| 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/Loads.h" |
| 21 | #include "llvm/Analysis/OverflowInstAnalysis.h" |
| 22 | #include "llvm/Analysis/ValueTracking.h" |
| 23 | #include "llvm/Analysis/VectorUtils.h" |
| 24 | #include "llvm/IR/BasicBlock.h" |
| 25 | #include "llvm/IR/Constant.h" |
| 26 | #include "llvm/IR/ConstantRange.h" |
| 27 | #include "llvm/IR/Constants.h" |
| 28 | #include "llvm/IR/DerivedTypes.h" |
| 29 | #include "llvm/IR/FMF.h" |
| 30 | #include "llvm/IR/IRBuilder.h" |
| 31 | #include "llvm/IR/InstrTypes.h" |
| 32 | #include "llvm/IR/Instruction.h" |
| 33 | #include "llvm/IR/Instructions.h" |
| 34 | #include "llvm/IR/IntrinsicInst.h" |
| 35 | #include "llvm/IR/Intrinsics.h" |
| 36 | #include "llvm/IR/Operator.h" |
| 37 | #include "llvm/IR/PatternMatch.h" |
| 38 | #include "llvm/IR/ProfDataUtils.h" |
| 39 | #include "llvm/IR/Type.h" |
| 40 | #include "llvm/IR/User.h" |
| 41 | #include "llvm/IR/Value.h" |
| 42 | #include "llvm/Support/Casting.h" |
| 43 | #include "llvm/Support/ErrorHandling.h" |
| 44 | #include "llvm/Support/KnownBits.h" |
| 45 | #include "llvm/Support/MathExtras.h" |
| 46 | #include "llvm/Transforms/InstCombine/InstCombiner.h" |
| 47 | #include <cassert> |
| 48 | #include <optional> |
| 49 | #include <utility> |
| 50 | |
| 51 | #define DEBUG_TYPE "instcombine" |
| 52 | #include "llvm/Transforms/Utils/InstructionWorklist.h" |
| 53 | |
| 54 | using namespace llvm; |
| 55 | using namespace PatternMatch; |
| 56 | |
| 57 | namespace llvm { |
| 58 | extern cl::opt<bool> ProfcheckDisableMetadataFixes; |
| 59 | } |
| 60 | |
| 61 | /// Replace a select operand based on an equality comparison with the identity |
| 62 | /// constant of a binop. |
| 63 | static Instruction *foldSelectBinOpIdentity(SelectInst &Sel, |
| 64 | const TargetLibraryInfo &TLI, |
| 65 | InstCombinerImpl &IC) { |
| 66 | // The select condition must be an equality compare with a constant operand. |
| 67 | Value *X; |
| 68 | Constant *C; |
| 69 | CmpPredicate Pred; |
| 70 | if (!match(V: Sel.getCondition(), P: m_Cmp(Pred, L: m_Value(V&: X), R: m_Constant(C)))) |
| 71 | return nullptr; |
| 72 | |
| 73 | bool IsEq; |
| 74 | if (ICmpInst::isEquality(P: Pred)) |
| 75 | IsEq = Pred == ICmpInst::ICMP_EQ; |
| 76 | else if (Pred == FCmpInst::FCMP_OEQ) |
| 77 | IsEq = true; |
| 78 | else if (Pred == FCmpInst::FCMP_UNE) |
| 79 | IsEq = false; |
| 80 | else |
| 81 | return nullptr; |
| 82 | |
| 83 | // A select operand must be a binop. |
| 84 | BinaryOperator *BO; |
| 85 | if (!match(V: Sel.getOperand(i_nocapture: IsEq ? 1 : 2), P: m_BinOp(I&: BO))) |
| 86 | return nullptr; |
| 87 | |
| 88 | // For absorbing values, we can fold to the compared value. |
| 89 | bool IsAbsorbingValue = false; |
| 90 | |
| 91 | // Last, match the compare variable operand with a binop operand. |
| 92 | Value *Y; |
| 93 | if (BO->isCommutative()) { |
| 94 | // Recognized 0 as an absorbing value for fmul, but we need to be careful |
| 95 | // about the sign. This could be more aggressive, by handling arbitrary sign |
| 96 | // bit operations as long as we know the fmul sign matches (and handling |
| 97 | // arbitrary opcodes). |
| 98 | if (match(V: BO, P: m_c_FMul(L: m_FAbs(Op0: m_Specific(V: X)), R: m_Value(V&: Y))) && |
| 99 | match(V: C, P: m_AnyZeroFP()) && |
| 100 | IC.fmulByZeroIsZero(MulVal: Y, FMF: BO->getFastMathFlags(), CtxI: &Sel)) |
| 101 | IsAbsorbingValue = true; |
| 102 | else if (!match(V: BO, P: m_c_BinOp(L: m_Value(V&: Y), R: m_Specific(V: X)))) |
| 103 | return nullptr; |
| 104 | } else { |
| 105 | if (!match(V: BO, P: m_BinOp(L: m_Value(V&: Y), R: m_Specific(V: X)))) |
| 106 | return nullptr; |
| 107 | } |
| 108 | |
| 109 | // The compare constant must be the identity constant for that binop. |
| 110 | // If this a floating-point compare with 0.0, any zero constant will do. |
| 111 | Type *Ty = BO->getType(); |
| 112 | |
| 113 | Value *FoldedVal; |
| 114 | if (IsAbsorbingValue) { |
| 115 | FoldedVal = C; |
| 116 | } else { |
| 117 | Constant *IdC = ConstantExpr::getBinOpIdentity(Opcode: BO->getOpcode(), Ty, AllowRHSConstant: true); |
| 118 | if (IdC != C) { |
| 119 | if (!IdC || !CmpInst::isFPPredicate(P: Pred)) |
| 120 | return nullptr; |
| 121 | |
| 122 | if (!match(V: IdC, P: m_AnyZeroFP()) || !match(V: C, P: m_AnyZeroFP())) |
| 123 | return nullptr; |
| 124 | } |
| 125 | |
| 126 | // +0.0 compares equal to -0.0, and so it does not behave as required for |
| 127 | // this transform. Bail out if we can not exclude that possibility. |
| 128 | if (const auto *FPO = dyn_cast<FPMathOperator>(Val: BO)) |
| 129 | if (!FPO->hasNoSignedZeros() && |
| 130 | !cannotBeNegativeZero(V: Y, |
| 131 | SQ: IC.getSimplifyQuery().getWithInstruction(I: &Sel))) |
| 132 | return nullptr; |
| 133 | |
| 134 | FoldedVal = Y; |
| 135 | } |
| 136 | |
| 137 | // BO = binop Y, X |
| 138 | // S = { select (cmp eq X, C), BO, ? } or { select (cmp ne X, C), ?, BO } |
| 139 | // => |
| 140 | // S = { select (cmp eq X, C), Y, ? } or { select (cmp ne X, C), ?, Y } |
| 141 | return IC.replaceOperand(I&: Sel, OpNum: IsEq ? 1 : 2, V: FoldedVal); |
| 142 | } |
| 143 | |
| 144 | /// This folds: |
| 145 | /// select (icmp eq (and X, C1)), TC, FC |
| 146 | /// iff C1 is a power 2 and the difference between TC and FC is a power-of-2. |
| 147 | /// To something like: |
| 148 | /// (shr (and (X, C1)), (log2(C1) - log2(TC-FC))) + FC |
| 149 | /// Or: |
| 150 | /// (shl (and (X, C1)), (log2(TC-FC) - log2(C1))) + FC |
| 151 | /// With some variations depending if FC is larger than TC, or the shift |
| 152 | /// isn't needed, or the bit widths don't match. |
| 153 | static Value *foldSelectICmpAnd(SelectInst &Sel, Value *CondVal, Value *TrueVal, |
| 154 | Value *FalseVal, Value *V, const APInt &AndMask, |
| 155 | bool CreateAnd, |
| 156 | InstCombiner::BuilderTy &Builder) { |
| 157 | const APInt *SelTC, *SelFC; |
| 158 | if (!match(V: TrueVal, P: m_APInt(Res&: SelTC)) || !match(V: FalseVal, P: m_APInt(Res&: SelFC))) |
| 159 | return nullptr; |
| 160 | |
| 161 | Type *SelType = Sel.getType(); |
| 162 | // In general, when both constants are non-zero, we would need an offset to |
| 163 | // replace the select. This would require more instructions than we started |
| 164 | // with. But there's one special-case that we handle here because it can |
| 165 | // simplify/reduce the instructions. |
| 166 | const APInt &TC = *SelTC; |
| 167 | const APInt &FC = *SelFC; |
| 168 | if (!TC.isZero() && !FC.isZero()) { |
| 169 | if (TC.getBitWidth() != AndMask.getBitWidth()) |
| 170 | return nullptr; |
| 171 | // If we have to create an 'and', then we must kill the cmp to not |
| 172 | // increase the instruction count. |
| 173 | if (CreateAnd && !CondVal->hasOneUse()) |
| 174 | return nullptr; |
| 175 | |
| 176 | // (V & AndMaskC) == 0 ? TC : FC --> TC | (V & AndMaskC) |
| 177 | // (V & AndMaskC) == 0 ? TC : FC --> TC ^ (V & AndMaskC) |
| 178 | // (V & AndMaskC) == 0 ? TC : FC --> TC + (V & AndMaskC) |
| 179 | // (V & AndMaskC) == 0 ? TC : FC --> TC - (V & AndMaskC) |
| 180 | Constant *TCC = ConstantInt::get(Ty: SelType, V: TC); |
| 181 | Constant *FCC = ConstantInt::get(Ty: SelType, V: FC); |
| 182 | Constant *MaskC = ConstantInt::get(Ty: SelType, V: AndMask); |
| 183 | for (auto Opc : {Instruction::Or, Instruction::Xor, Instruction::Add, |
| 184 | Instruction::Sub}) { |
| 185 | if (ConstantFoldBinaryOpOperands(Opcode: Opc, LHS: TCC, RHS: MaskC, DL: Sel.getDataLayout()) == |
| 186 | FCC) { |
| 187 | if (CreateAnd) |
| 188 | V = Builder.CreateAnd(LHS: V, RHS: MaskC); |
| 189 | return Builder.CreateBinOp(Opc, LHS: TCC, RHS: V); |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | return nullptr; |
| 194 | } |
| 195 | |
| 196 | // Make sure one of the select arms is a power-of-2. |
| 197 | if (!TC.isPowerOf2() && !FC.isPowerOf2()) |
| 198 | return nullptr; |
| 199 | |
| 200 | // Determine which shift is needed to transform result of the 'and' into the |
| 201 | // desired result. |
| 202 | const APInt &ValC = !TC.isZero() ? TC : FC; |
| 203 | unsigned ValZeros = ValC.logBase2(); |
| 204 | unsigned AndZeros = AndMask.logBase2(); |
| 205 | bool ShouldNotVal = !TC.isZero(); |
| 206 | bool NeedShift = ValZeros != AndZeros; |
| 207 | bool NeedZExtTrunc = |
| 208 | SelType->getScalarSizeInBits() != V->getType()->getScalarSizeInBits(); |
| 209 | |
| 210 | // If we would need to create an 'and' + 'shift' + 'xor' + cast to replace |
| 211 | // a 'select' + 'icmp', then this transformation would result in more |
| 212 | // instructions and potentially interfere with other folding. |
| 213 | if (CreateAnd + ShouldNotVal + NeedShift + NeedZExtTrunc > |
| 214 | 1 + CondVal->hasOneUse()) |
| 215 | return nullptr; |
| 216 | |
| 217 | // Insert the 'and' instruction on the input to the truncate. |
| 218 | if (CreateAnd) |
| 219 | V = Builder.CreateAnd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: AndMask)); |
| 220 | |
| 221 | // If types don't match, we can still convert the select by introducing a zext |
| 222 | // or a trunc of the 'and'. |
| 223 | if (ValZeros > AndZeros) { |
| 224 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
| 225 | V = Builder.CreateShl(LHS: V, RHS: ValZeros - AndZeros); |
| 226 | } else if (ValZeros < AndZeros) { |
| 227 | V = Builder.CreateLShr(LHS: V, RHS: AndZeros - ValZeros); |
| 228 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
| 229 | } else { |
| 230 | V = Builder.CreateZExtOrTrunc(V, DestTy: SelType); |
| 231 | } |
| 232 | |
| 233 | // Okay, now we know that everything is set up, we just don't know whether we |
| 234 | // have a icmp_ne or icmp_eq and whether the true or false val is the zero. |
| 235 | if (ShouldNotVal) |
| 236 | V = Builder.CreateXor(LHS: V, RHS: ValC); |
| 237 | |
| 238 | return V; |
| 239 | } |
| 240 | |
| 241 | /// We want to turn code that looks like this: |
| 242 | /// %C = or %A, %B |
| 243 | /// %D = select %cond, %C, %A |
| 244 | /// into: |
| 245 | /// %C = select %cond, %B, 0 |
| 246 | /// %D = or %A, %C |
| 247 | /// |
| 248 | /// Assuming that the specified instruction is an operand to the select, return |
| 249 | /// a bitmask indicating which operands of this instruction are foldable if they |
| 250 | /// equal the other incoming value of the select. |
| 251 | static unsigned getSelectFoldableOperands(BinaryOperator *I) { |
| 252 | switch (I->getOpcode()) { |
| 253 | case Instruction::Add: |
| 254 | case Instruction::FAdd: |
| 255 | case Instruction::Mul: |
| 256 | case Instruction::FMul: |
| 257 | case Instruction::And: |
| 258 | case Instruction::Or: |
| 259 | case Instruction::Xor: |
| 260 | return 3; // Can fold through either operand. |
| 261 | case Instruction::Sub: // Can only fold on the amount subtracted. |
| 262 | case Instruction::FSub: |
| 263 | case Instruction::FDiv: // Can only fold on the divisor amount. |
| 264 | case Instruction::Shl: // Can only fold on the shift amount. |
| 265 | case Instruction::LShr: |
| 266 | case Instruction::AShr: |
| 267 | return 1; |
| 268 | default: |
| 269 | return 0; // Cannot fold |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | /// We have (select c, TI, FI), and we know that TI and FI have the same opcode. |
| 274 | Instruction *InstCombinerImpl::foldSelectOpOp(SelectInst &SI, Instruction *TI, |
| 275 | Instruction *FI) { |
| 276 | // If this is a cast from the same type, merge. |
| 277 | Value *Cond = SI.getCondition(); |
| 278 | Type *CondTy = Cond->getType(); |
| 279 | if (TI->getNumOperands() == 1 && TI->isCast()) { |
| 280 | Type *FIOpndTy = FI->getOperand(i: 0)->getType(); |
| 281 | if (TI->getOperand(i: 0)->getType() != FIOpndTy) |
| 282 | return nullptr; |
| 283 | |
| 284 | // The select condition may be a vector. We may only change the operand |
| 285 | // type if the vector width remains the same (and matches the condition). |
| 286 | if (auto *CondVTy = dyn_cast<VectorType>(Val: CondTy)) { |
| 287 | if (!FIOpndTy->isVectorTy() || |
| 288 | CondVTy->getElementCount() != |
| 289 | cast<VectorType>(Val: FIOpndTy)->getElementCount()) |
| 290 | return nullptr; |
| 291 | |
| 292 | // TODO: If the backend knew how to deal with casts better, we could |
| 293 | // remove this limitation. For now, there's too much potential to create |
| 294 | // worse codegen by promoting the select ahead of size-altering casts |
| 295 | // (PR28160). |
| 296 | // |
| 297 | // Note that ValueTracking's matchSelectPattern() looks through casts |
| 298 | // without checking 'hasOneUse' when it matches min/max patterns, so this |
| 299 | // transform may end up happening anyway. |
| 300 | if (TI->getOpcode() != Instruction::BitCast && |
| 301 | (!TI->hasOneUse() || !FI->hasOneUse())) |
| 302 | return nullptr; |
| 303 | } else if (!TI->hasOneUse() || !FI->hasOneUse()) { |
| 304 | // TODO: The one-use restrictions for a scalar select could be eased if |
| 305 | // the fold of a select in visitLoadInst() was enhanced to match a pattern |
| 306 | // that includes a cast. |
| 307 | return nullptr; |
| 308 | } |
| 309 | |
| 310 | // Fold this by inserting a select from the input values. |
| 311 | Value *NewSI = |
| 312 | Builder.CreateSelect(C: Cond, True: TI->getOperand(i: 0), False: FI->getOperand(i: 0), |
| 313 | Name: SI.getName() + ".v" , MDFrom: &SI); |
| 314 | return CastInst::Create(Instruction::CastOps(TI->getOpcode()), S: NewSI, |
| 315 | Ty: TI->getType()); |
| 316 | } |
| 317 | |
| 318 | Value *OtherOpT, *OtherOpF; |
| 319 | bool MatchIsOpZero; |
| 320 | auto getCommonOp = [&](Instruction *TI, Instruction *FI, bool Commute, |
| 321 | bool Swapped = false) -> Value * { |
| 322 | assert(!(Commute && Swapped) && |
| 323 | "Commute and Swapped can't set at the same time" ); |
| 324 | if (!Swapped) { |
| 325 | if (TI->getOperand(i: 0) == FI->getOperand(i: 0)) { |
| 326 | OtherOpT = TI->getOperand(i: 1); |
| 327 | OtherOpF = FI->getOperand(i: 1); |
| 328 | MatchIsOpZero = true; |
| 329 | return TI->getOperand(i: 0); |
| 330 | } else if (TI->getOperand(i: 1) == FI->getOperand(i: 1)) { |
| 331 | OtherOpT = TI->getOperand(i: 0); |
| 332 | OtherOpF = FI->getOperand(i: 0); |
| 333 | MatchIsOpZero = false; |
| 334 | return TI->getOperand(i: 1); |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | if (!Commute && !Swapped) |
| 339 | return nullptr; |
| 340 | |
| 341 | // If we are allowing commute or swap of operands, then |
| 342 | // allow a cross-operand match. In that case, MatchIsOpZero |
| 343 | // means that TI's operand 0 (FI's operand 1) is the common op. |
| 344 | if (TI->getOperand(i: 0) == FI->getOperand(i: 1)) { |
| 345 | OtherOpT = TI->getOperand(i: 1); |
| 346 | OtherOpF = FI->getOperand(i: 0); |
| 347 | MatchIsOpZero = true; |
| 348 | return TI->getOperand(i: 0); |
| 349 | } else if (TI->getOperand(i: 1) == FI->getOperand(i: 0)) { |
| 350 | OtherOpT = TI->getOperand(i: 0); |
| 351 | OtherOpF = FI->getOperand(i: 1); |
| 352 | MatchIsOpZero = false; |
| 353 | return TI->getOperand(i: 1); |
| 354 | } |
| 355 | return nullptr; |
| 356 | }; |
| 357 | |
| 358 | if (TI->hasOneUse() || FI->hasOneUse()) { |
| 359 | // Cond ? -X : -Y --> -(Cond ? X : Y) |
| 360 | Value *X, *Y; |
| 361 | if (match(V: TI, P: m_FNeg(X: m_Value(V&: X))) && match(V: FI, P: m_FNeg(X: m_Value(V&: Y)))) { |
| 362 | // Intersect FMF from the fneg instructions and union those with the |
| 363 | // select. |
| 364 | FastMathFlags FMF = TI->getFastMathFlags(); |
| 365 | FMF &= FI->getFastMathFlags(); |
| 366 | FMF |= SI.getFastMathFlags(); |
| 367 | Value *NewSel = |
| 368 | Builder.CreateSelect(C: Cond, True: X, False: Y, Name: SI.getName() + ".v" , MDFrom: &SI); |
| 369 | if (auto *NewSelI = dyn_cast<Instruction>(Val: NewSel)) |
| 370 | NewSelI->setFastMathFlags(FMF); |
| 371 | Instruction *NewFNeg = UnaryOperator::CreateFNeg(V: NewSel); |
| 372 | NewFNeg->setFastMathFlags(FMF); |
| 373 | return NewFNeg; |
| 374 | } |
| 375 | |
| 376 | // Min/max intrinsic with a common operand can have the common operand |
| 377 | // pulled after the select. This is the same transform as below for binops, |
| 378 | // but specialized for intrinsic matching and without the restrictive uses |
| 379 | // clause. |
| 380 | auto *TII = dyn_cast<IntrinsicInst>(Val: TI); |
| 381 | auto *FII = dyn_cast<IntrinsicInst>(Val: FI); |
| 382 | if (TII && FII && TII->getIntrinsicID() == FII->getIntrinsicID()) { |
| 383 | if (match(V: TII, P: m_MaxOrMin(Op0: m_Value(), Op1: m_Value()))) { |
| 384 | if (Value *MatchOp = getCommonOp(TI, FI, true)) { |
| 385 | Value *NewSel = |
| 386 | Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, Name: "minmaxop" , MDFrom: &SI); |
| 387 | return CallInst::Create(Func: TII->getCalledFunction(), Args: {NewSel, MatchOp}); |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | // select c, (ldexp v, e0), (ldexp v, e1) -> ldexp v, (select c, e0, e1) |
| 392 | // select c, (ldexp v0, e), (ldexp v1, e) -> ldexp (select c, v0, v1), e |
| 393 | // |
| 394 | // select c, (ldexp v0, e0), (ldexp v1, e1) -> |
| 395 | // ldexp (select c, v0, v1), (select c, e0, e1) |
| 396 | if (TII->getIntrinsicID() == Intrinsic::ldexp) { |
| 397 | Value *LdexpVal0 = TII->getArgOperand(i: 0); |
| 398 | Value *LdexpExp0 = TII->getArgOperand(i: 1); |
| 399 | Value *LdexpVal1 = FII->getArgOperand(i: 0); |
| 400 | Value *LdexpExp1 = FII->getArgOperand(i: 1); |
| 401 | if (LdexpExp0->getType() == LdexpExp1->getType()) { |
| 402 | FPMathOperator *SelectFPOp = cast<FPMathOperator>(Val: &SI); |
| 403 | FastMathFlags FMF = cast<FPMathOperator>(Val: TII)->getFastMathFlags(); |
| 404 | FMF &= cast<FPMathOperator>(Val: FII)->getFastMathFlags(); |
| 405 | FMF |= SelectFPOp->getFastMathFlags(); |
| 406 | |
| 407 | Value *SelectVal = Builder.CreateSelect(C: Cond, True: LdexpVal0, False: LdexpVal1); |
| 408 | Value *SelectExp = Builder.CreateSelect(C: Cond, True: LdexpExp0, False: LdexpExp1); |
| 409 | |
| 410 | Value *NewLdexp = Builder.CreateIntrinsic( |
| 411 | RetTy: TII->getType(), ID: Intrinsic::ldexp, Args: {SelectVal, SelectExp}, FMFSource: FMF); |
| 412 | return replaceInstUsesWith(I&: SI, V: NewLdexp); |
| 413 | } |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | auto CreateCmpSel = [&](std::optional<CmpPredicate> P, |
| 418 | bool Swapped) -> CmpInst * { |
| 419 | if (!P) |
| 420 | return nullptr; |
| 421 | auto *MatchOp = getCommonOp(TI, FI, ICmpInst::isEquality(P: *P), |
| 422 | ICmpInst::isRelational(P: *P) && Swapped); |
| 423 | if (!MatchOp) |
| 424 | return nullptr; |
| 425 | Value *NewSel = Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, |
| 426 | Name: SI.getName() + ".v" , MDFrom: &SI); |
| 427 | return new ICmpInst(MatchIsOpZero ? *P |
| 428 | : ICmpInst::getSwappedCmpPredicate(Pred: *P), |
| 429 | MatchOp, NewSel); |
| 430 | }; |
| 431 | |
| 432 | // icmp with a common operand also can have the common operand |
| 433 | // pulled after the select. |
| 434 | CmpPredicate TPred, FPred; |
| 435 | if (match(V: TI, P: m_ICmp(Pred&: TPred, L: m_Value(), R: m_Value())) && |
| 436 | match(V: FI, P: m_ICmp(Pred&: FPred, L: m_Value(), R: m_Value()))) { |
| 437 | if (auto *R = |
| 438 | CreateCmpSel(CmpPredicate::getMatching(A: TPred, B: FPred), false)) |
| 439 | return R; |
| 440 | if (auto *R = |
| 441 | CreateCmpSel(CmpPredicate::getMatching( |
| 442 | A: TPred, B: ICmpInst::getSwappedCmpPredicate(Pred: FPred)), |
| 443 | true)) |
| 444 | return R; |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | // Only handle binary operators (including two-operand getelementptr) with |
| 449 | // one-use here. As with the cast case above, it may be possible to relax the |
| 450 | // one-use constraint, but that needs be examined carefully since it may not |
| 451 | // reduce the total number of instructions. |
| 452 | if (TI->getNumOperands() != 2 || FI->getNumOperands() != 2 || |
| 453 | !TI->isSameOperationAs(I: FI) || |
| 454 | (!isa<BinaryOperator>(Val: TI) && !isa<GetElementPtrInst>(Val: TI)) || |
| 455 | !TI->hasOneUse() || !FI->hasOneUse()) |
| 456 | return nullptr; |
| 457 | |
| 458 | // Figure out if the operations have any operands in common. |
| 459 | Value *MatchOp = getCommonOp(TI, FI, TI->isCommutative()); |
| 460 | if (!MatchOp) |
| 461 | return nullptr; |
| 462 | |
| 463 | // If the select condition is a vector, the operands of the original select's |
| 464 | // operands also must be vectors. This may not be the case for getelementptr |
| 465 | // for example. |
| 466 | if (CondTy->isVectorTy() && (!OtherOpT->getType()->isVectorTy() || |
| 467 | !OtherOpF->getType()->isVectorTy())) |
| 468 | return nullptr; |
| 469 | |
| 470 | // If we are sinking div/rem after a select, we may need to freeze the |
| 471 | // condition because div/rem may induce immediate UB with a poison operand. |
| 472 | // For example, the following transform is not safe if Cond can ever be poison |
| 473 | // because we can replace poison with zero and then we have div-by-zero that |
| 474 | // didn't exist in the original code: |
| 475 | // Cond ? x/y : x/z --> x / (Cond ? y : z) |
| 476 | auto *BO = dyn_cast<BinaryOperator>(Val: TI); |
| 477 | if (BO && BO->isIntDivRem() && !isGuaranteedNotToBePoison(V: Cond)) { |
| 478 | // A udiv/urem with a common divisor is safe because UB can only occur with |
| 479 | // div-by-zero, and that would be present in the original code. |
| 480 | if (BO->getOpcode() == Instruction::SDiv || |
| 481 | BO->getOpcode() == Instruction::SRem || MatchIsOpZero) |
| 482 | Cond = Builder.CreateFreeze(V: Cond); |
| 483 | } |
| 484 | |
| 485 | // If we reach here, they do have operations in common. |
| 486 | Value *NewSI = Builder.CreateSelect(C: Cond, True: OtherOpT, False: OtherOpF, |
| 487 | Name: SI.getName() + ".v" , MDFrom: &SI); |
| 488 | Value *Op0 = MatchIsOpZero ? MatchOp : NewSI; |
| 489 | Value *Op1 = MatchIsOpZero ? NewSI : MatchOp; |
| 490 | if (auto *BO = dyn_cast<BinaryOperator>(Val: TI)) { |
| 491 | BinaryOperator *NewBO = BinaryOperator::Create(Op: BO->getOpcode(), S1: Op0, S2: Op1); |
| 492 | NewBO->copyIRFlags(V: TI); |
| 493 | NewBO->andIRFlags(V: FI); |
| 494 | return NewBO; |
| 495 | } |
| 496 | if (auto *TGEP = dyn_cast<GetElementPtrInst>(Val: TI)) { |
| 497 | auto *FGEP = cast<GetElementPtrInst>(Val: FI); |
| 498 | Type *ElementType = TGEP->getSourceElementType(); |
| 499 | return GetElementPtrInst::Create( |
| 500 | PointeeType: ElementType, Ptr: Op0, IdxList: Op1, NW: TGEP->getNoWrapFlags() & FGEP->getNoWrapFlags()); |
| 501 | } |
| 502 | llvm_unreachable("Expected BinaryOperator or GEP" ); |
| 503 | return nullptr; |
| 504 | } |
| 505 | |
| 506 | /// This transforms patterns of the form: |
| 507 | /// select cond, intrinsic(x, ...), intrinsic(y, ...) |
| 508 | /// into: |
| 509 | /// intrinsic(select cond, x, y, ...) |
| 510 | Instruction *InstCombinerImpl::foldSelectIntrinsic(SelectInst &SI) { |
| 511 | auto *LHSIntrinsic = dyn_cast<IntrinsicInst>(Val: SI.getTrueValue()); |
| 512 | if (!LHSIntrinsic) |
| 513 | return nullptr; |
| 514 | auto *RHSIntrinsic = dyn_cast<IntrinsicInst>(Val: SI.getFalseValue()); |
| 515 | if (!RHSIntrinsic || |
| 516 | LHSIntrinsic->getIntrinsicID() != RHSIntrinsic->getIntrinsicID() || |
| 517 | !LHSIntrinsic->hasOneUse() || !RHSIntrinsic->hasOneUse()) |
| 518 | return nullptr; |
| 519 | |
| 520 | const Intrinsic::ID IID = LHSIntrinsic->getIntrinsicID(); |
| 521 | switch (IID) { |
| 522 | case Intrinsic::abs: |
| 523 | case Intrinsic::cttz: |
| 524 | case Intrinsic::ctlz: { |
| 525 | auto *TZ = cast<ConstantInt>(Val: LHSIntrinsic->getArgOperand(i: 1)); |
| 526 | auto *FZ = cast<ConstantInt>(Val: RHSIntrinsic->getArgOperand(i: 1)); |
| 527 | |
| 528 | Value *TV = LHSIntrinsic->getArgOperand(i: 0); |
| 529 | Value *FV = RHSIntrinsic->getArgOperand(i: 0); |
| 530 | |
| 531 | Value *NewSel = Builder.CreateSelect(C: SI.getCondition(), True: TV, False: FV, Name: "" , MDFrom: &SI); |
| 532 | Value *NewPoisonFlag = Builder.CreateAnd(LHS: TZ, RHS: FZ); |
| 533 | Value *NewCall = Builder.CreateBinaryIntrinsic(ID: IID, LHS: NewSel, RHS: NewPoisonFlag); |
| 534 | |
| 535 | return replaceInstUsesWith(I&: SI, V: NewCall); |
| 536 | } |
| 537 | case Intrinsic::ctpop: { |
| 538 | Value *TV = LHSIntrinsic->getArgOperand(i: 0); |
| 539 | Value *FV = RHSIntrinsic->getArgOperand(i: 0); |
| 540 | |
| 541 | Value *NewSel = Builder.CreateSelect(C: SI.getCondition(), True: TV, False: FV, Name: "" , MDFrom: &SI); |
| 542 | Value *NewCall = Builder.CreateUnaryIntrinsic(ID: IID, Op: NewSel); |
| 543 | |
| 544 | return replaceInstUsesWith(I&: SI, V: NewCall); |
| 545 | } |
| 546 | default: |
| 547 | return nullptr; |
| 548 | } |
| 549 | } |
| 550 | |
| 551 | static bool isSelect01(const APInt &C1I, const APInt &C2I) { |
| 552 | if (!C1I.isZero() && !C2I.isZero()) // One side must be zero. |
| 553 | return false; |
| 554 | return C1I.isOne() || C1I.isAllOnes() || C2I.isOne() || C2I.isAllOnes(); |
| 555 | } |
| 556 | |
| 557 | /// Try to fold the select into one of the operands to allow further |
| 558 | /// optimization. |
| 559 | Instruction *InstCombinerImpl::foldSelectIntoOp(SelectInst &SI, Value *TrueVal, |
| 560 | Value *FalseVal) { |
| 561 | // See the comment above getSelectFoldableOperands for a description of the |
| 562 | // transformation we are doing here. |
| 563 | auto TryFoldSelectIntoOp = [&](SelectInst &SI, Value *TrueVal, |
| 564 | Value *FalseVal, |
| 565 | bool Swapped) -> Instruction * { |
| 566 | auto *TVI = dyn_cast<BinaryOperator>(Val: TrueVal); |
| 567 | if (!TVI || !TVI->hasOneUse() || isa<Constant>(Val: FalseVal)) |
| 568 | return nullptr; |
| 569 | |
| 570 | unsigned SFO = getSelectFoldableOperands(I: TVI); |
| 571 | unsigned OpToFold = 0; |
| 572 | if ((SFO & 1) && FalseVal == TVI->getOperand(i_nocapture: 0)) |
| 573 | OpToFold = 1; |
| 574 | else if ((SFO & 2) && FalseVal == TVI->getOperand(i_nocapture: 1)) |
| 575 | OpToFold = 2; |
| 576 | |
| 577 | if (!OpToFold) |
| 578 | return nullptr; |
| 579 | |
| 580 | FastMathFlags FMF; |
| 581 | if (const auto *FPO = dyn_cast<FPMathOperator>(Val: &SI)) |
| 582 | FMF = FPO->getFastMathFlags(); |
| 583 | Constant *C = ConstantExpr::getBinOpIdentity( |
| 584 | Opcode: TVI->getOpcode(), Ty: TVI->getType(), AllowRHSConstant: true, NSZ: FMF.noSignedZeros()); |
| 585 | Value *OOp = TVI->getOperand(i_nocapture: 2 - OpToFold); |
| 586 | // Avoid creating select between 2 constants unless it's selecting |
| 587 | // between 0, 1 and -1. |
| 588 | const APInt *OOpC; |
| 589 | bool OOpIsAPInt = match(V: OOp, P: m_APInt(Res&: OOpC)); |
| 590 | if (isa<Constant>(Val: OOp) && |
| 591 | (!OOpIsAPInt || !isSelect01(C1I: C->getUniqueInteger(), C2I: *OOpC))) |
| 592 | return nullptr; |
| 593 | |
| 594 | // If the false value is a NaN then we have that the floating point math |
| 595 | // operation in the transformed code may not preserve the exact NaN |
| 596 | // bit-pattern -- e.g. `fadd sNaN, 0.0 -> qNaN`. |
| 597 | // This makes the transformation incorrect since the original program would |
| 598 | // have preserved the exact NaN bit-pattern. |
| 599 | // Avoid the folding if the false value might be a NaN. |
| 600 | if (isa<FPMathOperator>(Val: &SI) && |
| 601 | !computeKnownFPClass(V: FalseVal, FMF, InterestedClasses: fcNan, SQ: SQ.getWithInstruction(I: &SI)) |
| 602 | .isKnownNeverNaN()) |
| 603 | return nullptr; |
| 604 | |
| 605 | Value *NewSel = Builder.CreateSelect(C: SI.getCondition(), True: Swapped ? C : OOp, |
| 606 | False: Swapped ? OOp : C, Name: "" , MDFrom: &SI); |
| 607 | if (isa<FPMathOperator>(Val: &SI)) { |
| 608 | FastMathFlags NewSelFMF = FMF; |
| 609 | // We cannot propagate ninf from the original select, because OOp may be |
| 610 | // inf and the flag only guarantees that FalseVal (op OOp) is never |
| 611 | // infinity. |
