, Old, New, IC, Depth + 1); 1202 } 1203 } 1204 return Changed; 1205 } 1206 1207 /// If we have a select with an equality comparison, then we know the value in 1208 /// one of the arms of the select. See if substituting this value into an arm 1209 /// and simplifying the result yields the same value as the other arm. 1210 /// 1211 /// To make this transform safe, we must drop poison-generating flags 1212 /// (nsw, etc) if we simplified to a binop because the select may be guarding 1213 /// that poison from propagating. If the existing binop already had no 1214 /// poison-generating flags, then this transform can be done by instsimplify. 1215 /// 1216 /// Consider: 1217 /// %cmp = icmp eq i32 %x, 2147483647 1218 /// %add = add nsw i32 %x, 1 1219 /// %sel = select i1 %cmp, i32 -2147483648, i32 %add 1220 /// 1221 /// We can't replace %sel with %add unless we strip away the flags. 1222 /// TODO: Wrapping flags could be preserved in some cases with better analysis. 1223 Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, 1224 ICmpInst &Cmp) { 1225 if (!Cmp.isEquality()) 1226 return nullptr; 1227 1228 // Canonicalize the pattern to ICMP_EQ by swapping the select operands. 1229 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 1230 bool Swapped = false; 1231 if (Cmp.getPredicate() == ICmpInst::ICMP_NE) { 1232 std::swap(TrueVal, FalseVal); 1233 Swapped = true; 1234 } 1235 1236 // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. 1237 // Make sure Y cannot be undef though, as we might pick different values for 1238 // undef in the icmp and in f(Y). Additionally, take care to avoid replacing 1239 // X == Y ? X : Z with X == Y ? Y : Z, as that would lead to an infinite 1240 // replacement cycle. 1241 Value *CmpLHS = Cmp.getOperand(0), *CmpRHS = Cmp.getOperand(1); 1242 if (TrueVal != CmpLHS && 1243 isGuaranteedNotToBeUndefOrPoison(CmpRHS, SQ.AC, &Sel, &DT)) { 1244 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpLHS, CmpRHS, SQ, 1245 /* AllowRefinement */ true)) 1246 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1247 1248 // Even if TrueVal does not simplify, we can directly replace a use of 1249 // CmpLHS with CmpRHS, as long as the instruction is not used anywhere 1250 // else and is safe to speculatively execute (we may end up executing it 1251 // with different operands, which should not cause side-effects or trigger 1252 // undefined behavior). Only do this if CmpRHS is a constant, as 1253 // profitability is not clear for other cases. 1254 // FIXME: Support vectors. 1255 if (match(CmpRHS, m_ImmConstant()) && !match(CmpLHS, m_ImmConstant()) && 1256 !Cmp.getType()->isVectorTy()) 1257 if (replaceInInstruction(TrueVal, CmpLHS, CmpRHS, *this)) 1258 return &Sel; 1259 } 1260 if (TrueVal != CmpRHS && 1261 isGuaranteedNotToBeUndefOrPoison(CmpLHS, SQ.AC, &Sel, &DT)) 1262 if (Value *V = simplifyWithOpReplaced(TrueVal, CmpRHS, CmpLHS, SQ, 1263 /* AllowRefinement */ true)) 1264 return replaceOperand(Sel, Swapped ? 2 : 1, V); 1265 1266 auto *FalseInst = dyn_cast<Instruction>(FalseVal); 1267 if (!FalseInst) 1268 return nullptr; 1269 1270 // InstSimplify already performed this fold if it was possible subject to 1271 // current poison-generating flags. Try the transform again with 1272 // poison-generating flags temporarily dropped. 1273 bool WasNUW = false, WasNSW = false, WasExact = false, WasInBounds = false; 1274 if (auto *OBO = dyn_cast<OverflowingBinaryOperator>(FalseVal)) { 1275 WasNUW = OBO->hasNoUnsignedWrap(); 1276 WasNSW = OBO->hasNoSignedWrap(); 1277 FalseInst->setHasNoUnsignedWrap(false); 1278 FalseInst->setHasNoSignedWrap(false); 1279 } 1280 if (auto *PEO = dyn_cast<PossiblyExactOperator>(FalseVal)) { 1281 WasExact = PEO->isExact(); 1282 FalseInst->setIsExact(false); 1283 } 1284 if (auto *GEP = dyn_cast<GetElementPtrInst>(FalseVal)) { 1285 WasInBounds = GEP->isInBounds(); 1286 GEP->setIsInBounds(false); 1287 } 1288 1289 // Try each equivalence substitution possibility. 1290 // We have an 'EQ' comparison, so the select's false value will propagate. 1291 // Example: 1292 // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 1293 if (simplifyWithOpReplaced(FalseVal, CmpLHS, CmpRHS, SQ, 1294 /* AllowRefinement */ false) == TrueVal || 1295 simplifyWithOpReplaced(FalseVal, CmpRHS, CmpLHS, SQ, 1296 /* AllowRefinement */ false) == TrueVal) { 1297 return replaceInstUsesWith(Sel, FalseVal); 1298 } 1299 1300 // Restore poison-generating flags if the transform did not apply. 1301 if (WasNUW) 1302 FalseInst->setHasNoUnsignedWrap(); 1303 if (WasNSW) 1304 FalseInst->setHasNoSignedWrap(); 1305 if (WasExact) 1306 FalseInst->setIsExact(); 1307 if (WasInBounds) 1308 cast<GetElementPtrInst>(FalseInst)->setIsInBounds(); 1309 1310 return nullptr; 1311 } 1312 1313 // See if this is a pattern like: 1314 // %old_cmp1 = icmp slt i32 %x, C2 1315 // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high 1316 // %old_x_offseted = add i32 %x, C1 1317 // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 1318 // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement 1319 // This can be rewritten as more canonical pattern: 1320 // %new_cmp1 = icmp slt i32 %x, -C1 1321 // %new_cmp2 = icmp sge i32 %x, C0-C1 1322 // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x 1323 // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low 1324 // Iff -C1 s<= C2 s<= C0-C1 1325 // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) 1326 // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) 1327 static Value *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, 1328 InstCombiner::BuilderTy &Builder) { 1329 Value *X = Sel0.getTrueValue(); 1330 Value *Sel1 = Sel0.getFalseValue(); 1331 1332 // First match the condition of the outermost select. 1333 // Said condition must be one-use. 1334 if (!Cmp0.hasOneUse()) 1335 return nullptr; 1336 ICmpInst::Predicate Pred0 = Cmp0.getPredicate(); 1337 Value *Cmp00 = Cmp0.getOperand(0); 1338 Constant *C0; 1339 if (!match(Cmp0.getOperand(1), 1340 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))) 1341 return nullptr; 1342 1343 if (!isa<SelectInst>(Sel1)) { 1344 Pred0 = ICmpInst::getInversePredicate(Pred0); 1345 std::swap(X, Sel1); 1346 } 1347 1348 // Canonicalize Cmp0 into ult or uge. 1349 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1350 switch (Pred0) { 1351 case ICmpInst::Predicate::ICMP_ULT: 1352 case ICmpInst::Predicate::ICMP_UGE: 1353 // Although icmp ult %x, 0 is an unusual thing to try and should generally 1354 // have been simplified, it does not verify with undef inputs so ensure we 1355 // are not in a strange state. 1356 if (!match(C0, m_SpecificInt_ICMP( 1357 ICmpInst::Predicate::ICMP_NE, 1358 APInt::getZero(C0->getType()->getScalarSizeInBits())))) 1359 return nullptr; 1360 break; // Great! 1361 case ICmpInst::Predicate::ICMP_ULE: 1362 case ICmpInst::Predicate::ICMP_UGT: 1363 // We want to canonicalize it to 'ult' or 'uge', so we'll need to increment 1364 // C0, which again means it must not have any all-ones elements. 1365 if (!match(C0, 1366 m_SpecificInt_ICMP( 1367 ICmpInst::Predicate::ICMP_NE, 1368 APInt::getAllOnes(C0->getType()->getScalarSizeInBits())))) 1369 return nullptr; // Can't do, have all-ones element[s]. 1370 Pred0 = ICmpInst::getFlippedStrictnessPredicate(Pred0); 1371 C0 = InstCombiner::AddOne(C0); 1372 break; 1373 default: 1374 return nullptr; // Unknown predicate. 1375 } 1376 1377 // Now that we've canonicalized the ICmp, we know the X we expect; 1378 // the select in other hand should be one-use. 1379 if (!Sel1->hasOneUse()) 1380 return nullptr; 1381 1382 // If the types do not match, look through any truncs to the underlying 1383 // instruction. 1384 if (Cmp00->getType() != X->getType() && X->hasOneUse()) 1385 match(X, m_TruncOrSelf(m_Value(X))); 1386 1387 // We now can finish matching the condition of the outermost select: 1388 // it should either be the X itself, or an addition of some constant to X. 1389 Constant *C1; 1390 if (Cmp00 == X) 1391 C1 = ConstantInt::getNullValue(X->getType()); 1392 else if (!match(Cmp00, 1393 m_Add(m_Specific(X), 1394 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C1))))) 1395 return nullptr; 1396 1397 Value *Cmp1; 1398 ICmpInst::Predicate Pred1; 1399 Constant *C2; 1400 Value *ReplacementLow, *ReplacementHigh; 1401 if (!match(Sel1, m_Select(m_Value(Cmp1), m_Value(ReplacementLow), 1402 m_Value(ReplacementHigh))) || 1403 !match(Cmp1, 1404 m_ICmp(Pred1, m_Specific(X), 1405 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C2))))) 1406 return nullptr; 1407 1408 if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) 1409 return nullptr; // Not enough one-use instructions for the fold. 1410 // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of 1411 // two comparisons we'll need to build. 1412 1413 // Canonicalize Cmp1 into the form we expect. 1414 // FIXME: we shouldn't care about lanes that are 'undef' in the end? 1415 switch (Pred1) { 1416 case ICmpInst::Predicate::ICMP_SLT: 1417 break; 1418 case ICmpInst::Predicate::ICMP_SLE: 1419 // We'd have to increment C2 by one, and for that it must not have signed 1420 // max element, but then it would have been canonicalized to 'slt' before 1421 // we get here. So we can't do anything useful with 'sle'. 1422 return nullptr; 1423 case ICmpInst::Predicate::ICMP_SGT: 1424 // We want to canonicalize it to 'slt', so we'll need to increment C2, 1425 // which again means it must not have any signed max elements. 1426 if (!match(C2, 1427 m_SpecificInt_ICMP(ICmpInst::Predicate::ICMP_NE, 1428 APInt::getSignedMaxValue( 1429 C2->getType()->getScalarSizeInBits())))) 1430 return nullptr; // Can't do, have signed max element[s]. 1431 C2 = InstCombiner::AddOne(C2); 1432 [[fallthrough]]; 1433 case ICmpInst::Predicate::ICMP_SGE: 1434 // Also non-canonical, but here we don't need to change C2, 1435 // so we don't have any restrictions on C2, so we can just handle it. 1436 Pred1 = ICmpInst::Predicate::ICMP_SLT; 1437 std::swap(ReplacementLow, ReplacementHigh); 1438 break; 1439 default: 1440 return nullptr; // Unknown predicate. 1441 } 1442 assert(Pred1 == ICmpInst::Predicate::ICMP_SLT && 1443 "Unexpected predicate type."); 1444 1445 // The thresholds of this clamp-like pattern. 1446 auto *ThresholdLowIncl = ConstantExpr::getNeg(C1); 1447 auto *ThresholdHighExcl = ConstantExpr::getSub(C0, C1); 1448 1449 assert((Pred0 == ICmpInst::Predicate::ICMP_ULT || 1450 Pred0 == ICmpInst::Predicate::ICMP_UGE) && 1451 "Unexpected predicate type."); 1452 if (Pred0 == ICmpInst::Predicate::ICMP_UGE) 1453 std::swap(ThresholdLowIncl, ThresholdHighExcl); 1454 1455 // The fold has a precondition 1: C2 s>= ThresholdLow 1456 auto *Precond1 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SGE, C2, 1457 ThresholdLowIncl); 1458 if (!match(Precond1, m_One())) 1459 return nullptr; 1460 // The fold has a precondition 2: C2 s<= ThresholdHigh 1461 auto *Precond2 = ConstantExpr::getICmp(ICmpInst::Predicate::ICMP_SLE, C2, 1462 ThresholdHighExcl); 1463 if (!match(Precond2, m_One())) 1464 return nullptr; 1465 1466 // If we are matching from a truncated input, we need to sext the 1467 // ReplacementLow and ReplacementHigh values. Only do the transform if they 1468 // are free to extend due to being constants. 1469 if (X->getType() != Sel0.getType()) { 1470 Constant *LowC, *HighC; 1471 if (!match(ReplacementLow, m_ImmConstant(LowC)) || 1472 !match(ReplacementHigh, m_ImmConstant(HighC))) 1473 return nullptr; 1474 ReplacementLow = ConstantExpr::getSExt(LowC, X->getType()); 1475 ReplacementHigh = ConstantExpr::getSExt(HighC, X->getType()); 1476 } 1477 1478 // All good, finally emit the new pattern. 1479 Value *ShouldReplaceLow = Builder.CreateICmpSLT(X, ThresholdLowIncl); 1480 Value *ShouldReplaceHigh = Builder.CreateICmpSGE(X, ThresholdHighExcl); 1481 Value *MaybeReplacedLow = 1482 Builder.CreateSelect(ShouldReplaceLow, ReplacementLow, X); 1483 1484 // Create the final select. If we looked through a truncate above, we will 1485 // need to retruncate the result. 1486 Value *MaybeReplacedHigh = Builder.CreateSelect( 1487 ShouldReplaceHigh, ReplacementHigh, MaybeReplacedLow); 1488 return Builder.CreateTrunc(MaybeReplacedHigh, Sel0.getType()); 1489 } 1490 1491 // If we have 1492 // %cmp = icmp [canonical predicate] i32 %x, C0 1493 // %r = select i1 %cmp, i32 %y, i32 C1 1494 // Where C0 != C1 and %x may be different from %y, see if the constant that we 1495 // will have if we flip the strictness of the predicate (i.e. without changing 1496 // the result) is identical to the C1 in select. If it matches we can change 1497 // original comparison to one with swapped predicate, reuse the constant, 1498 // and swap the hands of select. 1499 static Instruction * 1500 tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, 1501 InstCombinerImpl &IC) { 1502 ICmpInst::Predicate Pred; 1503 Value *X; 1504 Constant *C0; 1505 if (!match(&Cmp, m_OneUse(m_ICmp( 1506 Pred, m_Value(X), 1507 m_CombineAnd(m_AnyIntegralConstant(), m_Constant(C0)))))) 1508 return nullptr; 1509 1510 // If comparison predicate is non-relational, we won't be able to do anything. 1511 if (ICmpInst::isEquality(Pred)) 1512 return nullptr; 1513 1514 // If comparison predicate is non-canonical, then we certainly won't be able 1515 // to make it canonical; canonicalizeCmpWithConstant() already tried. 1516 if (!InstCombiner::isCanonicalPredicate(Pred)) 1517 return nullptr; 1518 1519 // If the [input] type of comparison and select type are different, lets abort 1520 // for now. We could try to compare constants with trunc/[zs]ext though. 1521 if (C0->getType() != Sel.getType()) 1522 return nullptr; 1523 1524 // ULT with 'add' of a constant is canonical. See foldICmpAddConstant(). 1525 // FIXME: Are there more magic icmp predicate+constant pairs we must avoid? 1526 // Or should we just abandon this transform entirely? 1527 if (Pred == CmpInst::ICMP_ULT && match(X, m_Add(m_Value(), m_Constant()))) 1528 return nullptr; 1529 1530 1531 Value *SelVal0, *SelVal1; // We do not care which one is from where. 1532 match(&Sel, m_Select(m_Value(), m_Value(SelVal0), m_Value(SelVal1))); 1533 // At least one of these values we are selecting between must be a constant 1534 // else we'll never succeed. 1535 if (!match(SelVal0, m_AnyIntegralConstant()) && 1536 !match(SelVal1, m_AnyIntegralConstant())) 1537 return nullptr; 1538 1539 // Does this constant C match any of the `select` values? 1540 auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { 1541 return C->isElementWiseEqual(SelVal0) || C->isElementWiseEqual(SelVal1); 1542 }; 1543 1544 // If C0 *already* matches true/false value of select, we are done. 1545 if (MatchesSelectValue(C0)) 1546 return nullptr; 1547 1548 // Check the constant we'd have with flipped-strictness predicate. 1549 auto FlippedStrictness = 1550 InstCombiner::getFlippedStrictnessPredicateAndConstant(Pred, C0); 1551 if (!FlippedStrictness) 1552 return nullptr; 1553 1554 // If said constant doesn't match either, then there is no hope, 1555 if (!MatchesSelectValue(FlippedStrictness->second)) 1556 return nullptr; 1557 1558 // It matched! Lets insert the new comparison just before select. 1559 InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); 1560 IC.Builder.SetInsertPoint(&Sel); 1561 1562 Pred = ICmpInst::getSwappedPredicate(Pred); // Yes, swapped. 1563 Value *NewCmp = IC.Builder.CreateICmp(Pred, X, FlippedStrictness->second, 1564 Cmp.getName() + ".inv"); 1565 IC.replaceOperand(Sel, 0, NewCmp); 1566 Sel.swapValues(); 1567 Sel.swapProfMetadata(); 1568 1569 return &Sel; 1570 } 1571 1572 static Instruction *foldSelectZeroOrOnes(ICmpInst *Cmp, Value *TVal, 1573 Value *FVal, 1574 InstCombiner::BuilderTy &Builder) { 1575 if (!Cmp->hasOneUse()) 1576 return nullptr; 1577 1578 const APInt *CmpC; 1579 if (!match(Cmp->getOperand(1), m_APIntAllowUndef(CmpC))) 1580 return nullptr; 1581 1582 // (X u< 2) ? -X : -1 --> sext (X != 0) 1583 Value *X = Cmp->getOperand(0); 1584 if (Cmp->getPredicate() == ICmpInst::ICMP_ULT && *CmpC == 2 && 1585 match(TVal, m_Neg(m_Specific(X))) && match(FVal, m_AllOnes())) 1586 return new SExtInst(Builder.CreateIsNotNull(X), TVal->getType()); 1587 1588 // (X u> 1) ? -1 : -X --> sext (X != 0) 1589 if (Cmp->getPredicate() == ICmpInst::ICMP_UGT && *CmpC == 1 && 1590 match(FVal, m_Neg(m_Specific(X))) && match(TVal, m_AllOnes())) 1591 return new SExtInst(Builder.CreateIsNotNull(X), TVal->getType()); 1592 1593 return nullptr; 1594 } 1595 1596 static Value *foldSelectInstWithICmpConst(SelectInst &SI, ICmpInst *ICI) { 1597 const APInt *CmpC; 1598 Value *V; 1599 CmpInst::Predicate Pred; 1600 if (!match(ICI, m_ICmp(Pred, m_Value(V), m_APInt(CmpC)))) 1601 return nullptr; 1602 1603 BinaryOperator *BO; 1604 const APInt *C; 1605 CmpInst::Predicate CPred; 1606 if (match(&SI, m_Select(m_Specific(ICI), m_APInt(C), m_BinOp(BO)))) 1607 CPred = ICI->getPredicate(); 1608 else if (match(&SI, m_Select(m_Specific(ICI), m_BinOp(BO), m_APInt(C)))) 1609 CPred = ICI->getInversePredicate(); 1610 else 1611 return nullptr; 1612 1613 const APInt *BinOpC; 1614 if (!match(BO, m_BinOp(m_Specific(V), m_APInt(BinOpC)))) 1615 return nullptr; 1616 1617 ConstantRange R = ConstantRange::makeExactICmpRegion(CPred, *CmpC) 1618 .binaryOp(BO->getOpcode(), *BinOpC); 1619 if (R == *C) { 1620 BO->dropPoisonGeneratingFlags(); 1621 return BO; 1622 } 1623 return nullptr; 1624 } 1625 1626 /// Visit a SelectInst that has an ICmpInst as its first operand. 1627 Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, 1628 ICmpInst *ICI) { 1629 if (Instruction *NewSel = foldSelectValueEquivalence(SI, *ICI)) 1630 return NewSel; 1631 1632 if (Instruction *NewSPF = canonicalizeSPF(SI, *ICI, *this)) 1633 return NewSPF; 1634 1635 if (Value *V = foldSelectInstWithICmpConst(SI, ICI)) 1636 return replaceInstUsesWith(SI, V); 1637 1638 if (Value *V = canonicalizeClampLike(SI, *ICI, Builder)) 1639 return replaceInstUsesWith(SI, V); 1640 1641 if (Instruction *NewSel = 1642 tryToReuseConstantFromSelectInComparison(SI, *ICI, *this)) 1643 return NewSel; 1644 1645 bool Changed = adjustMinMax(SI, *ICI); 1646 1647 if (Value *V = foldSelectICmpAnd(SI, ICI, Builder)) 1648 return replaceInstUsesWith(SI, V); 1649 1650 // NOTE: if we wanted to, this is where to detect integer MIN/MAX 1651 Value *TrueVal = SI.getTrueValue(); 1652 Value *FalseVal = SI.getFalseValue(); 1653 ICmpInst::Predicate Pred = ICI->getPredicate(); 1654 Value *CmpLHS = ICI->getOperand(0); 1655 Value *CmpRHS = ICI->getOperand(1); 1656 if (CmpRHS != CmpLHS && isa<Constant>(CmpRHS)) { 1657 if (CmpLHS == TrueVal && Pred == ICmpInst::ICMP_EQ) { 1658 // Transform (X == C) ? X : Y -> (X == C) ? C : Y 1659 SI.setOperand(1, CmpRHS); 1660 Changed = true; 1661 } else if (CmpLHS == FalseVal && Pred == ICmpInst::ICMP_NE) { 1662 // Transform (X != C) ? Y : X -> (X != C) ? Y : C 1663 SI.setOperand(2, CmpRHS); 1664 Changed = true; 1665 } 1666 } 1667 1668 // Canonicalize a signbit condition to use zero constant by swapping: 1669 // (CmpLHS > -1) ? TV : FV --> (CmpLHS < 0) ? FV : TV 1670 // To avoid conflicts (infinite loops) with other canonicalizations, this is 1671 // not applied with any constant select arm. 1672 if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes()) && 1673 !match(TrueVal, m_Constant()) && !match(FalseVal, m_Constant()) && 1674 ICI->hasOneUse()) { 1675 InstCombiner::BuilderTy::InsertPointGuard Guard(Builder); 1676 Builder.SetInsertPoint(&SI); 1677 Value *IsNeg = Builder.CreateIsNeg(CmpLHS, ICI->getName()); 1678 replaceOperand(SI, 0, IsNeg); 1679 SI.swapValues(); 1680 SI.swapProfMetadata(); 1681 return &SI; 1682 } 1683 1684 // FIXME: This code is nearly duplicated in InstSimplify. Using/refactoring 1685 // decomposeBitTestICmp() might help. 1686 { 1687 unsigned BitWidth = 1688 DL.getTypeSizeInBits(TrueVal->getType()->getScalarType()); 1689 APInt MinSignedValue = APInt::getSignedMinValue(BitWidth); 1690 Value *X; 1691 const APInt *Y, *C; 1692 bool TrueWhenUnset; 1693 bool IsBitTest = false; 1694 if (ICmpInst::isEquality(Pred) && 1695 match(CmpLHS, m_And(m_Value(X), m_Power2(Y))) && 1696 match(CmpRHS, m_Zero())) { 1697 IsBitTest = true; 1698 TrueWhenUnset = Pred == ICmpInst::ICMP_EQ; 1699 } else if (Pred == ICmpInst::ICMP_SLT && match(CmpRHS, m_Zero())) { 1700 X = CmpLHS; 1701 Y = &MinSignedValue; 1702 IsBitTest = true; 1703 TrueWhenUnset = false; 1704 } else if (Pred == ICmpInst::ICMP_SGT && match(CmpRHS, m_AllOnes())) { 1705 X = CmpLHS; 1706 Y = &MinSignedValue; 1707 IsBitTest = true; 1708 TrueWhenUnset = true; 1709 } 1710 if (IsBitTest) { 1711 Value *V = nullptr; 1712 // (X & Y) == 0 ? X : X ^ Y --> X & ~Y 1713 if (TrueWhenUnset && TrueVal == X && 1714 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1715 V = Builder.CreateAnd(X, ~(*Y)); 1716 // (X & Y) != 0 ? X ^ Y : X --> X & ~Y 1717 else if (!TrueWhenUnset && FalseVal == X && 1718 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1719 V = Builder.CreateAnd(X, ~(*Y)); 1720 // (X & Y) == 0 ? X ^ Y : X --> X | Y 1721 else if (TrueWhenUnset && FalseVal == X && 1722 match(TrueVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1723 V = Builder.CreateOr(X, *Y); 1724 // (X & Y) != 0 ? X : X ^ Y --> X | Y 1725 else if (!TrueWhenUnset && TrueVal == X && 1726 match(FalseVal, m_Xor(m_Specific(X), m_APInt(C))) && *Y == *C) 1727 V = Builder.CreateOr(X, *Y); 1728 1729 if (V) 1730 return replaceInstUsesWith(SI, V); 1731 } 1732 } 1733 1734 if (Instruction *V = 1735 foldSelectICmpAndAnd(SI.getType(), ICI, TrueVal, FalseVal, Builder)) 1736 return V; 1737 1738 if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) 1739 return V; 1740 1741 if (Instruction *V = foldSelectZeroOrOnes(ICI, TrueVal, FalseVal, Builder)) 1742 return V; 1743 1744 if (Value *V = foldSelectICmpAndOr(ICI, TrueVal, FalseVal, Builder)) 1745 return replaceInstUsesWith(SI, V); 1746 1747 if (Value *V = foldSelectICmpLshrAshr(ICI, TrueVal, FalseVal, Builder)) 1748 return replaceInstUsesWith(SI, V); 1749 1750 if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, Builder)) 1751 return replaceInstUsesWith(SI, V); 1752 1753 if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) 1754 return replaceInstUsesWith(SI, V); 1755 1756 if (Value *V = canonicalizeSaturatedAdd(ICI, TrueVal, FalseVal, Builder)) 1757 return replaceInstUsesWith(SI, V); 1758 1759 return Changed ? &SI : nullptr; 1760 } 1761 1762 /// SI is a select whose condition is a PHI node (but the two may be in 1763 /// different blocks). See if the true/false values (V) are live in all of the 1764 /// predecessor blocks of the PHI. For example, cases like this can't be mapped: 1765 /// 1766 /// X = phi [ C1, BB1], [C2, BB2] 1767 /// Y = add 1768 /// Z = select X, Y, 0 1769 /// 1770 /// because Y is not live in BB1/BB2. 1771 static bool canSelectOperandBeMappingIntoPredBlock(const Value *V, 1772 const SelectInst &SI) { 1773 // If the value is a non-instruction value like a constant or argument, it 1774 // can always be mapped. 1775 const Instruction *I = dyn_cast<Instruction>(V); 1776 if (!I) return true; 1777 1778 // If V is a PHI node defined in the same block as the condition PHI, we can 1779 // map the arguments. 