| 612 | // Examples: -inf + +inf = NaN, -inf - -inf = NaN, 0 * inf = NaN |
| 613 | // Specifically, if the original select has both ninf and nnan, we can |
| 614 | // safely propagate the flag. |
| 615 | // Note: This property holds for fadd, fsub, and fmul, but does not |
| 616 | // hold for fdiv (e.g. A / Inf == 0.0). |
| 617 | bool CanInferFiniteOperandsFromResult = |
| 618 | TVI->getOpcode() == Instruction::FAdd || |
| 619 | TVI->getOpcode() == Instruction::FSub || |
| 620 | TVI->getOpcode() == Instruction::FMul; |
| 621 | NewSelFMF.setNoInfs(TVI->hasNoInfs() || |
| 622 | (CanInferFiniteOperandsFromResult && |
| 623 | NewSelFMF.noInfs() && NewSelFMF.noNaNs())); |
| 624 | cast<Instruction>(Val: NewSel)->setFastMathFlags(NewSelFMF); |
| 625 | } |
| 626 | NewSel->takeName(V: TVI); |
| 627 | BinaryOperator *BO = |
| 628 | BinaryOperator::Create(Op: TVI->getOpcode(), S1: FalseVal, S2: NewSel); |
| 629 | BO->copyIRFlags(V: TVI); |
| 630 | if (isa<FPMathOperator>(Val: &SI)) { |
| 631 | // Merge poison generating flags from the select. |
| 632 | BO->setHasNoNaNs(BO->hasNoNaNs() && FMF.noNaNs()); |
| 633 | BO->setHasNoInfs(BO->hasNoInfs() && FMF.noInfs()); |
| 634 | // Merge no-signed-zeros flag from the select. |
| 635 | // Otherwise we may produce zeros with different sign. |
| 636 | BO->setHasNoSignedZeros(BO->hasNoSignedZeros() && FMF.noSignedZeros()); |
| 637 | } |
| 638 | return BO; |
| 639 | }; |
| 640 | |
| 641 | if (Instruction *R = TryFoldSelectIntoOp(SI, TrueVal, FalseVal, false)) |
| 642 | return R; |
| 643 | |
| 644 | if (Instruction *R = TryFoldSelectIntoOp(SI, FalseVal, TrueVal, true)) |
| 645 | return R; |
| 646 | |
| 647 | return nullptr; |
| 648 | } |
| 649 | |
| 650 | static Value *canoncalizeSelectICmpMinMax(const ICmpInst *Cmp, Value *TVal, |
| 651 | Value *FVal, |
| 652 | InstCombiner::BuilderTy &Builder, |
| 653 | const SimplifyQuery &SQ) { |
| 654 | Value *CmpLHS = Cmp->getOperand(i_nocapture: 0); |
| 655 | Value *CmpRHS = Cmp->getOperand(i_nocapture: 1); |
| 656 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
| 657 | if (match(V: FVal, P: m_Zero())) { |
| 658 | std::swap(a&: TVal, b&: FVal); |
| 659 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
| 660 | } |
| 661 | if (!match(V: TVal, P: m_Zero())) |
| 662 | return nullptr; |
| 663 | |
| 664 | if (Pred == CmpInst::ICMP_SGT || Pred == CmpInst::ICMP_SGE) { |
| 665 | std::swap(a&: CmpLHS, b&: CmpRHS); |
| 666 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 667 | } |
| 668 | |
| 669 | // Handles: |
| 670 | // (X <= Y) ? 0 : (X - Y) |
| 671 | // (X <= Y) ? (Y - X) : 0 |
| 672 | // (X >= Y) ? 0 : (Y - X) |
| 673 | // (X >= Y) ? (X - Y) : 0 |
| 674 | if ((Pred == CmpInst::ICMP_SLT || Pred == CmpInst::ICMP_SLE) && |
| 675 | match(V: FVal, P: m_NSWSub(L: m_Specific(V: CmpLHS), R: m_Specific(V: CmpRHS))) && |
| 676 | isGuaranteedNotToBeUndef(V: CmpLHS, AC: SQ.AC, CtxI: SQ.CxtI, DT: SQ.DT)) { |
| 677 | Value *SMin = |
| 678 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::smin, LHS: CmpRHS, RHS: CmpLHS); |
| 679 | return Builder.CreateNSWSub(LHS: CmpLHS, RHS: SMin); |
| 680 | } |
| 681 | |
| 682 | return nullptr; |
| 683 | } |
| 684 | |
| 685 | /// Try to fold a select to a min/max intrinsic. Many cases are already handled |
| 686 | /// by matchDecomposedSelectPattern but here we handle the cases where more |
| 687 | /// extensive modification of the IR is required. |
| 688 | static Value *foldSelectICmpMinMax(const ICmpInst *Cmp, Value *TVal, |
| 689 | Value *FVal, |
| 690 | InstCombiner::BuilderTy &Builder, |
| 691 | const SimplifyQuery &SQ) { |
| 692 | Value *CmpLHS = Cmp->getOperand(i_nocapture: 0); |
| 693 | Value *CmpRHS = Cmp->getOperand(i_nocapture: 1); |
| 694 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
| 695 | |
| 696 | if (Value *V = canoncalizeSelectICmpMinMax(Cmp, TVal, FVal, Builder, SQ)) |
| 697 | return V; |
| 698 | |
| 699 | // (X > Y) ? X : (Y - 1) ==> MIN(X, Y - 1) |
| 700 | // (X < Y) ? X : (Y + 1) ==> MAX(X, Y + 1) |
| 701 | // This transformation is valid when overflow corresponding to the sign of |
| 702 | // the comparison is poison and we must drop the non-matching overflow flag. |
| 703 | if (CmpRHS == TVal) { |
| 704 | std::swap(a&: CmpLHS, b&: CmpRHS); |
| 705 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 706 | } |
| 707 | |
| 708 | // TODO: consider handling 'or disjoint' as well, though these would need to |
| 709 | // be converted to 'add' instructions. |
| 710 | if (!(CmpLHS == TVal && isa<Instruction>(Val: FVal))) |
| 711 | return nullptr; |
| 712 | |
| 713 | if (Pred == CmpInst::ICMP_SGT && |
| 714 | match(V: FVal, P: m_NSWAdd(L: m_Specific(V: CmpRHS), R: m_One()))) { |
| 715 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
| 716 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smax, LHS: TVal, RHS: FVal); |
| 717 | } |
| 718 | |
| 719 | if (Pred == CmpInst::ICMP_SLT && |
| 720 | match(V: FVal, P: m_NSWAdd(L: m_Specific(V: CmpRHS), R: m_AllOnes()))) { |
| 721 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
| 722 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smin, LHS: TVal, RHS: FVal); |
| 723 | } |
| 724 | |
| 725 | if (Pred == CmpInst::ICMP_UGT && |
| 726 | match(V: FVal, P: m_NUWAdd(L: m_Specific(V: CmpRHS), R: m_One()))) { |
| 727 | cast<Instruction>(Val: FVal)->setHasNoSignedWrap(false); |
| 728 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: TVal, RHS: FVal); |
| 729 | } |
| 730 | |
| 731 | // Note: We must use isKnownNonZero here because "sub nuw %x, 1" will be |
| 732 | // canonicalized to "add %x, -1" discarding the nuw flag. |
| 733 | if (Pred == CmpInst::ICMP_ULT && |
| 734 | match(V: FVal, P: m_Add(L: m_Specific(V: CmpRHS), R: m_AllOnes())) && |
| 735 | isKnownNonZero(V: CmpRHS, Q: SQ)) { |
| 736 | cast<Instruction>(Val: FVal)->setHasNoSignedWrap(false); |
| 737 | cast<Instruction>(Val: FVal)->setHasNoUnsignedWrap(false); |
| 738 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: TVal, RHS: FVal); |
| 739 | } |
| 740 | |
| 741 | return nullptr; |
| 742 | } |
| 743 | |
| 744 | /// We want to turn: |
| 745 | /// (select (icmp eq (and X, Y), 0), (and (lshr X, Z), 1), 1) |
| 746 | /// into: |
| 747 | /// zext (icmp ne i32 (and X, (or Y, (shl 1, Z))), 0) |
| 748 | /// Note: |
| 749 | /// Z may be 0 if lshr is missing. |
| 750 | /// Worst-case scenario is that we will replace 5 instructions with 5 different |
| 751 | /// instructions, but we got rid of select. |
| 752 | static Instruction *foldSelectICmpAndAnd(Type *SelType, const ICmpInst *Cmp, |
| 753 | Value *TVal, Value *FVal, |
| 754 | InstCombiner::BuilderTy &Builder) { |
| 755 | if (!(Cmp->hasOneUse() && Cmp->getOperand(i_nocapture: 0)->hasOneUse() && |
| 756 | Cmp->getPredicate() == ICmpInst::ICMP_EQ && |
| 757 | match(V: Cmp->getOperand(i_nocapture: 1), P: m_Zero()) && match(V: FVal, P: m_One()))) |
| 758 | return nullptr; |
| 759 | |
| 760 | // The TrueVal has general form of: and %B, 1 |
| 761 | Value *B; |
| 762 | if (!match(V: TVal, P: m_OneUse(SubPattern: m_And(L: m_Value(V&: B), R: m_One())))) |
| 763 | return nullptr; |
| 764 | |
| 765 | // Where %B may be optionally shifted: lshr %X, %Z. |
| 766 | Value *X, *Z; |
| 767 | const bool HasShift = match(V: B, P: m_OneUse(SubPattern: m_LShr(L: m_Value(V&: X), R: m_Value(V&: Z)))); |
| 768 | |
| 769 | // The shift must be valid. |
| 770 | // TODO: This restricts the fold to constant shift amounts. Is there a way to |
| 771 | // handle variable shifts safely? PR47012 |
| 772 | if (HasShift && |
| 773 | !match(V: Z, P: m_SpecificInt_ICMP(Predicate: CmpInst::ICMP_ULT, |
| 774 | Threshold: APInt(SelType->getScalarSizeInBits(), |
| 775 | SelType->getScalarSizeInBits())))) |
| 776 | return nullptr; |
| 777 | |
| 778 | if (!HasShift) |
| 779 | X = B; |
| 780 | |
| 781 | Value *Y; |
| 782 | if (!match(V: Cmp->getOperand(i_nocapture: 0), P: m_c_And(L: m_Specific(V: X), R: m_Value(V&: Y)))) |
| 783 | return nullptr; |
| 784 | |
| 785 | // ((X & Y) == 0) ? ((X >> Z) & 1) : 1 --> (X & (Y | (1 << Z))) != 0 |
| 786 | // ((X & Y) == 0) ? (X & 1) : 1 --> (X & (Y | 1)) != 0 |
| 787 | Constant *One = ConstantInt::get(Ty: SelType, V: 1); |
| 788 | Value *MaskB = HasShift ? Builder.CreateShl(LHS: One, RHS: Z) : One; |
| 789 | Value *FullMask = Builder.CreateOr(LHS: Y, RHS: MaskB); |
| 790 | Value *MaskedX = Builder.CreateAnd(LHS: X, RHS: FullMask); |
| 791 | Value *ICmpNeZero = Builder.CreateIsNotNull(Arg: MaskedX); |
| 792 | return new ZExtInst(ICmpNeZero, SelType); |
| 793 | } |
| 794 | |
| 795 | /// We want to turn: |
| 796 | /// (select (icmp eq (and X, C1), 0), 0, (shl [nsw/nuw] X, C2)); |
| 797 | /// iff C1 is a mask and the number of its leading zeros is equal to C2 |
| 798 | /// into: |
| 799 | /// shl X, C2 |
| 800 | static Value *foldSelectICmpAndZeroShl(const ICmpInst *Cmp, Value *TVal, |
| 801 | Value *FVal, |
| 802 | InstCombiner::BuilderTy &Builder) { |
| 803 | CmpPredicate Pred; |
| 804 | Value *AndVal; |
| 805 | if (!match(V: Cmp, P: m_ICmp(Pred, L: m_Value(V&: AndVal), R: m_Zero()))) |
| 806 | return nullptr; |
| 807 | |
| 808 | if (Pred == ICmpInst::ICMP_NE) { |
| 809 | Pred = ICmpInst::ICMP_EQ; |
| 810 | std::swap(a&: TVal, b&: FVal); |
| 811 | } |
| 812 | |
| 813 | Value *X; |
| 814 | const APInt *C2, *C1; |
| 815 | if (Pred != ICmpInst::ICMP_EQ || |
| 816 | !match(V: AndVal, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: C1))) || |
| 817 | !match(V: TVal, P: m_Zero()) || !match(V: FVal, P: m_Shl(L: m_Specific(V: X), R: m_APInt(Res&: C2)))) |
| 818 | return nullptr; |
| 819 | |
| 820 | if (!C1->isMask() || |
| 821 | C1->countLeadingZeros() != static_cast<unsigned>(C2->getZExtValue())) |
| 822 | return nullptr; |
| 823 | |
| 824 | auto *FI = dyn_cast<Instruction>(Val: FVal); |
| 825 | if (!FI) |
| 826 | return nullptr; |
| 827 | |
| 828 | FI->setHasNoSignedWrap(false); |
| 829 | FI->setHasNoUnsignedWrap(false); |
| 830 | return FVal; |
| 831 | } |
| 832 | |
| 833 | /// We want to turn: |
| 834 | /// (select (icmp sgt x, C), lshr (X, Y), ashr (X, Y)); iff C s>= -1 |
| 835 | /// (select (icmp slt x, C), ashr (X, Y), lshr (X, Y)); iff C s>= 0 |
| 836 | /// into: |
| 837 | /// ashr (X, Y) |
| 838 | static Value *foldSelectICmpLshrAshr(const ICmpInst *IC, Value *TrueVal, |
| 839 | Value *FalseVal, |
| 840 | InstCombiner::BuilderTy &Builder) { |
| 841 | ICmpInst::Predicate Pred = IC->getPredicate(); |
| 842 | Value *CmpLHS = IC->getOperand(i_nocapture: 0); |
| 843 | Value *CmpRHS = IC->getOperand(i_nocapture: 1); |
| 844 | if (!CmpRHS->getType()->isIntOrIntVectorTy()) |
| 845 | return nullptr; |
| 846 | |
| 847 | Value *X, *Y; |
| 848 | unsigned Bitwidth = CmpRHS->getType()->getScalarSizeInBits(); |
| 849 | if ((Pred != ICmpInst::ICMP_SGT || |
| 850 | !match(V: CmpRHS, P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_SGE, |
| 851 | Threshold: APInt::getAllOnes(numBits: Bitwidth)))) && |
| 852 | (Pred != ICmpInst::ICMP_SLT || |
| 853 | !match(V: CmpRHS, P: m_SpecificInt_ICMP(Predicate: ICmpInst::ICMP_SGE, |
| 854 | Threshold: APInt::getZero(numBits: Bitwidth))))) |
| 855 | return nullptr; |
| 856 | |
| 857 | // Canonicalize so that ashr is in FalseVal. |
| 858 | if (Pred == ICmpInst::ICMP_SLT) |
| 859 | std::swap(a&: TrueVal, b&: FalseVal); |
| 860 | |
| 861 | if (match(V: TrueVal, P: m_LShr(L: m_Value(V&: X), R: m_Value(V&: Y))) && |
| 862 | match(V: FalseVal, P: m_AShr(L: m_Specific(V: X), R: m_Specific(V: Y))) && |
| 863 | match(V: CmpLHS, P: m_Specific(V: X))) { |
| 864 | const auto *Ashr = cast<Instruction>(Val: FalseVal); |
| 865 | // if lshr is not exact and ashr is, this new ashr must not be exact. |
| 866 | bool IsExact = Ashr->isExact() && cast<Instruction>(Val: TrueVal)->isExact(); |
| 867 | return Builder.CreateAShr(LHS: X, RHS: Y, Name: IC->getName(), isExact: IsExact); |
| 868 | } |
| 869 | |
| 870 | return nullptr; |
| 871 | } |
| 872 | |
| 873 | /// We want to turn: |
| 874 | /// (select (icmp eq (and X, C1), 0), Y, (BinOp Y, C2)) |
| 875 | /// into: |
| 876 | /// IF C2 u>= C1 |
| 877 | /// (BinOp Y, (shl (and X, C1), C3)) |
| 878 | /// ELSE |
| 879 | /// (BinOp Y, (lshr (and X, C1), C3)) |
| 880 | /// iff: |
| 881 | /// 0 on the RHS is the identity value (i.e add, xor, shl, etc...) |
| 882 | /// C1 and C2 are both powers of 2 |
| 883 | /// where: |
| 884 | /// IF C2 u>= C1 |
| 885 | /// C3 = Log(C2) - Log(C1) |
| 886 | /// ELSE |
| 887 | /// C3 = Log(C1) - Log(C2) |
| 888 | /// |
| 889 | /// This transform handles cases where: |
| 890 | /// 1. The icmp predicate is inverted |
| 891 | /// 2. The select operands are reversed |
| 892 | /// 3. The magnitude of C2 and C1 are flipped |
| 893 | static Value *foldSelectICmpAndBinOp(Value *CondVal, Value *TrueVal, |
| 894 | Value *FalseVal, Value *V, |
| 895 | const APInt &AndMask, bool CreateAnd, |
| 896 | InstCombiner::BuilderTy &Builder) { |
| 897 | // Only handle integer compares. |
| 898 | if (!TrueVal->getType()->isIntOrIntVectorTy()) |
| 899 | return nullptr; |
| 900 | |
| 901 | unsigned C1Log = AndMask.logBase2(); |
| 902 | Value *Y; |
| 903 | BinaryOperator *BinOp; |
| 904 | const APInt *C2; |
| 905 | bool NeedXor; |
| 906 | if (match(V: FalseVal, P: m_BinOp(L: m_Specific(V: TrueVal), R: m_Power2(V&: C2)))) { |
| 907 | Y = TrueVal; |
| 908 | BinOp = cast<BinaryOperator>(Val: FalseVal); |
| 909 | NeedXor = false; |
| 910 | } else if (match(V: TrueVal, P: m_BinOp(L: m_Specific(V: FalseVal), R: m_Power2(V&: C2)))) { |
| 911 | Y = FalseVal; |
| 912 | BinOp = cast<BinaryOperator>(Val: TrueVal); |
| 913 | NeedXor = true; |
| 914 | } else { |
| 915 | return nullptr; |
| 916 | } |
| 917 | |
| 918 | // Check that 0 on RHS is identity value for this binop. |
| 919 | auto *IdentityC = |
| 920 | ConstantExpr::getBinOpIdentity(Opcode: BinOp->getOpcode(), Ty: BinOp->getType(), |
| 921 | /*AllowRHSConstant*/ true); |
| 922 | if (IdentityC == nullptr || !IdentityC->isNullValue()) |
| 923 | return nullptr; |
| 924 | |
| 925 | unsigned C2Log = C2->logBase2(); |
| 926 | |
| 927 | bool NeedShift = C1Log != C2Log; |
| 928 | bool NeedZExtTrunc = Y->getType()->getScalarSizeInBits() != |
| 929 | V->getType()->getScalarSizeInBits(); |
| 930 | |
| 931 | // the demanded bits for the created shl make the and redundant |
| 932 | if (AndMask.isOne() && C2->isSignBitSet()) |
| 933 | CreateAnd = false; |
| 934 | |
| 935 | // Make sure we don't create more instructions than we save. |
| 936 | if ((NeedShift + NeedXor + NeedZExtTrunc + CreateAnd) > |
| 937 | (CondVal->hasOneUse() + BinOp->hasOneUse())) |
| 938 | return nullptr; |
| 939 | |
| 940 | if (CreateAnd) { |
| 941 | // Insert the AND instruction on the input to the truncate. |
| 942 | V = Builder.CreateAnd(LHS: V, RHS: ConstantInt::get(Ty: V->getType(), V: AndMask)); |
| 943 | } |
| 944 | |
| 945 | if (C2Log > C1Log) { |
| 946 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
| 947 | V = Builder.CreateShl(LHS: V, RHS: C2Log - C1Log); |
| 948 | } else if (C1Log > C2Log) { |
| 949 | V = Builder.CreateLShr(LHS: V, RHS: C1Log - C2Log); |
| 950 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
| 951 | } else |
| 952 | V = Builder.CreateZExtOrTrunc(V, DestTy: Y->getType()); |
| 953 | |
| 954 | if (NeedXor) |
| 955 | V = Builder.CreateXor(LHS: V, RHS: *C2); |
| 956 | |
| 957 | auto *Res = Builder.CreateBinOp(Opc: BinOp->getOpcode(), LHS: Y, RHS: V); |
| 958 | if (auto *BO = dyn_cast<BinaryOperator>(Val: Res)) |
| 959 | BO->copyIRFlags(V: BinOp); |
| 960 | return Res; |
| 961 | } |
| 962 | |
| 963 | /// Canonicalize a set or clear of a masked set of constant bits to |
| 964 | /// select-of-constants form. |
| 965 | static Instruction *foldSetClearBits(SelectInst &Sel, |
| 966 | InstCombiner::BuilderTy &Builder) { |
| 967 | Value *Cond = Sel.getCondition(); |
| 968 | Value *T = Sel.getTrueValue(); |
| 969 | Value *F = Sel.getFalseValue(); |
| 970 | Type *Ty = Sel.getType(); |
| 971 | Value *X; |
| 972 | const APInt *NotC, *C; |
| 973 | |
| 974 | // Cond ? (X & ~C) : (X | C) --> (X & ~C) | (Cond ? 0 : C) |
| 975 | if (match(V: T, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: NotC))) && |
| 976 | match(V: F, P: m_OneUse(SubPattern: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C)))) && *NotC == ~(*C)) { |
| 977 | Constant *Zero = ConstantInt::getNullValue(Ty); |
| 978 | Constant *OrC = ConstantInt::get(Ty, V: *C); |
| 979 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Zero, False: OrC, Name: "masksel" , MDFrom: &Sel); |
| 980 | return BinaryOperator::CreateOr(V1: T, V2: NewSel); |
| 981 | } |
| 982 | |
| 983 | // Cond ? (X | C) : (X & ~C) --> (X & ~C) | (Cond ? C : 0) |
| 984 | if (match(V: F, P: m_And(L: m_Value(V&: X), R: m_APInt(Res&: NotC))) && |
| 985 | match(V: T, P: m_OneUse(SubPattern: m_Or(L: m_Specific(V: X), R: m_APInt(Res&: C)))) && *NotC == ~(*C)) { |
| 986 | Constant *Zero = ConstantInt::getNullValue(Ty); |
| 987 | Constant *OrC = ConstantInt::get(Ty, V: *C); |
| 988 | Value *NewSel = Builder.CreateSelect(C: Cond, True: OrC, False: Zero, Name: "masksel" , MDFrom: &Sel); |
| 989 | return BinaryOperator::CreateOr(V1: F, V2: NewSel); |
| 990 | } |
| 991 | |
| 992 | return nullptr; |
| 993 | } |
| 994 | |
| 995 | // select (x == 0), 0, x * y --> freeze(y) * x |
| 996 | // select (y == 0), 0, x * y --> freeze(x) * y |
| 997 | // select (x == 0), undef, x * y --> freeze(y) * x |
| 998 | // select (x == undef), 0, x * y --> freeze(y) * x |
| 999 | // Usage of mul instead of 0 will make the result more poisonous, |
| 1000 | // so the operand that was not checked in the condition should be frozen. |
| 1001 | // The latter folding is applied only when a constant compared with x is |
| 1002 | // is a vector consisting of 0 and undefs. If a constant compared with x |
| 1003 | // is a scalar undefined value or undefined vector then an expression |
| 1004 | // should be already folded into a constant. |
| 1005 | // |
| 1006 | // This also holds all operations such that Op(0) == 0 |
| 1007 | // e.g. Shl, Umin, etc |
| 1008 | static Instruction *foldSelectZeroOrFixedOp(SelectInst &SI, |
| 1009 | InstCombinerImpl &IC) { |
| 1010 | auto *CondVal = SI.getCondition(); |
| 1011 | auto *TrueVal = SI.getTrueValue(); |
| 1012 | auto *FalseVal = SI.getFalseValue(); |
| 1013 | Value *X, *Y; |
| 1014 | CmpPredicate Predicate; |
| 1015 | |
| 1016 | // Assuming that constant compared with zero is not undef (but it may be |
| 1017 | // a vector with some undef elements). Otherwise (when a constant is undef) |
| 1018 | // the select expression should be already simplified. |
| 1019 | if (!match(V: CondVal, P: m_ICmp(Pred&: Predicate, L: m_Value(V&: X), R: m_Zero())) || |
| 1020 | !ICmpInst::isEquality(P: Predicate)) |
| 1021 | return nullptr; |
| 1022 | |
| 1023 | if (Predicate == ICmpInst::ICMP_NE) |
| 1024 | std::swap(a&: TrueVal, b&: FalseVal); |
| 1025 | |
| 1026 | // Check that TrueVal is a constant instead of matching it with m_Zero() |
| 1027 | // to handle the case when it is a scalar undef value or a vector containing |
| 1028 | // non-zero elements that are masked by undef elements in the compare |
| 1029 | // constant. |
| 1030 | auto *TrueValC = dyn_cast<Constant>(Val: TrueVal); |
| 1031 | if (TrueValC == nullptr || !isa<Instruction>(Val: FalseVal)) |
| 1032 | return nullptr; |
| 1033 | |
| 1034 | bool FreezeY; |
| 1035 | if (match(V: FalseVal, P: m_c_Mul(L: m_Specific(V: X), R: m_Value(V&: Y))) || |
| 1036 | match(V: FalseVal, P: m_c_And(L: m_Specific(V: X), R: m_Value(V&: Y))) || |
| 1037 | match(V: FalseVal, P: m_FShl(Op0: m_Specific(V: X), Op1: m_Specific(V: X), Op2: m_Value(V&: Y))) || |
| 1038 | match(V: FalseVal, P: m_FShr(Op0: m_Specific(V: X), Op1: m_Specific(V: X), Op2: m_Value(V&: Y))) || |
| 1039 | match(V: FalseVal, |
| 1040 | P: m_c_Intrinsic<Intrinsic::umin>(Op0: m_Specific(V: X), Op1: m_Value(V&: Y)))) { |
| 1041 | FreezeY = true; |
| 1042 | } else if (match(V: FalseVal, P: m_IDiv(L: m_Specific(V: X), R: m_Value(V&: Y))) || |
| 1043 | match(V: FalseVal, P: m_IRem(L: m_Specific(V: X), R: m_Value(V&: Y)))) { |
| 1044 | FreezeY = false; |
| 1045 | } else { |
| 1046 | return nullptr; |
| 1047 | } |
| 1048 | |
| 1049 | auto *ZeroC = cast<Constant>(Val: cast<Instruction>(Val: CondVal)->getOperand(i: 1)); |
| 1050 | auto *MergedC = Constant::mergeUndefsWith(C: TrueValC, Other: ZeroC); |
| 1051 | // If X is compared with 0 then TrueVal could be either zero or undef. |
| 1052 | // m_Zero match vectors containing some undef elements, but for scalars |
| 1053 | // m_Undef should be used explicitly. |
| 1054 | if (!match(V: MergedC, P: m_Zero()) && !match(V: MergedC, P: m_Undef())) |
| 1055 | return nullptr; |
| 1056 | |
| 1057 | auto *FalseValI = cast<Instruction>(Val: FalseVal); |
| 1058 | if (FreezeY) { |
| 1059 | auto *FrY = IC.InsertNewInstBefore(New: new FreezeInst(Y, Y->getName() + ".fr" ), |
| 1060 | Old: FalseValI->getIterator()); |
| 1061 | IC.replaceOperand(I&: *FalseValI, |
| 1062 | OpNum: FalseValI->getOperand(i: 0) == Y |
| 1063 | ? 0 |
| 1064 | : (FalseValI->getOperand(i: 1) == Y ? 1 : 2), |
| 1065 | V: FrY); |
| 1066 | } |
| 1067 | return IC.replaceInstUsesWith(I&: SI, V: FalseValI); |
| 1068 | } |
| 1069 | |
| 1070 | /// Transform patterns such as (a > b) ? a - b : 0 into usub.sat(a, b). |
| 1071 | /// There are 8 commuted/swapped variants of this pattern. |
| 1072 | static Value * |
| 1073 | canonicalizeSaturatedSubtractUnsigned(const ICmpInst *ICI, const Value *TrueVal, |
| 1074 | const Value *FalseVal, |
| 1075 | InstCombiner::BuilderTy &Builder) { |
| 1076 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
| 1077 | Value *A = ICI->getOperand(i_nocapture: 0); |
| 1078 | Value *B = ICI->getOperand(i_nocapture: 1); |
| 1079 | |
| 1080 | // (b > a) ? 0 : a - b -> (b <= a) ? a - b : 0 |
| 1081 | // (a == 0) ? 0 : a - 1 -> (a != 0) ? a - 1 : 0 |
| 1082 | if (match(V: TrueVal, P: m_Zero())) { |
| 1083 | Pred = ICmpInst::getInversePredicate(pred: Pred); |
| 1084 | std::swap(a&: TrueVal, b&: FalseVal); |
| 1085 | } |
| 1086 | |
| 1087 | if (!match(V: FalseVal, P: m_Zero())) |
| 1088 | return nullptr; |
| 1089 | |
| 1090 | // ugt 0 is canonicalized to ne 0 and requires special handling |
| 1091 | // (a != 0) ? a + -1 : 0 -> usub.sat(a, 1) |
| 1092 | if (Pred == ICmpInst::ICMP_NE) { |
| 1093 | if (match(V: B, P: m_Zero()) && match(V: TrueVal, P: m_Add(L: m_Specific(V: A), R: m_AllOnes()))) |
| 1094 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: A, |
| 1095 | RHS: ConstantInt::get(Ty: A->getType(), V: 1)); |
| 1096 | return nullptr; |
| 1097 | } |
| 1098 | |
| 1099 | if (!ICmpInst::isUnsigned(Pred)) |
| 1100 | return nullptr; |
| 1101 | |
| 1102 | if (Pred == ICmpInst::ICMP_ULE || Pred == ICmpInst::ICMP_ULT) { |
| 1103 | // (b < a) ? a - b : 0 -> (a > b) ? a - b : 0 |
| 1104 | std::swap(a&: A, b&: B); |
| 1105 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 1106 | } |
| 1107 | |
| 1108 | assert((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && |
| 1109 | "Unexpected isUnsigned predicate!" ); |
| 1110 | |
| 1111 | // Ensure the sub is of the form: |
| 1112 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
| 1113 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
| 1114 | // Checking for both a-b and a+(-b) as a constant. |
| 1115 | bool IsNegative = false; |
| 1116 | const APInt *C; |
| 1117 | if (match(V: TrueVal, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))) || |
| 1118 | (match(V: A, P: m_APInt(Res&: C)) && |
| 1119 | match(V: TrueVal, P: m_Add(L: m_Specific(V: B), R: m_SpecificInt(V: -*C))))) |
| 1120 | IsNegative = true; |
| 1121 | else if (!match(V: TrueVal, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
| 1122 | !(match(V: B, P: m_APInt(Res&: C)) && |
| 1123 | match(V: TrueVal, P: m_Add(L: m_Specific(V: A), R: m_SpecificInt(V: -*C))))) |
| 1124 | return nullptr; |
| 1125 | |
| 1126 | // If we are adding a negate and the sub and icmp are used anywhere else, we |
| 1127 | // would end up with more instructions. |
| 1128 | if (IsNegative && !TrueVal->hasOneUse() && !ICI->hasOneUse()) |
| 1129 | return nullptr; |
| 1130 | |
| 1131 | // (a > b) ? a - b : 0 -> usub.sat(a, b) |
| 1132 | // (a > b) ? b - a : 0 -> -usub.sat(a, b) |
| 1133 | Value *Result = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: A, RHS: B); |
| 1134 | if (IsNegative) |
| 1135 | Result = Builder.CreateNeg(V: Result); |
| 1136 | return Result; |
| 1137 | } |
| 1138 | |
| 1139 | static Value * |
| 1140 | canonicalizeSaturatedSubtractSigned(const ICmpInst *ICI, const Value *TrueVal, |
| 1141 | const Value *FalseVal, |
| 1142 | InstCombiner::BuilderTy &Builder) { |
| 1143 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
| 1144 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
| 1145 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
| 1146 | |
| 1147 | // `A != B ? X : Y` --> `A == B ? Y : X` |
| 1148 | // This canonicalization allows us to handle more patterns with fewer checks. |
| 1149 | if (Pred == ICmpInst::ICMP_NE) { |
| 1150 | Pred = ICmpInst::ICMP_EQ; |
| 1151 | std::swap(a&: TrueVal, b&: FalseVal); |
| 1152 | } |
| 1153 | |
| 1154 | // `A == MIN_INT ? MAX_INT : 0 - A` --> `ssub_sat 0, A` |
| 1155 | if (Pred == ICmpInst::ICMP_EQ && match(V: CmpRHS, P: m_SignMask()) && |
| 1156 | match(V: TrueVal, P: m_MaxSignedValue()) && |
| 1157 | match(V: FalseVal, P: m_Neg(V: m_Specific(V: CmpLHS)))) { |
| 1158 | return Builder.CreateBinaryIntrinsic( |
| 1159 | ID: Intrinsic::ssub_sat, LHS: ConstantInt::getNullValue(Ty: CmpLHS->getType()), |
| 1160 | RHS: CmpLHS); |
| 1161 | } |
| 1162 | |
| 1163 | return nullptr; |
| 1164 | } |
| 1165 | |
| 1166 | static Value *canonicalizeSaturatedSubtract(const ICmpInst *ICI, |
| 1167 | const Value *TrueVal, |
| 1168 | const Value *FalseVal, |
| 1169 | InstCombiner::BuilderTy &Builder) { |
| 1170 | if (Value *V = canonicalizeSaturatedSubtractUnsigned(ICI, TrueVal, FalseVal, |
| 1171 | Builder)) |
| 1172 | return V; |
| 1173 | |
| 1174 | if (Value *V = |
| 1175 | canonicalizeSaturatedSubtractSigned(ICI, TrueVal, FalseVal, Builder)) |
| 1176 | return V; |
| 1177 | |
| 1178 | return nullptr; |
| 1179 | } |
| 1180 | |
| 1181 | static Value * |
| 1182 | canonicalizeSaturatedAddUnsigned(ICmpInst *Cmp, Value *TVal, Value *FVal, |
| 1183 | InstCombiner::BuilderTy &Builder) { |
| 1184 | |
| 1185 | // Match unsigned saturated add with constant. |
| 1186 | Value *Cmp0 = Cmp->getOperand(i_nocapture: 0); |
| 1187 | Value *Cmp1 = Cmp->getOperand(i_nocapture: 1); |
| 1188 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
| 1189 | Value *X; |
| 1190 | const APInt *C; |
| 1191 | |
| 1192 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
| 1193 | // There are 8 commuted variants. |
| 1194 | // Canonicalize -1 (saturated result) to true value of the select. |
| 1195 | if (match(V: FVal, P: m_AllOnes())) { |
| 1196 | std::swap(a&: TVal, b&: FVal); |
| 1197 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 1198 | } |
| 1199 | if (!match(V: TVal, P: m_AllOnes())) |
| 1200 | return nullptr; |
| 1201 | |
| 1202 | // uge -1 is canonicalized to eq -1 and requires special handling |
| 1203 | // (a == -1) ? -1 : a + 1 -> uadd.sat(a, 1) |
| 1204 | if (Pred == ICmpInst::ICMP_EQ) { |
| 1205 | if (match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_One())) && |
| 1206 | match(V: Cmp1, P: m_AllOnes())) { |
| 1207 | return Builder.CreateBinaryIntrinsic( |
| 1208 | ID: Intrinsic::uadd_sat, LHS: Cmp0, RHS: ConstantInt::get(Ty: Cmp0->getType(), V: 1)); |
| 1209 | } |
| 1210 | return nullptr; |
| 1211 | } |
| 1212 | |
| 1213 | if ((Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_UGT) && |
| 1214 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
| 1215 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: ~*C))) { |