1780 const PHINode *CondPHI = cast<PHINode>(SI.getCondition()); 1781 1782 if (const PHINode *VP = dyn_cast<PHINode>(I)) 1783 if (VP->getParent() == CondPHI->getParent()) 1784 return true; 1785 1786 // Otherwise, if the PHI and select are defined in the same block and if V is 1787 // defined in a different block, then we can transform it. 1788 if (SI.getParent() == CondPHI->getParent() && 1789 I->getParent() != CondPHI->getParent()) 1790 return true; 1791 1792 // Otherwise we have a 'hard' case and we can't tell without doing more 1793 // detailed dominator based analysis, punt. 1794 return false; 1795 } 1796 1797 /// We have an SPF (e.g. a min or max) of an SPF of the form: 1798 /// SPF2(SPF1(A, B), C) 1799 Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, 1800 SelectPatternFlavor SPF1, Value *A, 1801 Value *B, Instruction &Outer, 1802 SelectPatternFlavor SPF2, 1803 Value *C) { 1804 if (Outer.getType() != Inner->getType()) 1805 return nullptr; 1806 1807 if (C == A || C == B) { 1808 // MAX(MAX(A, B), B) -> MAX(A, B) 1809 // MIN(MIN(a, b), a) -> MIN(a, b) 1810 // TODO: This could be done in instsimplify. 1811 if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF1)) 1812 return replaceInstUsesWith(Outer, Inner); 1813 } 1814 1815 return nullptr; 1816 } 1817 1818 /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). 1819 /// This is even legal for FP. 1820 static Instruction *foldAddSubSelect(SelectInst &SI, 1821 InstCombiner::BuilderTy &Builder) { 1822 Value *CondVal = SI.getCondition(); 1823 Value *TrueVal = SI.getTrueValue(); 1824 Value *FalseVal = SI.getFalseValue(); 1825 auto *TI = dyn_cast<Instruction>(TrueVal); 1826 auto *FI = dyn_cast<Instruction>(FalseVal); 1827 if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) 1828 return nullptr; 1829 1830 Instruction *AddOp = nullptr, *SubOp = nullptr; 1831 if ((TI->getOpcode() == Instruction::Sub && 1832 FI->getOpcode() == Instruction::Add) || 1833 (TI->getOpcode() == Instruction::FSub && 1834 FI->getOpcode() == Instruction::FAdd)) { 1835 AddOp = FI; 1836 SubOp = TI; 1837 } else if ((FI->getOpcode() == Instruction::Sub && 1838 TI->getOpcode() == Instruction::Add) || 1839 (FI->getOpcode() == Instruction::FSub && 1840 TI->getOpcode() == Instruction::FAdd)) { 1841 AddOp = TI; 1842 SubOp = FI; 1843 } 1844 1845 if (AddOp) { 1846 Value *OtherAddOp = nullptr; 1847 if (SubOp->getOperand(0) == AddOp->getOperand(0)) { 1848 OtherAddOp = AddOp->getOperand(1); 1849 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) { 1850 OtherAddOp = AddOp->getOperand(0); 1851 } 1852 1853 if (OtherAddOp) { 1854 // So at this point we know we have (Y -> OtherAddOp): 1855 // select C, (add X, Y), (sub X, Z) 1856 Value *NegVal; // Compute -Z 1857 if (SI.getType()->isFPOrFPVectorTy()) { 1858 NegVal = Builder.CreateFNeg(SubOp->getOperand(1)); 1859 if (Instruction *NegInst = dyn_cast<Instruction>(NegVal)) { 1860 FastMathFlags Flags = AddOp->getFastMathFlags(); 1861 Flags &= SubOp->getFastMathFlags(); 1862 NegInst->setFastMathFlags(Flags); 1863 } 1864 } else { 1865 NegVal = Builder.CreateNeg(SubOp->getOperand(1)); 1866 } 1867 1868 Value *NewTrueOp = OtherAddOp; 1869 Value *NewFalseOp = NegVal; 1870 if (AddOp != TI) 1871 std::swap(NewTrueOp, NewFalseOp); 1872 Value *NewSel = Builder.CreateSelect(CondVal, NewTrueOp, NewFalseOp, 1873 SI.getName() + ".p", &SI); 1874 1875 if (SI.getType()->isFPOrFPVectorTy()) { 1876 Instruction *RI = 1877 BinaryOperator::CreateFAdd(SubOp->getOperand(0), NewSel); 1878 1879 FastMathFlags Flags = AddOp->getFastMathFlags(); 1880 Flags &= SubOp->getFastMathFlags(); 1881 RI->setFastMathFlags(Flags); 1882 return RI; 1883 } else 1884 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel); 1885 } 1886 } 1887 return nullptr; 1888 } 1889 1890 /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1891 /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1892 /// Along with a number of patterns similar to: 1893 /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1894 /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1895 static Instruction * 1896 foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { 1897 Value *CondVal = SI.getCondition(); 1898 Value *TrueVal = SI.getTrueValue(); 1899 Value *FalseVal = SI.getFalseValue(); 1900 1901 WithOverflowInst *II; 1902 if (!match(CondVal, m_ExtractValue<1>(m_WithOverflowInst(II))) || 1903 !match(FalseVal, m_ExtractValue<0>(m_Specific(II)))) 1904 return nullptr; 1905 1906 Value *X = II->getLHS(); 1907 Value *Y = II->getRHS(); 1908 1909 auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { 1910 Type *Ty = Limit->getType(); 1911 1912 ICmpInst::Predicate Pred; 1913 Value *TrueVal, *FalseVal, *Op; 1914 const APInt *C; 1915 if (!match(Limit, m_Select(m_ICmp(Pred, m_Value(Op), m_APInt(C)), 1916 m_Value(TrueVal), m_Value(FalseVal)))) 1917 return false; 1918 1919 auto IsZeroOrOne = [](const APInt &C) { return C.isZero() || C.isOne(); }; 1920 auto IsMinMax = [&](Value *Min, Value *Max) { 1921 APInt MinVal = APInt::getSignedMinValue(Ty->getScalarSizeInBits()); 1922 APInt MaxVal = APInt::getSignedMaxValue(Ty->getScalarSizeInBits()); 1923 return match(Min, m_SpecificInt(MinVal)) && 1924 match(Max, m_SpecificInt(MaxVal)); 1925 }; 1926 1927 if (Op != X && Op != Y) 1928 return false; 1929 1930 if (IsAdd) { 1931 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1932 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1933 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1934 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1935 if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1936 IsMinMax(TrueVal, FalseVal)) 1937 return true; 1938 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1939 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1940 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1941 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1942 if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1943 IsMinMax(FalseVal, TrueVal)) 1944 return true; 1945 } else { 1946 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1947 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1948 if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && 1949 IsMinMax(TrueVal, FalseVal)) 1950 return true; 1951 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1952 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1953 if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && 1954 IsMinMax(FalseVal, TrueVal)) 1955 return true; 1956 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1957 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1958 if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && 1959 IsMinMax(FalseVal, TrueVal)) 1960 return true; 1961 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1962 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1963 if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && 1964 IsMinMax(TrueVal, FalseVal)) 1965 return true; 1966 } 1967 1968 return false; 1969 }; 1970 1971 Intrinsic::ID NewIntrinsicID; 1972 if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && 1973 match(TrueVal, m_AllOnes())) 1974 // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y 1975 NewIntrinsicID = Intrinsic::uadd_sat; 1976 else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && 1977 match(TrueVal, m_Zero())) 1978 // X - Y overflows ? 0 : X - Y -> usub_sat X, Y 1979 NewIntrinsicID = Intrinsic::usub_sat; 1980 else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && 1981 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) 1982 // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1983 // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1984 // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1985 // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1986 // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1987 // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y 1988 // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1989 // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y 1990 NewIntrinsicID = Intrinsic::sadd_sat; 1991 else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && 1992 IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) 1993 // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1994 // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 1995 // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1996 // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1997 // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1998 // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y 1999 // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 2000 // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y 2001 NewIntrinsicID = Intrinsic::ssub_sat; 2002 else 2003 return nullptr; 2004 2005 Function *F = 2006 Intrinsic::getDeclaration(SI.getModule(), NewIntrinsicID, SI.getType()); 2007 return CallInst::Create(F, {X, Y}); 2008 } 2009 2010 Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { 2011 Constant *C; 2012 if (!match(Sel.getTrueValue(), m_Constant(C)) && 2013 !match(Sel.getFalseValue(), m_Constant(C))) 2014 return nullptr; 2015 2016 Instruction *ExtInst; 2017 if (!match(Sel.getTrueValue(), m_Instruction(ExtInst)) && 2018 !match(Sel.getFalseValue(), m_Instruction(ExtInst))) 2019 return nullptr; 2020 2021 auto ExtOpcode = ExtInst->getOpcode(); 2022 if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) 2023 return nullptr; 2024 2025 // If we are extending from a boolean type or if we can create a select that 2026 // has the same size operands as its condition, try to narrow the select. 2027 Value *X = ExtInst->getOperand(0); 2028 Type *SmallType = X->getType(); 2029 Value *Cond = Sel.getCondition(); 2030 auto *Cmp = dyn_cast<CmpInst>(Cond); 2031 if (!SmallType->isIntOrIntVectorTy(1) && 2032 (!Cmp || Cmp->getOperand(0)->getType() != SmallType)) 2033 return nullptr; 2034 2035 // If the constant is the same after truncation to the smaller type and 2036 // extension to the original type, we can narrow the select. 2037 Type *SelType = Sel.getType(); 2038 Constant *TruncC = ConstantExpr::getTrunc(C, SmallType); 2039 Constant *ExtC = ConstantExpr::getCast(ExtOpcode, TruncC, SelType); 2040 if (ExtC == C && ExtInst->hasOneUse()) { 2041 Value *TruncCVal = cast<Value>(TruncC); 2042 if (ExtInst == Sel.getFalseValue()) 2043 std::swap(X, TruncCVal); 2044 2045 // select Cond, (ext X), C --> ext(select Cond, X, C') 2046 // select Cond, C, (ext X) --> ext(select Cond, C', X) 2047 Value *NewSel = Builder.CreateSelect(Cond, X, TruncCVal, "narrow", &Sel); 2048 return CastInst::Create(Instruction::CastOps(ExtOpcode), NewSel, SelType); 2049 } 2050 2051 // If one arm of the select is the extend of the condition, replace that arm 2052 // with the extension of the appropriate known bool value. 2053 if (Cond == X) { 2054 if (ExtInst == Sel.getTrueValue()) { 2055 // select X, (sext X), C --> select X, -1, C 2056 // select X, (zext X), C --> select X, 1, C 2057 Constant *One = ConstantInt::getTrue(SmallType); 2058 Constant *AllOnesOrOne = ConstantExpr::getCast(ExtOpcode, One, SelType); 2059 return SelectInst::Create(Cond, AllOnesOrOne, C, "", nullptr, &Sel); 2060 } else { 2061 // select X, C, (sext X) --> select X, C, 0 2062 // select X, C, (zext X) --> select X, C, 0 2063 Constant *Zero = ConstantInt::getNullValue(SelType); 2064 return SelectInst::Create(Cond, C, Zero, "", nullptr, &Sel); 2065 } 2066 } 2067 2068 return nullptr; 2069 } 2070 2071 /// Try to transform a vector select with a constant condition vector into a 2072 /// shuffle for easier combining with other shuffles and insert/extract. 2073 static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { 2074 Value *CondVal = SI.getCondition(); 2075 Constant *CondC; 2076 auto *CondValTy = dyn_cast<FixedVectorType>(CondVal->getType()); 2077 if (!CondValTy || !match(CondVal, m_Constant(CondC))) 2078 return nullptr; 2079 2080 unsigned NumElts = CondValTy->getNumElements(); 2081 SmallVector<int, 16> Mask; 2082 Mask.reserve(NumElts); 2083 for (unsigned i = 0; i != NumElts; ++i) { 2084 Constant *Elt = CondC->getAggregateElement(i); 2085 if (!Elt) 2086 return nullptr; 2087 2088 if (Elt->isOneValue()) { 2089 // If the select condition element is true, choose from the 1st vector. 2090 Mask.push_back(i); 2091 } else if (Elt->isNullValue()) { 2092 // If the select condition element is false, choose from the 2nd vector. 2093 Mask.push_back(i + NumElts); 2094 } else if (isa<UndefValue>(Elt)) { 2095 // Undef in a select condition (choose one of the operands) does not mean 2096 // the same thing as undef in a shuffle mask (any value is acceptable), so 2097 // give up. 2098 return nullptr; 2099 } else { 2100 // Bail out on a constant expression. 2101 return nullptr; 2102 } 2103 } 2104 2105 return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); 2106 } 2107 2108 /// If we have a select of vectors with a scalar condition, try to convert that 2109 /// to a vector select by splatting the condition. A splat may get folded with 2110 /// other operations in IR and having all operands of a select be vector types 2111 /// is likely better for vector codegen. 2112 static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, 2113 InstCombinerImpl &IC) { 2114 auto *Ty = dyn_cast<VectorType>(Sel.getType()); 2115 if (!Ty) 2116 return nullptr; 2117 2118 // We can replace a single-use extract with constant index. 2119 Value *Cond = Sel.getCondition(); 2120 if (!match(Cond, m_OneUse(m_ExtractElt(m_Value(), m_ConstantInt())))) 2121 return nullptr; 2122 2123 // select (extelt V, Index), T, F --> select (splat V, Index), T, F 2124 // Splatting the extracted condition reduces code (we could directly create a 2125 // splat shuffle of the source vector to eliminate the intermediate step). 2126 return IC.replaceOperand( 2127 Sel, 0, IC.Builder.CreateVectorSplat(Ty->getElementCount(), Cond)); 2128 } 2129 2130 /// Reuse bitcasted operands between a compare and select: 2131 /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2132 /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) 2133 static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, 2134 InstCombiner::BuilderTy &Builder) { 2135 Value *Cond = Sel.getCondition(); 2136 Value *TVal = Sel.getTrueValue(); 2137 Value *FVal = Sel.getFalseValue(); 2138 2139 CmpInst::Predicate Pred; 2140 Value *A, *B; 2141 if (!match(Cond, m_Cmp(Pred, m_Value(A), m_Value(B)))) 2142 return nullptr; 2143 2144 // The select condition is a compare instruction. If the select's true/false 2145 // values are already the same as the compare operands, there's nothing to do. 2146 if (TVal == A || TVal == B || FVal == A || FVal == B) 2147 return nullptr; 2148 2149 Value *C, *D; 2150 if (!match(A, m_BitCast(m_Value(C))) || !match(B, m_BitCast(m_Value(D)))) 2151 return nullptr; 2152 2153 // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) 2154 Value *TSrc, *FSrc; 2155 if (!match(TVal, m_BitCast(m_Value(TSrc))) || 2156 !match(FVal, m_BitCast(m_Value(FSrc)))) 2157 return nullptr; 2158 2159 // If the select true/false values are *different bitcasts* of the same source 2160 // operands, make the select operands the same as the compare operands and 2161 // cast the result. This is the canonical select form for min/max. 2162 Value *NewSel; 2163 if (TSrc == C && FSrc == D) { 2164 // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> 2165 // bitcast (select (cmp A, B), A, B) 2166 NewSel = Builder.CreateSelect(Cond, A, B, "", &Sel); 2167 } else if (TSrc == D && FSrc == C) { 2168 // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> 2169 // bitcast (select (cmp A, B), B, A) 2170 NewSel = Builder.CreateSelect(Cond, B, A, "", &Sel); 2171 } else { 2172 return nullptr; 2173 } 2174 return CastInst::CreateBitOrPointerCast(NewSel, Sel.getType()); 2175 } 2176 2177 /// Try to eliminate select instructions that test the returned flag of cmpxchg 2178 /// instructions. 2179 /// 2180 /// If a select instruction tests the returned flag of a cmpxchg instruction and 2181 /// selects between the returned value of the cmpxchg instruction its compare 2182 /// operand, the result of the select will always be equal to its false value. 2183 /// For example: 2184 /// 2185 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2186 /// %1 = extractvalue { i64, i1 } %0, 1 2187 /// %2 = extractvalue { i64, i1 } %0, 0 2188 /// %3 = select i1 %1, i64 %compare, i64 %2 2189 /// ret i64 %3 2190 /// 2191 /// The returned value of the cmpxchg instruction (%2) is the original value 2192 /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %2 2193 /// must have been equal to %compare. Thus, the result of the select is always 2194 /// equal to %2, and the code can be simplified to: 2195 /// 2196 /// %0 = cmpxchg i64* %ptr, i64 %compare, i64 %new_value seq_cst seq_cst 2197 /// %1 = extractvalue { i64, i1 } %0, 0 2198 /// ret i64 %1 2199 /// 2200 static Value *foldSelectCmpXchg(SelectInst &SI) { 2201 // A helper that determines if V is an extractvalue instruction whose 2202 // aggregate operand is a cmpxchg instruction and whose single index is equal 2203 // to I. If such conditions are true, the helper returns the cmpxchg 2204 // instruction; otherwise, a nullptr is returned. 2205 auto isExtractFromCmpXchg = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { 2206 auto *Extract = dyn_cast<ExtractValueInst>(V); 2207 if (!Extract) 2208 return nullptr; 2209 if (Extract->getIndices()[0] != I) 2210 return nullptr; 2211 return dyn_cast<AtomicCmpXchgInst>(Extract->getAggregateOperand()); 2212 }; 2213 2214 // If the select has a single user, and this user is a select instruction that 2215 // we can simplify, skip the cmpxchg simplification for now. 2216 if (SI.hasOneUse()) 2217 if (auto *Select = dyn_cast<SelectInst>(SI.user_back())) 2218 if (Select->getCondition() == SI.getCondition()) 2219 if (Select->getFalseValue() == SI.getTrueValue() || 2220 Select->getTrueValue() == SI.getFalseValue()) 2221 return nullptr; 2222 2223 // Ensure the select condition is the returned flag of a cmpxchg instruction. 2224 auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); 2225 if (!CmpXchg) 2226 return nullptr; 2227 2228 // Check the true value case: The true value of the select is the returned 2229 // value of the same cmpxchg used by the condition, and the false value is the 2230 // cmpxchg instruction's compare operand. 2231 if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) 2232 if (X == CmpXchg && X->getCompareOperand() == SI.getFalseValue()) 2233 return SI.getFalseValue(); 2234 2235 // Check the false value case: The false value of the select is the returned 2236 // value of the same cmpxchg used by the condition, and the true value is the 2237 // cmpxchg instruction's compare operand. 2238 if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) 2239 if (X == CmpXchg && X->getCompareOperand() == SI.getTrueValue()) 2240 return SI.getFalseValue(); 2241 2242 return nullptr; 2243 } 2244 2245 /// Try to reduce a funnel/rotate pattern that includes a compare and select 2246 /// into a funnel shift intrinsic. Example: 2247 /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) 2248 /// --> call llvm.fshl.i32(a, a, b) 2249 /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) 2250 /// --> call llvm.fshl.i32(a, b, c) 2251 /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) 2252 /// --> call llvm.fshr.i32(a, b, c) 2253 static Instruction *foldSelectFunnelShift(SelectInst &Sel, 2254 InstCombiner::BuilderTy &Builder) { 2255 // This must be a power-of-2 type for a bitmasking transform to be valid. 2256 unsigned Width = Sel.getType()->getScalarSizeInBits(); 2257 if (!isPowerOf2_32(Width)) 2258 return nullptr; 2259 2260 BinaryOperator *Or0, *Or1; 2261 if (!match(Sel.getFalseValue(), m_OneUse(m_Or(m_BinOp(Or0), m_BinOp(Or1))))) 2262 return nullptr; 2263 2264 Value *SV0, *SV1, *SA0, *SA1; 2265 if (!match(Or0, m_OneUse(m_LogicalShift(m_Value(SV0), 2266 m_ZExtOrSelf(m_Value(SA0))))) || 2267 !match(Or1, m_OneUse(m_LogicalShift(m_Value(SV1), 2268 m_ZExtOrSelf(m_Value(SA1))))) || 2269 Or0->getOpcode() == Or1->getOpcode()) 2270 return nullptr; 2271 2272 // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). 2273 if (Or0->getOpcode() == BinaryOperator::LShr) { 2274 std::swap(Or0, Or1); 2275 std::swap(SV0, SV1); 2276 std::swap(SA0, SA1); 2277 } 2278 assert(Or0->getOpcode() == BinaryOperator::Shl && 2279 Or1->getOpcode() == BinaryOperator::LShr && 2280 "Illegal or(shift,shift) pair"); 2281 2282 // Check the shift amounts to see if they are an opposite pair. 2283 Value *ShAmt; 2284 if (match(SA1, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA0))))) 2285 ShAmt = SA0; 2286 else if (match(SA0, m_OneUse(m_Sub(m_SpecificInt(Width), m_Specific(SA1))))) 2287 ShAmt = SA1; 2288 else 2289 return nullptr; 2290 2291 // We should now have this pattern: 2292 // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) 2293 // The false value of the select must be a funnel-shift of the true value: 2294 // IsFShl -> TVal must be SV0 else TVal must be SV1. 2295 bool IsFshl = (ShAmt == SA0); 2296 Value *TVal = Sel.getTrueValue(); 2297 if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) 2298 return nullptr; 2299 2300 // Finally, see if the select is filtering out a shift-by-zero. 2301 Value *Cond = Sel.getCondition(); 2302 ICmpInst::Predicate Pred; 2303 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_Specific(ShAmt), m_ZeroInt()))) || 2304 Pred != ICmpInst::ICMP_EQ) 2305 return nullptr; 2306 2307 // If this is not a rotate then the select was blocking poison from the 2308 // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. 2309 if (SV0 != SV1) { 2310 if (IsFshl && !llvm::isGuaranteedNotToBePoison(SV1)) 2311 SV1 = Builder.CreateFreeze(SV1); 2312 else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(SV0)) 2313 SV0 = Builder.CreateFreeze(SV0); 2314 } 2315 2316 // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. 2317 // Convert to funnel shift intrinsic. 2318 Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; 2319 Function *F = Intrinsic::getDeclaration(Sel.getModule(), IID, Sel.getType()); 2320 ShAmt = Builder.CreateZExt(ShAmt, Sel.getType()); 2321 return CallInst::Create(F, { SV0, SV1, ShAmt }); 2322 } 2323 2324 static Instruction *foldSelectToCopysign(SelectInst &Sel, 2325 InstCombiner::BuilderTy &Builder) { 2326 Value *Cond = Sel.getCondition(); 2327 Value *TVal = Sel.getTrueValue(); 2328 Value *FVal = Sel.getFalseValue(); 2329 Type *SelType = Sel.getType(); 2330 2331 // Match select ?, TC, FC where the constants are equal but negated. 2332 // TODO: Generalize to handle a negated variable operand? 2333 const APFloat *TC, *FC; 2334 if (!match(TVal, m_APFloatAllowUndef(TC)) || 2335 !match(FVal, m_APFloatAllowUndef(FC)) || 2336 !abs(*TC).bitwiseIsEqual(abs(*FC))) 2337 return nullptr; 2338 2339 assert(TC != FC && "Expected equal select arms to simplify"); 2340 2341 Value *X; 2342 const APInt *C; 2343 bool IsTrueIfSignSet; 2344 ICmpInst::Predicate Pred; 2345 if (!match(Cond, m_OneUse(m_ICmp(Pred, m_BitCast(m_Value(X)), m_APInt(C)))) || 2346 !InstCombiner::isSignBitCheck(Pred, *C, IsTrueIfSignSet) || 2347 X->getType() != SelType) 2348 return nullptr; 2349 2350 // If needed, negate the value that will be the sign argument of the copysign: 2351 // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) 2352 // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) 2353 // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) 2354 // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) 2355 // Note: FMF from the select can not be propagated to the new instructions. 