| 1216 | // (X u> ~C) ? -1 : (X + C) --> uadd.sat(X, C) |
| 1217 | // (X u>= ~C)? -1 : (X + C) --> uadd.sat(X, C) |
| 1218 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
| 1219 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
| 1220 | } |
| 1221 | |
| 1222 | // Negative one does not work here because X u> -1 ? -1, X + -1 is not a |
| 1223 | // saturated add. |
| 1224 | if (Pred == ICmpInst::ICMP_UGT && |
| 1225 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
| 1226 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: ~*C - 1)) && !C->isAllOnes()) { |
| 1227 | // (X u> ~C - 1) ? -1 : (X + C) --> uadd.sat(X, C) |
| 1228 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
| 1229 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
| 1230 | } |
| 1231 | |
| 1232 | // Zero does not work here because X u>= 0 ? -1 : X -> is always -1, which is |
| 1233 | // not a saturated add. |
| 1234 | if (Pred == ICmpInst::ICMP_UGE && |
| 1235 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_APIntAllowPoison(Res&: C))) && |
| 1236 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: -*C)) && !C->isZero()) { |
| 1237 | // (X u >= -C) ? -1 : (X + C) --> uadd.sat(X, C) |
| 1238 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp0, |
| 1239 | RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
| 1240 | } |
| 1241 | |
| 1242 | // Canonicalize predicate to less-than or less-or-equal-than. |
| 1243 | if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) { |
| 1244 | std::swap(a&: Cmp0, b&: Cmp1); |
| 1245 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 1246 | } |
| 1247 | if (Pred != ICmpInst::ICMP_ULT && Pred != ICmpInst::ICMP_ULE) |
| 1248 | return nullptr; |
| 1249 | |
| 1250 | // Match unsigned saturated add of 2 variables with an unnecessary 'not'. |
| 1251 | // Strictness of the comparison is irrelevant. |
| 1252 | Value *Y; |
| 1253 | if (match(V: Cmp0, P: m_Not(V: m_Value(V&: X))) && |
| 1254 | match(V: FVal, P: m_c_Add(L: m_Specific(V: X), R: m_Value(V&: Y))) && Y == Cmp1) { |
| 1255 | // (~X u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
| 1256 | // (~X u< Y) ? -1 : (Y + X) --> uadd.sat(X, Y) |
| 1257 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: X, RHS: Y); |
| 1258 | } |
| 1259 | // The 'not' op may be included in the sum but not the compare. |
| 1260 | // Strictness of the comparison is irrelevant. |
| 1261 | X = Cmp0; |
| 1262 | Y = Cmp1; |
| 1263 | if (match(V: FVal, P: m_c_Add(L: m_NotForbidPoison(V: m_Specific(V: X)), R: m_Specific(V: Y)))) { |
| 1264 | // (X u< Y) ? -1 : (~X + Y) --> uadd.sat(~X, Y) |
| 1265 | // (X u< Y) ? -1 : (Y + ~X) --> uadd.sat(Y, ~X) |
| 1266 | BinaryOperator *BO = cast<BinaryOperator>(Val: FVal); |
| 1267 | return Builder.CreateBinaryIntrinsic( |
| 1268 | ID: Intrinsic::uadd_sat, LHS: BO->getOperand(i_nocapture: 0), RHS: BO->getOperand(i_nocapture: 1)); |
| 1269 | } |
| 1270 | // The overflow may be detected via the add wrapping round. |
| 1271 | // This is only valid for strict comparison! |
| 1272 | if (Pred == ICmpInst::ICMP_ULT && |
| 1273 | match(V: Cmp0, P: m_c_Add(L: m_Specific(V: Cmp1), R: m_Value(V&: Y))) && |
| 1274 | match(V: FVal, P: m_c_Add(L: m_Specific(V: Cmp1), R: m_Specific(V: Y)))) { |
| 1275 | // ((X + Y) u< X) ? -1 : (X + Y) --> uadd.sat(X, Y) |
| 1276 | // ((X + Y) u< Y) ? -1 : (X + Y) --> uadd.sat(X, Y) |
| 1277 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::uadd_sat, LHS: Cmp1, RHS: Y); |
| 1278 | } |
| 1279 | |
| 1280 | return nullptr; |
| 1281 | } |
| 1282 | |
| 1283 | static Value *canonicalizeSaturatedAddSigned(ICmpInst *Cmp, Value *TVal, |
| 1284 | Value *FVal, |
| 1285 | InstCombiner::BuilderTy &Builder) { |
| 1286 | // Match saturated add with constant. |
| 1287 | Value *Cmp0 = Cmp->getOperand(i_nocapture: 0); |
| 1288 | Value *Cmp1 = Cmp->getOperand(i_nocapture: 1); |
| 1289 | ICmpInst::Predicate Pred = Cmp->getPredicate(); |
| 1290 | |
| 1291 | // Canonicalize TVal to be the saturation constant. |
| 1292 | if (match(V: FVal, P: m_MaxSignedValue()) || match(V: FVal, P: m_SignMask())) { |
| 1293 | std::swap(a&: TVal, b&: FVal); |
| 1294 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 1295 | } |
| 1296 | |
| 1297 | const APInt *SatC; |
| 1298 | if (!match(V: TVal, P: m_APInt(Res&: SatC)) || |
| 1299 | !(SatC->isMaxSignedValue() || SatC->isSignMask())) |
| 1300 | return nullptr; |
| 1301 | |
| 1302 | bool IsMax = SatC->isMaxSignedValue(); |
| 1303 | |
| 1304 | // sge maximum signed value is canonicalized to eq maximum signed value and |
| 1305 | // requires special handling. sle minimum signed value is similarly |
| 1306 | // canonicalized to eq minimum signed value. |
| 1307 | if (Pred == ICmpInst::ICMP_EQ && Cmp1 == TVal) { |
| 1308 | // (a == INT_MAX) ? INT_MAX : a + 1 -> sadd.sat(a, 1) |
| 1309 | if (IsMax && match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_One()))) { |
| 1310 | return Builder.CreateBinaryIntrinsic( |
| 1311 | ID: Intrinsic::sadd_sat, LHS: Cmp0, RHS: ConstantInt::get(Ty: Cmp0->getType(), V: 1)); |
| 1312 | } |
| 1313 | |
| 1314 | // (a == INT_MIN) ? INT_MIN : a + -1 -> sadd.sat(a, -1) |
| 1315 | if (!IsMax && match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_AllOnes()))) { |
| 1316 | return Builder.CreateBinaryIntrinsic( |
| 1317 | ID: Intrinsic::sadd_sat, LHS: Cmp0, |
| 1318 | RHS: ConstantInt::getAllOnesValue(Ty: Cmp0->getType())); |
| 1319 | } |
| 1320 | return nullptr; |
| 1321 | } |
| 1322 | |
| 1323 | const APInt *C; |
| 1324 | |
| 1325 | // (X > Y) ? INT_MAX : (X + C) --> sadd.sat(X, C) |
| 1326 | // (X >= Y) ? INT_MAX : (X + C) --> sadd.sat(X, C) |
| 1327 | // where C > 0 and Y is INT_MAX - C or INT_MAX - C - 1 |
| 1328 | if (IsMax && (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) && |
| 1329 | isa<Constant>(Val: Cmp1) && |
| 1330 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_StrictlyPositive(V&: C)))) { |
| 1331 | // Normalize SGE to SGT for threshold comparison. |
| 1332 | if (Pred == ICmpInst::ICMP_SGE) { |
| 1333 | if (auto Flipped = getFlippedStrictnessPredicateAndConstant( |
| 1334 | Pred, C: cast<Constant>(Val: Cmp1))) { |
| 1335 | Pred = Flipped->first; |
| 1336 | Cmp1 = Flipped->second; |
| 1337 | } |
| 1338 | } |
| 1339 | // Check: X > INT_MAX - C or X > INT_MAX - C - 1 |
| 1340 | APInt Threshold = *SatC - *C; |
| 1341 | if (Pred == ICmpInst::ICMP_SGT && |
| 1342 | (match(V: Cmp1, P: m_SpecificIntAllowPoison(V: Threshold)) || |
| 1343 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: Threshold - 1)))) |
| 1344 | return Builder.CreateBinaryIntrinsic( |
| 1345 | ID: Intrinsic::sadd_sat, LHS: Cmp0, RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
| 1346 | } |
| 1347 | |
| 1348 | // (X < Y) ? INT_MIN : (X + C) --> sadd.sat(X, C) |
| 1349 | // (X <= Y) ? INT_MIN : (X + C) --> sadd.sat(X, C) |
| 1350 | // where C < 0 and Y is INT_MIN - C or INT_MIN - C + 1 |
| 1351 | if (!IsMax && (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) && |
| 1352 | isa<Constant>(Val: Cmp1) && |
| 1353 | match(V: FVal, P: m_Add(L: m_Specific(V: Cmp0), R: m_Negative(V&: C)))) { |
| 1354 | // Normalize SLE to SLT for threshold comparison. |
| 1355 | if (Pred == ICmpInst::ICMP_SLE) { |
| 1356 | if (auto Flipped = getFlippedStrictnessPredicateAndConstant( |
| 1357 | Pred, C: cast<Constant>(Val: Cmp1))) { |
| 1358 | Pred = Flipped->first; |
| 1359 | Cmp1 = Flipped->second; |
| 1360 | } |
| 1361 | } |
| 1362 | // Check: X < INT_MIN - C or X < INT_MIN - C + 1 |
| 1363 | // INT_MIN - C for negative C is like INT_MIN + |C| |
| 1364 | APInt Threshold = *SatC - *C; |
| 1365 | if (Pred == ICmpInst::ICMP_SLT && |
| 1366 | (match(V: Cmp1, P: m_SpecificIntAllowPoison(V: Threshold)) || |
| 1367 | match(V: Cmp1, P: m_SpecificIntAllowPoison(V: Threshold + 1)))) |
| 1368 | return Builder.CreateBinaryIntrinsic( |
| 1369 | ID: Intrinsic::sadd_sat, LHS: Cmp0, RHS: ConstantInt::get(Ty: Cmp0->getType(), V: *C)); |
| 1370 | } |
| 1371 | |
| 1372 | // Canonicalize predicate to less-than or less-or-equal-than. |
| 1373 | if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) { |
| 1374 | std::swap(a&: Cmp0, b&: Cmp1); |
| 1375 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 1376 | } |
| 1377 | |
| 1378 | if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_SLE) |
| 1379 | return nullptr; |
| 1380 | |
| 1381 | Value *X; |
| 1382 | |
| 1383 | // (INT_MAX - X s< Y) ? INT_MAX : (X + Y) --> sadd.sat(X, Y) |
| 1384 | // (INT_MAX - X s< Y) ? INT_MAX : (Y + X) --> sadd.sat(X, Y) |
| 1385 | if (IsMax && match(V: Cmp0, P: m_NSWSub(L: m_SpecificInt(V: *SatC), R: m_Value(V&: X))) && |
| 1386 | match(V: FVal, P: m_c_Add(L: m_Specific(V: X), R: m_Specific(V: Cmp1)))) { |
| 1387 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::sadd_sat, LHS: X, RHS: Cmp1); |
| 1388 | } |
| 1389 | |
| 1390 | // (INT_MIN - X s> Y) ? INT_MIN : (X + Y) --> sadd.sat(X, Y) |
| 1391 | // (INT_MIN - X s> Y) ? INT_MIN : (Y + X) --> sadd.sat(X, Y) |
| 1392 | // After swapping operands from the SGT/SGE canonicalization above, |
| 1393 | // this becomes (Y s< INT_MIN - X). |
| 1394 | if (!IsMax && match(V: Cmp1, P: m_NSWSub(L: m_SpecificInt(V: *SatC), R: m_Value(V&: X))) && |
| 1395 | match(V: FVal, P: m_c_Add(L: m_Specific(V: X), R: m_Specific(V: Cmp0)))) { |
| 1396 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::sadd_sat, LHS: X, RHS: Cmp0); |
| 1397 | } |
| 1398 | |
| 1399 | return nullptr; |
| 1400 | } |
| 1401 | |
| 1402 | static Value *canonicalizeSaturatedAdd(ICmpInst *Cmp, Value *TVal, Value *FVal, |
| 1403 | InstCombiner::BuilderTy &Builder) { |
| 1404 | if (!Cmp->hasOneUse()) |
| 1405 | return nullptr; |
| 1406 | |
| 1407 | if (Value *V = canonicalizeSaturatedAddUnsigned(Cmp, TVal, FVal, Builder)) |
| 1408 | return V; |
| 1409 | |
| 1410 | if (Value *V = canonicalizeSaturatedAddSigned(Cmp, TVal, FVal, Builder)) |
| 1411 | return V; |
| 1412 | |
| 1413 | return nullptr; |
| 1414 | } |
| 1415 | |
| 1416 | /// Try to match patterns with select and subtract as absolute difference. |
| 1417 | static Value *foldAbsDiff(ICmpInst *Cmp, Value *TVal, Value *FVal, |
| 1418 | InstCombiner::BuilderTy &Builder) { |
| 1419 | auto *TI = dyn_cast<Instruction>(Val: TVal); |
| 1420 | auto *FI = dyn_cast<Instruction>(Val: FVal); |
| 1421 | if (!TI || !FI) |
| 1422 | return nullptr; |
| 1423 | |
| 1424 | // Normalize predicate to gt/lt rather than ge/le. |
| 1425 | ICmpInst::Predicate Pred = Cmp->getStrictPredicate(); |
| 1426 | Value *A = Cmp->getOperand(i_nocapture: 0); |
| 1427 | Value *B = Cmp->getOperand(i_nocapture: 1); |
| 1428 | |
| 1429 | // Normalize "A - B" as the true value of the select. |
| 1430 | if (match(V: FI, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B)))) { |
| 1431 | std::swap(a&: FI, b&: TI); |
| 1432 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 1433 | } |
| 1434 | |
| 1435 | // With any pair of no-wrap subtracts: |
| 1436 | // (A > B) ? (A - B) : (B - A) --> abs(A - B) |
| 1437 | if (Pred == CmpInst::ICMP_SGT && |
| 1438 | match(V: TI, P: m_Sub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
| 1439 | match(V: FI, P: m_Sub(L: m_Specific(V: B), R: m_Specific(V: A))) && |
| 1440 | (TI->hasNoSignedWrap() || TI->hasNoUnsignedWrap()) && |
| 1441 | (FI->hasNoSignedWrap() || FI->hasNoUnsignedWrap())) { |
| 1442 | // The remaining subtract is not "nuw" any more. |
| 1443 | // If there's one use of the subtract (no other use than the use we are |
| 1444 | // about to replace), then we know that the sub is "nsw" in this context |
| 1445 | // even if it was only "nuw" before. If there's another use, then we can't |
| 1446 | // add "nsw" to the existing instruction because it may not be safe in the |
| 1447 | // other user's context. |
| 1448 | TI->setHasNoUnsignedWrap(false); |
| 1449 | if (!TI->hasNoSignedWrap()) |
| 1450 | TI->setHasNoSignedWrap(TI->hasOneUse()); |
| 1451 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: TI, RHS: Builder.getTrue()); |
| 1452 | } |
| 1453 | |
| 1454 | // Match: (A > B) ? (A - B) : (0 - (A - B)) --> abs(A - B) |
| 1455 | if (Pred == CmpInst::ICMP_SGT && |
| 1456 | match(V: TI, P: m_NSWSub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
| 1457 | match(V: FI, P: m_Neg(V: m_Specific(V: TI)))) { |
| 1458 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: TI, |
| 1459 | RHS: Builder.getFalse()); |
| 1460 | } |
| 1461 | |
| 1462 | // Match: (A < B) ? (0 - (A - B)) : (A - B) --> abs(A - B) |
| 1463 | if (Pred == CmpInst::ICMP_SLT && |
| 1464 | match(V: FI, P: m_NSWSub(L: m_Specific(V: A), R: m_Specific(V: B))) && |
| 1465 | match(V: TI, P: m_Neg(V: m_Specific(V: FI)))) { |
| 1466 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: FI, |
| 1467 | RHS: Builder.getFalse()); |
| 1468 | } |
| 1469 | |
| 1470 | // Match: (A > B) ? (0 - (B - A)) : (B - A) --> abs(B - A) |
| 1471 | if (Pred == CmpInst::ICMP_SGT && |
| 1472 | match(V: FI, P: m_NSWSub(L: m_Specific(V: B), R: m_Specific(V: A))) && |
| 1473 | match(V: TI, P: m_Neg(V: m_Specific(V: FI)))) { |
| 1474 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: FI, |
| 1475 | RHS: Builder.getFalse()); |
| 1476 | } |
| 1477 | |
| 1478 | // Match: (A < B) ? (B - A) : (0 - (B - A)) --> abs(B - A) |
| 1479 | if (Pred == CmpInst::ICMP_SLT && |
| 1480 | match(V: TI, P: m_NSWSub(L: m_Specific(V: B), R: m_Specific(V: A))) && |
| 1481 | match(V: FI, P: m_Neg(V: m_Specific(V: TI)))) { |
| 1482 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS: TI, |
| 1483 | RHS: Builder.getFalse()); |
| 1484 | } |
| 1485 | |
| 1486 | return nullptr; |
| 1487 | } |
| 1488 | |
| 1489 | /// Fold the following code sequence: |
| 1490 | /// \code |
| 1491 | /// int a = ctlz(x & -x); |
| 1492 | // x ? 31 - a : 32; |
| 1493 | /// \code |
| 1494 | /// |
| 1495 | /// into: |
| 1496 | /// cttz(x) |
| 1497 | static Instruction *foldSelectCtlzToCttz(ICmpInst *ICI, Value *TrueVal, |
| 1498 | Value *FalseVal, |
| 1499 | InstCombiner::BuilderTy &Builder) { |
| 1500 | unsigned BitWidth = TrueVal->getType()->getScalarSizeInBits(); |
| 1501 | if (!ICI->isEquality() || !match(V: ICI->getOperand(i_nocapture: 1), P: m_Zero())) |
| 1502 | return nullptr; |
| 1503 | |
| 1504 | if (ICI->getPredicate() == ICmpInst::ICMP_NE) |
| 1505 | std::swap(a&: TrueVal, b&: FalseVal); |
| 1506 | |
| 1507 | Value *Ctlz; |
| 1508 | if (match(V: FalseVal, |
| 1509 | P: m_Xor(L: m_Value(V&: Ctlz), R: m_SpecificIntAllowPoison(V: BitWidth - 1)))) { |
| 1510 | if (!isPowerOf2_32(Value: BitWidth)) |
| 1511 | return nullptr; |
| 1512 | } else if (!match(V: FalseVal, P: m_Sub(L: m_SpecificIntAllowPoison(V: BitWidth - 1), |
| 1513 | R: m_Value(V&: Ctlz)))) { |
| 1514 | return nullptr; |
| 1515 | } |
| 1516 | |
| 1517 | if (!match(V: Ctlz, P: m_Ctlz(Op0: m_Value(), Op1: m_Value()))) |
| 1518 | return nullptr; |
| 1519 | |
| 1520 | if (!match(V: TrueVal, P: m_SpecificInt(V: BitWidth))) |
| 1521 | return nullptr; |
| 1522 | |
| 1523 | Value *X = ICI->getOperand(i_nocapture: 0); |
| 1524 | auto *II = cast<IntrinsicInst>(Val: Ctlz); |
| 1525 | if (!match(V: II->getOperand(i_nocapture: 0), P: m_c_And(L: m_Specific(V: X), R: m_Neg(V: m_Specific(V: X))))) |
| 1526 | return nullptr; |
| 1527 | |
| 1528 | // The original select returns the constant bitwidth when x == 0, so the |
| 1529 | // result is defined there; the cttz must use is_zero_poison = false. |
| 1530 | Function *F = Intrinsic::getOrInsertDeclaration( |
| 1531 | M: II->getModule(), id: Intrinsic::cttz, OverloadTys: II->getType()); |
| 1532 | return CallInst::Create(Func: F, Args: {X, Builder.getFalse()}); |
| 1533 | } |
| 1534 | |
| 1535 | /// Attempt to fold a cttz/ctlz followed by a icmp plus select into a single |
| 1536 | /// call to cttz/ctlz with flag 'is_zero_poison' cleared. |
| 1537 | /// |
| 1538 | /// For example, we can fold the following code sequence: |
| 1539 | /// \code |
| 1540 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 true) |
| 1541 | /// %1 = icmp ne i32 %x, 0 |
| 1542 | /// %2 = select i1 %1, i32 %0, i32 32 |
| 1543 | /// \code |
| 1544 | /// |
| 1545 | /// into: |
| 1546 | /// %0 = tail call i32 @llvm.cttz.i32(i32 %x, i1 false) |
| 1547 | static Value *foldSelectCttzCtlz(ICmpInst *ICI, Value *TrueVal, Value *FalseVal, |
| 1548 | InstCombinerImpl &IC) { |
| 1549 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
| 1550 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
| 1551 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
| 1552 | |
| 1553 | // Check if the select condition compares a value for equality. |
| 1554 | if (!ICI->isEquality()) |
| 1555 | return nullptr; |
| 1556 | |
| 1557 | Value *SelectArg = FalseVal; |
| 1558 | Value *ValueOnZero = TrueVal; |
| 1559 | if (Pred == ICmpInst::ICMP_NE) |
| 1560 | std::swap(a&: SelectArg, b&: ValueOnZero); |
| 1561 | |
| 1562 | // Skip zero extend/truncate. |
| 1563 | Value *Count = nullptr; |
| 1564 | if (!match(V: SelectArg, P: m_ZExt(Op: m_Value(V&: Count))) && |
| 1565 | !match(V: SelectArg, P: m_Trunc(Op: m_Value(V&: Count)))) |
| 1566 | Count = SelectArg; |
| 1567 | |
| 1568 | // Check that 'Count' is a call to intrinsic cttz/ctlz. Also check that the |
| 1569 | // input to the cttz/ctlz is used as LHS for the compare instruction. |
| 1570 | Value *X; |
| 1571 | if (!match(V: Count, P: m_Cttz(Op0: m_Value(V&: X), Op1: m_Value())) && |
| 1572 | !match(V: Count, P: m_Ctlz(Op0: m_Value(V&: X), Op1: m_Value()))) |
| 1573 | return nullptr; |
| 1574 | |
| 1575 | // (X == 0) ? BitWidth : ctz(X) |
| 1576 | // (X == -1) ? BitWidth : ctz(~X) |
| 1577 | // (X == Y) ? BitWidth : ctz(X ^ Y) |
| 1578 | if ((X != CmpLHS || !match(V: CmpRHS, P: m_Zero())) && |
| 1579 | (!match(V: X, P: m_Not(V: m_Specific(V: CmpLHS))) || !match(V: CmpRHS, P: m_AllOnes())) && |
| 1580 | !match(V: X, P: m_c_Xor(L: m_Specific(V: CmpLHS), R: m_Specific(V: CmpRHS)))) |
| 1581 | return nullptr; |
| 1582 | |
| 1583 | IntrinsicInst *II = cast<IntrinsicInst>(Val: Count); |
| 1584 | |
| 1585 | // Check if the value propagated on zero is a constant number equal to the |
| 1586 | // sizeof in bits of 'Count'. |
| 1587 | unsigned SizeOfInBits = Count->getType()->getScalarSizeInBits(); |
| 1588 | if (match(V: ValueOnZero, P: m_SpecificInt(V: SizeOfInBits))) { |
| 1589 | // A range annotation on the intrinsic may no longer be valid. |
| 1590 | II->dropPoisonGeneratingAnnotations(); |
| 1591 | IC.addToWorklist(I: II); |
| 1592 | return SelectArg; |
| 1593 | } |
| 1594 | |
| 1595 | // The ValueOnZero is not the bitwidth. But if the cttz/ctlz (and optional |
| 1596 | // zext/trunc) have one use (ending at the select), the cttz/ctlz result will |
| 1597 | // not be used if the input is zero. Relax to 'zero is poison' for that case. |
| 1598 | if (II->hasOneUse() && SelectArg->hasOneUse() && |
| 1599 | !match(V: II->getArgOperand(i: 1), P: m_One())) { |
| 1600 | II->setArgOperand(i: 1, v: ConstantInt::getTrue(Context&: II->getContext())); |
| 1601 | // noundef attribute on the intrinsic may no longer be valid. |
| 1602 | II->dropUBImplyingAttrsAndMetadata(); |
| 1603 | IC.addToWorklist(I: II); |
| 1604 | } |
| 1605 | |
| 1606 | return nullptr; |
| 1607 | } |
| 1608 | |
| 1609 | static Value *canonicalizeSPF(ICmpInst &Cmp, Value *TrueVal, Value *FalseVal, |
| 1610 | InstCombinerImpl &IC) { |
| 1611 | Value *LHS, *RHS; |
| 1612 | // TODO: What to do with pointer min/max patterns? |
| 1613 | if (!TrueVal->getType()->isIntOrIntVectorTy()) |
| 1614 | return nullptr; |
| 1615 | |
| 1616 | SelectPatternFlavor SPF = |
| 1617 | matchDecomposedSelectPattern(CmpI: &Cmp, TrueVal, FalseVal, LHS, RHS).Flavor; |
| 1618 | if (SPF == SelectPatternFlavor::SPF_ABS || |
| 1619 | SPF == SelectPatternFlavor::SPF_NABS) { |
| 1620 | if (!Cmp.hasOneUse() && !RHS->hasOneUse()) |
| 1621 | return nullptr; // TODO: Relax this restriction. |
| 1622 | |
| 1623 | // Note that NSW flag can only be propagated for normal, non-negated abs! |
| 1624 | bool IntMinIsPoison = SPF == SelectPatternFlavor::SPF_ABS && |
| 1625 | match(V: RHS, P: m_NSWNeg(V: m_Specific(V: LHS))); |
| 1626 | Constant *IntMinIsPoisonC = |
| 1627 | ConstantInt::get(Ty: Type::getInt1Ty(C&: Cmp.getContext()), V: IntMinIsPoison); |
| 1628 | Value *Abs = |
| 1629 | IC.Builder.CreateBinaryIntrinsic(ID: Intrinsic::abs, LHS, RHS: IntMinIsPoisonC); |
| 1630 | |
| 1631 | if (SPF == SelectPatternFlavor::SPF_NABS) |
| 1632 | return IC.Builder.CreateNeg(V: Abs); // Always without NSW flag! |
| 1633 | return Abs; |
| 1634 | } |
| 1635 | |
| 1636 | if (SelectPatternResult::isMinOrMax(SPF)) { |
| 1637 | Intrinsic::ID IntrinsicID = getMinMaxIntrinsic(SPF); |
| 1638 | return IC.Builder.CreateBinaryIntrinsic(ID: IntrinsicID, LHS, RHS); |
| 1639 | } |
| 1640 | |
| 1641 | return nullptr; |
| 1642 | } |
| 1643 | |
| 1644 | bool InstCombinerImpl::replaceInInstruction(Value *V, Value *Old, Value *New, |
| 1645 | unsigned Depth) { |
| 1646 | // Conservatively limit replacement to two instructions upwards. |
| 1647 | if (Depth == 2) |
| 1648 | return false; |
| 1649 | |
| 1650 | assert(!isa<Constant>(Old) && "Only replace non-constant values" ); |
| 1651 | |
| 1652 | auto *I = dyn_cast<Instruction>(Val: V); |
| 1653 | if (!I || !I->hasOneUse() || |
| 1654 | !isSafeToSpeculativelyExecuteWithVariableReplaced(I)) |
| 1655 | return false; |
| 1656 | |
| 1657 | // Forbid potentially lane-crossing instructions. |
| 1658 | if (Old->getType()->isVectorTy() && !isNotCrossLaneOperation(I)) |
| 1659 | return false; |
| 1660 | |
| 1661 | bool Changed = false; |
| 1662 | for (Use &U : I->operands()) { |
| 1663 | if (U == Old) { |
| 1664 | replaceUse(U, NewValue: New); |
| 1665 | Worklist.add(I); |
| 1666 | Changed = true; |
| 1667 | } else { |
| 1668 | Changed |= replaceInInstruction(V: U, Old, New, Depth: Depth + 1); |
| 1669 | } |
| 1670 | } |
| 1671 | return Changed; |
| 1672 | } |
| 1673 | |
| 1674 | /// If we have a select with an equality comparison, then we know the value in |
| 1675 | /// one of the arms of the select. See if substituting this value into an arm |
| 1676 | /// and simplifying the result yields the same value as the other arm. |
| 1677 | /// |
| 1678 | /// To make this transform safe, we must drop poison-generating flags |
| 1679 | /// (nsw, etc) if we simplified to a binop because the select may be guarding |
| 1680 | /// that poison from propagating. If the existing binop already had no |
| 1681 | /// poison-generating flags, then this transform can be done by instsimplify. |
| 1682 | /// |
| 1683 | /// Consider: |
| 1684 | /// %cmp = icmp eq i32 %x, 2147483647 |
| 1685 | /// %add = add nsw i32 %x, 1 |
| 1686 | /// %sel = select i1 %cmp, i32 -2147483648, i32 %add |
| 1687 | /// |
| 1688 | /// We can't replace %sel with %add unless we strip away the flags. |
| 1689 | /// TODO: Wrapping flags could be preserved in some cases with better analysis. |
| 1690 | Instruction *InstCombinerImpl::foldSelectValueEquivalence(SelectInst &Sel, |
| 1691 | CmpInst &Cmp) { |
| 1692 | // Canonicalize the pattern to an equivalence on the predicate by swapping the |
| 1693 | // select operands. |
| 1694 | Value *TrueVal = Sel.getTrueValue(), *FalseVal = Sel.getFalseValue(); |
| 1695 | bool Swapped = false; |
| 1696 | if (Cmp.isEquivalence(/*Invert=*/true)) { |
| 1697 | std::swap(a&: TrueVal, b&: FalseVal); |
| 1698 | Swapped = true; |
| 1699 | } else if (!Cmp.isEquivalence()) { |
| 1700 | return nullptr; |
| 1701 | } |
| 1702 | |
| 1703 | Value *CmpLHS = Cmp.getOperand(i_nocapture: 0), *CmpRHS = Cmp.getOperand(i_nocapture: 1); |
| 1704 | auto ReplaceOldOpWithNewOp = [&](Value *OldOp, |
| 1705 | Value *NewOp) -> Instruction * { |
| 1706 | // In X == Y ? f(X) : Z, try to evaluate f(Y) and replace the operand. |
| 1707 | // Take care to avoid replacing X == Y ? X : Z with X == Y ? Y : Z, as that |
| 1708 | // would lead to an infinite replacement cycle. |
| 1709 | // If we will be able to evaluate f(Y) to a constant, we can allow undef, |
| 1710 | // otherwise Y cannot be undef as we might pick different values for undef |
| 1711 | // in the cmp and in f(Y). |
| 1712 | if (TrueVal == OldOp && (isa<Constant>(Val: OldOp) || !isa<Constant>(Val: NewOp))) |
| 1713 | return nullptr; |
| 1714 | |
| 1715 | if (Value *V = simplifyWithOpReplaced(V: TrueVal, Op: OldOp, RepOp: NewOp, Q: SQ, |
| 1716 | /* AllowRefinement=*/true)) { |
| 1717 | // Need some guarantees about the new simplified op to ensure we don't inf |
| 1718 | // loop. |
| 1719 | // If we simplify to a constant, replace if we aren't creating new undef. |
| 1720 | if (match(V, P: m_ImmConstant()) && |
| 1721 | isGuaranteedNotToBeUndef(V, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
| 1722 | return replaceOperand(I&: Sel, OpNum: Swapped ? 2 : 1, V); |
| 1723 | |
| 1724 | // If NewOp is a constant and OldOp is not replace iff NewOp doesn't |
| 1725 | // contain and undef elements. |
| 1726 | // Make sure that V is always simpler than TrueVal, otherwise we might |
| 1727 | // end up in an infinite loop. |
| 1728 | if (match(V: NewOp, P: m_ImmConstant()) || |
| 1729 | (isa<Instruction>(Val: TrueVal) && |
| 1730 | is_contained(Range: cast<Instruction>(Val: TrueVal)->operands(), Element: V))) { |
| 1731 | if (isGuaranteedNotToBeUndef(V: NewOp, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
| 1732 | return replaceOperand(I&: Sel, OpNum: Swapped ? 2 : 1, V); |
| 1733 | return nullptr; |
| 1734 | } |
| 1735 | } |
| 1736 | |
| 1737 | // Even if TrueVal does not simplify, we can directly replace a use of |
| 1738 | // CmpLHS with CmpRHS, as long as the instruction is not used anywhere |
| 1739 | // else and is safe to speculatively execute (we may end up executing it |
| 1740 | // with different operands, which should not cause side-effects or trigger |
| 1741 | // undefined behavior). Only do this if CmpRHS is a constant, as |
| 1742 | // profitability is not clear for other cases. |
| 1743 | if (OldOp == CmpLHS && match(V: NewOp, P: m_ImmConstant()) && |
| 1744 | !match(V: OldOp, P: m_Constant()) && |
| 1745 | isGuaranteedNotToBeUndef(V: NewOp, AC: SQ.AC, CtxI: &Sel, DT: &DT)) |
| 1746 | if (replaceInInstruction(V: TrueVal, Old: OldOp, New: NewOp)) |
| 1747 | return &Sel; |
| 1748 | return nullptr; |
| 1749 | }; |
| 1750 | |
| 1751 | bool CanReplaceCmpLHSWithRHS = canReplacePointersIfEqual(From: CmpLHS, To: CmpRHS, DL); |
| 1752 | if (CanReplaceCmpLHSWithRHS) { |
| 1753 | if (Instruction *R = ReplaceOldOpWithNewOp(CmpLHS, CmpRHS)) |
| 1754 | return R; |
| 1755 | } |
| 1756 | bool CanReplaceCmpRHSWithLHS = canReplacePointersIfEqual(From: CmpRHS, To: CmpLHS, DL); |
| 1757 | if (CanReplaceCmpRHSWithLHS) { |
| 1758 | if (Instruction *R = ReplaceOldOpWithNewOp(CmpRHS, CmpLHS)) |
| 1759 | return R; |
| 1760 | } |
| 1761 | |
| 1762 | auto *FalseInst = dyn_cast<Instruction>(Val: FalseVal); |
| 1763 | if (!FalseInst) |
| 1764 | return nullptr; |
| 1765 | |
| 1766 | // InstSimplify already performed this fold if it was possible subject to |
| 1767 | // current poison-generating flags. Check whether dropping poison-generating |
| 1768 | // flags enables the transform. |
| 1769 | |
| 1770 | // Try each equivalence substitution possibility. |
| 1771 | // We have an 'EQ' comparison, so the select's false value will propagate. |
| 1772 | // Example: |
| 1773 | // (X == 42) ? 43 : (X + 1) --> (X == 42) ? (X + 1) : (X + 1) --> X + 1 |
| 1774 | SmallVector<Instruction *> DropFlags; |
| 1775 | if ((CanReplaceCmpLHSWithRHS && |
| 1776 | simplifyWithOpReplaced(V: FalseVal, Op: CmpLHS, RepOp: CmpRHS, Q: SQ, |
| 1777 | /* AllowRefinement */ false, |
| 1778 | DropFlags: &DropFlags) == TrueVal) || |
| 1779 | (CanReplaceCmpRHSWithLHS && |
| 1780 | simplifyWithOpReplaced(V: FalseVal, Op: CmpRHS, RepOp: CmpLHS, Q: SQ, |