2356 if (IsTrueIfSignSet ^ TC->isNegative()) 2357 X = Builder.CreateFNeg(X); 2358 2359 // Canonicalize the magnitude argument as the positive constant since we do 2360 // not care about its sign. 2361 Value *MagArg = ConstantFP::get(SelType, abs(*TC)); 2362 Function *F = Intrinsic::getDeclaration(Sel.getModule(), Intrinsic::copysign, 2363 Sel.getType()); 2364 return CallInst::Create(F, { MagArg, X }); 2365 } 2366 2367 Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { 2368 if (!isa<VectorType>(Sel.getType())) 2369 return nullptr; 2370 2371 Value *Cond = Sel.getCondition(); 2372 Value *TVal = Sel.getTrueValue(); 2373 Value *FVal = Sel.getFalseValue(); 2374 Value *C, *X, *Y; 2375 2376 if (match(Cond, m_VecReverse(m_Value(C)))) { 2377 auto createSelReverse = [&](Value *C, Value *X, Value *Y) { 2378 Value *V = Builder.CreateSelect(C, X, Y, Sel.getName(), &Sel); 2379 if (auto *I = dyn_cast<Instruction>(V)) 2380 I->copyIRFlags(&Sel); 2381 Module *M = Sel.getModule(); 2382 Function *F = Intrinsic::getDeclaration( 2383 M, Intrinsic::experimental_vector_reverse, V->getType()); 2384 return CallInst::Create(F, V); 2385 }; 2386 2387 if (match(TVal, m_VecReverse(m_Value(X)))) { 2388 // select rev(C), rev(X), rev(Y) --> rev(select C, X, Y) 2389 if (match(FVal, m_VecReverse(m_Value(Y))) && 2390 (Cond->hasOneUse() || TVal->hasOneUse() || FVal->hasOneUse())) 2391 return createSelReverse(C, X, Y); 2392 2393 // select rev(C), rev(X), FValSplat --> rev(select C, X, FValSplat) 2394 if ((Cond->hasOneUse() || TVal->hasOneUse()) && isSplatValue(FVal)) 2395 return createSelReverse(C, X, FVal); 2396 } 2397 // select rev(C), TValSplat, rev(Y) --> rev(select C, TValSplat, Y) 2398 else if (isSplatValue(TVal) && match(FVal, m_VecReverse(m_Value(Y))) && 2399 (Cond->hasOneUse() || FVal->hasOneUse())) 2400 return createSelReverse(C, TVal, Y); 2401 } 2402 2403 auto *VecTy = dyn_cast<FixedVectorType>(Sel.getType()); 2404 if (!VecTy) 2405 return nullptr; 2406 2407 unsigned NumElts = VecTy->getNumElements(); 2408 APInt UndefElts(NumElts, 0); 2409 APInt AllOnesEltMask(APInt::getAllOnes(NumElts)); 2410 if (Value *V = SimplifyDemandedVectorElts(&Sel, AllOnesEltMask, UndefElts)) { 2411 if (V != &Sel) 2412 return replaceInstUsesWith(Sel, V); 2413 return &Sel; 2414 } 2415 2416 // A select of a "select shuffle" with a common operand can be rearranged 2417 // to select followed by "select shuffle". Because of poison, this only works 2418 // in the case of a shuffle with no undefined mask elements. 2419 ArrayRef<int> Mask; 2420 if (match(TVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2421 !is_contained(Mask, UndefMaskElem) && 2422 cast<ShuffleVectorInst>(TVal)->isSelect()) { 2423 if (X == FVal) { 2424 // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) 2425 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2426 return new ShuffleVectorInst(X, NewSel, Mask); 2427 } 2428 if (Y == FVal) { 2429 // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y 2430 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2431 return new ShuffleVectorInst(NewSel, Y, Mask); 2432 } 2433 } 2434 if (match(FVal, m_OneUse(m_Shuffle(m_Value(X), m_Value(Y), m_Mask(Mask)))) && 2435 !is_contained(Mask, UndefMaskElem) && 2436 cast<ShuffleVectorInst>(FVal)->isSelect()) { 2437 if (X == TVal) { 2438 // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) 2439 Value *NewSel = Builder.CreateSelect(Cond, X, Y, "sel", &Sel); 2440 return new ShuffleVectorInst(X, NewSel, Mask); 2441 } 2442 if (Y == TVal) { 2443 // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y 2444 Value *NewSel = Builder.CreateSelect(Cond, Y, X, "sel", &Sel); 2445 return new ShuffleVectorInst(NewSel, Y, Mask); 2446 } 2447 } 2448 2449 return nullptr; 2450 } 2451 2452 static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, 2453 const DominatorTree &DT, 2454 InstCombiner::BuilderTy &Builder) { 2455 // Find the block's immediate dominator that ends with a conditional branch 2456 // that matches select's condition (maybe inverted). 2457 auto *IDomNode = DT[BB]->getIDom(); 2458 if (!IDomNode) 2459 return nullptr; 2460 BasicBlock *IDom = IDomNode->getBlock(); 2461 2462 Value *Cond = Sel.getCondition(); 2463 Value *IfTrue, *IfFalse; 2464 BasicBlock *TrueSucc, *FalseSucc; 2465 if (match(IDom->getTerminator(), 2466 m_Br(m_Specific(Cond), m_BasicBlock(TrueSucc), 2467 m_BasicBlock(FalseSucc)))) { 2468 IfTrue = Sel.getTrueValue(); 2469 IfFalse = Sel.getFalseValue(); 2470 } else if (match(IDom->getTerminator(), 2471 m_Br(m_Not(m_Specific(Cond)), m_BasicBlock(TrueSucc), 2472 m_BasicBlock(FalseSucc)))) { 2473 IfTrue = Sel.getFalseValue(); 2474 IfFalse = Sel.getTrueValue(); 2475 } else 2476 return nullptr; 2477 2478 // Make sure the branches are actually different. 2479 if (TrueSucc == FalseSucc) 2480 return nullptr; 2481 2482 // We want to replace select %cond, %a, %b with a phi that takes value %a 2483 // for all incoming edges that are dominated by condition `%cond == true`, 2484 // and value %b for edges dominated by condition `%cond == false`. If %a 2485 // or %b are also phis from the same basic block, we can go further and take 2486 // their incoming values from the corresponding blocks. 2487 BasicBlockEdge TrueEdge(IDom, TrueSucc); 2488 BasicBlockEdge FalseEdge(IDom, FalseSucc); 2489 DenseMap<BasicBlock *, Value *> Inputs; 2490 for (auto *Pred : predecessors(BB)) { 2491 // Check implication. 2492 BasicBlockEdge Incoming(Pred, BB); 2493 if (DT.dominates(TrueEdge, Incoming)) 2494 Inputs[Pred] = IfTrue->DoPHITranslation(BB, Pred); 2495 else if (DT.dominates(FalseEdge, Incoming)) 2496 Inputs[Pred] = IfFalse->DoPHITranslation(BB, Pred); 2497 else 2498 return nullptr; 2499 // Check availability. 2500 if (auto *Insn = dyn_cast<Instruction>(Inputs[Pred])) 2501 if (!DT.dominates(Insn, Pred->getTerminator())) 2502 return nullptr; 2503 } 2504 2505 Builder.SetInsertPoint(&*BB->begin()); 2506 auto *PN = Builder.CreatePHI(Sel.getType(), Inputs.size()); 2507 for (auto *Pred : predecessors(BB)) 2508 PN->addIncoming(Inputs[Pred], Pred); 2509 PN->takeName(&Sel); 2510 return PN; 2511 } 2512 2513 static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, 2514 InstCombiner::BuilderTy &Builder) { 2515 // Try to replace this select with Phi in one of these blocks. 2516 SmallSetVector<BasicBlock *, 4> CandidateBlocks; 2517 CandidateBlocks.insert(Sel.getParent()); 2518 for (Value *V : Sel.operands()) 2519 if (auto *I = dyn_cast<Instruction>(V)) 2520 CandidateBlocks.insert(I->getParent()); 2521 2522 for (BasicBlock *BB : CandidateBlocks) 2523 if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) 2524 return PN; 2525 return nullptr; 2526 } 2527 2528 static Value *foldSelectWithFrozenICmp(SelectInst &Sel, InstCombiner::BuilderTy &Builder) { 2529 FreezeInst *FI = dyn_cast<FreezeInst>(Sel.getCondition()); 2530 if (!FI) 2531 return nullptr; 2532 2533 Value *Cond = FI->getOperand(0); 2534 Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); 2535 2536 // select (freeze(x == y)), x, y --> y 2537 // select (freeze(x != y)), x, y --> x 2538 // The freeze should be only used by this select. Otherwise, remaining uses of 2539 // the freeze can observe a contradictory value. 2540 // c = freeze(x == y) ; Let's assume that y = poison & x = 42; c is 0 or 1 2541 // a = select c, x, y ; 2542 // f(a, c) ; f(poison, 1) cannot happen, but if a is folded 2543 // ; to y, this can happen. 2544 CmpInst::Predicate Pred; 2545 if (FI->hasOneUse() && 2546 match(Cond, m_c_ICmp(Pred, m_Specific(TrueVal), m_Specific(FalseVal))) && 2547 (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)) { 2548 return Pred == ICmpInst::ICMP_EQ ? FalseVal : TrueVal; 2549 } 2550 2551 return nullptr; 2552 } 2553 2554 Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op, 2555 SelectInst &SI, 2556 bool IsAnd) { 2557 Value *CondVal = SI.getCondition(); 2558 Value *A = SI.getTrueValue(); 2559 Value *B = SI.getFalseValue(); 2560 2561 assert(Op->getType()->isIntOrIntVectorTy(1) && 2562 "Op must be either i1 or vector of i1."); 2563 2564 std::optional<bool> Res = isImpliedCondition(Op, CondVal, DL, IsAnd); 2565 if (!Res) 2566 return nullptr; 2567 2568 Value *Zero = Constant::getNullValue(A->getType()); 2569 Value *One = Constant::getAllOnesValue(A->getType()); 2570 2571 if (*Res == true) { 2572 if (IsAnd) 2573 // select op, (select cond, A, B), false => select op, A, false 2574 // and op, (select cond, A, B) => select op, A, false 2575 // if op = true implies condval = true. 2576 return SelectInst::Create(Op, A, Zero); 2577 else 2578 // select op, true, (select cond, A, B) => select op, true, A 2579 // or op, (select cond, A, B) => select op, true, A 2580 // if op = false implies condval = true. 2581 return SelectInst::Create(Op, One, A); 2582 } else { 2583 if (IsAnd) 2584 // select op, (select cond, A, B), false => select op, B, false 2585 // and op, (select cond, A, B) => select op, B, false 2586 // if op = true implies condval = false. 2587 return SelectInst::Create(Op, B, Zero); 2588 else 2589 // select op, true, (select cond, A, B) => select op, true, B 2590 // or op, (select cond, A, B) => select op, true, B 2591 // if op = false implies condval = false. 2592 return SelectInst::Create(Op, One, B); 2593 } 2594 } 2595 2596 // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need 2597 // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. 2598 static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI, 2599 InstCombinerImpl &IC) { 2600 Value *CondVal = SI.getCondition(); 2601 2602 bool ChangedFMF = false; 2603 for (bool Swap : {false, true}) { 2604 Value *TrueVal = SI.getTrueValue(); 2605 Value *X = SI.getFalseValue(); 2606 CmpInst::Predicate Pred; 2607 2608 if (Swap) 2609 std::swap(TrueVal, X); 2610 2611 if (!match(CondVal, m_FCmp(Pred, m_Specific(X), m_AnyZeroFP()))) 2612 continue; 2613 2614 // fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false 2615 // fold (X > +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true 2616 if (match(TrueVal, m_FSub(m_PosZeroFP(), m_Specific(X)))) { 2617 if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { 2618 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI); 2619 return IC.replaceInstUsesWith(SI, Fabs); 2620 } 2621 if (Swap && (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { 2622 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI); 2623 return IC.replaceInstUsesWith(SI, Fabs); 2624 } 2625 } 2626 2627 if (!match(TrueVal, m_FNeg(m_Specific(X)))) 2628 return nullptr; 2629 2630 // Forward-propagate nnan and ninf from the fneg to the select. 2631 // If all inputs are not those values, then the select is not either. 2632 // Note: nsz is defined differently, so it may not be correct to propagate. 2633 FastMathFlags FMF = cast<FPMathOperator>(TrueVal)->getFastMathFlags(); 2634 if (FMF.noNaNs() && !SI.hasNoNaNs()) { 2635 SI.setHasNoNaNs(true); 2636 ChangedFMF = true; 2637 } 2638 if (FMF.noInfs() && !SI.hasNoInfs()) { 2639 SI.setHasNoInfs(true); 2640 ChangedFMF = true; 2641 } 2642 2643 // With nsz, when 'Swap' is false: 2644 // fold (X < +/-0.0) ? -X : X or (X <= +/-0.0) ? -X : X to fabs(X) 2645 // fold (X > +/-0.0) ? -X : X or (X >= +/-0.0) ? -X : X to -fabs(x) 2646 // when 'Swap' is true: 2647 // fold (X > +/-0.0) ? X : -X or (X >= +/-0.0) ? X : -X to fabs(X) 2648 // fold (X < +/-0.0) ? X : -X or (X <= +/-0.0) ? X : -X to -fabs(X) 2649 // 2650 // Note: We require "nnan" for this fold because fcmp ignores the signbit 2651 // of NAN, but IEEE-754 specifies the signbit of NAN values with 2652 // fneg/fabs operations. 2653 if (!SI.hasNoSignedZeros() || !SI.hasNoNaNs()) 2654 return nullptr; 2655 2656 if (Swap) 2657 Pred = FCmpInst::getSwappedPredicate(Pred); 2658 2659 bool IsLTOrLE = Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || 2660 Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE; 2661 bool IsGTOrGE = Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || 2662 Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE; 2663 2664 if (IsLTOrLE) { 2665 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI); 2666 return IC.replaceInstUsesWith(SI, Fabs); 2667 } 2668 if (IsGTOrGE) { 2669 Value *Fabs = IC.Builder.CreateUnaryIntrinsic(Intrinsic::fabs, X, &SI); 2670 Instruction *NewFNeg = UnaryOperator::CreateFNeg(Fabs); 2671 NewFNeg->setFastMathFlags(SI.getFastMathFlags()); 2672 return NewFNeg; 2673 } 2674 } 2675 2676 return ChangedFMF ? &SI : nullptr; 2677 } 2678 2679 // Match the following IR pattern: 2680 // %x.lowbits = and i8 %x, %lowbitmask 2681 // %x.lowbits.are.zero = icmp eq i8 %x.lowbits, 0 2682 // %x.biased = add i8 %x, %bias 2683 // %x.biased.highbits = and i8 %x.biased, %highbitmask 2684 // %x.roundedup = select i1 %x.lowbits.are.zero, i8 %x, i8 %x.biased.highbits 2685 // Define: 2686 // %alignment = add i8 %lowbitmask, 1 2687 // Iff 1. an %alignment is a power-of-two (aka, %lowbitmask is a low bit mask) 2688 // and 2. %bias is equal to either %lowbitmask or %alignment, 2689 // and 3. %highbitmask is equal to ~%lowbitmask (aka, to -%alignment) 2690 // then this pattern can be transformed into: 2691 // %x.offset = add i8 %x, %lowbitmask 2692 // %x.roundedup = and i8 %x.offset, %highbitmask 2693 static Value * 2694 foldRoundUpIntegerWithPow2Alignment(SelectInst &SI, 2695 InstCombiner::BuilderTy &Builder) { 2696 Value *Cond = SI.getCondition(); 2697 Value *X = SI.getTrueValue(); 2698 Value *XBiasedHighBits = SI.getFalseValue(); 2699 2700 ICmpInst::Predicate Pred; 2701 Value *XLowBits; 2702 if (!match(Cond, m_ICmp(Pred, m_Value(XLowBits), m_ZeroInt())) || 2703 !ICmpInst::isEquality(Pred)) 2704 return nullptr; 2705 2706 if (Pred == ICmpInst::Predicate::ICMP_NE) 2707 std::swap(X, XBiasedHighBits); 2708 2709 // FIXME: we could support non non-splats here. 2710 2711 const APInt *LowBitMaskCst; 2712 if (!match(XLowBits, m_And(m_Specific(X), m_APIntAllowUndef(LowBitMaskCst)))) 2713 return nullptr; 2714 2715 // Match even if the AND and ADD are swapped. 2716 const APInt *BiasCst, *HighBitMaskCst; 2717 if (!match(XBiasedHighBits, 2718 m_And(m_Add(m_Specific(X), m_APIntAllowUndef(BiasCst)), 2719 m_APIntAllowUndef(HighBitMaskCst))) && 2720 !match(XBiasedHighBits, 2721 m_Add(m_And(m_Specific(X), m_APIntAllowUndef(HighBitMaskCst)), 2722 m_APIntAllowUndef(BiasCst)))) 2723 return nullptr; 2724 2725 if (!LowBitMaskCst->isMask()) 2726 return nullptr; 2727 2728 APInt InvertedLowBitMaskCst = ~*LowBitMaskCst; 2729 if (InvertedLowBitMaskCst != *HighBitMaskCst) 2730 return nullptr; 2731 2732 APInt AlignmentCst = *LowBitMaskCst + 1; 2733 2734 if (*BiasCst != AlignmentCst && *BiasCst != *LowBitMaskCst) 2735 return nullptr; 2736 2737 if (!XBiasedHighBits->hasOneUse()) { 2738 if (*BiasCst == *LowBitMaskCst) 2739 return XBiasedHighBits; 2740 return nullptr; 2741 } 2742 2743 // FIXME: could we preserve undef's here? 2744 Type *Ty = X->getType(); 2745 Value *XOffset = Builder.CreateAdd(X, ConstantInt::get(Ty, *LowBitMaskCst), 2746 X->getName() + ".biased"); 2747 Value *R = Builder.CreateAnd(XOffset, ConstantInt::get(Ty, *HighBitMaskCst)); 2748 R->takeName(&SI); 2749 return R; 2750 } 2751 2752 namespace { 2753 struct DecomposedSelect { 2754 Value *Cond = nullptr; 2755 Value *TrueVal = nullptr; 2756 Value *FalseVal = nullptr; 2757 }; 2758 } // namespace 2759 2760 /// Look for patterns like 2761 /// %outer.cond = select i1 %inner.cond, i1 %alt.cond, i1 false 2762 /// %inner.sel = select i1 %inner.cond, i8 %inner.sel.t, i8 %inner.sel.f 2763 /// %outer.sel = select i1 %outer.cond, i8 %outer.sel.t, i8 %inner.sel 2764 /// and rewrite it as 2765 /// %inner.sel = select i1 %cond.alternative, i8 %sel.outer.t, i8 %sel.inner.t 2766 /// %sel.outer = select i1 %cond.inner, i8 %inner.sel, i8 %sel.inner.f 2767 static Instruction *foldNestedSelects(SelectInst &OuterSelVal, 2768 InstCombiner::BuilderTy &Builder) { 2769 // We must start with a `select`. 2770 DecomposedSelect OuterSel; 2771 match(&OuterSelVal, 2772 m_Select(m_Value(OuterSel.Cond), m_Value(OuterSel.TrueVal), 2773 m_Value(OuterSel.FalseVal))); 2774 2775 // Canonicalize inversion of the outermost `select`'s condition. 2776 if (match(OuterSel.Cond, m_Not(m_Value(OuterSel.Cond)))) 2777 std::swap(OuterSel.TrueVal, OuterSel.FalseVal); 2778 2779 // The condition of the outermost select must be an `and`/`or`. 2780 if (!match(OuterSel.Cond, m_c_LogicalOp(m_Value(), m_Value()))) 2781 return nullptr; 2782 2783 // Depending on the logical op, inner select might be in different hand. 2784 bool IsAndVariant = match(OuterSel.Cond, m_LogicalAnd()); 2785 Value *InnerSelVal = IsAndVariant ? OuterSel.FalseVal : OuterSel.TrueVal; 2786 2787 // Profitability check - avoid increasing instruction count. 2788 if (none_of(ArrayRef<Value *>({OuterSelVal.getCondition(), InnerSelVal}), 2789 [](Value *V) { return V->hasOneUse(); })) 2790 return nullptr; 2791 2792 // The appropriate hand of the outermost `select` must be a select itself. 2793 DecomposedSelect InnerSel; 2794 if (!match(InnerSelVal, 2795 m_Select(m_Value(InnerSel.Cond), m_Value(InnerSel.TrueVal), 2796 m_Value(InnerSel.FalseVal)))) 2797 return nullptr; 2798 2799 // Canonicalize inversion of the innermost `select`'s condition. 2800 if (match(InnerSel.Cond, m_Not(m_Value(InnerSel.Cond)))) 2801 std::swap(InnerSel.TrueVal, InnerSel.FalseVal); 2802 2803 Value *AltCond = nullptr; 2804 auto matchOuterCond = [OuterSel, &AltCond](auto m_InnerCond) { 2805 return match(OuterSel.Cond, m_c_LogicalOp(m_InnerCond, m_Value(AltCond))); 2806 }; 2807 2808 // Finally, match the condition that was driving the outermost `select`, 2809 // it should be a logical operation between the condition that was driving 2810 // the innermost `select` (after accounting for the possible inversions 2811 // of the condition), and some other condition. 2812 if (matchOuterCond(m_Specific(InnerSel.Cond))) { 2813 // Done! 2814 } else if (Value * NotInnerCond; matchOuterCond(m_CombineAnd( 2815 m_Not(m_Specific(InnerSel.Cond)), m_Value(NotInnerCond)))) { 2816 // Done! 2817 std::swap(InnerSel.TrueVal, InnerSel.FalseVal); 2818 InnerSel.Cond = NotInnerCond; 2819 } else // Not the pattern we were looking for. 2820 return nullptr; 2821 2822 Value *SelInner = Builder.CreateSelect( 2823 AltCond, IsAndVariant ? OuterSel.TrueVal : InnerSel.FalseVal, 2824 IsAndVariant ? InnerSel.TrueVal : OuterSel.FalseVal); 2825 SelInner->takeName(InnerSelVal); 2826 return SelectInst::Create(InnerSel.Cond, 2827 IsAndVariant ? SelInner : InnerSel.TrueVal, 2828 !IsAndVariant ? SelInner : InnerSel.FalseVal); 2829 } 2830 2831 Instruction *InstCombinerImpl::foldSelectOfBools(SelectInst &SI) { 2832 Value *CondVal = SI.getCondition(); 2833 Value *TrueVal = SI.getTrueValue(); 2834 Value *FalseVal = SI.getFalseValue(); 2835 Type *SelType = SI.getType(); 2836 2837 // Avoid potential infinite loops by checking for non-constant condition. 2838 // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()? 2839 // Scalar select must have simplified? 2840 if (!SelType->isIntOrIntVectorTy(1) || isa<Constant>(CondVal) || 2841 TrueVal->getType() != CondVal->getType()) 2842 return nullptr; 2843 2844 auto *One = ConstantInt::getTrue(SelType); 2845 auto *Zero = ConstantInt::getFalse(SelType); 2846 Value *A, *B, *C, *D; 2847 2848 // Folding select to and/or i1 isn't poison safe in general. impliesPoison 2849 // checks whether folding it does not convert a well-defined value into 2850 // poison. 2851 if (match(TrueVal, m_One())) { 2852 if (impliesPoison(FalseVal, CondVal)) { 2853 // Change: A = select B, true, C --> A = or B, C 2854 return BinaryOperator::CreateOr(CondVal, FalseVal); 2855 } 2856 2857 if (auto *LHS = dyn_cast<FCmpInst>(CondVal)) 2858 if (auto *RHS = dyn_cast<FCmpInst>(FalseVal)) 2859 if (Value *V = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ false, 2860 /*IsSelectLogical*/ true)) 2861 return replaceInstUsesWith(SI, V); 2862 2863 // (A && B) || (C && B) --> (A || C) && B 2864 if (match(CondVal, m_LogicalAnd(m_Value(A), m_Value(B))) && 2865 match(FalseVal, m_LogicalAnd(m_Value(C), m_Value(D))) && 2866 (CondVal->hasOneUse() || FalseVal->hasOneUse())) { 2867 bool CondLogicAnd = isa<SelectInst>(CondVal); 2868 bool FalseLogicAnd = isa<SelectInst>(FalseVal); 2869 auto AndFactorization = [&](Value *Common, Value *InnerCond, 2870 Value *InnerVal, 2871 bool SelFirst = false) -> Instruction * { 2872 Value *InnerSel = Builder.CreateSelect(InnerCond, One, InnerVal); 2873 if (SelFirst) 2874 std::swap(Common, InnerSel); 2875 if (FalseLogicAnd || (CondLogicAnd && Common == A)) 2876 return SelectInst::Create(Common, InnerSel, Zero); 2877 else 2878 return BinaryOperator::CreateAnd(Common, InnerSel); 2879 }; 2880 2881 if (A == C) 2882 return AndFactorization(A, B, D); 2883 if (A == D) 2884 return AndFactorization(A, B, C); 2885 if (B == C) 2886 return AndFactorization(B, A, D); 2887 if (B == D) 2888 return AndFactorization(B, A, C, CondLogicAnd && FalseLogicAnd); 2889 } 2890 } 2891 2892 if (match(FalseVal, m_Zero())) { 2893 if (impliesPoison(TrueVal, CondVal)) { 2894 // Change: A = select B, C, false --> A = and B, C 2895 return BinaryOperator::CreateAnd(CondVal, TrueVal); 2896 } 2897 2898 if (auto *LHS = dyn_cast<FCmpInst>(CondVal)) 2899 if (auto *RHS = dyn_cast<FCmpInst>(TrueVal)) 2900 if (Value *V = foldLogicOfFCmps(LHS, RHS, /*IsAnd*/ true, 2901 /*IsSelectLogical*/ true)) 2902 return replaceInstUsesWith(SI, V); 2903 2904 // (A || B) && (C || B) --> (A && C) || B 2905 if (match(CondVal, m_LogicalOr(m_Value(A), m_Value(B))) && 2906 match(TrueVal, m_LogicalOr(m_Value(C), m_Value(D))) && 2907 (CondVal->hasOneUse() || TrueVal->hasOneUse())) { 2908 bool CondLogicOr = isa<SelectInst>(CondVal); 2909 bool TrueLogicOr = isa<SelectInst>(TrueVal); 2910 auto OrFactorization = [&](Value *Common, Value *InnerCond, 2911 Value *InnerVal, 2912 bool SelFirst = false) -> Instruction * { 2913 Value *InnerSel = Builder.CreateSelect(InnerCond, InnerVal, Zero); 2914 if (SelFirst) 2915 std::swap(Common, InnerSel); 2916 if (TrueLogicOr || (CondLogicOr && Common == A)) 2917 return SelectInst::Create(Common, One, InnerSel); 2918 else 2919 return BinaryOperator::CreateOr(Common, InnerSel); 2920 }; 2921 2922 if (A == C) 2923 return OrFactorization(A, B, D); 2924 if (A == D) 2925 return OrFactorization(A, B, C); 2926 if (B == C) 2927 return OrFactorization(B, A, D); 2928 if (B == D) 2929 return OrFactorization(B, A, C, CondLogicOr && TrueLogicOr); 2930 } 2931 } 2932 2933 // We match the "full" 0 or 1 constant here to avoid a potential infinite 2934 // loop with vectors that may have undefined/poison elements. 2935 // select a, false, b -> select !a, b, false 2936 if (match(TrueVal, m_Specific(Zero))) { 2937 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2938 return SelectInst::Create(NotCond, FalseVal, Zero); 2939 } 2940 // select a, b, true -> select !a, true, b 2941 if (match(FalseVal, m_Specific(One))) { 2942 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 2943 return SelectInst::Create(NotCond, One, TrueVal); 2944 } 2945 2946 // DeMorgan in select form: !a && !b --> !