| 1781 | /* AllowRefinement */ false, |
| 1782 | DropFlags: &DropFlags) == TrueVal)) { |
| 1783 | for (Instruction *I : DropFlags) { |
| 1784 | I->dropPoisonGeneratingAnnotations(); |
| 1785 | Worklist.add(I); |
| 1786 | } |
| 1787 | |
| 1788 | return replaceInstUsesWith(I&: Sel, V: FalseVal); |
| 1789 | } |
| 1790 | |
| 1791 | Constant *CmpC; |
| 1792 | if (FalseVal->getType()->isIntOrIntVectorTy(BitWidth: 1) && |
| 1793 | match(V: FalseVal, P: m_NUWTrunc(Op: m_Specific(V: CmpLHS))) && |
| 1794 | match(V: CmpRHS, P: m_ImmConstant(C&: CmpC)) && |
| 1795 | ConstantFoldCompareInstOperands( |
| 1796 | Predicate: ICmpInst::Predicate::ICMP_NE, LHS: CmpC, |
| 1797 | RHS: ConstantInt::getNullValue(Ty: CmpLHS->getType()), DL) == TrueVal) { |
| 1798 | return new ICmpInst(CmpInst::Predicate::ICMP_NE, CmpLHS, |
| 1799 | ConstantInt::getNullValue(Ty: CmpLHS->getType())); |
| 1800 | } |
| 1801 | |
| 1802 | return nullptr; |
| 1803 | } |
| 1804 | |
| 1805 | /// Fold the following code sequence: |
| 1806 | /// \code |
| 1807 | /// %XeqZ = icmp eq i64 %X, %Z |
| 1808 | /// %YeqZ = icmp eq i64 %Y, %Z |
| 1809 | /// %XeqY = icmp eq i64 %X, %Y |
| 1810 | /// %not.YeqZ = xor i1 %YeqZ, true |
| 1811 | /// %and = select i1 %not.YeqZ, i1 %XeqY, i1 false |
| 1812 | /// %equal = select i1 %XeqZ, i1 %YeqZ, i1 %and |
| 1813 | /// \code |
| 1814 | /// |
| 1815 | /// into: |
| 1816 | /// %equal = icmp eq i64 %X, %Y |
| 1817 | Instruction *InstCombinerImpl::foldSelectEqualityTest(SelectInst &Sel) { |
| 1818 | Value *X, *Y, *Z; |
| 1819 | Value *XeqY, *XeqZ = Sel.getCondition(), *YeqZ = Sel.getTrueValue(); |
| 1820 | |
| 1821 | if (!match(V: XeqZ, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: X), R: m_Value(V&: Z)))) |
| 1822 | return nullptr; |
| 1823 | |
| 1824 | if (!match(V: YeqZ, |
| 1825 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: Y), R: m_Specific(V: Z)))) |
| 1826 | std::swap(a&: X, b&: Z); |
| 1827 | |
| 1828 | if (!match(V: YeqZ, |
| 1829 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Value(V&: Y), R: m_Specific(V: Z)))) |
| 1830 | return nullptr; |
| 1831 | |
| 1832 | if (!match(V: Sel.getFalseValue(), |
| 1833 | P: m_c_LogicalAnd(L: m_Not(V: m_Specific(V: YeqZ)), R: m_Value(V&: XeqY)))) |
| 1834 | return nullptr; |
| 1835 | |
| 1836 | if (!match(V: XeqY, |
| 1837 | P: m_c_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: X), R: m_Specific(V: Y)))) |
| 1838 | return nullptr; |
| 1839 | |
| 1840 | cast<ICmpInst>(Val: XeqY)->setSameSign(false); |
| 1841 | return replaceInstUsesWith(I&: Sel, V: XeqY); |
| 1842 | } |
| 1843 | |
| 1844 | // See if this is a pattern like: |
| 1845 | // %old_cmp1 = icmp slt i32 %x, C2 |
| 1846 | // %old_replacement = select i1 %old_cmp1, i32 %target_low, i32 %target_high |
| 1847 | // %old_x_offseted = add i32 %x, C1 |
| 1848 | // %old_cmp0 = icmp ult i32 %old_x_offseted, C0 |
| 1849 | // %r = select i1 %old_cmp0, i32 %x, i32 %old_replacement |
| 1850 | // This can be rewritten as more canonical pattern: |
| 1851 | // %new_cmp1 = icmp slt i32 %x, -C1 |
| 1852 | // %new_cmp2 = icmp sge i32 %x, C0-C1 |
| 1853 | // %new_clamped_low = select i1 %new_cmp1, i32 %target_low, i32 %x |
| 1854 | // %r = select i1 %new_cmp2, i32 %target_high, i32 %new_clamped_low |
| 1855 | // Iff -C1 s<= C2 s<= C0-C1 |
| 1856 | // Also ULT predicate can also be UGT iff C0 != -1 (+invert result) |
| 1857 | // SLT predicate can also be SGT iff C2 != INT_MAX (+invert res.) |
| 1858 | static Value *canonicalizeClampLike(SelectInst &Sel0, ICmpInst &Cmp0, |
| 1859 | InstCombiner::BuilderTy &Builder, |
| 1860 | InstCombiner &IC) { |
| 1861 | Value *X = Sel0.getTrueValue(); |
| 1862 | Value *Sel1 = Sel0.getFalseValue(); |
| 1863 | |
| 1864 | // First match the condition of the outermost select. |
| 1865 | // Said condition must be one-use. |
| 1866 | if (!Cmp0.hasOneUse()) |
| 1867 | return nullptr; |
| 1868 | ICmpInst::Predicate Pred0 = Cmp0.getPredicate(); |
| 1869 | Value *Cmp00 = Cmp0.getOperand(i_nocapture: 0); |
| 1870 | Constant *C0; |
| 1871 | if (!match(V: Cmp0.getOperand(i_nocapture: 1), |
| 1872 | P: m_CombineAnd(Ps: m_AnyIntegralConstant(), Ps: m_Constant(C&: C0)))) |
| 1873 | return nullptr; |
| 1874 | |
| 1875 | if (!isa<SelectInst>(Val: Sel1)) { |
| 1876 | Pred0 = ICmpInst::getInversePredicate(pred: Pred0); |
| 1877 | std::swap(a&: X, b&: Sel1); |
| 1878 | } |
| 1879 | |
| 1880 | // Canonicalize Cmp0 into ult or uge. |
| 1881 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
| 1882 | switch (Pred0) { |
| 1883 | case ICmpInst::Predicate::ICMP_ULT: |
| 1884 | case ICmpInst::Predicate::ICMP_UGE: |
| 1885 | // Although icmp ult %x, 0 is an unusual thing to try and should generally |
| 1886 | // have been simplified, it does not verify with undef inputs so ensure we |
| 1887 | // are not in a strange state. |
| 1888 | if (!match(V: C0, P: m_SpecificInt_ICMP( |
| 1889 | Predicate: ICmpInst::Predicate::ICMP_NE, |
| 1890 | Threshold: APInt::getZero(numBits: C0->getType()->getScalarSizeInBits())))) |
| 1891 | return nullptr; |
| 1892 | break; // Great! |
| 1893 | case ICmpInst::Predicate::ICMP_ULE: |
| 1894 | case ICmpInst::Predicate::ICMP_UGT: |
| 1895 | // We want to canonicalize it to 'ult' or 'uge', so we'll need to increment |
| 1896 | // C0, which again means it must not have any all-ones elements. |
| 1897 | if (!match(V: C0, |
| 1898 | P: m_SpecificInt_ICMP( |
| 1899 | Predicate: ICmpInst::Predicate::ICMP_NE, |
| 1900 | Threshold: APInt::getAllOnes(numBits: C0->getType()->getScalarSizeInBits())))) |
| 1901 | return nullptr; // Can't do, have all-ones element[s]. |
| 1902 | Pred0 = ICmpInst::getFlippedStrictnessPredicate(pred: Pred0); |
| 1903 | C0 = InstCombiner::AddOne(C: C0); |
| 1904 | break; |
| 1905 | default: |
| 1906 | return nullptr; // Unknown predicate. |
| 1907 | } |
| 1908 | |
| 1909 | // Now that we've canonicalized the ICmp, we know the X we expect; |
| 1910 | // the select in other hand should be one-use. |
| 1911 | if (!Sel1->hasOneUse()) |
| 1912 | return nullptr; |
| 1913 | |
| 1914 | // If the types do not match, look through any truncs to the underlying |
| 1915 | // instruction. |
| 1916 | if (Cmp00->getType() != X->getType() && X->hasOneUse()) |
| 1917 | match(V: X, P: m_TruncOrSelf(Op: m_Value(V&: X))); |
| 1918 | |
| 1919 | // We now can finish matching the condition of the outermost select: |
| 1920 | // it should either be the X itself, or an addition of some constant to X. |
| 1921 | Constant *C1; |
| 1922 | if (Cmp00 == X) |
| 1923 | C1 = ConstantInt::getNullValue(Ty: X->getType()); |
| 1924 | else if (!match(V: Cmp00, |
| 1925 | P: m_Add(L: m_Specific(V: X), |
| 1926 | R: m_CombineAnd(Ps: m_AnyIntegralConstant(), Ps: m_Constant(C&: C1))))) |
| 1927 | return nullptr; |
| 1928 | |
| 1929 | Value *Cmp1; |
| 1930 | CmpPredicate Pred1; |
| 1931 | Constant *C2; |
| 1932 | Value *ReplacementLow, *ReplacementHigh; |
| 1933 | if (!match(V: Sel1, P: m_Select(C: m_Value(V&: Cmp1), L: m_Value(V&: ReplacementLow), |
| 1934 | R: m_Value(V&: ReplacementHigh))) || |
| 1935 | !match(V: Cmp1, |
| 1936 | P: m_ICmp(Pred&: Pred1, L: m_Specific(V: X), |
| 1937 | R: m_CombineAnd(Ps: m_AnyIntegralConstant(), Ps: m_Constant(C&: C2))))) |
| 1938 | return nullptr; |
| 1939 | |
| 1940 | if (!Cmp1->hasOneUse() && (Cmp00 == X || !Cmp00->hasOneUse())) |
| 1941 | return nullptr; // Not enough one-use instructions for the fold. |
| 1942 | // FIXME: this restriction could be relaxed if Cmp1 can be reused as one of |
| 1943 | // two comparisons we'll need to build. |
| 1944 | |
| 1945 | // Canonicalize Cmp1 into the form we expect. |
| 1946 | // FIXME: we shouldn't care about lanes that are 'undef' in the end? |
| 1947 | switch (Pred1) { |
| 1948 | case ICmpInst::Predicate::ICMP_SLT: |
| 1949 | break; |
| 1950 | case ICmpInst::Predicate::ICMP_SLE: |
| 1951 | // We'd have to increment C2 by one, and for that it must not have signed |
| 1952 | // max element, but then it would have been canonicalized to 'slt' before |
| 1953 | // we get here. So we can't do anything useful with 'sle'. |
| 1954 | return nullptr; |
| 1955 | case ICmpInst::Predicate::ICMP_SGT: |
| 1956 | // We want to canonicalize it to 'slt', so we'll need to increment C2, |
| 1957 | // which again means it must not have any signed max elements. |
| 1958 | if (!match(V: C2, |
| 1959 | P: m_SpecificInt_ICMP(Predicate: ICmpInst::Predicate::ICMP_NE, |
| 1960 | Threshold: APInt::getSignedMaxValue( |
| 1961 | numBits: C2->getType()->getScalarSizeInBits())))) |
| 1962 | return nullptr; // Can't do, have signed max element[s]. |
| 1963 | C2 = InstCombiner::AddOne(C: C2); |
| 1964 | [[fallthrough]]; |
| 1965 | case ICmpInst::Predicate::ICMP_SGE: |
| 1966 | // Also non-canonical, but here we don't need to change C2, |
| 1967 | // so we don't have any restrictions on C2, so we can just handle it. |
| 1968 | Pred1 = ICmpInst::Predicate::ICMP_SLT; |
| 1969 | std::swap(a&: ReplacementLow, b&: ReplacementHigh); |
| 1970 | break; |
| 1971 | default: |
| 1972 | return nullptr; // Unknown predicate. |
| 1973 | } |
| 1974 | assert(Pred1 == ICmpInst::Predicate::ICMP_SLT && |
| 1975 | "Unexpected predicate type." ); |
| 1976 | |
| 1977 | // The thresholds of this clamp-like pattern. |
| 1978 | auto *ThresholdLowIncl = ConstantExpr::getNeg(C: C1); |
| 1979 | auto *ThresholdHighExcl = ConstantExpr::getSub(C1: C0, C2: C1); |
| 1980 | |
| 1981 | assert((Pred0 == ICmpInst::Predicate::ICMP_ULT || |
| 1982 | Pred0 == ICmpInst::Predicate::ICMP_UGE) && |
| 1983 | "Unexpected predicate type." ); |
| 1984 | if (Pred0 == ICmpInst::Predicate::ICMP_UGE) |
| 1985 | std::swap(a&: ThresholdLowIncl, b&: ThresholdHighExcl); |
| 1986 | |
| 1987 | // The fold has a precondition 1: C2 s>= ThresholdLow |
| 1988 | auto *Precond1 = ConstantFoldCompareInstOperands( |
| 1989 | Predicate: ICmpInst::Predicate::ICMP_SGE, LHS: C2, RHS: ThresholdLowIncl, DL: IC.getDataLayout()); |
| 1990 | if (!Precond1 || !match(V: Precond1, P: m_One())) |
| 1991 | return nullptr; |
| 1992 | // The fold has a precondition 2: C2 s<= ThresholdHigh |
| 1993 | auto *Precond2 = ConstantFoldCompareInstOperands( |
| 1994 | Predicate: ICmpInst::Predicate::ICMP_SLE, LHS: C2, RHS: ThresholdHighExcl, DL: IC.getDataLayout()); |
| 1995 | if (!Precond2 || !match(V: Precond2, P: m_One())) |
| 1996 | return nullptr; |
| 1997 | |
| 1998 | // If we are matching from a truncated input, we need to sext the |
| 1999 | // ReplacementLow and ReplacementHigh values. Only do the transform if they |
| 2000 | // are free to extend due to being constants. |
| 2001 | if (X->getType() != Sel0.getType()) { |
| 2002 | Constant *LowC, *HighC; |
| 2003 | if (!match(V: ReplacementLow, P: m_ImmConstant(C&: LowC)) || |
| 2004 | !match(V: ReplacementHigh, P: m_ImmConstant(C&: HighC))) |
| 2005 | return nullptr; |
| 2006 | const DataLayout &DL = Sel0.getDataLayout(); |
| 2007 | ReplacementLow = |
| 2008 | ConstantFoldCastOperand(Opcode: Instruction::SExt, C: LowC, DestTy: X->getType(), DL); |
| 2009 | ReplacementHigh = |
| 2010 | ConstantFoldCastOperand(Opcode: Instruction::SExt, C: HighC, DestTy: X->getType(), DL); |
| 2011 | assert(ReplacementLow && ReplacementHigh && |
| 2012 | "Constant folding of ImmConstant cannot fail" ); |
| 2013 | } |
| 2014 | |
| 2015 | // All good, finally emit the new pattern. |
| 2016 | Value *ShouldReplaceLow = Builder.CreateICmpSLT(LHS: X, RHS: ThresholdLowIncl); |
| 2017 | Value *ShouldReplaceHigh = Builder.CreateICmpSGE(LHS: X, RHS: ThresholdHighExcl); |
| 2018 | Value *MaybeReplacedLow = |
| 2019 | Builder.CreateSelect(C: ShouldReplaceLow, True: ReplacementLow, False: X); |
| 2020 | |
| 2021 | // Create the final select. If we looked through a truncate above, we will |
| 2022 | // need to retruncate the result. |
| 2023 | Value *MaybeReplacedHigh = Builder.CreateSelect( |
| 2024 | C: ShouldReplaceHigh, True: ReplacementHigh, False: MaybeReplacedLow); |
| 2025 | return Builder.CreateTrunc(V: MaybeReplacedHigh, DestTy: Sel0.getType()); |
| 2026 | } |
| 2027 | |
| 2028 | // If we have |
| 2029 | // %cmp = icmp [canonical predicate] i32 %x, C0 |
| 2030 | // %r = select i1 %cmp, i32 %y, i32 C1 |
| 2031 | // Where C0 != C1 and %x may be different from %y, see if the constant that we |
| 2032 | // will have if we flip the strictness of the predicate (i.e. without changing |
| 2033 | // the result) is identical to the C1 in select. If it matches we can change |
| 2034 | // original comparison to one with swapped predicate, reuse the constant, |
| 2035 | // and swap the hands of select. |
| 2036 | static Instruction * |
| 2037 | tryToReuseConstantFromSelectInComparison(SelectInst &Sel, ICmpInst &Cmp, |
| 2038 | InstCombinerImpl &IC) { |
| 2039 | CmpPredicate Pred; |
| 2040 | Value *X; |
| 2041 | Constant *C0; |
| 2042 | if (!match(V: &Cmp, P: m_OneUse(SubPattern: m_ICmp( |
| 2043 | Pred, L: m_Value(V&: X), |
| 2044 | R: m_CombineAnd(Ps: m_AnyIntegralConstant(), Ps: m_Constant(C&: C0)))))) |
| 2045 | return nullptr; |
| 2046 | |
| 2047 | // If comparison predicate is non-relational, we won't be able to do anything. |
| 2048 | if (ICmpInst::isEquality(P: Pred)) |
| 2049 | return nullptr; |
| 2050 | |
| 2051 | // If comparison predicate is non-canonical, then we certainly won't be able |
| 2052 | // to make it canonical; canonicalizeCmpWithConstant() already tried. |
| 2053 | if (!InstCombiner::isCanonicalPredicate(Pred)) |
| 2054 | return nullptr; |
| 2055 | |
| 2056 | // If the [input] type of comparison and select type are different, lets abort |
| 2057 | // for now. We could try to compare constants with trunc/[zs]ext though. |
| 2058 | if (C0->getType() != Sel.getType()) |
| 2059 | return nullptr; |
| 2060 | |
| 2061 | // ULT with 'add' of a constant is canonical. See foldICmpAddConstant(). |
| 2062 | // FIXME: Are there more magic icmp predicate+constant pairs we must avoid? |
| 2063 | // Or should we just abandon this transform entirely? |
| 2064 | if (Pred == CmpInst::ICMP_ULT && match(V: X, P: m_Add(L: m_Value(), R: m_Constant()))) |
| 2065 | return nullptr; |
| 2066 | |
| 2067 | |
| 2068 | Value *SelVal0, *SelVal1; // We do not care which one is from where. |
| 2069 | match(V: &Sel, P: m_Select(C: m_Value(), L: m_Value(V&: SelVal0), R: m_Value(V&: SelVal1))); |
| 2070 | // At least one of these values we are selecting between must be a constant |
| 2071 | // else we'll never succeed. |
| 2072 | if (!match(V: SelVal0, P: m_AnyIntegralConstant()) && |
| 2073 | !match(V: SelVal1, P: m_AnyIntegralConstant())) |
| 2074 | return nullptr; |
| 2075 | |
| 2076 | // Does this constant C match any of the `select` values? |
| 2077 | auto MatchesSelectValue = [SelVal0, SelVal1](Constant *C) { |
| 2078 | return C->isElementWiseEqual(Y: SelVal0) || C->isElementWiseEqual(Y: SelVal1); |
| 2079 | }; |
| 2080 | |
| 2081 | // If C0 *already* matches true/false value of select, we are done. |
| 2082 | if (MatchesSelectValue(C0)) |
| 2083 | return nullptr; |
| 2084 | |
| 2085 | // Check the constant we'd have with flipped-strictness predicate. |
| 2086 | auto FlippedStrictness = getFlippedStrictnessPredicateAndConstant(Pred, C: C0); |
| 2087 | if (!FlippedStrictness) |
| 2088 | return nullptr; |
| 2089 | |
| 2090 | // If said constant doesn't match either, then there is no hope, |
| 2091 | if (!MatchesSelectValue(FlippedStrictness->second)) |
| 2092 | return nullptr; |
| 2093 | |
| 2094 | // It matched! Lets insert the new comparison just before select. |
| 2095 | InstCombiner::BuilderTy::InsertPointGuard Guard(IC.Builder); |
| 2096 | IC.Builder.SetInsertPoint(&Sel); |
| 2097 | |
| 2098 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); // Yes, swapped. |
| 2099 | Value *NewCmp = IC.Builder.CreateICmp(P: Pred, LHS: X, RHS: FlippedStrictness->second, |
| 2100 | Name: Cmp.getName() + ".inv" ); |
| 2101 | IC.replaceOperand(I&: Sel, OpNum: 0, V: NewCmp); |
| 2102 | Sel.swapValues(); |
| 2103 | Sel.swapProfMetadata(); |
| 2104 | |
| 2105 | return &Sel; |
| 2106 | } |
| 2107 | |
| 2108 | static Instruction *foldSelectZeroOrOnes(ICmpInst *Cmp, Value *TVal, |
| 2109 | Value *FVal, |
| 2110 | InstCombiner::BuilderTy &Builder) { |
| 2111 | if (!Cmp->hasOneUse()) |
| 2112 | return nullptr; |
| 2113 | |
| 2114 | const APInt *CmpC; |
| 2115 | if (!match(V: Cmp->getOperand(i_nocapture: 1), P: m_APIntAllowPoison(Res&: CmpC))) |
| 2116 | return nullptr; |
| 2117 | |
| 2118 | // (X u< 2) ? -X : -1 --> sext (X != 0) |
| 2119 | Value *X = Cmp->getOperand(i_nocapture: 0); |
| 2120 | if (Cmp->getPredicate() == ICmpInst::ICMP_ULT && *CmpC == 2 && |
| 2121 | match(V: TVal, P: m_Neg(V: m_Specific(V: X))) && match(V: FVal, P: m_AllOnes())) |
| 2122 | return new SExtInst(Builder.CreateIsNotNull(Arg: X), TVal->getType()); |
| 2123 | |
| 2124 | // (X u> 1) ? -1 : -X --> sext (X != 0) |
| 2125 | if (Cmp->getPredicate() == ICmpInst::ICMP_UGT && *CmpC == 1 && |
| 2126 | match(V: FVal, P: m_Neg(V: m_Specific(V: X))) && match(V: TVal, P: m_AllOnes())) |
| 2127 | return new SExtInst(Builder.CreateIsNotNull(Arg: X), TVal->getType()); |
| 2128 | |
| 2129 | return nullptr; |
| 2130 | } |
| 2131 | |
| 2132 | static Value *foldSelectInstWithICmpConst(SelectInst &SI, ICmpInst *ICI, |
| 2133 | InstCombiner::BuilderTy &Builder) { |
| 2134 | const APInt *CmpC; |
| 2135 | Value *V; |
| 2136 | CmpPredicate Pred; |
| 2137 | if (!match(V: ICI, P: m_ICmp(Pred, L: m_Value(V), R: m_APInt(Res&: CmpC)))) |
| 2138 | return nullptr; |
| 2139 | |
| 2140 | // Match clamp away from min/max value as a max/min operation. |
| 2141 | Value *TVal = SI.getTrueValue(); |
| 2142 | Value *FVal = SI.getFalseValue(); |
| 2143 | if (Pred == ICmpInst::ICMP_EQ && V == FVal) { |
| 2144 | // (V == UMIN) ? UMIN+1 : V --> umax(V, UMIN+1) |
| 2145 | if (CmpC->isMinValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC + 1))) |
| 2146 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umax, LHS: V, RHS: TVal); |
| 2147 | // (V == UMAX) ? UMAX-1 : V --> umin(V, UMAX-1) |
| 2148 | if (CmpC->isMaxValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC - 1))) |
| 2149 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::umin, LHS: V, RHS: TVal); |
| 2150 | // (V == SMIN) ? SMIN+1 : V --> smax(V, SMIN+1) |
| 2151 | if (CmpC->isMinSignedValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC + 1))) |
| 2152 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smax, LHS: V, RHS: TVal); |
| 2153 | // (V == SMAX) ? SMAX-1 : V --> smin(V, SMAX-1) |
| 2154 | if (CmpC->isMaxSignedValue() && match(V: TVal, P: m_SpecificInt(V: *CmpC - 1))) |
| 2155 | return Builder.CreateBinaryIntrinsic(ID: Intrinsic::smin, LHS: V, RHS: TVal); |
| 2156 | } |
| 2157 | |
| 2158 | // Fold icmp(X) ? f(X) : C to f(X) when f(X) is guaranteed to be equal to C |
| 2159 | // for all X in the exact range of the inverse predicate. |
| 2160 | Instruction *Op; |
| 2161 | const APInt *C; |
| 2162 | CmpInst::Predicate CPred; |
| 2163 | if (match(V: &SI, P: m_Select(C: m_Specific(V: ICI), L: m_APInt(Res&: C), R: m_Instruction(I&: Op)))) |
| 2164 | CPred = ICI->getPredicate(); |
| 2165 | else if (match(V: &SI, P: m_Select(C: m_Specific(V: ICI), L: m_Instruction(I&: Op), R: m_APInt(Res&: C)))) |
| 2166 | CPred = ICI->getInversePredicate(); |
| 2167 | else |
| 2168 | return nullptr; |
| 2169 | |
| 2170 | ConstantRange InvDomCR = ConstantRange::makeExactICmpRegion(Pred: CPred, Other: *CmpC); |
| 2171 | const APInt *OpC; |
| 2172 | if (match(V: Op, P: m_BinOp(L: m_Specific(V), R: m_APInt(Res&: OpC)))) { |
| 2173 | ConstantRange R = InvDomCR.binaryOp( |
| 2174 | BinOp: static_cast<Instruction::BinaryOps>(Op->getOpcode()), Other: *OpC); |
| 2175 | if (R == *C) { |
| 2176 | Op->dropPoisonGeneratingFlags(); |
| 2177 | return Op; |
| 2178 | } |
| 2179 | } |
| 2180 | if (auto *MMI = dyn_cast<MinMaxIntrinsic>(Val: Op); |
| 2181 | MMI && MMI->getLHS() == V && match(V: MMI->getRHS(), P: m_APInt(Res&: OpC))) { |
| 2182 | ConstantRange R = ConstantRange::intrinsic(IntrinsicID: MMI->getIntrinsicID(), |
| 2183 | Ops: {InvDomCR, ConstantRange(*OpC)}); |
| 2184 | if (R == *C) { |
| 2185 | MMI->dropPoisonGeneratingAnnotations(); |
| 2186 | return MMI; |
| 2187 | } |
| 2188 | } |
| 2189 | |
| 2190 | return nullptr; |
| 2191 | } |
| 2192 | |
| 2193 | /// `A == MIN_INT ? B != MIN_INT : A < B` --> `A < B` |
| 2194 | /// `A == MAX_INT ? B != MAX_INT : A > B` --> `A > B` |
| 2195 | static Instruction *foldSelectWithExtremeEqCond(Value *CmpLHS, Value *CmpRHS, |
| 2196 | Value *TrueVal, |
| 2197 | Value *FalseVal) { |
| 2198 | Type *Ty = CmpLHS->getType(); |
| 2199 | |
| 2200 | if (Ty->isPtrOrPtrVectorTy()) |
| 2201 | return nullptr; |
| 2202 | |
| 2203 | CmpPredicate Pred; |
| 2204 | Value *B; |
| 2205 | |
| 2206 | if (!match(V: FalseVal, P: m_c_ICmp(Pred, L: m_Specific(V: CmpLHS), R: m_Value(V&: B)))) |
| 2207 | return nullptr; |
| 2208 | |
| 2209 | Value *TValRHS; |
| 2210 | if (!match(V: TrueVal, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, L: m_Specific(V: B), |
| 2211 | R: m_Value(V&: TValRHS)))) |
| 2212 | return nullptr; |
| 2213 | |
| 2214 | APInt C; |
| 2215 | unsigned BitWidth = Ty->getScalarSizeInBits(); |
| 2216 | |
| 2217 | if (ICmpInst::isLT(P: Pred)) { |
| 2218 | C = CmpInst::isSigned(Pred) ? APInt::getSignedMinValue(numBits: BitWidth) |
| 2219 | : APInt::getMinValue(numBits: BitWidth); |
| 2220 | } else if (ICmpInst::isGT(P: Pred)) { |
| 2221 | C = CmpInst::isSigned(Pred) ? APInt::getSignedMaxValue(numBits: BitWidth) |
| 2222 | : APInt::getMaxValue(numBits: BitWidth); |
| 2223 | } else { |
| 2224 | return nullptr; |
| 2225 | } |
| 2226 | |
| 2227 | if (!match(V: CmpRHS, P: m_SpecificInt(V: C)) || !match(V: TValRHS, P: m_SpecificInt(V: C))) |
| 2228 | return nullptr; |
| 2229 | |
| 2230 | return new ICmpInst(Pred, CmpLHS, B); |
| 2231 | } |
| 2232 | |
| 2233 | static Instruction *foldSelectICmpEq(SelectInst &SI, ICmpInst *ICI, |
| 2234 | InstCombinerImpl &IC) { |
| 2235 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
| 2236 | if (!ICmpInst::isEquality(P: Pred)) |
| 2237 | return nullptr; |
| 2238 | |
| 2239 | Value *TrueVal = SI.getTrueValue(); |
| 2240 | Value *FalseVal = SI.getFalseValue(); |
| 2241 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
| 2242 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
| 2243 | |
| 2244 | if (Pred == ICmpInst::ICMP_NE) |
| 2245 | std::swap(a&: TrueVal, b&: FalseVal); |
| 2246 | |
| 2247 | if (Instruction *Res = |
| 2248 | foldSelectWithExtremeEqCond(CmpLHS, CmpRHS, TrueVal, FalseVal)) |
| 2249 | return Res; |
| 2250 | |
| 2251 | return nullptr; |
| 2252 | } |
| 2253 | |
| 2254 | /// Fold `X Pred C1 ? X BOp C2 : C1 BOp C2` to `min/max(X, C1) BOp C2`. |
| 2255 | /// This allows for better canonicalization. |
| 2256 | Value *InstCombinerImpl::foldSelectWithConstOpToBinOp(ICmpInst *Cmp, |
| 2257 | Value *TrueVal, |
| 2258 | Value *FalseVal) { |
| 2259 | Constant *C1, *C2, *C3; |
| 2260 | Value *X; |
| 2261 | CmpPredicate Predicate; |
| 2262 | |
| 2263 | if (!match(V: Cmp, P: m_ICmp(Pred&: Predicate, L: m_Value(V&: X), R: m_Constant(C&: C1)))) |
| 2264 | return nullptr; |
| 2265 | |
| 2266 | if (!ICmpInst::isRelational(P: Predicate)) |
| 2267 | return nullptr; |
| 2268 | |
| 2269 | if (match(V: TrueVal, P: m_Constant())) { |
| 2270 | std::swap(a&: FalseVal, b&: TrueVal); |
| 2271 | Predicate = ICmpInst::getInversePredicate(pred: Predicate); |
| 2272 | } |
| 2273 | |
| 2274 | if (!match(V: FalseVal, P: m_Constant(C&: C3)) || !TrueVal->hasOneUse()) |
| 2275 | return nullptr; |
| 2276 | |
| 2277 | bool IsIntrinsic; |
| 2278 | unsigned Opcode; |
| 2279 | if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Val: TrueVal)) { |
| 2280 | Opcode = BOp->getOpcode(); |
| 2281 | IsIntrinsic = false; |
| 2282 | |
| 2283 | // This fold causes some regressions and is primarily intended for |
| 2284 | // add and sub. So we early exit for div and rem to minimize the |
| 2285 | // regressions. |
| 2286 | if (Instruction::isIntDivRem(Opcode)) |
| 2287 | return nullptr; |
| 2288 | |
| 2289 | if (!match(V: BOp, P: m_BinOp(L: m_Specific(V: X), R: m_Constant(C&: C2)))) |
| 2290 | return nullptr; |
| 2291 | |
| 2292 | } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: TrueVal)) { |
| 2293 | if (!match(V: II, P: m_MaxOrMin(Op0: m_Specific(V: X), Op1: m_Constant(C&: C2)))) |
| 2294 | return nullptr; |
| 2295 | Opcode = II->getIntrinsicID(); |
| 2296 | IsIntrinsic = true; |
| 2297 | } else { |
| 2298 | return nullptr; |
| 2299 | } |
| 2300 | |
| 2301 | Value *RHS; |
| 2302 | SelectPatternFlavor SPF; |
| 2303 | const DataLayout &DL = Cmp->getDataLayout(); |
| 2304 | auto Flipped = getFlippedStrictnessPredicateAndConstant(Pred: Predicate, C: C1); |
| 2305 | |
| 2306 | auto FoldBinaryOpOrIntrinsic = [&](Constant *LHS, Constant *RHS) { |
| 2307 | return IsIntrinsic |
| 2308 | ? ConstantFoldIntrinsic(ID: Opcode, Ops: {LHS, RHS}, Ty: LHS->getType(), DL) |
| 2309 | : ConstantFoldBinaryOpOperands(Opcode, LHS, RHS, DL); |
| 2310 | }; |
| 2311 | |
| 2312 | if (C3 == FoldBinaryOpOrIntrinsic(C1, C2)) { |
| 2313 | SPF = getSelectPattern(Pred: Predicate).Flavor; |
| 2314 | RHS = C1; |
| 2315 | } else if (Flipped && C3 == FoldBinaryOpOrIntrinsic(Flipped->second, C2)) { |
| 2316 | SPF = getSelectPattern(Pred: Flipped->first).Flavor; |
| 2317 | RHS = Flipped->second; |
| 2318 | } else { |
| 2319 | return nullptr; |
| 2320 | } |
| 2321 | |
| 2322 | Intrinsic::ID MinMaxID = getMinMaxIntrinsic(SPF); |
| 2323 | Value *MinMax = Builder.CreateBinaryIntrinsic(ID: MinMaxID, LHS: X, RHS); |
| 2324 | if (IsIntrinsic) |
| 2325 | return Builder.CreateBinaryIntrinsic(ID: Opcode, LHS: MinMax, RHS: C2); |
| 2326 | |
| 2327 | const auto BinOpc = Instruction::BinaryOps(Opcode); |
| 2328 | Value *BinOp = Builder.CreateBinOp(Opc: BinOpc, LHS: MinMax, RHS: C2); |
| 2329 | |
| 2330 | // If we can attach no-wrap flags to the new instruction, do so if the |
| 2331 | // old instruction had them and C1 BinOp C2 does not overflow. |
| 2332 | if (Instruction *BinOpInst = dyn_cast<Instruction>(Val: BinOp)) { |
| 2333 | if (BinOpc == Instruction::Add || BinOpc == Instruction::Sub || |
| 2334 | BinOpc == Instruction::Mul) { |
| 2335 | Instruction *OldBinOp = cast<BinaryOperator>(Val: TrueVal); |
| 2336 | if (OldBinOp->hasNoSignedWrap() && |
| 2337 | willNotOverflow(Opcode: BinOpc, LHS: RHS, RHS: C2, CxtI: *BinOpInst, /*IsSigned=*/true)) |
| 2338 | BinOpInst->setHasNoSignedWrap(); |
| 2339 | if (OldBinOp->hasNoUnsignedWrap() && |
| 2340 | willNotOverflow(Opcode: BinOpc, LHS: RHS, RHS: C2, CxtI: *BinOpInst, /*IsSigned=*/false)) |
| 2341 | BinOpInst->setHasNoUnsignedWrap(); |
| 2342 | } |
| 2343 | } |
| 2344 | return BinOp; |
| 2345 | } |
| 2346 | |
| 2347 | /// Folds: |
| 2348 | /// %a_sub = call @llvm.usub.sat(x, IntConst1) |
| 2349 | /// %b_sub = call @llvm.usub.sat(y, IntConst2) |
| 2350 | /// %or = or %a_sub, %b_sub |
| 2351 | /// %cmp = icmp eq %or, 0 |
| 2352 | /// %sel = select %cmp, 0, MostSignificantBit |
| 2353 | /// into: |
| 2354 | /// %a_sub' = usub.sat(x, IntConst1 - MostSignificantBit) |
| 2355 | /// %b_sub' = usub.sat(y, IntConst2 - MostSignificantBit) |
| 2356 | /// %or = or %a_sub', %b_sub' |
| 2357 | /// %and = and %or, MostSignificantBit |
| 2358 | /// Likewise, for vector arguments as well. |
| 2359 | static Instruction *foldICmpUSubSatWithAndForMostSignificantBitCmp( |
| 2360 | SelectInst &SI, ICmpInst *ICI, InstCombiner::BuilderTy &Builder) { |
| 2361 | if (!SI.hasOneUse() || !ICI->hasOneUse()) |
| 2362 | return nullptr; |
| 2363 | CmpPredicate Pred; |
| 2364 | Value *A, *B; |
| 2365 | const APInt *Constant1, *Constant2; |
| 2366 | if (!match(V: SI.getCondition(), |
| 2367 | P: m_ICmp(Pred, |
| 2368 | L: m_OneUse(SubPattern: m_Or(L: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>( |