(a || b) 2947 // select !a, !b, false --> not (select a, true, b) 2948 if (match(&SI, m_LogicalAnd(m_Not(m_Value(A)), m_Not(m_Value(B)))) && 2949 (CondVal->hasOneUse() || TrueVal->hasOneUse()) && 2950 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) 2951 return BinaryOperator::CreateNot(Builder.CreateSelect(A, One, B)); 2952 2953 // DeMorgan in select form: !a || !b --> !(a && b) 2954 // select !a, true, !b --> not (select a, b, false) 2955 if (match(&SI, m_LogicalOr(m_Not(m_Value(A)), m_Not(m_Value(B)))) && 2956 (CondVal->hasOneUse() || FalseVal->hasOneUse()) && 2957 !match(A, m_ConstantExpr()) && !match(B, m_ConstantExpr())) 2958 return BinaryOperator::CreateNot(Builder.CreateSelect(A, B, Zero)); 2959 2960 // select (select a, true, b), true, b -> select a, true, b 2961 if (match(CondVal, m_Select(m_Value(A), m_One(), m_Value(B))) && 2962 match(TrueVal, m_One()) && match(FalseVal, m_Specific(B))) 2963 return replaceOperand(SI, 0, A); 2964 // select (select a, b, false), b, false -> select a, b, false 2965 if (match(CondVal, m_Select(m_Value(A), m_Value(B), m_Zero())) && 2966 match(TrueVal, m_Specific(B)) && match(FalseVal, m_Zero())) 2967 return replaceOperand(SI, 0, A); 2968 2969 // ~(A & B) & (A | B) --> A ^ B 2970 if (match(&SI, m_c_LogicalAnd(m_Not(m_LogicalAnd(m_Value(A), m_Value(B))), 2971 m_c_LogicalOr(m_Deferred(A), m_Deferred(B))))) 2972 return BinaryOperator::CreateXor(A, B); 2973 2974 // select (~a | c), a, b -> and a, (or c, freeze(b)) 2975 if (match(CondVal, m_c_Or(m_Not(m_Specific(TrueVal)), m_Value(C))) && 2976 CondVal->hasOneUse()) { 2977 FalseVal = Builder.CreateFreeze(FalseVal); 2978 return BinaryOperator::CreateAnd(TrueVal, Builder.CreateOr(C, FalseVal)); 2979 } 2980 // select (~c & b), a, b -> and b, (or freeze(a), c) 2981 if (match(CondVal, m_c_And(m_Not(m_Value(C)), m_Specific(FalseVal))) && 2982 CondVal->hasOneUse()) { 2983 TrueVal = Builder.CreateFreeze(TrueVal); 2984 return BinaryOperator::CreateAnd(FalseVal, Builder.CreateOr(C, TrueVal)); 2985 } 2986 2987 if (match(FalseVal, m_Zero()) || match(TrueVal, m_One())) { 2988 Use *Y = nullptr; 2989 bool IsAnd = match(FalseVal, m_Zero()) ? true : false; 2990 Value *Op1 = IsAnd ? TrueVal : FalseVal; 2991 if (isCheckForZeroAndMulWithOverflow(CondVal, Op1, IsAnd, Y)) { 2992 auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr"); 2993 InsertNewInstBefore(FI, *cast<Instruction>(Y->getUser())); 2994 replaceUse(*Y, FI); 2995 return replaceInstUsesWith(SI, Op1); 2996 } 2997 2998 if (auto *Op1SI = dyn_cast<SelectInst>(Op1)) 2999 if (auto *I = foldAndOrOfSelectUsingImpliedCond(CondVal, *Op1SI, 3000 /* IsAnd */ IsAnd)) 3001 return I; 3002 3003 if (auto *ICmp0 = dyn_cast<ICmpInst>(CondVal)) 3004 if (auto *ICmp1 = dyn_cast<ICmpInst>(Op1)) 3005 if (auto *V = foldAndOrOfICmps(ICmp0, ICmp1, SI, IsAnd, 3006 /* IsLogical */ true)) 3007 return replaceInstUsesWith(SI, V); 3008 } 3009 3010 // select (a || b), c, false -> select a, c, false 3011 // select c, (a || b), false -> select c, a, false 3012 // if c implies that b is false. 3013 if (match(CondVal, m_LogicalOr(m_Value(A), m_Value(B))) && 3014 match(FalseVal, m_Zero())) { 3015 std::optional<bool> Res = isImpliedCondition(TrueVal, B, DL); 3016 if (Res && *Res == false) 3017 return replaceOperand(SI, 0, A); 3018 } 3019 if (match(TrueVal, m_LogicalOr(m_Value(A), m_Value(B))) && 3020 match(FalseVal, m_Zero())) { 3021 std::optional<bool> Res = isImpliedCondition(CondVal, B, DL); 3022 if (Res && *Res == false) 3023 return replaceOperand(SI, 1, A); 3024 } 3025 // select c, true, (a && b) -> select c, true, a 3026 // select (a && b), true, c -> select a, true, c 3027 // if c = false implies that b = true 3028 if (match(TrueVal, m_One()) && 3029 match(FalseVal, m_LogicalAnd(m_Value(A), m_Value(B)))) { 3030 std::optional<bool> Res = isImpliedCondition(CondVal, B, DL, false); 3031 if (Res && *Res == true) 3032 return replaceOperand(SI, 2, A); 3033 } 3034 if (match(CondVal, m_LogicalAnd(m_Value(A), m_Value(B))) && 3035 match(TrueVal, m_One())) { 3036 std::optional<bool> Res = isImpliedCondition(FalseVal, B, DL, false); 3037 if (Res && *Res == true) 3038 return replaceOperand(SI, 0, A); 3039 } 3040 3041 if (match(TrueVal, m_One())) { 3042 Value *C; 3043 3044 // (C && A) || (!C && B) --> sel C, A, B 3045 // (A && C) || (!C && B) --> sel C, A, B 3046 // (C && A) || (B && !C) --> sel C, A, B 3047 // (A && C) || (B && !C) --> sel C, A, B (may require freeze) 3048 if (match(FalseVal, m_c_LogicalAnd(m_Not(m_Value(C)), m_Value(B))) && 3049 match(CondVal, m_c_LogicalAnd(m_Specific(C), m_Value(A)))) { 3050 auto *SelCond = dyn_cast<SelectInst>(CondVal); 3051 auto *SelFVal = dyn_cast<SelectInst>(FalseVal); 3052 bool MayNeedFreeze = SelCond && SelFVal && 3053 match(SelFVal->getTrueValue(), 3054 m_Not(m_Specific(SelCond->getTrueValue()))); 3055 if (MayNeedFreeze) 3056 C = Builder.CreateFreeze(C); 3057 return SelectInst::Create(C, A, B); 3058 } 3059 3060 // (!C && A) || (C && B) --> sel C, B, A 3061 // (A && !C) || (C && B) --> sel C, B, A 3062 // (!C && A) || (B && C) --> sel C, B, A 3063 // (A && !C) || (B && C) --> sel C, B, A (may require freeze) 3064 if (match(CondVal, m_c_LogicalAnd(m_Not(m_Value(C)), m_Value(A))) && 3065 match(FalseVal, m_c_LogicalAnd(m_Specific(C), m_Value(B)))) { 3066 auto *SelCond = dyn_cast<SelectInst>(CondVal); 3067 auto *SelFVal = dyn_cast<SelectInst>(FalseVal); 3068 bool MayNeedFreeze = SelCond && SelFVal && 3069 match(SelCond->getTrueValue(), 3070 m_Not(m_Specific(SelFVal->getTrueValue()))); 3071 if (MayNeedFreeze) 3072 C = Builder.CreateFreeze(C); 3073 return SelectInst::Create(C, B, A); 3074 } 3075 } 3076 3077 return nullptr; 3078 } 3079 3080 Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { 3081 Value *CondVal = SI.getCondition(); 3082 Value *TrueVal = SI.getTrueValue(); 3083 Value *FalseVal = SI.getFalseValue(); 3084 Type *SelType = SI.getType(); 3085 3086 if (Value *V = simplifySelectInst(CondVal, TrueVal, FalseVal, 3087 SQ.getWithInstruction(&SI))) 3088 return replaceInstUsesWith(SI, V); 3089 3090 if (Instruction *I = canonicalizeSelectToShuffle(SI)) 3091 return I; 3092 3093 if (Instruction *I = canonicalizeScalarSelectOfVecs(SI, *this)) 3094 return I; 3095 3096 // If the type of select is not an integer type or if the condition and 3097 // the selection type are not both scalar nor both vector types, there is no 3098 // point in attempting to match these patterns. 3099 Type *CondType = CondVal->getType(); 3100 if (!isa<Constant>(CondVal) && SelType->isIntOrIntVectorTy() && 3101 CondType->isVectorTy() == SelType->isVectorTy()) { 3102 if (Value *S = simplifyWithOpReplaced(TrueVal, CondVal, 3103 ConstantInt::getTrue(CondType), SQ, 3104 /* AllowRefinement */ true)) 3105 return replaceOperand(SI, 1, S); 3106 3107 if (Value *S = simplifyWithOpReplaced(FalseVal, CondVal, 3108 ConstantInt::getFalse(CondType), SQ, 3109 /* AllowRefinement */ true)) 3110 return replaceOperand(SI, 2, S); 3111 3112 // Handle patterns involving sext/zext + not explicitly, 3113 // as simplifyWithOpReplaced() only looks past one instruction. 3114 Value *NotCond; 3115 3116 // select a, sext(!a), b -> select !a, b, 0 3117 // select a, zext(!a), b -> select !a, b, 0 3118 if (match(TrueVal, m_ZExtOrSExt(m_CombineAnd(m_Value(NotCond), 3119 m_Not(m_Specific(CondVal)))))) 3120 return SelectInst::Create(NotCond, FalseVal, 3121 Constant::getNullValue(SelType)); 3122 3123 // select a, b, zext(!a) -> select !a, 1, b 3124 if (match(FalseVal, m_ZExt(m_CombineAnd(m_Value(NotCond), 3125 m_Not(m_Specific(CondVal)))))) 3126 return SelectInst::Create(NotCond, ConstantInt::get(SelType, 1), TrueVal); 3127 3128 // select a, b, sext(!a) -> select !a, -1, b 3129 if (match(FalseVal, m_SExt(m_CombineAnd(m_Value(NotCond), 3130 m_Not(m_Specific(CondVal)))))) 3131 return SelectInst::Create(NotCond, Constant::getAllOnesValue(SelType), 3132 TrueVal); 3133 } 3134 3135 if (Instruction *R = foldSelectOfBools(SI)) 3136 return R; 3137 3138 // Selecting between two integer or vector splat integer constants? 3139 // 3140 // Note that we don't handle a scalar select of vectors: 3141 // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> 3142 // because that may need 3 instructions to splat the condition value: 3143 // extend, insertelement, shufflevector. 3144 // 3145 // Do not handle i1 TrueVal and FalseVal otherwise would result in 3146 // zext/sext i1 to i1. 3147 if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(1) && 3148 CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { 3149 // select C, 1, 0 -> zext C to int 3150 if (match(TrueVal, m_One()) && match(FalseVal, m_Zero())) 3151 return new ZExtInst(CondVal, SelType); 3152 3153 // select C, -1, 0 -> sext C to int 3154 if (match(TrueVal, m_AllOnes()) && match(FalseVal, m_Zero())) 3155 return new SExtInst(CondVal, SelType); 3156 3157 // select C, 0, 1 -> zext !C to int 3158 if (match(TrueVal, m_Zero()) && match(FalseVal, m_One())) { 3159 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 3160 return new ZExtInst(NotCond, SelType); 3161 } 3162 3163 // select C, 0, -1 -> sext !C to int 3164 if (match(TrueVal, m_Zero()) && match(FalseVal, m_AllOnes())) { 3165 Value *NotCond = Builder.CreateNot(CondVal, "not." + CondVal->getName()); 3166 return new SExtInst(NotCond, SelType); 3167 } 3168 } 3169 3170 if (auto *FCmp = dyn_cast<FCmpInst>(CondVal)) { 3171 Value *Cmp0 = FCmp->getOperand(0), *Cmp1 = FCmp->getOperand(1); 3172 // Are we selecting a value based on a comparison of the two values? 3173 if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || 3174 (Cmp0 == FalseVal && Cmp1 == TrueVal)) { 3175 // Canonicalize to use ordered comparisons by swapping the select 3176 // operands. 3177 // 3178 // e.g. 3179 // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X 3180 if (FCmp->hasOneUse() && FCmpInst::isUnordered(FCmp->getPredicate())) { 3181 FCmpInst::Predicate InvPred = FCmp->getInversePredicate(); 3182 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 3183 // FIXME: The FMF should propagate from the select, not the fcmp. 3184 Builder.setFastMathFlags(FCmp->getFastMathFlags()); 3185 Value *NewCond = Builder.CreateFCmp(InvPred, Cmp0, Cmp1, 3186 FCmp->getName() + ".inv"); 3187 Value *NewSel = Builder.CreateSelect(NewCond, FalseVal, TrueVal); 3188 return replaceInstUsesWith(SI, NewSel); 3189 } 3190 } 3191 } 3192 3193 if (isa<FPMathOperator>(SI)) { 3194 // TODO: Try to forward-propagate FMF from select arms to the select. 3195 3196 // Canonicalize select of FP values where NaN and -0.0 are not valid as 3197 // minnum/maxnum intrinsics. 3198 if (SI.hasNoNaNs() && SI.hasNoSignedZeros()) { 3199 Value *X, *Y; 3200 if (match(&SI, m_OrdFMax(m_Value(X), m_Value(Y)))) 3201 return replaceInstUsesWith( 3202 SI, Builder.CreateBinaryIntrinsic(Intrinsic::maxnum, X, Y, &SI)); 3203 3204 if (match(&SI, m_OrdFMin(m_Value(X), m_Value(Y)))) 3205 return replaceInstUsesWith( 3206 SI, Builder.CreateBinaryIntrinsic(Intrinsic::minnum, X, Y, &SI)); 3207 } 3208 } 3209 3210 // Fold selecting to fabs. 3211 if (Instruction *Fabs = foldSelectWithFCmpToFabs(SI, *this)) 3212 return Fabs; 3213 3214 // See if we are selecting two values based on a comparison of the two values. 