| 2369 | Ops: m_Value(V&: A), Ops: m_APInt(Res&: Constant1))), |
| 2370 | R: m_OneUse(SubPattern: m_Intrinsic<Intrinsic::usub_sat>( |
| 2371 | Ops: m_Value(V&: B), Ops: m_APInt(Res&: Constant2))))), |
| 2372 | R: m_Zero()))) |
| 2373 | return nullptr; |
| 2374 | |
| 2375 | Value *TrueVal = SI.getTrueValue(); |
| 2376 | Value *FalseVal = SI.getFalseValue(); |
| 2377 | if (!((Pred == ICmpInst::ICMP_EQ && match(V: TrueVal, P: m_Zero()) && |
| 2378 | match(V: FalseVal, P: m_SignMask())) || |
| 2379 | (Pred == ICmpInst::ICMP_NE && match(V: TrueVal, P: m_SignMask()) && |
| 2380 | match(V: FalseVal, P: m_Zero())))) |
| 2381 | return nullptr; |
| 2382 | |
| 2383 | auto *Ty = A->getType(); |
| 2384 | unsigned BW = Constant1->getBitWidth(); |
| 2385 | APInt MostSignificantBit = APInt::getSignMask(BitWidth: BW); |
| 2386 | |
| 2387 | // Anything over MSB is negative |
| 2388 | if (Constant1->isNonNegative() || Constant2->isNonNegative()) |
| 2389 | return nullptr; |
| 2390 | |
| 2391 | APInt AdjAP1 = *Constant1 - MostSignificantBit + 1; |
| 2392 | APInt AdjAP2 = *Constant2 - MostSignificantBit + 1; |
| 2393 | |
| 2394 | auto *Adj1 = ConstantInt::get(Ty, V: AdjAP1); |
| 2395 | auto *Adj2 = ConstantInt::get(Ty, V: AdjAP2); |
| 2396 | |
| 2397 | Value *NewA = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: A, RHS: Adj1); |
| 2398 | Value *NewB = Builder.CreateBinaryIntrinsic(ID: Intrinsic::usub_sat, LHS: B, RHS: Adj2); |
| 2399 | Value *Or = Builder.CreateOr(LHS: NewA, RHS: NewB); |
| 2400 | Constant *MSBConst = ConstantInt::get(Ty, V: MostSignificantBit); |
| 2401 | return BinaryOperator::CreateAnd(V1: Or, V2: MSBConst); |
| 2402 | } |
| 2403 | |
| 2404 | /// Visit a SelectInst that has an ICmpInst as its first operand. |
| 2405 | Instruction *InstCombinerImpl::foldSelectInstWithICmp(SelectInst &SI, |
| 2406 | ICmpInst *ICI) { |
| 2407 | if (Value *V = |
| 2408 | canonicalizeSPF(Cmp&: *ICI, TrueVal: SI.getTrueValue(), FalseVal: SI.getFalseValue(), IC&: *this)) |
| 2409 | return replaceInstUsesWith(I&: SI, V); |
| 2410 | |
| 2411 | if (Value *V = foldSelectInstWithICmpConst(SI, ICI, Builder)) |
| 2412 | return replaceInstUsesWith(I&: SI, V); |
| 2413 | |
| 2414 | if (Value *V = canonicalizeClampLike(Sel0&: SI, Cmp0&: *ICI, Builder, IC&: *this)) |
| 2415 | return replaceInstUsesWith(I&: SI, V); |
| 2416 | |
| 2417 | if (Instruction *NewSel = |
| 2418 | tryToReuseConstantFromSelectInComparison(Sel&: SI, Cmp&: *ICI, IC&: *this)) |
| 2419 | return NewSel; |
| 2420 | if (Instruction *Folded = |
| 2421 | foldICmpUSubSatWithAndForMostSignificantBitCmp(SI, ICI, Builder)) |
| 2422 | return Folded; |
| 2423 | |
| 2424 | // NOTE: if we wanted to, this is where to detect integer MIN/MAX |
| 2425 | bool Changed = false; |
| 2426 | Value *TrueVal = SI.getTrueValue(); |
| 2427 | Value *FalseVal = SI.getFalseValue(); |
| 2428 | ICmpInst::Predicate Pred = ICI->getPredicate(); |
| 2429 | Value *CmpLHS = ICI->getOperand(i_nocapture: 0); |
| 2430 | Value *CmpRHS = ICI->getOperand(i_nocapture: 1); |
| 2431 | |
| 2432 | if (Instruction *NewSel = foldSelectICmpEq(SI, ICI, IC&: *this)) |
| 2433 | return NewSel; |
| 2434 | |
| 2435 | // Canonicalize a signbit condition to use zero constant by swapping: |
| 2436 | // (CmpLHS > -1) ? TV : FV --> (CmpLHS < 0) ? FV : TV |
| 2437 | // To avoid conflicts (infinite loops) with other canonicalizations, this is |
| 2438 | // not applied with any constant select arm. |
| 2439 | if (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: m_AllOnes()) && |
| 2440 | !match(V: TrueVal, P: m_Constant()) && !match(V: FalseVal, P: m_Constant()) && |
| 2441 | ICI->hasOneUse()) { |
| 2442 | InstCombiner::BuilderTy::InsertPointGuard Guard(Builder); |
| 2443 | Builder.SetInsertPoint(&SI); |
| 2444 | Value *IsNeg = Builder.CreateIsNeg(Arg: CmpLHS, Name: ICI->getName()); |
| 2445 | replaceOperand(I&: SI, OpNum: 0, V: IsNeg); |
| 2446 | SI.swapValues(); |
| 2447 | SI.swapProfMetadata(); |
| 2448 | return &SI; |
| 2449 | } |
| 2450 | |
| 2451 | if (Value *V = foldSelectICmpMinMax(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder, SQ)) |
| 2452 | return replaceInstUsesWith(I&: SI, V); |
| 2453 | |
| 2454 | if (Instruction *V = |
| 2455 | foldSelectICmpAndAnd(SelType: SI.getType(), Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
| 2456 | return V; |
| 2457 | |
| 2458 | if (Value *V = foldSelectICmpAndZeroShl(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
| 2459 | return replaceInstUsesWith(I&: SI, V); |
| 2460 | |
| 2461 | if (Instruction *V = foldSelectCtlzToCttz(ICI, TrueVal, FalseVal, Builder)) |
| 2462 | return V; |
| 2463 | |
| 2464 | if (Instruction *V = foldSelectZeroOrOnes(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
| 2465 | return V; |
| 2466 | |
| 2467 | if (Value *V = foldSelectICmpLshrAshr(IC: ICI, TrueVal, FalseVal, Builder)) |
| 2468 | return replaceInstUsesWith(I&: SI, V); |
| 2469 | |
| 2470 | if (Value *V = foldSelectCttzCtlz(ICI, TrueVal, FalseVal, IC&: *this)) |
| 2471 | return replaceInstUsesWith(I&: SI, V); |
| 2472 | |
| 2473 | if (Value *V = canonicalizeSaturatedSubtract(ICI, TrueVal, FalseVal, Builder)) |
| 2474 | return replaceInstUsesWith(I&: SI, V); |
| 2475 | |
| 2476 | if (Value *V = canonicalizeSaturatedAdd(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
| 2477 | return replaceInstUsesWith(I&: SI, V); |
| 2478 | |
| 2479 | if (Value *V = foldAbsDiff(Cmp: ICI, TVal: TrueVal, FVal: FalseVal, Builder)) |
| 2480 | return replaceInstUsesWith(I&: SI, V); |
| 2481 | |
| 2482 | if (Value *V = foldSelectWithConstOpToBinOp(Cmp: ICI, TrueVal, FalseVal)) |
| 2483 | return replaceInstUsesWith(I&: SI, V); |
| 2484 | |
| 2485 | return Changed ? &SI : nullptr; |
| 2486 | } |
| 2487 | |
| 2488 | /// We have an SPF (e.g. a min or max) of an SPF of the form: |
| 2489 | /// SPF2(SPF1(A, B), C) |
| 2490 | Instruction *InstCombinerImpl::foldSPFofSPF(Instruction *Inner, |
| 2491 | SelectPatternFlavor SPF1, Value *A, |
| 2492 | Value *B, Instruction &Outer, |
| 2493 | SelectPatternFlavor SPF2, |
| 2494 | Value *C) { |
| 2495 | if (Outer.getType() != Inner->getType()) |
| 2496 | return nullptr; |
| 2497 | |
| 2498 | if (C == A || C == B) { |
| 2499 | // MAX(MAX(A, B), B) -> MAX(A, B) |
| 2500 | // MIN(MIN(a, b), a) -> MIN(a, b) |
| 2501 | // TODO: This could be done in instsimplify. |
| 2502 | if (SPF1 == SPF2 && SelectPatternResult::isMinOrMax(SPF: SPF1)) |
| 2503 | return replaceInstUsesWith(I&: Outer, V: Inner); |
| 2504 | } |
| 2505 | |
| 2506 | return nullptr; |
| 2507 | } |
| 2508 | |
| 2509 | /// Turn select C, (X + Y), (X - Y) --> (X + (select C, Y, (-Y))). |
| 2510 | /// This is even legal for FP. |
| 2511 | static Instruction *foldAddSubSelect(SelectInst &SI, |
| 2512 | InstCombiner::BuilderTy &Builder) { |
| 2513 | Value *CondVal = SI.getCondition(); |
| 2514 | Value *TrueVal = SI.getTrueValue(); |
| 2515 | Value *FalseVal = SI.getFalseValue(); |
| 2516 | auto *TI = dyn_cast<Instruction>(Val: TrueVal); |
| 2517 | auto *FI = dyn_cast<Instruction>(Val: FalseVal); |
| 2518 | if (!TI || !FI || !TI->hasOneUse() || !FI->hasOneUse()) |
| 2519 | return nullptr; |
| 2520 | |
| 2521 | Instruction *AddOp = nullptr, *SubOp = nullptr; |
| 2522 | if ((TI->getOpcode() == Instruction::Sub && |
| 2523 | FI->getOpcode() == Instruction::Add) || |
| 2524 | (TI->getOpcode() == Instruction::FSub && |
| 2525 | FI->getOpcode() == Instruction::FAdd)) { |
| 2526 | AddOp = FI; |
| 2527 | SubOp = TI; |
| 2528 | } else if ((FI->getOpcode() == Instruction::Sub && |
| 2529 | TI->getOpcode() == Instruction::Add) || |
| 2530 | (FI->getOpcode() == Instruction::FSub && |
| 2531 | TI->getOpcode() == Instruction::FAdd)) { |
| 2532 | AddOp = TI; |
| 2533 | SubOp = FI; |
| 2534 | } |
| 2535 | |
| 2536 | if (AddOp) { |
| 2537 | Value *OtherAddOp = nullptr; |
| 2538 | if (SubOp->getOperand(i: 0) == AddOp->getOperand(i: 0)) { |
| 2539 | OtherAddOp = AddOp->getOperand(i: 1); |
| 2540 | } else if (SubOp->getOperand(i: 0) == AddOp->getOperand(i: 1)) { |
| 2541 | OtherAddOp = AddOp->getOperand(i: 0); |
| 2542 | } |
| 2543 | |
| 2544 | if (OtherAddOp) { |
| 2545 | // So at this point we know we have (Y -> OtherAddOp): |
| 2546 | // select C, (add X, Y), (sub X, Z) |
| 2547 | Value *NegVal; // Compute -Z |
| 2548 | if (SI.getType()->isFPOrFPVectorTy()) { |
| 2549 | NegVal = Builder.CreateFNeg(V: SubOp->getOperand(i: 1)); |
| 2550 | if (Instruction *NegInst = dyn_cast<Instruction>(Val: NegVal)) { |
| 2551 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
| 2552 | Flags &= SubOp->getFastMathFlags(); |
| 2553 | NegInst->setFastMathFlags(Flags); |
| 2554 | } |
| 2555 | } else { |
| 2556 | NegVal = Builder.CreateNeg(V: SubOp->getOperand(i: 1)); |
| 2557 | } |
| 2558 | |
| 2559 | Value *NewTrueOp = OtherAddOp; |
| 2560 | Value *NewFalseOp = NegVal; |
| 2561 | if (AddOp != TI) |
| 2562 | std::swap(a&: NewTrueOp, b&: NewFalseOp); |
| 2563 | Value *NewSel = Builder.CreateSelect(C: CondVal, True: NewTrueOp, False: NewFalseOp, |
| 2564 | Name: SI.getName() + ".p" , MDFrom: &SI); |
| 2565 | |
| 2566 | if (SI.getType()->isFPOrFPVectorTy()) { |
| 2567 | Instruction *RI = |
| 2568 | BinaryOperator::CreateFAdd(V1: SubOp->getOperand(i: 0), V2: NewSel); |
| 2569 | |
| 2570 | FastMathFlags Flags = AddOp->getFastMathFlags(); |
| 2571 | Flags &= SubOp->getFastMathFlags(); |
| 2572 | RI->setFastMathFlags(Flags); |
| 2573 | return RI; |
| 2574 | } else |
| 2575 | return BinaryOperator::CreateAdd(V1: SubOp->getOperand(i: 0), V2: NewSel); |
| 2576 | } |
| 2577 | } |
| 2578 | return nullptr; |
| 2579 | } |
| 2580 | |
| 2581 | /// Turn X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
| 2582 | /// And X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
| 2583 | /// Along with a number of patterns similar to: |
| 2584 | /// X + Y overflows ? (X < 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2585 | /// X - Y overflows ? (X > 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2586 | static Instruction * |
| 2587 | foldOverflowingAddSubSelect(SelectInst &SI, InstCombiner::BuilderTy &Builder) { |
| 2588 | Value *CondVal = SI.getCondition(); |
| 2589 | Value *TrueVal = SI.getTrueValue(); |
| 2590 | Value *FalseVal = SI.getFalseValue(); |
| 2591 | |
| 2592 | WithOverflowInst *II; |
| 2593 | if (!match(V: CondVal, P: m_ExtractValue<1>(V: m_WithOverflowInst(I&: II))) || |
| 2594 | !match(V: FalseVal, P: m_ExtractValue<0>(V: m_Specific(V: II)))) |
| 2595 | return nullptr; |
| 2596 | |
| 2597 | Value *X = II->getLHS(); |
| 2598 | Value *Y = II->getRHS(); |
| 2599 | |
| 2600 | auto IsSignedSaturateLimit = [&](Value *Limit, bool IsAdd) { |
| 2601 | Type *Ty = Limit->getType(); |
| 2602 | |
| 2603 | CmpPredicate Pred; |
| 2604 | Value *TrueVal, *FalseVal, *Op; |
| 2605 | const APInt *C; |
| 2606 | if (!match(V: Limit, P: m_Select(C: m_ICmp(Pred, L: m_Value(V&: Op), R: m_APInt(Res&: C)), |
| 2607 | L: m_Value(V&: TrueVal), R: m_Value(V&: FalseVal)))) |
| 2608 | return false; |
| 2609 | |
| 2610 | auto IsZeroOrOne = [](const APInt &C) { return C.isZero() || C.isOne(); }; |
| 2611 | auto IsMinMax = [&](Value *Min, Value *Max) { |
| 2612 | APInt MinVal = APInt::getSignedMinValue(numBits: Ty->getScalarSizeInBits()); |
| 2613 | APInt MaxVal = APInt::getSignedMaxValue(numBits: Ty->getScalarSizeInBits()); |
| 2614 | return match(V: Min, P: m_SpecificInt(V: MinVal)) && |
| 2615 | match(V: Max, P: m_SpecificInt(V: MaxVal)); |
| 2616 | }; |
| 2617 | |
| 2618 | if (Op != X && Op != Y) |
| 2619 | return false; |
| 2620 | |
| 2621 | if (IsAdd) { |
| 2622 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2623 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2624 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2625 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2626 | if (Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
| 2627 | IsMinMax(TrueVal, FalseVal)) |
| 2628 | return true; |
| 2629 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2630 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2631 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2632 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2633 | if (Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
| 2634 | IsMinMax(FalseVal, TrueVal)) |
| 2635 | return true; |
| 2636 | } else { |
| 2637 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2638 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2639 | if (Op == X && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C + 1) && |
| 2640 | IsMinMax(TrueVal, FalseVal)) |
| 2641 | return true; |
| 2642 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2643 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2644 | if (Op == X && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 2) && |
| 2645 | IsMinMax(FalseVal, TrueVal)) |
| 2646 | return true; |
| 2647 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2648 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2649 | if (Op == Y && Pred == ICmpInst::ICMP_SLT && IsZeroOrOne(*C) && |
| 2650 | IsMinMax(FalseVal, TrueVal)) |
| 2651 | return true; |
| 2652 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2653 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2654 | if (Op == Y && Pred == ICmpInst::ICMP_SGT && IsZeroOrOne(*C + 1) && |
| 2655 | IsMinMax(TrueVal, FalseVal)) |
| 2656 | return true; |
| 2657 | } |
| 2658 | |
| 2659 | return false; |
| 2660 | }; |
| 2661 | |
| 2662 | Intrinsic::ID NewIntrinsicID; |
| 2663 | if (II->getIntrinsicID() == Intrinsic::uadd_with_overflow && |
| 2664 | match(V: TrueVal, P: m_AllOnes())) |
| 2665 | // X + Y overflows ? -1 : X + Y -> uadd_sat X, Y |
| 2666 | NewIntrinsicID = Intrinsic::uadd_sat; |
| 2667 | else if (II->getIntrinsicID() == Intrinsic::usub_with_overflow && |
| 2668 | match(V: TrueVal, P: m_Zero())) |
| 2669 | // X - Y overflows ? 0 : X - Y -> usub_sat X, Y |
| 2670 | NewIntrinsicID = Intrinsic::usub_sat; |
| 2671 | else if (II->getIntrinsicID() == Intrinsic::sadd_with_overflow && |
| 2672 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/true)) |
| 2673 | // X + Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2674 | // X + Y overflows ? (X <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2675 | // X + Y overflows ? (X >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2676 | // X + Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2677 | // X + Y overflows ? (Y <s 0 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2678 | // X + Y overflows ? (Y <s 1 ? INTMIN : INTMAX) : X + Y --> sadd_sat X, Y |
| 2679 | // X + Y overflows ? (Y >s 0 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2680 | // X + Y overflows ? (Y >s -1 ? INTMAX : INTMIN) : X + Y --> sadd_sat X, Y |
| 2681 | NewIntrinsicID = Intrinsic::sadd_sat; |
| 2682 | else if (II->getIntrinsicID() == Intrinsic::ssub_with_overflow && |
| 2683 | IsSignedSaturateLimit(TrueVal, /*IsAdd=*/false)) |
| 2684 | // X - Y overflows ? (X <s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2685 | // X - Y overflows ? (X <s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2686 | // X - Y overflows ? (X >s -1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2687 | // X - Y overflows ? (X >s -2 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2688 | // X - Y overflows ? (Y <s 0 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2689 | // X - Y overflows ? (Y <s 1 ? INTMAX : INTMIN) : X - Y --> ssub_sat X, Y |
| 2690 | // X - Y overflows ? (Y >s 0 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2691 | // X - Y overflows ? (Y >s -1 ? INTMIN : INTMAX) : X - Y --> ssub_sat X, Y |
| 2692 | NewIntrinsicID = Intrinsic::ssub_sat; |
| 2693 | else |
| 2694 | return nullptr; |
| 2695 | |
| 2696 | Function *F = Intrinsic::getOrInsertDeclaration(M: SI.getModule(), |
| 2697 | id: NewIntrinsicID, OverloadTys: SI.getType()); |
| 2698 | return CallInst::Create(Func: F, Args: {X, Y}); |
| 2699 | } |
| 2700 | |
| 2701 | Instruction *InstCombinerImpl::foldSelectExtConst(SelectInst &Sel) { |
| 2702 | Constant *C; |
| 2703 | if (!match(V: Sel.getTrueValue(), P: m_Constant(C)) && |
| 2704 | !match(V: Sel.getFalseValue(), P: m_Constant(C))) |
| 2705 | return nullptr; |
| 2706 | |
| 2707 | Instruction *ExtInst; |
| 2708 | if (!match(V: Sel.getTrueValue(), P: m_Instruction(I&: ExtInst)) && |
| 2709 | !match(V: Sel.getFalseValue(), P: m_Instruction(I&: ExtInst))) |
| 2710 | return nullptr; |
| 2711 | |
| 2712 | auto ExtOpcode = ExtInst->getOpcode(); |
| 2713 | if (ExtOpcode != Instruction::ZExt && ExtOpcode != Instruction::SExt) |
| 2714 | return nullptr; |
| 2715 | |
| 2716 | // If we are extending from a boolean type or if we can create a select that |
| 2717 | // has the same size operands as its condition, try to narrow the select. |
| 2718 | Value *X = ExtInst->getOperand(i: 0); |
| 2719 | Type *SmallType = X->getType(); |
| 2720 | Value *Cond = Sel.getCondition(); |
| 2721 | if (!SmallType->isIntOrIntVectorTy(BitWidth: 1) && |
| 2722 | (!isa<CmpInst, TruncInst>(Val: Cond) || |
| 2723 | cast<Instruction>(Val: Cond)->getOperand(i: 0)->getType() != SmallType)) |
| 2724 | return nullptr; |
| 2725 | |
| 2726 | // If the constant is the same after truncation to the smaller type and |
| 2727 | // extension to the original type, we can narrow the select. |
| 2728 | Type *SelType = Sel.getType(); |
| 2729 | Constant *TruncC = getLosslessInvCast(C, InvCastTo: SmallType, CastOp: ExtOpcode, DL); |
| 2730 | if (TruncC && ExtInst->hasOneUse()) { |
| 2731 | Value *TruncCVal = cast<Value>(Val: TruncC); |
| 2732 | if (ExtInst == Sel.getFalseValue()) |
| 2733 | std::swap(a&: X, b&: TruncCVal); |
| 2734 | |
| 2735 | // select Cond, (ext X), C --> ext(select Cond, X, C') |
| 2736 | // select Cond, C, (ext X) --> ext(select Cond, C', X) |
| 2737 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: TruncCVal, Name: "narrow" , MDFrom: &Sel); |
| 2738 | return CastInst::Create(Instruction::CastOps(ExtOpcode), S: NewSel, Ty: SelType); |
| 2739 | } |
| 2740 | |
| 2741 | return nullptr; |
| 2742 | } |
| 2743 | |
| 2744 | /// Try to transform a vector select with a constant condition vector into a |
| 2745 | /// shuffle for easier combining with other shuffles and insert/extract. |
| 2746 | static Instruction *canonicalizeSelectToShuffle(SelectInst &SI) { |
| 2747 | Value *CondVal = SI.getCondition(); |
| 2748 | Constant *CondC; |
| 2749 | auto *CondValTy = dyn_cast<FixedVectorType>(Val: CondVal->getType()); |
| 2750 | if (!CondValTy || !match(V: CondVal, P: m_Constant(C&: CondC))) |
| 2751 | return nullptr; |
| 2752 | |
| 2753 | unsigned NumElts = CondValTy->getNumElements(); |
| 2754 | SmallVector<int, 16> Mask; |
| 2755 | Mask.reserve(N: NumElts); |
| 2756 | for (unsigned i = 0; i != NumElts; ++i) { |
| 2757 | Constant *Elt = CondC->getAggregateElement(Elt: i); |
| 2758 | if (!Elt) |
| 2759 | return nullptr; |
| 2760 | |
| 2761 | if (Elt->isOneValue()) { |
| 2762 | // If the select condition element is true, choose from the 1st vector. |
| 2763 | Mask.push_back(Elt: i); |
| 2764 | } else if (Elt->isNullValue()) { |
| 2765 | // If the select condition element is false, choose from the 2nd vector. |
| 2766 | Mask.push_back(Elt: i + NumElts); |
| 2767 | } else if (isa<UndefValue>(Val: Elt)) { |
| 2768 | // Undef in a select condition (choose one of the operands) does not mean |
| 2769 | // the same thing as undef in a shuffle mask (any value is acceptable), so |
| 2770 | // give up. |
| 2771 | return nullptr; |
| 2772 | } else { |
| 2773 | // Bail out on a constant expression. |
| 2774 | return nullptr; |
| 2775 | } |
| 2776 | } |
| 2777 | |
| 2778 | return new ShuffleVectorInst(SI.getTrueValue(), SI.getFalseValue(), Mask); |
| 2779 | } |
| 2780 | |
| 2781 | /// If we have a select of vectors with a scalar condition, try to convert that |
| 2782 | /// to a vector select by splatting the condition. A splat may get folded with |
| 2783 | /// other operations in IR and having all operands of a select be vector types |
| 2784 | /// is likely better for vector codegen. |
| 2785 | static Instruction *canonicalizeScalarSelectOfVecs(SelectInst &Sel, |
| 2786 | InstCombinerImpl &IC) { |
| 2787 | auto *Ty = dyn_cast<VectorType>(Val: Sel.getType()); |
| 2788 | if (!Ty) |
| 2789 | return nullptr; |
| 2790 | |
| 2791 | // We can replace a single-use extract with constant index. |
| 2792 | Value *Cond = Sel.getCondition(); |
| 2793 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_ExtractElt(Val: m_Value(), Idx: m_ConstantInt())))) |
| 2794 | return nullptr; |
| 2795 | |
| 2796 | // select (extelt V, Index), T, F --> select (splat V, Index), T, F |
| 2797 | // Splatting the extracted condition reduces code (we could directly create a |
| 2798 | // splat shuffle of the source vector to eliminate the intermediate step). |
| 2799 | return IC.replaceOperand( |
| 2800 | I&: Sel, OpNum: 0, V: IC.Builder.CreateVectorSplat(EC: Ty->getElementCount(), V: Cond)); |
| 2801 | } |
| 2802 | |
| 2803 | /// Reuse bitcasted operands between a compare and select: |
| 2804 | /// select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
| 2805 | /// bitcast (select (cmp (bitcast C), (bitcast D)), (bitcast C), (bitcast D)) |
| 2806 | static Instruction *foldSelectCmpBitcasts(SelectInst &Sel, |
| 2807 | InstCombiner::BuilderTy &Builder) { |
| 2808 | Value *Cond = Sel.getCondition(); |
| 2809 | Value *TVal = Sel.getTrueValue(); |
| 2810 | Value *FVal = Sel.getFalseValue(); |
| 2811 | |
| 2812 | CmpPredicate Pred; |
| 2813 | Value *A, *B; |
| 2814 | if (!match(V: Cond, P: m_Cmp(Pred, L: m_Value(V&: A), R: m_Value(V&: B)))) |
| 2815 | return nullptr; |
| 2816 | |
| 2817 | // The select condition is a compare instruction. If the select's true/false |
| 2818 | // values are already the same as the compare operands, there's nothing to do. |
| 2819 | if (TVal == A || TVal == B || FVal == A || FVal == B) |
| 2820 | return nullptr; |
| 2821 | |
| 2822 | Value *C, *D; |
| 2823 | if (!match(V: A, P: m_BitCast(Op: m_Value(V&: C))) || !match(V: B, P: m_BitCast(Op: m_Value(V&: D)))) |
| 2824 | return nullptr; |
| 2825 | |
| 2826 | // select (cmp (bitcast C), (bitcast D)), (bitcast TSrc), (bitcast FSrc) |
| 2827 | Value *TSrc, *FSrc; |
| 2828 | if (!match(V: TVal, P: m_BitCast(Op: m_Value(V&: TSrc))) || |
| 2829 | !match(V: FVal, P: m_BitCast(Op: m_Value(V&: FSrc)))) |
| 2830 | return nullptr; |
| 2831 | |
| 2832 | // If the select true/false values are *different bitcasts* of the same source |
| 2833 | // operands, make the select operands the same as the compare operands and |
| 2834 | // cast the result. This is the canonical select form for min/max. |
| 2835 | Value *NewSel; |
| 2836 | if (TSrc == C && FSrc == D) { |
| 2837 | // select (cmp (bitcast C), (bitcast D)), (bitcast' C), (bitcast' D) --> |
| 2838 | // bitcast (select (cmp A, B), A, B) |
| 2839 | NewSel = Builder.CreateSelect(C: Cond, True: A, False: B, Name: "" , MDFrom: &Sel); |
| 2840 | } else if (TSrc == D && FSrc == C) { |
| 2841 | // select (cmp (bitcast C), (bitcast D)), (bitcast' D), (bitcast' C) --> |
| 2842 | // bitcast (select (cmp A, B), B, A) |
| 2843 | NewSel = Builder.CreateSelect(C: Cond, True: B, False: A, Name: "" , MDFrom: &Sel); |
| 2844 | } else { |
| 2845 | return nullptr; |
| 2846 | } |
| 2847 | return new BitCastInst(NewSel, Sel.getType()); |
| 2848 | } |
| 2849 | |
| 2850 | /// Try to eliminate select instructions that test the returned flag of cmpxchg |
| 2851 | /// instructions. |
| 2852 | /// |
| 2853 | /// If a select instruction tests the returned flag of a cmpxchg instruction and |
| 2854 | /// selects between the returned value of the cmpxchg instruction its compare |
| 2855 | /// operand, the result of the select will always be equal to its false value. |
| 2856 | /// For example: |
| 2857 | /// |
| 2858 | /// %cmpxchg = cmpxchg ptr %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
| 2859 | /// %val = extractvalue { i64, i1 } %cmpxchg, 0 |
| 2860 | /// %success = extractvalue { i64, i1 } %cmpxchg, 1 |
| 2861 | /// %sel = select i1 %success, i64 %compare, i64 %val |
| 2862 | /// ret i64 %sel |
| 2863 | /// |
| 2864 | /// The returned value of the cmpxchg instruction (%val) is the original value |
| 2865 | /// located at %ptr prior to any update. If the cmpxchg operation succeeds, %val |
| 2866 | /// must have been equal to %compare. Thus, the result of the select is always |
| 2867 | /// equal to %val, and the code can be simplified to: |
| 2868 | /// |
| 2869 | /// %cmpxchg = cmpxchg ptr %ptr, i64 %compare, i64 %new_value seq_cst seq_cst |
| 2870 | /// %val = extractvalue { i64, i1 } %cmpxchg, 0 |
| 2871 | /// ret i64 %val |
| 2872 | /// |
| 2873 | static Value *foldSelectCmpXchg(SelectInst &SI) { |
| 2874 | // A helper that determines if V is an extractvalue instruction whose |
| 2875 | // aggregate operand is a cmpxchg instruction and whose single index is equal |
| 2876 | // to I. If such conditions are true, the helper returns the cmpxchg |
| 2877 | // instruction; otherwise, a nullptr is returned. |
| 2878 | auto = [](Value *V, unsigned I) -> AtomicCmpXchgInst * { |
| 2879 | // When extracting the value loaded by a cmpxchg, allow peeking through a |
| 2880 | // bitcast. These are inserted for floating-point cmpxchg, for example: |
| 2881 | // %bc = bitcast float %compare to i32 |
| 2882 | // %cmpxchg = cmpxchg ptr %ptr, i32 %bc, i32 %new_value seq_cst seq_cst |
| 2883 | // %val = extractvalue { i32, i1 } %cmpxchg, 0 |
| 2884 | // %success = extractvalue { i32, i1 } %cmpxchg, 1 |
| 2885 | // %val.bc = bitcast i32 %val to float |
| 2886 | // %sel = select i1 %success, float %compare, float %val.bc |
| 2887 | if (auto *BI = dyn_cast<BitCastInst>(Val: V); BI && I == 0) |
| 2888 | V = BI->getOperand(i_nocapture: 0); |
| 2889 | auto * = dyn_cast<ExtractValueInst>(Val: V); |
| 2890 | if (!Extract) |
| 2891 | return nullptr; |
| 2892 | if (Extract->getIndices()[0] != I) |
| 2893 | return nullptr; |
| 2894 | return dyn_cast<AtomicCmpXchgInst>(Val: Extract->getAggregateOperand()); |
| 2895 | }; |
| 2896 | |
| 2897 | // Check if the compare value of a cmpxchg matches another value. |
| 2898 | auto isCompareSameAsValue = [](Value *CmpVal, Value *SelVal) { |
| 2899 | // The values match if they are the same or %CmpVal = bitcast %SelVal (see |
| 2900 | // above). |
| 2901 | if (CmpVal == SelVal || match(V: CmpVal, P: m_BitCast(Op: m_Specific(V: SelVal)))) |
| 2902 | return true; |
| 2903 | // For FP constants, the value may have been bitcast to Int directly. |
| 2904 | auto *IntC = dyn_cast<ConstantInt>(Val: CmpVal); |
| 2905 | auto *FpC = dyn_cast<ConstantFP>(Val: SelVal); |
| 2906 | return IntC && FpC && IntC->getValue() == FpC->getValue().bitcastToAPInt(); |
| 2907 | }; |
| 2908 | |
| 2909 | // If the select has a single user, and this user is a select instruction that |
| 2910 | // we can simplify, skip the cmpxchg simplification for now. |
| 2911 | if (SI.hasOneUse()) |
| 2912 | if (auto *Select = dyn_cast<SelectInst>(Val: SI.user_back())) |
| 2913 | if (Select->getCondition() == SI.getCondition()) |
| 2914 | if (Select->getFalseValue() == SI.getTrueValue() || |
| 2915 | Select->getTrueValue() == SI.getFalseValue()) |
| 2916 | return nullptr; |
| 2917 | |
| 2918 | // Ensure the select condition is the returned flag of a cmpxchg instruction. |
| 2919 | auto *CmpXchg = isExtractFromCmpXchg(SI.getCondition(), 1); |
| 2920 | if (!CmpXchg) |
| 2921 | return nullptr; |
| 2922 | |
| 2923 | // Check the true value case: The true value of the select is the returned |
| 2924 | // value of the same cmpxchg used by the condition, and the false value is the |
| 2925 | // cmpxchg instruction's compare operand. |