3215 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) 3216 if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) 3217 return Result; 3218 3219 if (Instruction *Add = foldAddSubSelect(SI, Builder)) 3220 return Add; 3221 if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) 3222 return Add; 3223 if (Instruction *Or = foldSetClearBits(SI, Builder)) 3224 return Or; 3225 if (Instruction *Mul = foldSelectZeroOrMul(SI, *this)) 3226 return Mul; 3227 3228 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) 3229 auto *TI = dyn_cast<Instruction>(TrueVal); 3230 auto *FI = dyn_cast<Instruction>(FalseVal); 3231 if (TI && FI && TI->getOpcode() == FI->getOpcode()) 3232 if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) 3233 return IV; 3234 3235 if (Instruction *I = foldSelectExtConst(SI)) 3236 return I; 3237 3238 // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0)) 3239 // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx)) 3240 auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base, 3241 bool Swap) -> GetElementPtrInst * { 3242 Value *Ptr = Gep->getPointerOperand(); 3243 if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base || 3244 !Gep->hasOneUse()) 3245 return nullptr; 3246 Value *Idx = Gep->getOperand(1); 3247 if (isa<VectorType>(CondVal->getType()) && !isa<VectorType>(Idx->getType())) 3248 return nullptr; 3249 Type *ElementType = Gep->getResultElementType(); 3250 Value *NewT = Idx; 3251 Value *NewF = Constant::getNullValue(Idx->getType()); 3252 if (Swap) 3253 std::swap(NewT, NewF); 3254 Value *NewSI = 3255 Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI); 3256 return GetElementPtrInst::Create(ElementType, Ptr, {NewSI}); 3257 }; 3258 if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal)) 3259 if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false)) 3260 return NewGep; 3261 if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal)) 3262 if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true)) 3263 return NewGep; 3264 3265 // See if we can fold the select into one of our operands. 3266 if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { 3267 if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) 3268 return FoldI; 3269 3270 Value *LHS, *RHS; 3271 Instruction::CastOps CastOp; 3272 SelectPatternResult SPR = matchSelectPattern(&SI, LHS, RHS, &CastOp); 3273 auto SPF = SPR.Flavor; 3274 if (SPF) { 3275 Value *LHS2, *RHS2; 3276 if (SelectPatternFlavor SPF2 = matchSelectPattern(LHS, LHS2, RHS2).Flavor) 3277 if (Instruction *R = foldSPFofSPF(cast<Instruction>(LHS), SPF2, LHS2, 3278 RHS2, SI, SPF, RHS)) 3279 return R; 3280 if (SelectPatternFlavor SPF2 = matchSelectPattern(RHS, LHS2, RHS2).Flavor) 3281 if (Instruction *R = foldSPFofSPF(cast<Instruction>(RHS), SPF2, LHS2, 3282 RHS2, SI, SPF, LHS)) 3283 return R; 3284 } 3285 3286 if (SelectPatternResult::isMinOrMax(SPF)) { 3287 // Canonicalize so that 3288 // - type casts are outside select patterns. 3289 // - float clamp is transformed to min/max pattern 3290 3291 bool IsCastNeeded = LHS->getType() != SelType; 3292 Value *CmpLHS = cast<CmpInst>(CondVal)->getOperand(0); 3293 Value *CmpRHS = cast<CmpInst>(CondVal)->getOperand(1); 3294 if (IsCastNeeded || 3295 (LHS->getType()->isFPOrFPVectorTy() && 3296 ((CmpLHS != LHS && CmpLHS != RHS) || 3297 (CmpRHS != LHS && CmpRHS != RHS)))) { 3298 CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, SPR.Ordered); 3299 3300 Value *Cmp; 3301 if (CmpInst::isIntPredicate(MinMaxPred)) { 3302 Cmp = Builder.CreateICmp(MinMaxPred, LHS, RHS); 3303 } else { 3304 IRBuilder<>::FastMathFlagGuard FMFG(Builder); 3305 auto FMF = 3306 cast<FPMathOperator>(SI.getCondition())->getFastMathFlags(); 3307 Builder.setFastMathFlags(FMF); 3308 Cmp = Builder.CreateFCmp(MinMaxPred, LHS, RHS); 3309 } 3310 3311 Value *NewSI = Builder.CreateSelect(Cmp, LHS, RHS, SI.getName(), &SI); 3312 if (!IsCastNeeded) 3313 return replaceInstUsesWith(SI, NewSI); 3314 3315 Value *NewCast = Builder.CreateCast(CastOp, NewSI, SelType); 3316 return replaceInstUsesWith(SI, NewCast); 3317 } 3318 } 3319 } 3320 3321 // See if we can fold the select into a phi node if the condition is a select. 3322 if (auto *PN = dyn_cast<PHINode>(SI.getCondition())) 3323 // The true/false values have to be live in the PHI predecessor's blocks. 3324 if (canSelectOperandBeMappingIntoPredBlock(TrueVal, SI) && 3325 canSelectOperandBeMappingIntoPredBlock(FalseVal, SI)) 3326 if (Instruction *NV = foldOpIntoPhi(SI, PN)) 3327 return NV; 3328 3329 if (SelectInst *TrueSI = dyn_cast<SelectInst>(TrueVal)) { 3330 if (TrueSI->getCondition()->getType() == CondVal->getType()) { 3331 // select(C, select(C, a, b), c) -> select(C, a, c) 3332 if (TrueSI->getCondition() == CondVal) { 3333 if (SI.getTrueValue() == TrueSI->getTrueValue()) 3334 return nullptr; 3335 return replaceOperand(SI, 1, TrueSI->getTrueValue()); 3336 } 3337 // select(C0, select(C1, a, b), b) -> select(C0&C1, a, b) 3338 // We choose this as normal form to enable folding on the And and 3339 // shortening paths for the values (this helps getUnderlyingObjects() for 3340 // example). 3341 if (TrueSI->getFalseValue() == FalseVal && TrueSI->hasOneUse()) { 3342 Value *And = Builder.CreateLogicalAnd(CondVal, TrueSI->getCondition()); 3343 replaceOperand(SI, 0, And); 3344 replaceOperand(SI, 1, TrueSI->getTrueValue()); 3345 return &SI; 3346 } 3347 } 3348 } 3349 if (SelectInst *FalseSI = dyn_cast<SelectInst>(FalseVal)) { 3350 if (FalseSI->getCondition()->getType() == CondVal->getType()) { 3351 // select(C, a, select(C, b, c)) -> select(C, a, c) 3352 if (FalseSI->getCondition() == CondVal) { 3353 if (SI.getFalseValue() == FalseSI->getFalseValue()) 3354 return nullptr; 3355 return replaceOperand(SI, 2, FalseSI->getFalseValue()); 3356 } 3357 // select(C0, a, select(C1, a, b)) -> select(C0|C1, a, b) 3358 if (FalseSI->getTrueValue() == TrueVal && FalseSI->hasOneUse()) { 3359 Value *Or = Builder.CreateLogicalOr(CondVal, FalseSI->getCondition()); 3360 replaceOperand(SI, 0, Or); 3361 replaceOperand(SI, 2, FalseSI->getFalseValue()); 3362 return &SI; 3363 } 3364 } 3365 } 3366 3367 auto canMergeSelectThroughBinop = [](BinaryOperator *BO) { 3368 // The select might be preventing a division by 0. 3369 switch (BO->getOpcode()) { 3370 default: 3371 return true; 3372 case Instruction::SRem: 3373 case Instruction::URem: 3374 case Instruction::SDiv: 3375 case Instruction::UDiv: 3376 return false; 3377 } 3378 }; 3379 3380 // Try to simplify a binop sandwiched between 2 selects with the same 3381 // condition. 3382 // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) 3383 BinaryOperator *TrueBO; 3384 if (match(TrueVal, m_OneUse(m_BinOp(TrueBO))) && 3385 canMergeSelectThroughBinop(TrueBO)) { 3386 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(0))) { 3387 if (TrueBOSI->getCondition() == CondVal) { 3388 replaceOperand(*TrueBO, 0, TrueBOSI->getTrueValue()); 3389 Worklist.push(TrueBO); 3390 return &SI; 3391 } 3392 } 3393 if (auto *TrueBOSI = dyn_cast<SelectInst>(TrueBO->getOperand(1))) { 3394 if (TrueBOSI->getCondition() == CondVal) { 3395 replaceOperand(*TrueBO, 1, TrueBOSI->getTrueValue()); 3396 Worklist.push(TrueBO); 3397 return &SI; 3398 } 3399 } 3400 } 3401 3402 // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) 3403 BinaryOperator *FalseBO; 3404 if (match(FalseVal, m_OneUse(m_BinOp(FalseBO))) && 3405 canMergeSelectThroughBinop(FalseBO)) { 3406 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(0))) { 3407 if (FalseBOSI->getCondition() == CondVal) { 3408 replaceOperand(*FalseBO, 0, FalseBOSI->getFalseValue()); 3409 Worklist.push(FalseBO); 3410 return &SI; 3411 } 3412 } 3413 if (auto *FalseBOSI = dyn_cast<SelectInst>(FalseBO->getOperand(1))) { 3414 if (FalseBOSI->getCondition() == CondVal) { 3415 replaceOperand(*FalseBO, 1, FalseBOSI->getFalseValue()); 3416 Worklist.push(FalseBO); 3417 return &SI; 3418 } 3419 } 3420 } 3421 3422 Value *NotCond; 3423 if (match(CondVal, m_Not(m_Value(NotCond))) && 3424 !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { 3425 replaceOperand(SI, 0, NotCond); 3426 SI.swapValues(); 3427 SI.swapProfMetadata(); 3428 return &SI; 3429 } 3430 3431 if (Instruction *I = foldVectorSelect(SI)) 3432 return I; 3433 3434 // If we can compute the condition, there's no need for a select. 3435 // Like the above fold, we are attempting to reduce compile-time cost by 3436 // putting this fold here with limitations rather than in InstSimplify. 3437 // The motivation for this call into value tracking is to take advantage of 3438 // the assumption cache, so make sure that is populated. 3439 if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { 3440 KnownBits Known(1); 3441 computeKnownBits(CondVal, Known, 0, &SI); 3442 if (Known.One.isOne()) 3443 return replaceInstUsesWith(SI, TrueVal); 3444 if (Known.Zero.isOne()) 3445 return replaceInstUsesWith(SI, FalseVal); 3446 } 3447 3448 if (Instruction *BitCastSel = foldSelectCmpBitcasts(SI, Builder)) 3449 return BitCastSel; 3450 3451 // Simplify selects that test the returned flag of cmpxchg instructions. 3452 if (Value *V = foldSelectCmpXchg(SI)) 3453 return replaceInstUsesWith(SI, V); 3454 3455 if (Instruction *Select = foldSelectBinOpIdentity(SI, TLI, *this)) 3456 return Select; 3457 3458 if (Instruction *Funnel = foldSelectFunnelShift(SI, Builder)) 3459 return Funnel; 3460 3461 if (Instruction *Copysign = foldSelectToCopysign(SI, Builder)) 3462 return Copysign; 3463 3464 if (Instruction *PN = foldSelectToPhi(SI, DT, Builder)) 3465 return replaceInstUsesWith(SI, PN); 3466 3467 if (Value *Fr = foldSelectWithFrozenICmp(SI, Builder)) 3468 return replaceInstUsesWith(SI, Fr); 3469 3470 if (Value *V = foldRoundUpIntegerWithPow2Alignment(SI, Builder)) 3471 return replaceInstUsesWith(SI, V); 3472 3473 // select(mask, mload(,,mask,0), 0) -> mload(,,mask,0) 3474 // Load inst is intentionally not checked for hasOneUse() 3475 if (match(FalseVal, m_Zero()) && 3476 (match(TrueVal, m_MaskedLoad(m_Value(), m_Value(), m_Specific(CondVal), 3477 m_CombineOr(m_Undef(), m_Zero()))) || 3478 match(TrueVal, m_MaskedGather(m_Value(), m_Value(), m_Specific(CondVal), 3479 m_CombineOr(m_Undef(), m_Zero()))))) { 3480 auto *MaskedInst = cast<IntrinsicInst>(TrueVal); 3481 if (isa<UndefValue>(MaskedInst->getArgOperand(3))) 3482 MaskedInst->setArgOperand(3, FalseVal /* Zero */); 3483 return replaceInstUsesWith(SI, MaskedInst); 3484 } 3485 3486 Value *Mask; 3487 if (match(TrueVal, m_Zero()) && 3488 (match(FalseVal, m_MaskedLoad(m_Value(), m_Value(), m_Value(Mask), 3489 m_CombineOr(m_Undef(), m_Zero()))) || 3490 match(FalseVal, m_MaskedGather(m_Value(), m_Value(), m_Value(Mask), 3491 m_CombineOr(m_Undef(), m_Zero())))) && 3492 (CondVal->getType() == Mask->getType())) { 3493 // We can remove the select by ensuring the load zeros all lanes the 3494 // select would have. We determine this by proving there is no overlap 3495 // between the load and select masks. 3496 // (i.e (load_mask & select_mask) == 0 == no overlap) 3497 bool CanMergeSelectIntoLoad = false; 3498 if (Value *V = simplifyAndInst(CondVal, Mask, SQ.getWithInstruction(&SI))) 3499 CanMergeSelectIntoLoad = match(V, m_Zero()); 3500 3501 if (CanMergeSelectIntoLoad) { 3502 auto *MaskedInst = cast<IntrinsicInst>(FalseVal); 3503 if (isa<UndefValue>(MaskedInst->getArgOperand(3))) 3504 MaskedInst->setArgOperand(3, TrueVal /* Zero */); 3505 return replaceInstUsesWith(SI, MaskedInst); 3506 } 3507 } 3508 3509 if (Instruction *I = foldNestedSelects(SI, Builder)) 3510 return I; 3511 3512 // Match logical variants of the pattern, 3513 // and transform them iff that gets rid of inversions. 3514 // (~x) | y --> ~(x & (~y)) 3515 // (~x) & y --> ~(x | (~y)) 3516 if (sinkNotIntoOtherHandOfLogicalOp(SI)) 3517 return &SI; 3518 3519 return nullptr; 3520 } 3521