| 2926 | if (auto *X = isExtractFromCmpXchg(SI.getTrueValue(), 0)) |
| 2927 | if (X == CmpXchg && |
| 2928 | isCompareSameAsValue(X->getCompareOperand(), SI.getFalseValue())) |
| 2929 | return SI.getFalseValue(); |
| 2930 | |
| 2931 | // Check the false value case: The false value of the select is the returned |
| 2932 | // value of the same cmpxchg used by the condition, and the true value is the |
| 2933 | // cmpxchg instruction's compare operand. |
| 2934 | if (auto *X = isExtractFromCmpXchg(SI.getFalseValue(), 0)) |
| 2935 | if (X == CmpXchg && |
| 2936 | isCompareSameAsValue(X->getCompareOperand(), SI.getTrueValue())) |
| 2937 | return SI.getFalseValue(); |
| 2938 | |
| 2939 | return nullptr; |
| 2940 | } |
| 2941 | |
| 2942 | /// Try to reduce a funnel/rotate pattern that includes a compare and select |
| 2943 | /// into a funnel shift intrinsic. Example: |
| 2944 | /// rotl32(a, b) --> (b == 0 ? a : ((a >> (32 - b)) | (a << b))) |
| 2945 | /// --> call llvm.fshl.i32(a, a, b) |
| 2946 | /// fshl32(a, b, c) --> (c == 0 ? a : ((b >> (32 - c)) | (a << c))) |
| 2947 | /// --> call llvm.fshl.i32(a, b, c) |
| 2948 | /// fshr32(a, b, c) --> (c == 0 ? b : ((a >> (32 - c)) | (b << c))) |
| 2949 | /// --> call llvm.fshr.i32(a, b, c) |
| 2950 | static Instruction *foldSelectFunnelShift(SelectInst &Sel, |
| 2951 | InstCombiner::BuilderTy &Builder) { |
| 2952 | // This must be a power-of-2 type for a bitmasking transform to be valid. |
| 2953 | unsigned Width = Sel.getType()->getScalarSizeInBits(); |
| 2954 | if (!isPowerOf2_32(Value: Width)) |
| 2955 | return nullptr; |
| 2956 | |
| 2957 | BinaryOperator *Or0, *Or1; |
| 2958 | if (!match(V: Sel.getFalseValue(), P: m_OneUse(SubPattern: m_Or(L: m_BinOp(I&: Or0), R: m_BinOp(I&: Or1))))) |
| 2959 | return nullptr; |
| 2960 | |
| 2961 | Value *SV0, *SV1, *SA0, *SA1; |
| 2962 | if (!match(V: Or0, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: SV0), |
| 2963 | R: m_ZExtOrSelf(Op: m_Value(V&: SA0))))) || |
| 2964 | !match(V: Or1, P: m_OneUse(SubPattern: m_LogicalShift(L: m_Value(V&: SV1), |
| 2965 | R: m_ZExtOrSelf(Op: m_Value(V&: SA1))))) || |
| 2966 | Or0->getOpcode() == Or1->getOpcode()) |
| 2967 | return nullptr; |
| 2968 | |
| 2969 | // Canonicalize to or(shl(SV0, SA0), lshr(SV1, SA1)). |
| 2970 | if (Or0->getOpcode() == BinaryOperator::LShr) { |
| 2971 | std::swap(a&: Or0, b&: Or1); |
| 2972 | std::swap(a&: SV0, b&: SV1); |
| 2973 | std::swap(a&: SA0, b&: SA1); |
| 2974 | } |
| 2975 | assert(Or0->getOpcode() == BinaryOperator::Shl && |
| 2976 | Or1->getOpcode() == BinaryOperator::LShr && |
| 2977 | "Illegal or(shift,shift) pair" ); |
| 2978 | |
| 2979 | // Check the shift amounts to see if they are an opposite pair. |
| 2980 | Value *ShAmt; |
| 2981 | if (match(V: SA1, P: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: Width), R: m_Specific(V: SA0))))) |
| 2982 | ShAmt = SA0; |
| 2983 | else if (match(V: SA0, P: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: Width), R: m_Specific(V: SA1))))) |
| 2984 | ShAmt = SA1; |
| 2985 | else |
| 2986 | return nullptr; |
| 2987 | |
| 2988 | // We should now have this pattern: |
| 2989 | // select ?, TVal, (or (shl SV0, SA0), (lshr SV1, SA1)) |
| 2990 | // The false value of the select must be a funnel-shift of the true value: |
| 2991 | // IsFShl -> TVal must be SV0 else TVal must be SV1. |
| 2992 | bool IsFshl = (ShAmt == SA0); |
| 2993 | Value *TVal = Sel.getTrueValue(); |
| 2994 | if ((IsFshl && TVal != SV0) || (!IsFshl && TVal != SV1)) |
| 2995 | return nullptr; |
| 2996 | |
| 2997 | // Finally, see if the select is filtering out a shift-by-zero. |
| 2998 | Value *Cond = Sel.getCondition(); |
| 2999 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, L: m_Specific(V: ShAmt), |
| 3000 | R: m_ZeroInt())))) |
| 3001 | return nullptr; |
| 3002 | |
| 3003 | // If this is not a rotate then the select was blocking poison from the |
| 3004 | // 'shift-by-zero' non-TVal, but a funnel shift won't - so freeze it. |
| 3005 | if (SV0 != SV1) { |
| 3006 | if (IsFshl && !llvm::isGuaranteedNotToBePoison(V: SV1)) |
| 3007 | SV1 = Builder.CreateFreeze(V: SV1); |
| 3008 | else if (!IsFshl && !llvm::isGuaranteedNotToBePoison(V: SV0)) |
| 3009 | SV0 = Builder.CreateFreeze(V: SV0); |
| 3010 | } |
| 3011 | |
| 3012 | // This is a funnel/rotate that avoids shift-by-bitwidth UB in a suboptimal way. |
| 3013 | // Convert to funnel shift intrinsic. |
| 3014 | Intrinsic::ID IID = IsFshl ? Intrinsic::fshl : Intrinsic::fshr; |
| 3015 | Function *F = |
| 3016 | Intrinsic::getOrInsertDeclaration(M: Sel.getModule(), id: IID, OverloadTys: Sel.getType()); |
| 3017 | ShAmt = Builder.CreateZExt(V: ShAmt, DestTy: Sel.getType()); |
| 3018 | return CallInst::Create(Func: F, Args: { SV0, SV1, ShAmt }); |
| 3019 | } |
| 3020 | |
| 3021 | static Instruction *foldSelectToCopysign(SelectInst &Sel, |
| 3022 | InstCombiner::BuilderTy &Builder) { |
| 3023 | Value *Cond = Sel.getCondition(); |
| 3024 | Value *TVal = Sel.getTrueValue(); |
| 3025 | Value *FVal = Sel.getFalseValue(); |
| 3026 | Type *SelType = Sel.getType(); |
| 3027 | |
| 3028 | // Match select ?, TC, FC where the constants are equal but negated. |
| 3029 | // TODO: Generalize to handle a negated variable operand? |
| 3030 | const APFloat *TC, *FC; |
| 3031 | if (!match(V: TVal, P: m_APFloatAllowPoison(Res&: TC)) || |
| 3032 | !match(V: FVal, P: m_APFloatAllowPoison(Res&: FC)) || |
| 3033 | !abs(X: *TC).bitwiseIsEqual(RHS: abs(X: *FC))) |
| 3034 | return nullptr; |
| 3035 | |
| 3036 | assert(TC != FC && "Expected equal select arms to simplify" ); |
| 3037 | |
| 3038 | Value *X; |
| 3039 | const APInt *C; |
| 3040 | bool IsTrueIfSignSet; |
| 3041 | CmpPredicate Pred; |
| 3042 | if (!match(V: Cond, P: m_OneUse(SubPattern: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Value(V&: X)), |
| 3043 | R: m_APInt(Res&: C)))) || |
| 3044 | !isSignBitCheck(Pred, RHS: *C, TrueIfSigned&: IsTrueIfSignSet) || X->getType() != SelType) |
| 3045 | return nullptr; |
| 3046 | |
| 3047 | // If needed, negate the value that will be the sign argument of the copysign: |
| 3048 | // (bitcast X) < 0 ? -TC : TC --> copysign(TC, X) |
| 3049 | // (bitcast X) < 0 ? TC : -TC --> copysign(TC, -X) |
| 3050 | // (bitcast X) >= 0 ? -TC : TC --> copysign(TC, -X) |
| 3051 | // (bitcast X) >= 0 ? TC : -TC --> copysign(TC, X) |
| 3052 | // Note: FMF from the select can not be propagated to the new instructions. |
| 3053 | if (IsTrueIfSignSet ^ TC->isNegative()) |
| 3054 | X = Builder.CreateFNeg(V: X); |
| 3055 | |
| 3056 | // Canonicalize the magnitude argument as the positive constant since we do |
| 3057 | // not care about its sign. |
| 3058 | Value *MagArg = ConstantFP::get(Ty: SelType, V: abs(X: *TC)); |
| 3059 | Function *F = Intrinsic::getOrInsertDeclaration( |
| 3060 | M: Sel.getModule(), id: Intrinsic::copysign, OverloadTys: Sel.getType()); |
| 3061 | return CallInst::Create(Func: F, Args: { MagArg, X }); |
| 3062 | } |
| 3063 | |
| 3064 | Instruction *InstCombinerImpl::foldVectorSelect(SelectInst &Sel) { |
| 3065 | if (!isa<VectorType>(Val: Sel.getType())) |
| 3066 | return nullptr; |
| 3067 | |
| 3068 | Value *Cond = Sel.getCondition(); |
| 3069 | Value *TVal = Sel.getTrueValue(); |
| 3070 | Value *FVal = Sel.getFalseValue(); |
| 3071 | Value *C, *X, *Y; |
| 3072 | |
| 3073 | if (match(V: Cond, P: m_VecReverse(Op0: m_Value(V&: C)))) { |
| 3074 | auto createSelReverse = [&](Value *C, Value *X, Value *Y) { |
| 3075 | Value *V = Builder.CreateSelect(C, True: X, False: Y, Name: Sel.getName(), MDFrom: &Sel); |
| 3076 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
| 3077 | I->copyIRFlags(V: &Sel); |
| 3078 | Module *M = Sel.getModule(); |
| 3079 | Function *F = Intrinsic::getOrInsertDeclaration( |
| 3080 | M, id: Intrinsic::vector_reverse, OverloadTys: V->getType()); |
| 3081 | return CallInst::Create(Func: F, Args: V); |
| 3082 | }; |
| 3083 | |
| 3084 | if (match(V: TVal, P: m_VecReverse(Op0: m_Value(V&: X)))) { |
| 3085 | // select rev(C), rev(X), rev(Y) --> rev(select C, X, Y) |
| 3086 | if (match(V: FVal, P: m_VecReverse(Op0: m_Value(V&: Y))) && |
| 3087 | (Cond->hasOneUse() || TVal->hasOneUse() || FVal->hasOneUse())) |
| 3088 | return createSelReverse(C, X, Y); |
| 3089 | |
| 3090 | // select rev(C), rev(X), FValSplat --> rev(select C, X, FValSplat) |
| 3091 | if ((Cond->hasOneUse() || TVal->hasOneUse()) && isSplatValue(V: FVal)) |
| 3092 | return createSelReverse(C, X, FVal); |
| 3093 | } |
| 3094 | // select rev(C), TValSplat, rev(Y) --> rev(select C, TValSplat, Y) |
| 3095 | else if (isSplatValue(V: TVal) && match(V: FVal, P: m_VecReverse(Op0: m_Value(V&: Y))) && |
| 3096 | (Cond->hasOneUse() || FVal->hasOneUse())) |
| 3097 | return createSelReverse(C, TVal, Y); |
| 3098 | } |
| 3099 | |
| 3100 | auto *VecTy = dyn_cast<FixedVectorType>(Val: Sel.getType()); |
| 3101 | if (!VecTy) |
| 3102 | return nullptr; |
| 3103 | |
| 3104 | unsigned NumElts = VecTy->getNumElements(); |
| 3105 | APInt PoisonElts(NumElts, 0); |
| 3106 | APInt AllOnesEltMask(APInt::getAllOnes(numBits: NumElts)); |
| 3107 | if (Value *V = SimplifyDemandedVectorElts(V: &Sel, DemandedElts: AllOnesEltMask, PoisonElts)) { |
| 3108 | if (V != &Sel) |
| 3109 | return replaceInstUsesWith(I&: Sel, V); |
| 3110 | return &Sel; |
| 3111 | } |
| 3112 | |
| 3113 | // A select of a "select shuffle" with a common operand can be rearranged |
| 3114 | // to select followed by "select shuffle". Because of poison, this only works |
| 3115 | // in the case of a shuffle with no undefined mask elements. |
| 3116 | ArrayRef<int> Mask; |
| 3117 | if (match(V: TVal, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: X), v2: m_Value(V&: Y), mask: m_Mask(Mask)))) && |
| 3118 | !is_contained(Range&: Mask, Element: PoisonMaskElem) && |
| 3119 | cast<ShuffleVectorInst>(Val: TVal)->isSelect()) { |
| 3120 | if (X == FVal) { |
| 3121 | // select Cond, (shuf_sel X, Y), X --> shuf_sel X, (select Cond, Y, X) |
| 3122 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Y, False: X, Name: "sel" , MDFrom: &Sel); |
| 3123 | return new ShuffleVectorInst(X, NewSel, Mask); |
| 3124 | } |
| 3125 | if (Y == FVal) { |
| 3126 | // select Cond, (shuf_sel X, Y), Y --> shuf_sel (select Cond, X, Y), Y |
| 3127 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: Y, Name: "sel" , MDFrom: &Sel); |
| 3128 | return new ShuffleVectorInst(NewSel, Y, Mask); |
| 3129 | } |
| 3130 | } |
| 3131 | if (match(V: FVal, P: m_OneUse(SubPattern: m_Shuffle(v1: m_Value(V&: X), v2: m_Value(V&: Y), mask: m_Mask(Mask)))) && |
| 3132 | !is_contained(Range&: Mask, Element: PoisonMaskElem) && |
| 3133 | cast<ShuffleVectorInst>(Val: FVal)->isSelect()) { |
| 3134 | if (X == TVal) { |
| 3135 | // select Cond, X, (shuf_sel X, Y) --> shuf_sel X, (select Cond, X, Y) |
| 3136 | Value *NewSel = Builder.CreateSelect(C: Cond, True: X, False: Y, Name: "sel" , MDFrom: &Sel); |
| 3137 | return new ShuffleVectorInst(X, NewSel, Mask); |
| 3138 | } |
| 3139 | if (Y == TVal) { |
| 3140 | // select Cond, Y, (shuf_sel X, Y) --> shuf_sel (select Cond, Y, X), Y |
| 3141 | Value *NewSel = Builder.CreateSelect(C: Cond, True: Y, False: X, Name: "sel" , MDFrom: &Sel); |
| 3142 | return new ShuffleVectorInst(NewSel, Y, Mask); |
| 3143 | } |
| 3144 | } |
| 3145 | |
| 3146 | return nullptr; |
| 3147 | } |
| 3148 | |
| 3149 | static Instruction *foldSelectToPhiImpl(SelectInst &Sel, BasicBlock *BB, |
| 3150 | const DominatorTree &DT, |
| 3151 | InstCombiner::BuilderTy &Builder) { |
| 3152 | // Find the block's immediate dominator that ends with a conditional branch |
| 3153 | // that matches select's condition (maybe inverted). |
| 3154 | auto *IDomNode = DT[BB]->getIDom(); |
| 3155 | if (!IDomNode) |
| 3156 | return nullptr; |
| 3157 | BasicBlock *IDom = IDomNode->getBlock(); |
| 3158 | |
| 3159 | Value *Cond = Sel.getCondition(); |
| 3160 | Value *IfTrue, *IfFalse; |
| 3161 | BasicBlock *TrueSucc, *FalseSucc; |
| 3162 | if (match(V: IDom->getTerminator(), |
| 3163 | P: m_Br(C: m_Specific(V: Cond), T: m_BasicBlock(V&: TrueSucc), |
| 3164 | F: m_BasicBlock(V&: FalseSucc)))) { |
| 3165 | IfTrue = Sel.getTrueValue(); |
| 3166 | IfFalse = Sel.getFalseValue(); |
| 3167 | } else if (match(V: IDom->getTerminator(), |
| 3168 | P: m_Br(C: m_Not(V: m_Specific(V: Cond)), T: m_BasicBlock(V&: TrueSucc), |
| 3169 | F: m_BasicBlock(V&: FalseSucc)))) { |
| 3170 | IfTrue = Sel.getFalseValue(); |
| 3171 | IfFalse = Sel.getTrueValue(); |
| 3172 | } else |
| 3173 | return nullptr; |
| 3174 | |
| 3175 | // Make sure the branches are actually different. |
| 3176 | if (TrueSucc == FalseSucc) |
| 3177 | return nullptr; |
| 3178 | |
| 3179 | // We want to replace select %cond, %a, %b with a phi that takes value %a |
| 3180 | // for all incoming edges that are dominated by condition `%cond == true`, |
| 3181 | // and value %b for edges dominated by condition `%cond == false`. If %a |
| 3182 | // or %b are also phis from the same basic block, we can go further and take |
| 3183 | // their incoming values from the corresponding blocks. |
| 3184 | BasicBlockEdge TrueEdge(IDom, TrueSucc); |
| 3185 | BasicBlockEdge FalseEdge(IDom, FalseSucc); |
| 3186 | DenseMap<BasicBlock *, Value *> Inputs; |
| 3187 | for (auto *Pred : predecessors(BB)) { |
| 3188 | // Check implication. |
| 3189 | BasicBlockEdge Incoming(Pred, BB); |
| 3190 | if (DT.dominates(BBE1: TrueEdge, BBE2: Incoming)) |
| 3191 | Inputs[Pred] = IfTrue->DoPHITranslation(CurBB: BB, PredBB: Pred); |
| 3192 | else if (DT.dominates(BBE1: FalseEdge, BBE2: Incoming)) |
| 3193 | Inputs[Pred] = IfFalse->DoPHITranslation(CurBB: BB, PredBB: Pred); |
| 3194 | else |
| 3195 | return nullptr; |
| 3196 | // Check availability. |
| 3197 | if (auto *Insn = dyn_cast<Instruction>(Val: Inputs[Pred])) |
| 3198 | if (!DT.dominates(Def: Insn, User: Pred->getTerminator())) |
| 3199 | return nullptr; |
| 3200 | } |
| 3201 | |
| 3202 | Builder.SetInsertPoint(TheBB: BB, IP: BB->begin()); |
| 3203 | auto *PN = Builder.CreatePHI(Ty: Sel.getType(), NumReservedValues: Inputs.size()); |
| 3204 | for (auto *Pred : predecessors(BB)) |
| 3205 | PN->addIncoming(V: Inputs[Pred], BB: Pred); |
| 3206 | PN->takeName(V: &Sel); |
| 3207 | return PN; |
| 3208 | } |
| 3209 | |
| 3210 | static Instruction *foldSelectToPhi(SelectInst &Sel, const DominatorTree &DT, |
| 3211 | InstCombiner::BuilderTy &Builder) { |
| 3212 | // Try to replace this select with Phi in one of these blocks. |
| 3213 | SmallSetVector<BasicBlock *, 4> CandidateBlocks; |
| 3214 | CandidateBlocks.insert(X: Sel.getParent()); |
| 3215 | for (Value *V : Sel.operands()) |
| 3216 | if (auto *I = dyn_cast<Instruction>(Val: V)) |
| 3217 | CandidateBlocks.insert(X: I->getParent()); |
| 3218 | |
| 3219 | for (BasicBlock *BB : CandidateBlocks) |
| 3220 | if (auto *PN = foldSelectToPhiImpl(Sel, BB, DT, Builder)) |
| 3221 | return PN; |
| 3222 | return nullptr; |
| 3223 | } |
| 3224 | |
| 3225 | /// Tries to reduce a pattern that arises when calculating the remainder of the |
| 3226 | /// Euclidean division. When the divisor is a power of two and is guaranteed not |
| 3227 | /// to be negative, a signed remainder can be folded with a bitwise and. |
| 3228 | /// |
| 3229 | /// (x % n) < 0 ? (x % n) + n : (x % n) |
| 3230 | /// -> x & (n - 1) |
| 3231 | static Instruction *foldSelectWithSRem(SelectInst &SI, InstCombinerImpl &IC, |
| 3232 | IRBuilderBase &Builder) { |
| 3233 | Value *CondVal = SI.getCondition(); |
| 3234 | Value *TrueVal = SI.getTrueValue(); |
| 3235 | Value *FalseVal = SI.getFalseValue(); |
| 3236 | |
| 3237 | CmpPredicate Pred; |
| 3238 | Value *Op, *RemRes, *Remainder; |
| 3239 | const APInt *C; |
| 3240 | bool TrueIfSigned = false; |
| 3241 | |
| 3242 | if (!(match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: RemRes), R: m_APInt(Res&: C))) && |
| 3243 | isSignBitCheck(Pred, RHS: *C, TrueIfSigned))) |
| 3244 | return nullptr; |
| 3245 | |
| 3246 | // If the sign bit is not set, we have a SGE/SGT comparison, and the operands |
| 3247 | // of the select are inverted. |
| 3248 | if (!TrueIfSigned) |
| 3249 | std::swap(a&: TrueVal, b&: FalseVal); |
| 3250 | |
| 3251 | auto FoldToBitwiseAnd = [&](Value *Remainder) -> Instruction * { |
| 3252 | Value *Add = Builder.CreateAdd( |
| 3253 | LHS: Remainder, RHS: Constant::getAllOnesValue(Ty: RemRes->getType())); |
| 3254 | return BinaryOperator::CreateAnd(V1: Op, V2: Add); |
| 3255 | }; |
| 3256 | |
| 3257 | // Match the general case: |
| 3258 | // %rem = srem i32 %x, %n |
| 3259 | // %cnd = icmp slt i32 %rem, 0 |
| 3260 | // %add = add i32 %rem, %n |
| 3261 | // %sel = select i1 %cnd, i32 %add, i32 %rem |
| 3262 | if (match(V: TrueVal, P: m_c_Add(L: m_Specific(V: RemRes), R: m_Value(V&: Remainder))) && |
| 3263 | match(V: RemRes, P: m_SRem(L: m_Value(V&: Op), R: m_Specific(V: Remainder))) && |
| 3264 | IC.isKnownToBeAPowerOfTwo(V: Remainder, /*OrZero=*/true) && |
| 3265 | FalseVal == RemRes) |
| 3266 | return FoldToBitwiseAnd(Remainder); |
| 3267 | |
| 3268 | // Match the case where the one arm has been replaced by constant 1: |
| 3269 | // %rem = srem i32 %n, 2 |
| 3270 | // %cnd = icmp slt i32 %rem, 0 |
| 3271 | // %sel = select i1 %cnd, i32 1, i32 %rem |
| 3272 | if (match(V: TrueVal, P: m_One()) && |
| 3273 | match(V: RemRes, P: m_SRem(L: m_Value(V&: Op), R: m_SpecificInt(V: 2))) && |
| 3274 | FalseVal == RemRes) |
| 3275 | return FoldToBitwiseAnd(ConstantInt::get(Ty: RemRes->getType(), V: 2)); |
| 3276 | |
| 3277 | return nullptr; |
| 3278 | } |
| 3279 | |
| 3280 | /// Given that \p CondVal is known to be \p CondIsTrue, try to simplify \p SI. |
| 3281 | static Value *simplifyNestedSelectsUsingImpliedCond(SelectInst &SI, |
| 3282 | Value *CondVal, |
| 3283 | bool CondIsTrue, |
| 3284 | const DataLayout &DL) { |
| 3285 | Value *InnerCondVal = SI.getCondition(); |
| 3286 | Value *InnerTrueVal = SI.getTrueValue(); |
| 3287 | Value *InnerFalseVal = SI.getFalseValue(); |
| 3288 | assert(CondVal->getType() == InnerCondVal->getType() && |
| 3289 | "The type of inner condition must match with the outer." ); |
| 3290 | if (auto Implied = isImpliedCondition(LHS: CondVal, RHS: InnerCondVal, DL, LHSIsTrue: CondIsTrue)) |
| 3291 | return *Implied ? InnerTrueVal : InnerFalseVal; |
| 3292 | return nullptr; |
| 3293 | } |
| 3294 | |
| 3295 | Instruction *InstCombinerImpl::foldAndOrOfSelectUsingImpliedCond(Value *Op, |
| 3296 | SelectInst &SI, |
| 3297 | bool IsAnd) { |
| 3298 | assert(Op->getType()->isIntOrIntVectorTy(1) && |
| 3299 | "Op must be either i1 or vector of i1." ); |
| 3300 | if (SI.getCondition()->getType() != Op->getType()) |
| 3301 | return nullptr; |
| 3302 | if (Value *V = simplifyNestedSelectsUsingImpliedCond(SI, CondVal: Op, CondIsTrue: IsAnd, DL)) |
| 3303 | return createSelectInstWithUnknownProfile( |
| 3304 | C: Op, S1: IsAnd ? V : ConstantInt::getTrue(Ty: Op->getType()), |
| 3305 | S2: IsAnd ? ConstantInt::getFalse(Ty: Op->getType()) : V); |
| 3306 | return nullptr; |
| 3307 | } |
| 3308 | |
| 3309 | // Canonicalize select with fcmp to fabs(). -0.0 makes this tricky. We need |
| 3310 | // fast-math-flags (nsz) or fsub with +0.0 (not fneg) for this to work. |
| 3311 | static Instruction *foldSelectWithFCmpToFabs(SelectInst &SI, |
| 3312 | InstCombinerImpl &IC) { |
| 3313 | Value *CondVal = SI.getCondition(); |
| 3314 | |
| 3315 | bool ChangedFMF = false; |
| 3316 | for (bool Swap : {false, true}) { |
| 3317 | Value *TrueVal = SI.getTrueValue(); |
| 3318 | Value *X = SI.getFalseValue(); |
| 3319 | CmpPredicate Pred; |
| 3320 | |
| 3321 | if (Swap) |
| 3322 | std::swap(a&: TrueVal, b&: X); |
| 3323 | |
| 3324 | if (!match(V: CondVal, P: m_FCmp(Pred, L: m_Specific(V: X), R: m_AnyZeroFP()))) |
| 3325 | continue; |
| 3326 | |
| 3327 | // fold (X <= +/-0.0) ? (0.0 - X) : X to fabs(X), when 'Swap' is false |
| 3328 | // fold (X > +/-0.0) ? X : (0.0 - X) to fabs(X), when 'Swap' is true |
| 3329 | // Note: We require "nnan" for this fold because fcmp ignores the signbit |
| 3330 | // of NAN, but IEEE-754 specifies the signbit of NAN values with |
| 3331 | // fneg/fabs operations. |
| 3332 | if (match(V: TrueVal, P: m_FSub(L: m_PosZeroFP(), R: m_Specific(V: X))) && |
| 3333 | (cast<FPMathOperator>(Val: CondVal)->hasNoNaNs() || SI.hasNoNaNs() || |
| 3334 | (SI.hasOneUse() && canIgnoreSignBitOfNaN(U: *SI.use_begin())) || |
| 3335 | isKnownNeverNaN(V: X, SQ: IC.getSimplifyQuery().getWithInstruction( |
| 3336 | I: cast<Instruction>(Val: CondVal))))) { |
| 3337 | if (!Swap && (Pred == FCmpInst::FCMP_OLE || Pred == FCmpInst::FCMP_ULE)) { |
| 3338 | Value *Fabs = IC.Builder.CreateFAbs(V: X, FMFSource: &SI); |
| 3339 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
| 3340 | } |
| 3341 | if (Swap && (Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_UGT)) { |
| 3342 | Value *Fabs = IC.Builder.CreateFAbs(V: X, FMFSource: &SI); |
| 3343 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
| 3344 | } |
| 3345 | } |
| 3346 | |
| 3347 | if (!match(V: TrueVal, P: m_FNeg(X: m_Specific(V: X)))) |
| 3348 | return nullptr; |
| 3349 | |
| 3350 | // Forward-propagate nnan and ninf from the fcmp to the select. |
| 3351 | // If all inputs are not those values, then the select is not either. |
| 3352 | // Note: nsz is defined differently, so it may not be correct to propagate. |
| 3353 | FastMathFlags FMF = cast<FPMathOperator>(Val: CondVal)->getFastMathFlags(); |
| 3354 | if (FMF.noNaNs() && !SI.hasNoNaNs()) { |
| 3355 | SI.setHasNoNaNs(true); |
| 3356 | ChangedFMF = true; |
| 3357 | } |
| 3358 | if (FMF.noInfs() && !SI.hasNoInfs()) { |
| 3359 | SI.setHasNoInfs(true); |
| 3360 | ChangedFMF = true; |
| 3361 | } |
| 3362 | // Forward-propagate nnan from the fneg to the select. |
| 3363 | // The nnan flag can be propagated iff fneg is selected when X is NaN. |
| 3364 | if (!SI.hasNoNaNs() && cast<FPMathOperator>(Val: TrueVal)->hasNoNaNs() && |
| 3365 | (Swap ? FCmpInst::isOrdered(predicate: Pred) : FCmpInst::isUnordered(predicate: Pred))) { |
| 3366 | SI.setHasNoNaNs(true); |
| 3367 | ChangedFMF = true; |
| 3368 | } |
| 3369 | |
| 3370 | // With nsz, when 'Swap' is false: |
| 3371 | // fold (X < +/-0.0) ? -X : X or (X <= +/-0.0) ? -X : X to fabs(X) |
| 3372 | // fold (X > +/-0.0) ? -X : X or (X >= +/-0.0) ? -X : X to -fabs(x) |
| 3373 | // when 'Swap' is true: |
| 3374 | // fold (X > +/-0.0) ? X : -X or (X >= +/-0.0) ? X : -X to fabs(X) |
| 3375 | // fold (X < +/-0.0) ? X : -X or (X <= +/-0.0) ? X : -X to -fabs(X) |
| 3376 | // |
| 3377 | // Note: We require "nnan" for this fold because fcmp ignores the signbit |
| 3378 | // of NAN, but IEEE-754 specifies the signbit of NAN values with |
| 3379 | // fneg/fabs operations. |
| 3380 | if (!SI.hasNoSignedZeros() && |
| 3381 | (!SI.hasOneUse() || !canIgnoreSignBitOfZero(U: *SI.use_begin()))) |
| 3382 | return nullptr; |
| 3383 | if (!SI.hasNoNaNs() && |
| 3384 | (!SI.hasOneUse() || !canIgnoreSignBitOfNaN(U: *SI.use_begin()))) |
| 3385 | return nullptr; |
| 3386 | |
| 3387 | if (Swap) |
| 3388 | Pred = FCmpInst::getSwappedPredicate(pred: Pred); |
| 3389 | |
| 3390 | bool IsLTOrLE = Pred == FCmpInst::FCMP_OLT || Pred == FCmpInst::FCMP_OLE || |
| 3391 | Pred == FCmpInst::FCMP_ULT || Pred == FCmpInst::FCMP_ULE; |
| 3392 | bool IsGTOrGE = Pred == FCmpInst::FCMP_OGT || Pred == FCmpInst::FCMP_OGE || |
| 3393 | Pred == FCmpInst::FCMP_UGT || Pred == FCmpInst::FCMP_UGE; |
| 3394 | |
| 3395 | if (IsLTOrLE) { |
| 3396 | Value *Fabs = IC.Builder.CreateFAbs(V: X, FMFSource: &SI); |
| 3397 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
| 3398 | } |
| 3399 | if (IsGTOrGE) { |
| 3400 | Value *Fabs = IC.Builder.CreateFAbs(V: X, FMFSource: &SI); |
| 3401 | Instruction *NewFNeg = UnaryOperator::CreateFNeg(V: Fabs); |
| 3402 | NewFNeg->setFastMathFlags(SI.getFastMathFlags()); |
| 3403 | return NewFNeg; |
| 3404 | } |
| 3405 | } |
| 3406 | |
| 3407 | // Match select with (icmp slt (bitcast X to int), 0) |
| 3408 | // or (icmp sgt (bitcast X to int), -1) |
| 3409 | |
| 3410 | for (bool Swap : {false, true}) { |
| 3411 | Value *TrueVal = SI.getTrueValue(); |
| 3412 | Value *X = SI.getFalseValue(); |
| 3413 | |
| 3414 | if (Swap) |
| 3415 | std::swap(a&: TrueVal, b&: X); |
| 3416 | |
| 3417 | CmpPredicate Pred; |
| 3418 | const APInt *C; |
| 3419 | bool TrueIfSigned; |
| 3420 | if (!match(V: CondVal, |
| 3421 | P: m_ICmp(Pred, L: m_ElementWiseBitCast(Op: m_Specific(V: X)), R: m_APInt(Res&: C))) || |
| 3422 | !isSignBitCheck(Pred, RHS: *C, TrueIfSigned)) |
| 3423 | continue; |
| 3424 | if (!match(V: TrueVal, P: m_FNeg(X: m_Specific(V: X)))) |
| 3425 | return nullptr; |
| 3426 | if (Swap == TrueIfSigned && !CondVal->hasOneUse() && !TrueVal->hasOneUse()) |
| 3427 | return nullptr; |
| 3428 | |
| 3429 | // Fold (IsNeg ? -X : X) or (!IsNeg ? X : -X) to fabs(X) |
| 3430 | // Fold (IsNeg ? X : -X) or (!IsNeg ? -X : X) to -fabs(X) |
| 3431 | Value *Fabs = IC.Builder.CreateFAbs(V: X, FMFSource: &SI); |
| 3432 | if (Swap != TrueIfSigned) |
| 3433 | return IC.replaceInstUsesWith(I&: SI, V: Fabs); |
| 3434 | return UnaryOperator::CreateFNegFMF(Op: Fabs, FMFSource: &SI); |
| 3435 | } |
| 3436 | |
| 3437 | return ChangedFMF ? &SI : nullptr; |
| 3438 | } |
| 3439 | |
| 3440 | // Fold a select of an ordered fcmp using fabs of a NaN-scrubbed value: |
| 3441 | // %s = select i1 (isnotnan T %x), T %x, T %y |
| 3442 | // %a = call T @llvm.fabs.T(T %s) |
| 3443 | // %c = fcmp <ordered-pred> T %a, %k |
| 3444 | // %r = select i1 %c, T %s, T %y |
| 3445 | // => |
| 3446 | // %a2 = call T @llvm.fabs.T(T %x) |
| 3447 | // %c2 = fcmp <ordered-pred> T %a2, %k |
| 3448 | // %r2 = select i1 %c2, T %x, T %y |
| 3449 | static Instruction * |
| 3450 | foldSelectOfOrderedFAbsCmpOfNaNScrubbedValue(SelectInst &SI, |
| 3451 | InstCombinerImpl &IC) { |
| 3452 | Instruction *OuterCmpI; |
| 3453 | Value *Cmp0, *Cmp1; |
| 3454 | if (!match(V: SI.getCondition(), |
| 3455 | P: m_OneUse(SubPattern: m_Instruction(I&: OuterCmpI, |
| 3456 | P: m_FCmp(L: m_Value(V&: Cmp0), R: m_Value(V&: Cmp1)))))) |
| 3457 | return nullptr; |
| 3458 | |
| 3459 | auto *OuterCmp = cast<FCmpInst>(Val: OuterCmpI); |
| 3460 | CmpInst::Predicate Pred = OuterCmp->getPredicate(); |
| 3461 | if (!FCmpInst::isOrdered(predicate: Pred)) |
| 3462 | return nullptr; |
| 3463 | |
| 3464 | Value *Y = SI.getFalseValue(); |
| 3465 | Value *InnerSel = SI.getTrueValue(); |
| 3466 | |
| 3467 | // Match a select that returns X when X is not NaN, and Y otherwise: |
| 3468 | // select (fcmp ord X, 0.0), X, Y |
| 3469 | Value *X; |
| 3470 | if (!match(V: InnerSel, |
| 3471 | P: m_Select(C: m_OneUse(SubPattern: m_SpecificFCmp(MatchPred: FCmpInst::FCMP_ORD, L: m_Value(V&: X), |
| 3472 | R: m_AnyZeroFP())), |
| 3473 | L: m_Deferred(V: X), R: m_Specific(V: Y)))) |
| 3474 | return nullptr; |
| 3475 | |
| 3476 | Instruction *FAbsI; |
| 3477 | auto MatchFAbsOfInnerSel = [&](Value *V) { |
| 3478 | return match(V, |
| 3479 | P: m_OneUse(SubPattern: m_Instruction(I&: FAbsI, P: m_FAbs(Op0: m_Specific(V: InnerSel))))); |
| 3480 | }; |
| 3481 | |
| 3482 | if (!MatchFAbsOfInnerSel(Cmp0)) { |
| 3483 | if (!MatchFAbsOfInnerSel(Cmp1)) |
| 3484 | return nullptr; |
| 3485 | |
| 3486 | std::swap(a&: Cmp0, b&: Cmp1); |
| 3487 | Pred = CmpInst::getSwappedPredicate(pred: Pred); |
| 3488 | } |
| 3489 | |
| 3490 | FastMathFlags FAbsFMF = FAbsI->getFastMathFlags(); |
| 3491 | FastMathFlags CmpFMF = OuterCmp->getFastMathFlags(); |
| 3492 | |
| 3493 | FastMathFlags CommonRewriteFMF = |
| 3494 | FastMathFlags::intersectRewrite(LHS: FAbsFMF, RHS: CmpFMF); |
| 3495 | |
| 3496 | // unionValue with FastMathFlags() drops all rewriter based flags |
| 3497 | FastMathFlags NewFAbsFMF = |
| 3498 | CommonRewriteFMF | FastMathFlags::unionValue(LHS: FAbsFMF, RHS: FastMathFlags()); |
| 3499 | FastMathFlags NewCmpFMF = |
| 3500 | CommonRewriteFMF | FastMathFlags::unionValue(LHS: CmpFMF, RHS: FastMathFlags()); |
| 3501 | |
| 3502 | // When X is NaN, the old code evaluated fabs(Y), while the new code evaluates |
| 3503 | // fabs(X). Do not preserve nnan on either newly-created instruction. |
| 3504 | NewFAbsFMF.setNoNaNs(false); |
| 3505 | NewCmpFMF.setNoNaNs(false); |
| 3506 | |
| 3507 | Value *NewAbs = IC.Builder.CreateFAbs(V: X, FMFSource: FMFSource(NewFAbsFMF)); |
| 3508 | Value *NewCmp = |
| 3509 | IC.Builder.CreateFCmpFMF(P: Pred, LHS: NewAbs, RHS: Cmp1, FMFSource: FMFSource(NewCmpFMF)); |
| 3510 | Value *NewSel = IC.Builder.CreateSelectFMF(C: NewCmp, True: X, False: Y, FMFSource: &SI); |
| 3511 | return IC.replaceInstUsesWith(I&: SI, V: NewSel); |
| 3512 | } |
| 3513 | |
| 3514 | // Match the following IR pattern: |
| 3515 | // %x.lowbits = and i8 %x, %lowbitmask |
| 3516 | // %x.lowbits.are.zero = icmp eq i8 %x.lowbits, 0 |
| 3517 | // %x.biased = add i8 %x, %bias |
| 3518 | // %x.biased.highbits = and i8 %x.biased, %highbitmask |
| 3519 | // %x.roundedup = select i1 %x.lowbits.are.zero, i8 %x, i8 %x.biased.highbits |
| 3520 | // Define: |
| 3521 | // %alignment = add i8 %lowbitmask, 1 |
| 3522 | // Iff 1. an %alignment is a power-of-two (aka, %lowbitmask is a low bit mask) |
| 3523 | // and 2. %bias is equal to either %lowbitmask or %alignment, |
| 3524 | // and 3. %highbitmask is equal to ~%lowbitmask (aka, to -%alignment) |
| 3525 | // then this pattern can be transformed into: |
| 3526 | // %x.offset = add i8 %x, %lowbitmask |
| 3527 | // %x.roundedup = and i8 %x.offset, %highbitmask |
| 3528 | static Value * |
| 3529 | foldRoundUpIntegerWithPow2Alignment(SelectInst &SI, |
| 3530 | InstCombiner::BuilderTy &Builder) { |
| 3531 | Value *Cond = SI.getCondition(); |
| 3532 | Value *X = SI.getTrueValue(); |
| 3533 | Value *XBiasedHighBits = SI.getFalseValue(); |
| 3534 | |
| 3535 | CmpPredicate Pred; |
| 3536 | Value *XLowBits; |
| 3537 | if (!match(V: Cond, P: m_ICmp(Pred, L: m_Value(V&: XLowBits), R: m_ZeroInt())) || |
| 3538 | !ICmpInst::isEquality(P: Pred)) |
| 3539 | return nullptr; |
| 3540 | |
| 3541 | if (Pred == ICmpInst::Predicate::ICMP_NE) |
| 3542 | std::swap(a&: X, b&: XBiasedHighBits); |
| 3543 | |
| 3544 | // FIXME: we could support non non-splats here. |
| 3545 | |
| 3546 | const APInt *LowBitMaskCst; |
| 3547 | if (!match(V: XLowBits, P: m_And(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: LowBitMaskCst)))) |
| 3548 | return nullptr; |
| 3549 | |
| 3550 | // Match even if the AND and ADD are swapped. |
| 3551 | const APInt *BiasCst, *HighBitMaskCst; |
| 3552 | if (!match(V: XBiasedHighBits, |
| 3553 | P: m_And(L: m_Add(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: BiasCst)), |
| 3554 | R: m_APIntAllowPoison(Res&: HighBitMaskCst))) && |
| 3555 | !match(V: XBiasedHighBits, |
| 3556 | P: m_Add(L: m_And(L: m_Specific(V: X), R: m_APIntAllowPoison(Res&: HighBitMaskCst)), |
| 3557 | R: m_APIntAllowPoison(Res&: BiasCst)))) |
| 3558 | return nullptr; |
| 3559 | |
| 3560 | if (!LowBitMaskCst->isMask()) |
| 3561 | return nullptr; |
| 3562 | |
| 3563 | APInt InvertedLowBitMaskCst = ~*LowBitMaskCst; |
| 3564 | if (InvertedLowBitMaskCst != *HighBitMaskCst) |
| 3565 | return nullptr; |
| 3566 | |
| 3567 | APInt AlignmentCst = *LowBitMaskCst + 1; |
| 3568 | |
| 3569 | if (*BiasCst != AlignmentCst && *BiasCst != *LowBitMaskCst) |
| 3570 | return nullptr; |
| 3571 | |
| 3572 | if (!XBiasedHighBits->hasOneUse()) { |
| 3573 | // We can't directly return XBiasedHighBits if it is more poisonous. |
| 3574 | if (*BiasCst == *LowBitMaskCst && impliesPoison(ValAssumedPoison: XBiasedHighBits, V: X)) |
| 3575 | return XBiasedHighBits; |
| 3576 | return nullptr; |
| 3577 | } |
| 3578 | |
| 3579 | // FIXME: could we preserve undef's here? |
| 3580 | Type *Ty = X->getType(); |
| 3581 | Value *XOffset = Builder.CreateAdd(LHS: X, RHS: ConstantInt::get(Ty, V: *LowBitMaskCst), |
| 3582 | Name: X->getName() + ".biased" ); |
| 3583 | Value *R = Builder.CreateAnd(LHS: XOffset, RHS: ConstantInt::get(Ty, V: *HighBitMaskCst)); |
| 3584 | R->takeName(V: &SI); |
| 3585 | return R; |
| 3586 | } |
| 3587 | |
| 3588 | namespace { |
| 3589 | struct DecomposedSelect { |
| 3590 | Value *Cond = nullptr; |
| 3591 | Value *TrueVal = nullptr; |
| 3592 | Value *FalseVal = nullptr; |
| 3593 | }; |
| 3594 | } // namespace |
| 3595 | |
| 3596 | /// Folds patterns like: |
| 3597 | /// select c2 (select c1 a b) (select c1 b a) |
| 3598 | /// into: |
| 3599 | /// select (xor c1 c2) b a |
| 3600 | static Instruction * |
| 3601 | foldSelectOfSymmetricSelect(SelectInst &OuterSelVal, |
| 3602 | InstCombiner::BuilderTy &Builder) { |
| 3603 | |
| 3604 | Value *OuterCond, *InnerCond, *InnerTrueVal, *InnerFalseVal; |
| 3605 | if (!match( |
| 3606 | V: &OuterSelVal, |
| 3607 | P: m_Select(C: m_Value(V&: OuterCond), |
| 3608 | L: m_OneUse(SubPattern: m_Select(C: m_Value(V&: InnerCond), L: m_Value(V&: InnerTrueVal), |
| 3609 | R: m_Value(V&: InnerFalseVal))), |
| 3610 | R: m_OneUse(SubPattern: m_Select(C: m_Deferred(V: InnerCond), |
| 3611 | L: m_Deferred(V: InnerFalseVal), |
| 3612 | R: m_Deferred(V: InnerTrueVal)))))) |
| 3613 | return nullptr; |
| 3614 | |
| 3615 | if (OuterCond->getType() != InnerCond->getType()) |
| 3616 | return nullptr; |
| 3617 | |
| 3618 | Value *Xor = Builder.CreateXor(LHS: InnerCond, RHS: OuterCond); |
| 3619 | return SelectInst::Create(C: Xor, S1: InnerFalseVal, S2: InnerTrueVal); |
| 3620 | } |
| 3621 | |
| 3622 | /// Look for patterns like |
| 3623 | /// %outer.cond = select i1 %inner.cond, i1 %alt.cond, i1 false |
| 3624 | /// %inner.sel = select i1 %inner.cond, i8 %inner.sel.t, i8 %inner.sel.f |
| 3625 | /// %outer.sel = select i1 %outer.cond, i8 %outer.sel.t, i8 %inner.sel |
| 3626 | /// and rewrite it as |
| 3627 | /// %inner.sel = select i1 %cond.alternative, i8 %sel.outer.t, i8 %sel.inner.t |
| 3628 | /// %sel.outer = select i1 %cond.inner, i8 %inner.sel, i8 %sel.inner.f |
| 3629 | static Instruction *foldNestedSelects(SelectInst &OuterSelVal, |
| 3630 | InstCombiner::BuilderTy &Builder) { |
| 3631 | // We must start with a `select`. |
| 3632 | DecomposedSelect OuterSel; |
| 3633 | match(V: &OuterSelVal, |
| 3634 | P: m_Select(C: m_Value(V&: OuterSel.Cond), L: m_Value(V&: OuterSel.TrueVal), |
| 3635 | R: m_Value(V&: OuterSel.FalseVal))); |
| 3636 | |
| 3637 | // Canonicalize inversion of the outermost `select`'s condition. |
| 3638 | if (match(V: OuterSel.Cond, P: m_Not(V: m_Value(V&: OuterSel.Cond)))) |
| 3639 | std::swap(a&: OuterSel.TrueVal, b&: OuterSel.FalseVal); |
| 3640 | |
| 3641 | // The condition of the outermost select must be an `and`/`or`. |
| 3642 | if (!match(V: OuterSel.Cond, P: m_c_LogicalOp(L: m_Value(), R: m_Value()))) |
| 3643 | return nullptr; |
| 3644 | |
| 3645 | // Depending on the logical op, inner select might be in different hand. |
| 3646 | bool IsAndVariant = match(V: OuterSel.Cond, P: m_LogicalAnd()); |
| 3647 | Value *InnerSelVal = IsAndVariant ? OuterSel.FalseVal : OuterSel.TrueVal; |
| 3648 | |
| 3649 | // Profitability check - avoid increasing instruction count. |
| 3650 | if (none_of(Range: ArrayRef<Value *>({OuterSelVal.getCondition(), InnerSelVal}), |
| 3651 | P: match_fn(P: m_OneUse(SubPattern: m_Value())))) |
| 3652 | return nullptr; |
| 3653 | |
| 3654 | // The appropriate hand of the outermost `select` must be a select itself. |
| 3655 | DecomposedSelect InnerSel; |
| 3656 | if (!match(V: InnerSelVal, |
| 3657 | P: m_Select(C: m_Value(V&: InnerSel.Cond), L: m_Value(V&: InnerSel.TrueVal), |
| 3658 | R: m_Value(V&: InnerSel.FalseVal)))) |
| 3659 | return nullptr; |
| 3660 | |
| 3661 | // Canonicalize inversion of the innermost `select`'s condition. |
| 3662 | if (match(V: InnerSel.Cond, P: m_Not(V: m_Value(V&: InnerSel.Cond)))) |
| 3663 | std::swap(a&: InnerSel.TrueVal, b&: InnerSel.FalseVal); |
| 3664 | |
| 3665 | Value *AltCond = nullptr; |
| 3666 | auto matchOuterCond = [OuterSel, IsAndVariant, &AltCond](auto m_InnerCond) { |
| 3667 | // An unsimplified select condition can match both LogicalAnd and LogicalOr |
| 3668 | // (select true, true, false). Since below we assume that LogicalAnd implies |
| 3669 | // InnerSel match the FVal and vice versa for LogicalOr, we can't match the |
| 3670 | // alternative pattern here. |
| 3671 | return IsAndVariant ? match(OuterSel.Cond, |
| 3672 | m_c_LogicalAnd(m_InnerCond, m_Value(V&: AltCond))) |
| 3673 | : match(OuterSel.Cond, |
| 3674 | m_c_LogicalOr(m_InnerCond, m_Value(V&: AltCond))); |
| 3675 | }; |
| 3676 | |
| 3677 | // Finally, match the condition that was driving the outermost `select`, |
| 3678 | // it should be a logical operation between the condition that was driving |
| 3679 | // the innermost `select` (after accounting for the possible inversions |
| 3680 | // of the condition), and some other condition. |
| 3681 | if (matchOuterCond(m_Specific(V: InnerSel.Cond))) { |
| 3682 | // Done! |
| 3683 | } else if (Value * NotInnerCond; matchOuterCond(m_CombineAnd( |
| 3684 | Ps: m_Not(V: m_Specific(V: InnerSel.Cond)), Ps: m_Value(V&: NotInnerCond)))) { |
| 3685 | // Done! |
| 3686 | std::swap(a&: InnerSel.TrueVal, b&: InnerSel.FalseVal); |
| 3687 | InnerSel.Cond = NotInnerCond; |
| 3688 | } else // Not the pattern we were looking for. |
| 3689 | return nullptr; |
| 3690 | |
| 3691 | Value *SelInner = Builder.CreateSelect( |
| 3692 | C: AltCond, True: IsAndVariant ? OuterSel.TrueVal : InnerSel.FalseVal, |
| 3693 | False: IsAndVariant ? InnerSel.TrueVal : OuterSel.FalseVal); |
| 3694 | SelInner->takeName(V: InnerSelVal); |
| 3695 | return SelectInst::Create(C: InnerSel.Cond, |
| 3696 | S1: IsAndVariant ? SelInner : InnerSel.TrueVal, |
| 3697 | S2: !IsAndVariant ? SelInner : InnerSel.FalseVal); |
| 3698 | } |
| 3699 | |
| 3700 | /// Return true if V is poison or \p Expected given that ValAssumedPoison is |
| 3701 | /// already poison. For example, if ValAssumedPoison is `icmp samesign X, 10` |
| 3702 | /// and V is `icmp ne X, 5`, impliesPoisonOrCond returns true. |
| 3703 | static bool impliesPoisonOrCond(const Value *ValAssumedPoison, const Value *V, |
| 3704 | bool Expected, const SimplifyQuery &SQ) { |
| 3705 | if (impliesPoison(ValAssumedPoison, V)) |
| 3706 | return true; |
| 3707 | |
| 3708 | // Handle the case that ValAssumedPoison is `icmp samesign pred X, C1` and V |
| 3709 | // is `icmp pred X, C2`, where C1 is well-defined. |
| 3710 | if (auto *ICmp = dyn_cast<ICmpInst>(Val: ValAssumedPoison)) { |
| 3711 | Value *LHS = ICmp->getOperand(i_nocapture: 0); |
| 3712 | const APInt *RHSC1; |
| 3713 | const APInt *RHSC2; |
| 3714 | CmpPredicate Pred; |
| 3715 | if (ICmp->hasSameSign() && |
| 3716 | match(V: ICmp->getOperand(i_nocapture: 1), P: m_APIntForbidPoison(Res&: RHSC1)) && |
| 3717 | match(V, P: m_ICmp(Pred, L: m_Specific(V: LHS), R: m_APIntAllowPoison(Res&: RHSC2)))) { |
| 3718 | unsigned BitWidth = RHSC1->getBitWidth(); |
| 3719 | ConstantRange CRX = |
| 3720 | RHSC1->isNonNegative() |
| 3721 | ? ConstantRange(APInt::getSignedMinValue(numBits: BitWidth), |
| 3722 | APInt::getZero(numBits: BitWidth)) |
| 3723 | : ConstantRange(APInt::getZero(numBits: BitWidth), |
| 3724 | APInt::getSignedMinValue(numBits: BitWidth)); |
| 3725 | return CRX.icmp(Pred: Expected ? Pred : ICmpInst::getInverseCmpPredicate(Pred), |
| 3726 | Other: *RHSC2); |
| 3727 | } |
| 3728 | } |
| 3729 | Value *A; |
| 3730 | if (match(V: ValAssumedPoison, P: m_NUWTrunc(Op: m_Value(V&: A))) && |
| 3731 | isGuaranteedNotToBePoison(V: A)) { |
| 3732 | assert(ValAssumedPoison->getType()->isIntOrIntVectorTy(1)); |
| 3733 | return computeKnownBits( |
| 3734 | V: A, Q: SQ.getWithInstruction(I: cast<Instruction>(Val: ValAssumedPoison))) |
| 3735 | .getMaxValue() == 1; |
| 3736 | } |
| 3737 | |
| 3738 | return false; |
| 3739 | } |
| 3740 | |
| 3741 | Instruction *InstCombinerImpl::foldSelectOfBools(SelectInst &SI) { |
| 3742 | Value *CondVal = SI.getCondition(); |
| 3743 | Value *TrueVal = SI.getTrueValue(); |
| 3744 | Value *FalseVal = SI.getFalseValue(); |
| 3745 | Type *SelType = SI.getType(); |
| 3746 | |
| 3747 | // Avoid potential infinite loops by checking for non-constant condition. |
| 3748 | // TODO: Can we assert instead by improving canonicalizeSelectToShuffle()? |
| 3749 | // Scalar select must have simplified? |
| 3750 | if (!SelType->isIntOrIntVectorTy(BitWidth: 1) || isa<Constant>(Val: CondVal) || |
| 3751 | TrueVal->getType() != CondVal->getType()) |
| 3752 | return nullptr; |
| 3753 | |
| 3754 | auto *One = ConstantInt::getTrue(Ty: SelType); |
| 3755 | auto *Zero = ConstantInt::getFalse(Ty: SelType); |
| 3756 | Value *A, *B, *C, *D; |
| 3757 | |
| 3758 | // Folding select to and/or i1 isn't poison safe in general. impliesPoison |
| 3759 | // checks whether folding it does not convert a well-defined value into |
| 3760 | // poison. |
| 3761 | if (match(V: TrueVal, P: m_One())) { |
| 3762 | if (impliesPoisonOrCond(ValAssumedPoison: FalseVal, V: CondVal, /*Expected=*/false, SQ)) { |
| 3763 | // Change: A = select B, true, C --> A = or B, C |
| 3764 | return BinaryOperator::CreateOr(V1: CondVal, V2: FalseVal); |
| 3765 | } |
| 3766 | |
| 3767 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: A), L: m_One(), R: m_Value(V&: B)))) && |
| 3768 | impliesPoisonOrCond(ValAssumedPoison: FalseVal, V: B, /*Expected=*/false, SQ)) { |
| 3769 | // (A || B) || C --> A || (B | C) |
| 3770 | Value *LOr = Builder.CreateLogicalOr(Cond1: A, Cond2: Builder.CreateOr(LHS: B, RHS: FalseVal)); |
| 3771 | if (auto *I = dyn_cast<Instruction>(Val: LOr)) { |
| 3772 | setExplicitlyUnknownBranchWeightsIfProfiled(I&: *I, DEBUG_TYPE); |
| 3773 | } |
| 3774 | return replaceInstUsesWith(I&: SI, V: LOr); |
| 3775 | } |
| 3776 | |
| 3777 | // (A && B) || (C && B) --> (A || C) && B |
| 3778 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3779 | match(V: FalseVal, P: m_LogicalAnd(L: m_Value(V&: C), R: m_Value(V&: D))) && |
| 3780 | (CondVal->hasOneUse() || FalseVal->hasOneUse())) { |
| 3781 | bool CondLogicAnd = isa<SelectInst>(Val: CondVal); |
| 3782 | bool FalseLogicAnd = isa<SelectInst>(Val: FalseVal); |
| 3783 | auto AndFactorization = [&](Value *Common, Value *InnerCond, |
| 3784 | Value *InnerVal, |
| 3785 | bool SelFirst = false) -> Instruction * { |
| 3786 | Value *InnerSel = Builder.CreateSelectWithUnknownProfile( |
| 3787 | C: InnerCond, True: One, False: InnerVal, DEBUG_TYPE); |
| 3788 | if (SelFirst) |
| 3789 | std::swap(a&: Common, b&: InnerSel); |
| 3790 | if (FalseLogicAnd || (CondLogicAnd && Common == A)) |
| 3791 | return createSelectInstWithUnknownProfile(C: Common, S1: InnerSel, S2: Zero); |
| 3792 | else |
| 3793 | return BinaryOperator::CreateAnd(V1: Common, V2: InnerSel); |
| 3794 | }; |
| 3795 | |
| 3796 | if (A == C) |
| 3797 | return AndFactorization(A, B, D); |
| 3798 | if (A == D) |
| 3799 | return AndFactorization(A, B, C); |
| 3800 | if (B == C) |
| 3801 | return AndFactorization(B, A, D); |
| 3802 | if (B == D) |
| 3803 | return AndFactorization(B, A, C, CondLogicAnd && FalseLogicAnd); |
| 3804 | } |
| 3805 | } |
| 3806 | |
| 3807 | if (match(V: FalseVal, P: m_Zero())) { |
| 3808 | if (impliesPoisonOrCond(ValAssumedPoison: TrueVal, V: CondVal, /*Expected=*/true, SQ)) { |
| 3809 | // Change: A = select B, C, false --> A = and B, C |
| 3810 | return BinaryOperator::CreateAnd(V1: CondVal, V2: TrueVal); |
| 3811 | } |
| 3812 | |
| 3813 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_Select(C: m_Value(V&: A), L: m_Value(V&: B), R: m_Zero()))) && |
| 3814 | impliesPoisonOrCond(ValAssumedPoison: TrueVal, V: B, /*Expected=*/true, SQ)) { |
| 3815 | // (A && B) && C --> A && (B & C) |
| 3816 | Value *LAnd = Builder.CreateLogicalAnd(Cond1: A, Cond2: Builder.CreateAnd(LHS: B, RHS: TrueVal)); |
| 3817 | if (auto *I = dyn_cast<Instruction>(Val: LAnd)) { |
| 3818 | setExplicitlyUnknownBranchWeightsIfProfiled(I&: *I, DEBUG_TYPE); |
| 3819 | } |
| 3820 | return replaceInstUsesWith(I&: SI, V: LAnd); |
| 3821 | } |
| 3822 | |
| 3823 | // (A || B) && (C || B) --> (A && C) || B |
| 3824 | if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3825 | match(V: TrueVal, P: m_LogicalOr(L: m_Value(V&: C), R: m_Value(V&: D))) && |
| 3826 | (CondVal->hasOneUse() || TrueVal->hasOneUse())) { |
| 3827 | bool CondLogicOr = isa<SelectInst>(Val: CondVal); |
| 3828 | bool TrueLogicOr = isa<SelectInst>(Val: TrueVal); |
| 3829 | auto OrFactorization = [&](Value *Common, Value *InnerCond, |
| 3830 | Value *InnerVal, |
| 3831 | bool SelFirst = false) -> Instruction * { |
| 3832 | Value *InnerSel = Builder.CreateSelectWithUnknownProfile( |
| 3833 | C: InnerCond, True: InnerVal, False: Zero, DEBUG_TYPE); |
| 3834 | if (SelFirst) |
| 3835 | std::swap(a&: Common, b&: InnerSel); |
| 3836 | if (TrueLogicOr || (CondLogicOr && Common == A)) |
| 3837 | return createSelectInstWithUnknownProfile(C: Common, S1: One, S2: InnerSel); |
| 3838 | else |
| 3839 | return BinaryOperator::CreateOr(V1: Common, V2: InnerSel); |
| 3840 | }; |
| 3841 | |
| 3842 | if (A == C) |
| 3843 | return OrFactorization(A, B, D); |
| 3844 | if (A == D) |
| 3845 | return OrFactorization(A, B, C); |
| 3846 | if (B == C) |
| 3847 | return OrFactorization(B, A, D); |
| 3848 | if (B == D) |
| 3849 | return OrFactorization(B, A, C, CondLogicOr && TrueLogicOr); |
| 3850 | } |
| 3851 | } |
| 3852 | |
| 3853 | // We match the "full" 0 or 1 constant here to avoid a potential infinite |
| 3854 | // loop with vectors that may have undefined/poison elements. |
| 3855 | // select a, false, b -> select !a, b, false |
| 3856 | if (match(V: TrueVal, P: m_Specific(V: Zero))) { |
| 3857 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
| 3858 | Instruction *MDFrom = ProfcheckDisableMetadataFixes ? nullptr : &SI; |
| 3859 | SelectInst *NewSI = |
| 3860 | SelectInst::Create(C: NotCond, S1: FalseVal, S2: Zero, NameStr: "" , InsertBefore: nullptr, MDFrom); |
| 3861 | NewSI->swapProfMetadata(); |
| 3862 | return NewSI; |
| 3863 | } |
| 3864 | // select a, b, true -> select !a, true, b |
| 3865 | if (match(V: FalseVal, P: m_Specific(V: One))) { |
| 3866 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
| 3867 | Instruction *MDFrom = ProfcheckDisableMetadataFixes ? nullptr : &SI; |
| 3868 | SelectInst *NewSI = |
| 3869 | SelectInst::Create(C: NotCond, S1: One, S2: TrueVal, NameStr: "" , InsertBefore: nullptr, MDFrom); |
| 3870 | NewSI->swapProfMetadata(); |
| 3871 | return NewSI; |
| 3872 | } |
| 3873 | |
| 3874 | // DeMorgan in select form: !a && !b --> !(a || b) |
| 3875 | // select !a, !b, false --> not (select a, true, b) |
| 3876 | if (match(V: &SI, P: m_LogicalAnd(L: m_Not(V: m_Value(V&: A)), R: m_Not(V: m_Value(V&: B)))) && |
| 3877 | (CondVal->hasOneUse() || TrueVal->hasOneUse()) && |
| 3878 | !match(V: A, P: m_ConstantExpr()) && !match(V: B, P: m_ConstantExpr())) { |
| 3879 | Instruction *MDFrom = ProfcheckDisableMetadataFixes ? nullptr : &SI; |
| 3880 | SelectInst *NewSI = |
| 3881 | cast<SelectInst>(Val: Builder.CreateSelect(C: A, True: One, False: B, Name: "" , MDFrom)); |
| 3882 | NewSI->swapProfMetadata(); |
| 3883 | return BinaryOperator::CreateNot(Op: NewSI); |
| 3884 | } |
| 3885 | |
| 3886 | // DeMorgan in select form: !a || !b --> !(a && b) |
| 3887 | // select !a, true, !b --> not (select a, b, false) |
| 3888 | if (match(V: &SI, P: m_LogicalOr(L: m_Not(V: m_Value(V&: A)), R: m_Not(V: m_Value(V&: B)))) && |
| 3889 | (CondVal->hasOneUse() || FalseVal->hasOneUse()) && |
| 3890 | !match(V: A, P: m_ConstantExpr()) && !match(V: B, P: m_ConstantExpr())) { |
| 3891 | Instruction *MDFrom = ProfcheckDisableMetadataFixes ? nullptr : &SI; |
| 3892 | SelectInst *NewSI = |
| 3893 | cast<SelectInst>(Val: Builder.CreateSelect(C: A, True: B, False: Zero, Name: "" , MDFrom)); |
| 3894 | NewSI->swapProfMetadata(); |
| 3895 | return BinaryOperator::CreateNot(Op: NewSI); |
| 3896 | } |
| 3897 | |
| 3898 | // select (select a, true, b), true, b -> select a, true, b |
| 3899 | if (match(V: CondVal, P: m_Select(C: m_Value(V&: A), L: m_One(), R: m_Value(V&: B))) && |
| 3900 | match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_Specific(V: B))) |
| 3901 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
| 3902 | // select (select a, b, false), b, false -> select a, b, false |
| 3903 | if (match(V: CondVal, P: m_Select(C: m_Value(V&: A), L: m_Value(V&: B), R: m_Zero())) && |
| 3904 | match(V: TrueVal, P: m_Specific(V: B)) && match(V: FalseVal, P: m_Zero())) |
| 3905 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
| 3906 | |
| 3907 | // ~(A & B) & (A | B) --> A ^ B |
| 3908 | if (match(V: &SI, P: m_c_LogicalAnd(L: m_Not(V: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))), |
| 3909 | R: m_c_LogicalOr(L: m_Deferred(V: A), R: m_Deferred(V: B))))) |
| 3910 | return BinaryOperator::CreateXor(V1: A, V2: B); |
| 3911 | |
| 3912 | // select (~a | c), a, b -> select a, (select c, true, b), false |
| 3913 | if (match(V: CondVal, |
| 3914 | P: m_OneUse(SubPattern: m_c_Or(L: m_Not(V: m_Specific(V: TrueVal)), R: m_Value(V&: C))))) { |
| 3915 | // TODO(#183864): We could improve the profile if P(~a | c) < 0.5, which |
| 3916 | // implies strong bounds on both operands (P(a) is high, P(c) is low). |
| 3917 | Value *OrV = |
| 3918 | Builder.CreateSelectWithUnknownProfile(C, True: One, False: FalseVal, DEBUG_TYPE); |
| 3919 | return createSelectInstWithUnknownProfile(C: TrueVal, S1: OrV, S2: Zero); |
| 3920 | } |
| 3921 | // select (c & b), a, b -> select b, (select ~c, true, a), false |
| 3922 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: C), R: m_Specific(V: FalseVal))))) { |
| 3923 | if (Value *NotC = getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder)) { |
| 3924 | Value *OrV = Builder.CreateSelectWithUnknownProfile(C: NotC, True: One, False: TrueVal, |
| 3925 | DEBUG_TYPE); |
| 3926 | return createSelectInstWithUnknownProfile(C: FalseVal, S1: OrV, S2: Zero); |
| 3927 | } |
| 3928 | } |
| 3929 | // select (a | c), a, b -> select a, true, (select ~c, b, false) |
| 3930 | if (match(V: CondVal, P: m_OneUse(SubPattern: m_c_Or(L: m_Specific(V: TrueVal), R: m_Value(V&: C))))) { |
| 3931 | if (Value *NotC = getFreelyInverted(V: C, WillInvertAllUses: C->hasOneUse(), Builder: &Builder)) { |
| 3932 | // TODO(#183864): We could improve the profile if P(a | c) < 0.5, which |
| 3933 | // implies strong bounds on both operands (both P(a) and P(c) are low). |
| 3934 | Value *AndV = Builder.CreateSelectWithUnknownProfile(C: NotC, True: FalseVal, False: Zero, |
| 3935 | DEBUG_TYPE); |
| 3936 | return createSelectInstWithUnknownProfile(C: TrueVal, S1: One, S2: AndV); |
| 3937 | } |
| 3938 | } |
| 3939 | // select (c & ~b), a, b -> select b, true, (select c, a, false) |
| 3940 | if (match(V: CondVal, |
| 3941 | P: m_OneUse(SubPattern: m_c_And(L: m_Value(V&: C), R: m_Not(V: m_Specific(V: FalseVal)))))) { |
| 3942 | Value *AndV = |
| 3943 | Builder.CreateSelectWithUnknownProfile(C, True: TrueVal, False: Zero, DEBUG_TYPE); |
| 3944 | return createSelectInstWithUnknownProfile(C: FalseVal, S1: One, S2: AndV); |
| 3945 | } |
| 3946 | |
| 3947 | if (match(V: FalseVal, P: m_Zero()) || match(V: TrueVal, P: m_One())) { |
| 3948 | Use *Y = nullptr; |
| 3949 | bool IsAnd = match(V: FalseVal, P: m_Zero()) ? true : false; |
| 3950 | Value *Op1 = IsAnd ? TrueVal : FalseVal; |
| 3951 | if (isCheckForZeroAndMulWithOverflow(Op0: CondVal, Op1, IsAnd, Y)) { |
| 3952 | auto *FI = new FreezeInst(*Y, (*Y)->getName() + ".fr" ); |
| 3953 | InsertNewInstBefore(New: FI, Old: cast<Instruction>(Val: Y->getUser())->getIterator()); |
| 3954 | replaceUse(U&: *Y, NewValue: FI); |
| 3955 | return replaceInstUsesWith(I&: SI, V: Op1); |
| 3956 | } |
| 3957 | |
| 3958 | if (auto *V = foldBooleanAndOr(LHS: CondVal, RHS: Op1, I&: SI, IsAnd, |
| 3959 | /*IsLogical=*/true)) |
| 3960 | return replaceInstUsesWith(I&: SI, V); |
| 3961 | } |
| 3962 | |
| 3963 | // select (a || b), c, false -> select a, c, false |
| 3964 | // select c, (a || b), false -> select c, a, false |
| 3965 | // if c implies that b is false. |
| 3966 | if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3967 | match(V: FalseVal, P: m_Zero())) { |
| 3968 | std::optional<bool> Res = isImpliedCondition(LHS: TrueVal, RHS: B, DL); |
| 3969 | if (Res && *Res == false) |
| 3970 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
| 3971 | } |
| 3972 | if (match(V: TrueVal, P: m_LogicalOr(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3973 | match(V: FalseVal, P: m_Zero())) { |
| 3974 | std::optional<bool> Res = isImpliedCondition(LHS: CondVal, RHS: B, DL); |
| 3975 | if (Res && *Res == false) |
| 3976 | return replaceOperand(I&: SI, OpNum: 1, V: A); |
| 3977 | } |
| 3978 | // select c, true, (a && b) -> select c, true, a |
| 3979 | // select (a && b), true, c -> select a, true, c |
| 3980 | // if c = false implies that b = true |
| 3981 | if (match(V: TrueVal, P: m_One()) && |
| 3982 | match(V: FalseVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B)))) { |
| 3983 | std::optional<bool> Res = isImpliedCondition(LHS: CondVal, RHS: B, DL, LHSIsTrue: false); |
| 3984 | if (Res && *Res == true) |
| 3985 | return replaceOperand(I&: SI, OpNum: 2, V: A); |
| 3986 | } |
| 3987 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: A), R: m_Value(V&: B))) && |
| 3988 | match(V: TrueVal, P: m_One())) { |
| 3989 | std::optional<bool> Res = isImpliedCondition(LHS: FalseVal, RHS: B, DL, LHSIsTrue: false); |
| 3990 | if (Res && *Res == true) |
| 3991 | return replaceOperand(I&: SI, OpNum: 0, V: A); |
| 3992 | } |
| 3993 | |
| 3994 | if (match(V: TrueVal, P: m_One())) { |
| 3995 | // (C && A) || (!C && B) --> select C, A, B (and similar cases) |
| 3996 | if (auto *V = FoldOrOfLogicalAnds(Op0: CondVal, Op1: FalseVal)) { |
| 3997 | return V; |
| 3998 | } |
| 3999 | } |
| 4000 | |
| 4001 | return nullptr; |
| 4002 | } |
| 4003 | |
| 4004 | // Return true if we can safely remove the select instruction for std::bit_ceil |
| 4005 | // pattern. |
| 4006 | static bool isSafeToRemoveBitCeilSelect(ICmpInst::Predicate Pred, Value *Cond0, |
| 4007 | const APInt *Cond1, Value *CtlzOp, |
| 4008 | unsigned BitWidth, |
| 4009 | bool &ShouldDropNoWrap) { |
| 4010 | // The challenge in recognizing std::bit_ceil(X) is that the operand is used |
| 4011 | // for the CTLZ proper and select condition, each possibly with some |
| 4012 | // operation like add and sub. |
| 4013 | // |
| 4014 | // Our aim is to make sure that -ctlz & (BitWidth - 1) == 0 even when the |
| 4015 | // select instruction would select 1, which allows us to get rid of the select |
| 4016 | // instruction. |
| 4017 | // |
| 4018 | // To see if we can do so, we do some symbolic execution with ConstantRange. |
| 4019 | // Specifically, we compute the range of values that Cond0 could take when |
| 4020 | // Cond == false. Then we successively transform the range until we obtain |
| 4021 | // the range of values that CtlzOp could take. |
| 4022 | // |
| 4023 | // Conceptually, we follow the def-use chain backward from Cond0 while |
| 4024 | // transforming the range for Cond0 until we meet the common ancestor of Cond0 |
| 4025 | // and CtlzOp. Then we follow the def-use chain forward until we obtain the |
| 4026 | // range for CtlzOp. That said, we only follow at most one ancestor from |
| 4027 | // Cond0. Likewise, we only follow at most one ancestor from CtrlOp. |
| 4028 | |
| 4029 | ConstantRange CR = ConstantRange::makeExactICmpRegion( |
| 4030 | Pred: CmpInst::getInversePredicate(pred: Pred), Other: *Cond1); |
| 4031 | |
| 4032 | ShouldDropNoWrap = false; |
| 4033 | |
| 4034 | // Match the operation that's used to compute CtlzOp from CommonAncestor. If |
| 4035 | // CtlzOp == CommonAncestor, return true as no operation is needed. If a |
| 4036 | // match is found, execute the operation on CR, update CR, and return true. |
| 4037 | // Otherwise, return false. |
| 4038 | auto MatchForward = [&](Value *CommonAncestor) { |
| 4039 | const APInt *C = nullptr; |
| 4040 | if (CtlzOp == CommonAncestor) |
| 4041 | return true; |
| 4042 | if (match(V: CtlzOp, P: m_Add(L: m_Specific(V: CommonAncestor), R: m_APInt(Res&: C)))) { |
| 4043 | ShouldDropNoWrap = true; |
| 4044 | CR = CR.add(Other: *C); |
| 4045 | return true; |
| 4046 | } |
| 4047 | if (match(V: CtlzOp, P: m_Sub(L: m_APInt(Res&: C), R: m_Specific(V: CommonAncestor)))) { |
| 4048 | ShouldDropNoWrap = true; |
| 4049 | CR = ConstantRange(*C).sub(Other: CR); |
| 4050 | return true; |
| 4051 | } |
| 4052 | if (match(V: CtlzOp, P: m_Not(V: m_Specific(V: CommonAncestor)))) { |
| 4053 | CR = CR.binaryNot(); |
| 4054 | return true; |
| 4055 | } |
| 4056 | return false; |
| 4057 | }; |
| 4058 | |
| 4059 | const APInt *C = nullptr; |
| 4060 | Value *CommonAncestor; |
| 4061 | if (MatchForward(Cond0)) { |
| 4062 | // Cond0 is either CtlzOp or CtlzOp's parent. CR has been updated. |
| 4063 | } else if (match(V: Cond0, P: m_Add(L: m_Value(V&: CommonAncestor), R: m_APInt(Res&: C)))) { |
| 4064 | CR = CR.sub(Other: *C); |
| 4065 | if (!MatchForward(CommonAncestor)) |
| 4066 | return false; |
| 4067 | // Cond0's parent is either CtlzOp or CtlzOp's parent. CR has been updated. |
| 4068 | } else { |
| 4069 | return false; |
| 4070 | } |
| 4071 | |
| 4072 | // Return true if all the values in the range are either 0 or negative (if |
| 4073 | // treated as signed). We do so by evaluating: |
| 4074 | // |
| 4075 | // CR - 1 u>= (1 << BitWidth) - 1. |
| 4076 | APInt IntMax = APInt::getSignMask(BitWidth) - 1; |
| 4077 | CR = CR.sub(Other: APInt(BitWidth, 1)); |
| 4078 | return CR.icmp(Pred: ICmpInst::ICMP_UGE, Other: IntMax); |
| 4079 | } |
| 4080 | |
| 4081 | // Transform the std::bit_ceil(X) pattern like: |
| 4082 | // |
| 4083 | // %dec = add i32 %x, -1 |
| 4084 | // %ctlz = tail call i32 @llvm.ctlz.i32(i32 %dec, i1 false) |
| 4085 | // %sub = sub i32 32, %ctlz |
| 4086 | // %shl = shl i32 1, %sub |
| 4087 | // %ugt = icmp ugt i32 %x, 1 |
| 4088 | // %sel = select i1 %ugt, i32 %shl, i32 1 |
| 4089 | // |
| 4090 | // into: |
| 4091 | // |
| 4092 | // %dec = add i32 %x, -1 |
| 4093 | // %ctlz = tail call i32 @llvm.ctlz.i32(i32 %dec, i1 false) |
| 4094 | // %neg = sub i32 0, %ctlz |
| 4095 | // %masked = and i32 %ctlz, 31 |
| 4096 | // %shl = shl i32 1, %sub |
| 4097 | // |
| 4098 | // Note that the select is optimized away while the shift count is masked with |
| 4099 | // 31. We handle some variations of the input operand like std::bit_ceil(X + |
| 4100 | // 1). |
| 4101 | static Instruction *foldBitCeil(SelectInst &SI, IRBuilderBase &Builder, |
| 4102 | InstCombinerImpl &IC) { |
| 4103 | Type *SelType = SI.getType(); |
| 4104 | unsigned BitWidth = SelType->getScalarSizeInBits(); |
| 4105 | if (!isPowerOf2_32(Value: BitWidth)) |
| 4106 | return nullptr; |
| 4107 | |
| 4108 | Value *FalseVal = SI.getFalseValue(); |
| 4109 | Value *TrueVal = SI.getTrueValue(); |
| 4110 | CmpPredicate Pred; |
| 4111 | const APInt *Cond1; |
| 4112 | Value *Cond0, *Ctlz, *CtlzOp; |
| 4113 | if (!match(V: SI.getCondition(), P: m_ICmp(Pred, L: m_Value(V&: Cond0), R: m_APInt(Res&: Cond1)))) |
| 4114 | return nullptr; |
| 4115 | |
| 4116 | if (match(V: TrueVal, P: m_One())) { |
| 4117 | std::swap(a&: FalseVal, b&: TrueVal); |
| 4118 | Pred = CmpInst::getInversePredicate(pred: Pred); |
| 4119 | } |
| 4120 | |
| 4121 | bool ShouldDropNoWrap; |
| 4122 | |
| 4123 | if (!match(V: FalseVal, P: m_One()) || |
| 4124 | !match(V: TrueVal, |
| 4125 | P: m_OneUse(SubPattern: m_Shl(L: m_One(), R: m_OneUse(SubPattern: m_Sub(L: m_SpecificInt(V: BitWidth), |
| 4126 | R: m_Value(V&: Ctlz)))))) || |
| 4127 | !match(V: Ctlz, P: m_Ctlz(Op0: m_Value(V&: CtlzOp), Op1: m_Value())) || |
| 4128 | !isSafeToRemoveBitCeilSelect(Pred, Cond0, Cond1, CtlzOp, BitWidth, |
| 4129 | ShouldDropNoWrap)) |
| 4130 | return nullptr; |
| 4131 | |
| 4132 | if (ShouldDropNoWrap) { |
| 4133 | cast<Instruction>(Val: CtlzOp)->setHasNoUnsignedWrap(false); |
| 4134 | cast<Instruction>(Val: CtlzOp)->setHasNoSignedWrap(false); |
| 4135 | } |
| 4136 | |
| 4137 | // Build 1 << (-CTLZ & (BitWidth-1)). The negation likely corresponds to a |
| 4138 | // single hardware instruction as opposed to BitWidth - CTLZ, where BitWidth |
| 4139 | // is an integer constant. Masking with BitWidth-1 comes free on some |
| 4140 | // hardware as part of the shift instruction. |
| 4141 | |
| 4142 | // Drop range attributes and re-infer them in the next iteration. |
| 4143 | cast<Instruction>(Val: Ctlz)->dropPoisonGeneratingAnnotations(); |
| 4144 | IC.addToWorklist(I: cast<Instruction>(Val: Ctlz)); |
| 4145 | Value *Neg = Builder.CreateNeg(V: Ctlz); |
| 4146 | Value *Masked = |
| 4147 | Builder.CreateAnd(LHS: Neg, RHS: ConstantInt::get(Ty: SelType, V: BitWidth - 1)); |
| 4148 | return BinaryOperator::Create(Op: Instruction::Shl, S1: ConstantInt::get(Ty: SelType, V: 1), |
| 4149 | S2: Masked); |
| 4150 | } |
| 4151 | |
| 4152 | // This function tries to fold the following operations: |
| 4153 | // (x < y) ? -1 : zext(x != y) |
| 4154 | // (x < y) ? -1 : zext(x > y) |
| 4155 | // (x > y) ? 1 : sext(x != y) |
| 4156 | // (x > y) ? 1 : sext(x < y) |
| 4157 | // (x == y) ? 0 : (x > y ? 1 : -1) |
| 4158 | // (x == y) ? 0 : (x < y ? -1 : 1) |
| 4159 | // Special case: x == C ? 0 : (x > C - 1 ? 1 : -1) |
| 4160 | // Special case: x == C ? 0 : (x < C + 1 ? -1 : 1) |
| 4161 | // Into ucmp/scmp(x, y), where signedness is determined by the signedness |
| 4162 | // of the comparison in the original sequence. |
| 4163 | Instruction *InstCombinerImpl::foldSelectToCmp(SelectInst &SI) { |
| 4164 | Value *TV = SI.getTrueValue(); |
| 4165 | Value *FV = SI.getFalseValue(); |
| 4166 | |
| 4167 | CmpPredicate Pred; |
| 4168 | Value *LHS, *RHS; |
| 4169 | if (!match(V: SI.getCondition(), P: m_ICmp(Pred, L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) |
| 4170 | return nullptr; |
| 4171 | |
| 4172 | if (!LHS->getType()->isIntOrIntVectorTy()) |
| 4173 | return nullptr; |
| 4174 | |
| 4175 | // If there is no -1, 0 or 1 at TV, then invert the select statement and try |
| 4176 | // to canonicalize to one of the forms above |
| 4177 | if (!isa<Constant>(Val: TV)) { |
| 4178 | if (!isa<Constant>(Val: FV)) |
| 4179 | return nullptr; |
| 4180 | Pred = ICmpInst::getInverseCmpPredicate(Pred); |
| 4181 | std::swap(a&: TV, b&: FV); |
| 4182 | } |
| 4183 | |
| 4184 | if (ICmpInst::isNonStrictPredicate(predicate: Pred)) { |
| 4185 | if (Constant *C = dyn_cast<Constant>(Val: RHS)) { |
| 4186 | auto FlippedPredAndConst = |
| 4187 | getFlippedStrictnessPredicateAndConstant(Pred, C); |
| 4188 | if (!FlippedPredAndConst) |
| 4189 | return nullptr; |
| 4190 | Pred = FlippedPredAndConst->first; |
| 4191 | RHS = FlippedPredAndConst->second; |
| 4192 | } else { |
| 4193 | return nullptr; |
| 4194 | } |
| 4195 | } |
| 4196 | |
| 4197 | // Try to swap operands and the predicate. We need to be careful when doing |
| 4198 | // so because two of the patterns have opposite predicates, so use the |
| 4199 | // constant inside select to determine if swapping operands would be |
| 4200 | // beneficial to us. |
| 4201 | if ((ICmpInst::isGT(P: Pred) && match(V: TV, P: m_AllOnes())) || |
| 4202 | (ICmpInst::isLT(P: Pred) && match(V: TV, P: m_One()))) { |
| 4203 | Pred = ICmpInst::getSwappedPredicate(pred: Pred); |
| 4204 | std::swap(a&: LHS, b&: RHS); |
| 4205 | } |
| 4206 | bool IsSigned = ICmpInst::isSigned(Pred); |
| 4207 | |
| 4208 | bool Replace = false; |
| 4209 | CmpPredicate ExtendedCmpPredicate; |
| 4210 | // (x < y) ? -1 : zext(x != y) |
| 4211 | // (x < y) ? -1 : zext(x > y) |
| 4212 | if (ICmpInst::isLT(P: Pred) && match(V: TV, P: m_AllOnes()) && |
| 4213 | match(V: FV, P: m_ZExt(Op: m_c_ICmp(Pred&: ExtendedCmpPredicate, L: m_Specific(V: LHS), |
| 4214 | R: m_Specific(V: RHS)))) && |
| 4215 | (ExtendedCmpPredicate == ICmpInst::ICMP_NE || |
| 4216 | ICmpInst::getSwappedPredicate(pred: ExtendedCmpPredicate) == Pred)) |
| 4217 | Replace = true; |
| 4218 | |
| 4219 | // (x > y) ? 1 : sext(x != y) |
| 4220 | // (x > y) ? 1 : sext(x < y) |
| 4221 | if (ICmpInst::isGT(P: Pred) && match(V: TV, P: m_One()) && |
| 4222 | match(V: FV, P: m_SExt(Op: m_c_ICmp(Pred&: ExtendedCmpPredicate, L: m_Specific(V: LHS), |
| 4223 | R: m_Specific(V: RHS)))) && |
| 4224 | (ExtendedCmpPredicate == ICmpInst::ICMP_NE || |
| 4225 | ICmpInst::getSwappedPredicate(pred: ExtendedCmpPredicate) == Pred)) |
| 4226 | Replace = true; |
| 4227 | |
| 4228 | // (x == y) ? 0 : (x > y ? 1 : -1) |
| 4229 | CmpPredicate FalseBranchSelectPredicate; |
| 4230 | const APInt *InnerTV, *InnerFV; |
| 4231 | if (Pred == ICmpInst::ICMP_EQ && match(V: TV, P: m_Zero()) && |
| 4232 | match(V: FV, P: m_Select(C: m_c_ICmp(Pred&: FalseBranchSelectPredicate, L: m_Specific(V: LHS), |
| 4233 | R: m_Specific(V: RHS)), |
| 4234 | L: m_APInt(Res&: InnerTV), R: m_APInt(Res&: InnerFV)))) { |
| 4235 | if (!ICmpInst::isGT(P: FalseBranchSelectPredicate)) { |
| 4236 | FalseBranchSelectPredicate = |
| 4237 | ICmpInst::getSwappedPredicate(pred: FalseBranchSelectPredicate); |
| 4238 | std::swap(a&: LHS, b&: RHS); |
| 4239 | } |
| 4240 | |
| 4241 | if (!InnerTV->isOne()) { |
| 4242 | std::swap(a&: InnerTV, b&: InnerFV); |
| 4243 | std::swap(a&: LHS, b&: RHS); |
| 4244 | } |
| 4245 | |
| 4246 | if (ICmpInst::isGT(P: FalseBranchSelectPredicate) && InnerTV->isOne() && |
| 4247 | InnerFV->isAllOnes()) { |
| 4248 | IsSigned = ICmpInst::isSigned(Pred: FalseBranchSelectPredicate); |
| 4249 | Replace = true; |
| 4250 | } |
| 4251 | } |
| 4252 | |
| 4253 | // Special cases with constants: x == C ? 0 : (x > C-1 ? 1 : -1) |
| 4254 | if (Pred == ICmpInst::ICMP_EQ && match(V: TV, P: m_Zero())) { |
| 4255 | const APInt *C; |
| 4256 | if (match(V: RHS, P: m_APInt(Res&: C))) { |
| 4257 | CmpPredicate InnerPred; |
| 4258 | Value *InnerRHS; |
| 4259 | const APInt *InnerTV, *InnerFV; |
| 4260 | if (match(V: FV, |
| 4261 | P: m_Select(C: m_ICmp(Pred&: InnerPred, L: m_Specific(V: LHS), R: m_Value(V&: InnerRHS)), |
| 4262 | L: m_APInt(Res&: InnerTV), R: m_APInt(Res&: InnerFV)))) { |
| 4263 | |
| 4264 | // x == C ? 0 : (x > C-1 ? 1 : -1) |
| 4265 | if (ICmpInst::isGT(P: InnerPred) && InnerTV->isOne() && |
| 4266 | InnerFV->isAllOnes()) { |
| 4267 | IsSigned = ICmpInst::isSigned(Pred: InnerPred); |
| 4268 | bool CanSubOne = IsSigned ? !C->isMinSignedValue() : !C->isMinValue(); |
| 4269 | if (CanSubOne) { |
| 4270 | APInt Cminus1 = *C - 1; |
| 4271 | if (match(V: InnerRHS, P: m_SpecificInt(V: Cminus1))) |
| 4272 | Replace = true; |
| 4273 | } |
| 4274 | } |
| 4275 | |
| 4276 | // x == C ? 0 : (x < C+1 ? -1 : 1) |
| 4277 | if (ICmpInst::isLT(P: InnerPred) && InnerTV->isAllOnes() && |
| 4278 | InnerFV->isOne()) { |
| 4279 | IsSigned = ICmpInst::isSigned(Pred: InnerPred); |
| 4280 | bool CanAddOne = IsSigned ? !C->isMaxSignedValue() : !C->isMaxValue(); |
| 4281 | if (CanAddOne) { |
| 4282 | APInt Cplus1 = *C + 1; |
| 4283 | if (match(V: InnerRHS, P: m_SpecificInt(V: Cplus1))) |
| 4284 | Replace = true; |
| 4285 | } |
| 4286 | } |
| 4287 | } |
| 4288 | } |
| 4289 | } |
| 4290 | |
| 4291 | Intrinsic::ID IID = IsSigned ? Intrinsic::scmp : Intrinsic::ucmp; |
| 4292 | if (Replace) |
| 4293 | return replaceInstUsesWith( |
| 4294 | I&: SI, V: Builder.CreateIntrinsic(RetTy: SI.getType(), ID: IID, Args: {LHS, RHS})); |
| 4295 | return nullptr; |
| 4296 | } |
| 4297 | |
| 4298 | bool InstCombinerImpl::fmulByZeroIsZero(Value *MulVal, FastMathFlags FMF, |
| 4299 | const Instruction *CtxI) const { |
| 4300 | KnownFPClass Known = |
| 4301 | computeKnownFPClass(V: MulVal, FMF, InterestedClasses: fcNegative, SQ: SQ.getWithInstruction(I: CtxI)); |
| 4302 | |
| 4303 | return Known.isKnownNeverNaN() && Known.isKnownNeverInfinity() && |
| 4304 | (FMF.noSignedZeros() || Known.signBitIsZeroOrNaN()); |
| 4305 | } |
| 4306 | |
| 4307 | static bool matchFMulByZeroIfResultEqZero(InstCombinerImpl &IC, Value *Cmp0, |
| 4308 | Value *Cmp1, Value *TrueVal, |
| 4309 | Value *FalseVal, Instruction &CtxI, |
| 4310 | bool SelectIsNSZ) { |
| 4311 | Value *MulRHS; |
| 4312 | if (match(V: Cmp1, P: m_PosZeroFP()) && |
| 4313 | match(V: TrueVal, P: m_c_FMul(L: m_Specific(V: Cmp0), R: m_Value(V&: MulRHS)))) { |
| 4314 | FastMathFlags FMF = cast<FPMathOperator>(Val: TrueVal)->getFastMathFlags(); |
| 4315 | // nsz must be on the select, it must be ignored on the multiply. We |
| 4316 | // need nnan and ninf on the multiply for the other value. |
| 4317 | FMF.setNoSignedZeros(SelectIsNSZ); |
| 4318 | return IC.fmulByZeroIsZero(MulVal: MulRHS, FMF, CtxI: &CtxI); |
| 4319 | } |
| 4320 | |
| 4321 | return false; |
| 4322 | } |
| 4323 | |
| 4324 | /// Check whether the KnownBits of a select arm may be affected by the |
| 4325 | /// select condition. |
| 4326 | static bool hasAffectedValue(Value *V, SmallPtrSetImpl<Value *> &Affected, |
| 4327 | unsigned Depth) { |
| 4328 | if (Depth == MaxAnalysisRecursionDepth) |
| 4329 | return false; |
| 4330 | |
| 4331 | // Ignore the case where the select arm itself is affected. These cases |
| 4332 | // are handled more efficiently by other optimizations. |
| 4333 | if (Depth != 0 && Affected.contains(Ptr: V)) |
| 4334 | return true; |
| 4335 | |
| 4336 | if (auto *I = dyn_cast<Instruction>(Val: V)) { |
| 4337 | if (isa<PHINode>(Val: I)) { |
| 4338 | if (Depth == MaxAnalysisRecursionDepth - 1) |
| 4339 | return false; |
| 4340 | Depth = MaxAnalysisRecursionDepth - 2; |
| 4341 | } |
| 4342 | return any_of(Range: I->operands(), P: [&](Value *Op) { |
| 4343 | return Op->getType()->isIntOrIntVectorTy() && |
| 4344 | hasAffectedValue(V: Op, Affected, Depth: Depth + 1); |
| 4345 | }); |
| 4346 | } |
| 4347 | |
| 4348 | return false; |
| 4349 | } |
| 4350 | |
| 4351 | // This transformation enables the possibility of transforming fcmp + sel into |
| 4352 | // a fmaxnum/fminnum intrinsic. |
| 4353 | static Value *foldSelectIntoAddConstant(SelectInst &SI, |
| 4354 | InstCombiner::BuilderTy &Builder) { |
| 4355 | // Do this transformation only when select instruction gives NaN and NSZ |
| 4356 | // guarantee. |
| 4357 | auto *SIFOp = dyn_cast<FPMathOperator>(Val: &SI); |
| 4358 | if (!SIFOp || !SIFOp->hasNoSignedZeros() || !SIFOp->hasNoNaNs()) |
| 4359 | return nullptr; |
| 4360 | |
| 4361 | auto TryFoldIntoAddConstant = |
| 4362 | [&Builder, &SI](CmpInst::Predicate Pred, Value *X, Value *Z, |
| 4363 | Instruction *FAdd, Constant *C, bool Swapped) -> Value * { |
| 4364 | // Only these relational predicates can be transformed into maxnum/minnum |
| 4365 | // intrinsic. |
| 4366 | if (!CmpInst::isRelational(P: Pred) || !match(V: Z, P: m_AnyZeroFP())) |
| 4367 | return nullptr; |
| 4368 | |
| 4369 | if (!match(V: FAdd, P: m_FAdd(L: m_Specific(V: X), R: m_Specific(V: C)))) |
| 4370 | return nullptr; |
| 4371 | |
| 4372 | Value *NewSelect = Builder.CreateSelect(C: SI.getCondition(), True: Swapped ? Z : X, |
| 4373 | False: Swapped ? X : Z, Name: "" , MDFrom: &SI); |
| 4374 | NewSelect->takeName(V: &SI); |
| 4375 | |
| 4376 | Value *NewFAdd = Builder.CreateFAdd(L: NewSelect, R: C); |
| 4377 | NewFAdd->takeName(V: FAdd); |
| 4378 | |
| 4379 | // Propagate FastMath flags |
| 4380 | FastMathFlags SelectFMF = SI.getFastMathFlags(); |
| 4381 | FastMathFlags FAddFMF = FAdd->getFastMathFlags(); |
| 4382 | FastMathFlags NewFMF = FastMathFlags::intersectRewrite(LHS: SelectFMF, RHS: FAddFMF) | |
| 4383 | FastMathFlags::unionValue(LHS: SelectFMF, RHS: FAddFMF); |
| 4384 | cast<Instruction>(Val: NewFAdd)->setFastMathFlags(NewFMF); |
| 4385 | cast<Instruction>(Val: NewSelect)->setFastMathFlags(NewFMF); |
| 4386 | |
| 4387 | return NewFAdd; |
| 4388 | }; |
| 4389 | |
| 4390 | // select((fcmp Pred, X, 0), (fadd X, C), C) |
| 4391 | // => fadd((select (fcmp Pred, X, 0), X, 0), C) |
| 4392 | // |
| 4393 | // Pred := OGT, OGE, OLT, OLE, UGT, UGE, ULT, and ULE |
| 4394 | Instruction *FAdd; |
| 4395 | Constant *C; |
| 4396 | Value *X, *Z; |
| 4397 | CmpPredicate Pred; |
| 4398 | |
| 4399 | // Note: OneUse check for `Cmp` is necessary because it makes sure that other |
| 4400 | // InstCombine folds don't undo this transformation and cause an infinite |
| 4401 | // loop. Furthermore, it could also increase the operation count. |
| 4402 | if (match(V: &SI, P: m_Select(C: m_OneUse(SubPattern: m_FCmp(Pred, L: m_Value(V&: X), R: m_Value(V&: Z))), |
| 4403 | L: m_OneUse(SubPattern: m_Instruction(I&: FAdd)), R: m_Constant(C)))) |
| 4404 | return TryFoldIntoAddConstant(Pred, X, Z, FAdd, C, /*Swapped=*/false); |
| 4405 | |
| 4406 | if (match(V: &SI, P: m_Select(C: m_OneUse(SubPattern: m_FCmp(Pred, L: m_Value(V&: X), R: m_Value(V&: Z))), |
| 4407 | L: m_Constant(C), R: m_OneUse(SubPattern: m_Instruction(I&: FAdd))))) |
| 4408 | return TryFoldIntoAddConstant(Pred, X, Z, FAdd, C, /*Swapped=*/true); |
| 4409 | |
| 4410 | return nullptr; |
| 4411 | } |
| 4412 | |
| 4413 | static Value *foldSelectBitTest(SelectInst &Sel, Value *CondVal, Value *TrueVal, |
| 4414 | Value *FalseVal, |
| 4415 | InstCombiner::BuilderTy &Builder, |
| 4416 | const SimplifyQuery &SQ) { |
| 4417 | // If this is a vector select, we need a vector compare. |
| 4418 | Type *SelType = Sel.getType(); |
| 4419 | if (SelType->isVectorTy() != CondVal->getType()->isVectorTy()) |
| 4420 | return nullptr; |
| 4421 | |
| 4422 | Value *V; |
| 4423 | APInt AndMask; |
| 4424 | bool CreateAnd = false; |
| 4425 | CmpPredicate Pred; |
| 4426 | Value *CmpLHS, *CmpRHS; |
| 4427 | |
| 4428 | if (match(V: CondVal, P: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS)))) { |
| 4429 | if (ICmpInst::isEquality(P: Pred)) { |
| 4430 | if (!match(V: CmpRHS, P: m_Zero())) |
| 4431 | return nullptr; |
| 4432 | |
| 4433 | V = CmpLHS; |
| 4434 | const APInt *AndRHS; |
| 4435 | if (!match(V: CmpLHS, P: m_And(L: m_Value(), R: m_Power2(V&: AndRHS)))) |
| 4436 | return nullptr; |
| 4437 | |
| 4438 | AndMask = *AndRHS; |
| 4439 | } else if (auto Res = decomposeBitTestICmp(LHS: CmpLHS, RHS: CmpRHS, Pred)) { |
| 4440 | assert(ICmpInst::isEquality(Res->Pred) && "Not equality test?" ); |
| 4441 | AndMask = Res->Mask; |
| 4442 | V = Res->X; |
| 4443 | KnownBits Known = computeKnownBits(V, Q: SQ.getWithInstruction(I: &Sel)); |
| 4444 | AndMask &= Known.getMaxValue(); |
| 4445 | if (!AndMask.isPowerOf2()) |
| 4446 | return nullptr; |
| 4447 | |
| 4448 | Pred = Res->Pred; |
| 4449 | CreateAnd = true; |
| 4450 | } else { |
| 4451 | return nullptr; |
| 4452 | } |
| 4453 | } else if (auto *Trunc = dyn_cast<TruncInst>(Val: CondVal)) { |
| 4454 | V = Trunc->getOperand(i_nocapture: 0); |
| 4455 | AndMask = APInt(V->getType()->getScalarSizeInBits(), 1); |
| 4456 | Pred = ICmpInst::ICMP_NE; |
| 4457 | CreateAnd = !Trunc->hasNoUnsignedWrap(); |
| 4458 | } else { |
| 4459 | return nullptr; |
| 4460 | } |
| 4461 | |
| 4462 | if (Pred == ICmpInst::ICMP_NE) |
| 4463 | std::swap(a&: TrueVal, b&: FalseVal); |
| 4464 | |
| 4465 | if (Value *X = foldSelectICmpAnd(Sel, CondVal, TrueVal, FalseVal, V, AndMask, |
| 4466 | CreateAnd, Builder)) |
| 4467 | return X; |
| 4468 | |
| 4469 | if (Value *X = foldSelectICmpAndBinOp(CondVal, TrueVal, FalseVal, V, AndMask, |
| 4470 | CreateAnd, Builder)) |
| 4471 | return X; |
| 4472 | |
| 4473 | return nullptr; |
| 4474 | } |
| 4475 | |
| 4476 | /// This function makes the following folds: |
| 4477 | /// select C, (sub 0, X), (xor X, -1) |
| 4478 | /// -> sub (sext !C), X |
| 4479 | /// select C, (xor X, -1), (sub 0, X) |
| 4480 | /// -> sub (sext C), X |
| 4481 | static Instruction *foldSelectNegNot(SelectInst &SI, |
| 4482 | InstCombiner::BuilderTy &Builder) { |
| 4483 | auto *CondVal = SI.getCondition(); |
| 4484 | auto *TrueVal = SI.getTrueValue(); |
| 4485 | auto *FalseVal = SI.getFalseValue(); |
| 4486 | auto *SelTy = SI.getType(); |
| 4487 | |
| 4488 | if (!SelTy->isIntOrIntVectorTy() || SelTy->isIntOrIntVectorTy(BitWidth: 1)) |
| 4489 | return nullptr; |
| 4490 | |
| 4491 | if (CondVal->getType()->isVectorTy() != SelTy->isVectorTy()) |
| 4492 | return nullptr; |
| 4493 | |
| 4494 | auto matchNegNot = [&](Value *Neg, Value *Not, Value *&X) -> bool { |
| 4495 | return match(V: Neg, P: m_OneUse(SubPattern: m_Neg(V: m_Value(V&: X)))) && |
| 4496 | match(V: Not, P: m_OneUse(SubPattern: m_Not(V: m_Specific(V: X)))); |
| 4497 | }; |
| 4498 | |
| 4499 | Value *X; |
| 4500 | Value *Mask; |
| 4501 | |
| 4502 | // select C, (sub 0, X), (xor X, -1) -> sub (sext !C), X |
| 4503 | if (matchNegNot(TrueVal, FalseVal, X)) { |
| 4504 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
| 4505 | Mask = Builder.CreateSExt(V: NotCond, DestTy: SelTy); |
| 4506 | return BinaryOperator::CreateSub(V1: Mask, V2: X); |
| 4507 | } |
| 4508 | |
| 4509 | // select C, (xor X, -1), (sub 0, X) -> sub (sext C), X |
| 4510 | if (matchNegNot(FalseVal, TrueVal, X)) { |
| 4511 | Mask = Builder.CreateSExt(V: CondVal, DestTy: SelTy); |
| 4512 | return BinaryOperator::CreateSub(V1: Mask, V2: X); |
| 4513 | } |
| 4514 | |
| 4515 | return nullptr; |
| 4516 | } |
| 4517 | |
| 4518 | /// Fold select (A & Shift == 0 | B & Shift == 0), 0, Shift -> Shift & A & B |
| 4519 | /// where Shift is known to be a power of two. |
| 4520 | static Instruction *foldSelectAndOrPowerOfTwo(SelectInst &SI, |
| 4521 | InstCombiner::BuilderTy &Builder, |
| 4522 | const SimplifyQuery &SQ) { |
| 4523 | Value *Cond = SI.getCondition(); |
| 4524 | |
| 4525 | if (!Cond->hasOneUse()) |
| 4526 | return nullptr; |
| 4527 | |
| 4528 | Value *TrueVal = SI.getTrueValue(); |
| 4529 | Value *FalseVal = SI.getFalseValue(); |
| 4530 | |
| 4531 | Value *A, *B, *Shift; |
| 4532 | |
| 4533 | bool Case1 = |
| 4534 | match(V: TrueVal, P: m_Zero()) && match(V: FalseVal, P: m_Value(V&: Shift)) && |
| 4535 | match(V: Cond, P: m_Or(L: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, |
| 4536 | L: m_c_And(L: m_Specific(V: Shift), R: m_Value(V&: A)), |
| 4537 | R: m_Zero()), |
| 4538 | R: m_SpecificICmp(MatchPred: ICmpInst::ICMP_EQ, |
| 4539 | L: m_c_And(L: m_Specific(V: Shift), R: m_Value(V&: B)), |
| 4540 | R: m_Zero()))); |
| 4541 | |
| 4542 | bool Case2 = |
| 4543 | match(V: FalseVal, P: m_Zero()) && match(V: TrueVal, P: m_Value(V&: Shift)) && |
| 4544 | match(V: Cond, P: m_And(L: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, |
| 4545 | L: m_c_And(L: m_Specific(V: Shift), R: m_Value(V&: A)), |
| 4546 | R: m_Zero()), |
| 4547 | R: m_SpecificICmp(MatchPred: ICmpInst::ICMP_NE, |
| 4548 | L: m_c_And(L: m_Specific(V: Shift), R: m_Value(V&: B)), |
| 4549 | R: m_Zero()))); |
| 4550 | |
| 4551 | if ((Case1 || Case2) && isKnownToBeAPowerOfTwo(V: Shift, /*OrZero=*/true, |
| 4552 | Q: SQ.getWithInstruction(I: &SI))) { |
| 4553 | Value *And1 = Builder.CreateAnd(LHS: Shift, RHS: A); |
| 4554 | return BinaryOperator::CreateAnd(V1: And1, V2: B); |
| 4555 | } |
| 4556 | |
| 4557 | return nullptr; |
| 4558 | } |
| 4559 | |
| 4560 | // Return true if no use can observe the sign of zero of the select result, |
| 4561 | // looking through phis, selects and the loop back edge to the select itself. |
| 4562 | static bool isSelectZeroSignInsignificant(SelectInst &SI) { |
| 4563 | // Bound the number of uses to look through to keep the compile time in |
| 4564 | // check. |
| 4565 | constexpr unsigned MaxUsesToLookThrough = 16; |
| 4566 | unsigned NumUses = 0; |
| 4567 | SmallPtrSet<Instruction *, 4> Visited; |
| 4568 | SmallVector<Instruction *> Worklist(1, &SI); |
| 4569 | while (!Worklist.empty()) { |
| 4570 | for (Use &U : Worklist.pop_back_val()->uses()) { |
| 4571 | if (++NumUses > MaxUsesToLookThrough) |
| 4572 | return false; |
| 4573 | auto *User = cast<Instruction>(Val: U.getUser()); |
| 4574 | if (User == &SI) |
| 4575 | continue; |
| 4576 | if (canIgnoreSignBitOfZero(U)) |
| 4577 | continue; |
| 4578 | if (isa<PHINode, SelectInst>(Val: User)) { |
| 4579 | if (Visited.insert(Ptr: User).second) |
| 4580 | Worklist.push_back(Elt: User); |
| 4581 | continue; |
| 4582 | } |
| 4583 | return false; |
| 4584 | } |
| 4585 | } |
| 4586 | return true; |
| 4587 | } |
| 4588 | |
| 4589 | Instruction *InstCombinerImpl::visitSelectInst(SelectInst &SI) { |
| 4590 | Value *CondVal = SI.getCondition(); |
| 4591 | Value *TrueVal = SI.getTrueValue(); |
| 4592 | Value *FalseVal = SI.getFalseValue(); |
| 4593 | Type *SelType = SI.getType(); |
| 4594 | |
| 4595 | FastMathFlags FMF; |
| 4596 | if (auto *FPMO = dyn_cast_if_present<FPMathOperator>(Val: &SI)) |
| 4597 | FMF = FPMO->getFastMathFlags(); |
| 4598 | |
| 4599 | if (Value *V = simplifySelectInst(Cond: CondVal, TrueVal, FalseVal, FMF, |
| 4600 | Q: SQ.getWithInstruction(I: &SI))) |
| 4601 | return replaceInstUsesWith(I&: SI, V); |
| 4602 | |
| 4603 | if (Instruction *I = canonicalizeSelectToShuffle(SI)) |
| 4604 | return I; |
| 4605 | |
| 4606 | if (Instruction *I = canonicalizeScalarSelectOfVecs(Sel&: SI, IC&: *this)) |
| 4607 | return I; |
| 4608 | |
| 4609 | // Fold: select (icmp ult X, 2), X, ctpop(X) --> ctpop(X) |
| 4610 | // ctpop(0)==0 and ctpop(1)==1, so the guard is always redundant. |
| 4611 | if (match(V: FalseVal, P: m_Ctpop(Op0: m_Specific(V: TrueVal))) && |
| 4612 | match(V: CondVal, P: m_SpecificICmp(MatchPred: ICmpInst::ICMP_ULT, L: m_Specific(V: TrueVal), |
| 4613 | R: m_SpecificInt(V: 2)))) { |
| 4614 | cast<Instruction>(Val: FalseVal)->dropPoisonGeneratingAnnotations(); |
| 4615 | addToWorklist(I: cast<Instruction>(Val: FalseVal)); |
| 4616 | return replaceInstUsesWith(I&: SI, V: FalseVal); |
| 4617 | } |
| 4618 | |
| 4619 | // If the type of select is not an integer type or if the condition and |
| 4620 | // the selection type are not both scalar nor both vector types, there is no |
| 4621 | // point in attempting to match these patterns. |
| 4622 | Type *CondType = CondVal->getType(); |
| 4623 | if (!isa<Constant>(Val: CondVal) && SelType->isIntOrIntVectorTy() && |
| 4624 | CondType->isVectorTy() == SelType->isVectorTy()) { |
| 4625 | if (Value *S = simplifyWithOpReplaced(V: TrueVal, Op: CondVal, |
| 4626 | RepOp: ConstantInt::getTrue(Ty: CondType), Q: SQ, |
| 4627 | /* AllowRefinement */ true)) |
| 4628 | return replaceOperand(I&: SI, OpNum: 1, V: S); |
| 4629 | |
| 4630 | if (Value *S = simplifyWithOpReplaced(V: FalseVal, Op: CondVal, |
| 4631 | RepOp: ConstantInt::getFalse(Ty: CondType), Q: SQ, |
| 4632 | /* AllowRefinement */ true)) |
| 4633 | return replaceOperand(I&: SI, OpNum: 2, V: S); |
| 4634 | |
| 4635 | if (replaceInInstruction(V: TrueVal, Old: CondVal, |
| 4636 | New: ConstantInt::getTrue(Ty: CondType)) || |
| 4637 | replaceInInstruction(V: FalseVal, Old: CondVal, |
| 4638 | New: ConstantInt::getFalse(Ty: CondType))) |
| 4639 | return &SI; |
| 4640 | } |
| 4641 | |
| 4642 | if (Instruction *R = foldSelectOfBools(SI)) |
| 4643 | return R; |
| 4644 | |
| 4645 | // Selecting between two integer or vector splat integer constants? |
| 4646 | // |
| 4647 | // Note that we don't handle a scalar select of vectors: |
| 4648 | // select i1 %c, <2 x i8> <1, 1>, <2 x i8> <0, 0> |
| 4649 | // because that may need 3 instructions to splat the condition value: |
| 4650 | // extend, insertelement, shufflevector. |
| 4651 | // |
| 4652 | // Do not handle i1 TrueVal and FalseVal otherwise would result in |
| 4653 | // zext/sext i1 to i1. |
| 4654 | if (SelType->isIntOrIntVectorTy() && !SelType->isIntOrIntVectorTy(BitWidth: 1) && |
| 4655 | CondVal->getType()->isVectorTy() == SelType->isVectorTy()) { |
| 4656 | // select C, 1, 0 -> zext C to int |
| 4657 | if (match(V: TrueVal, P: m_One()) && match(V: FalseVal, P: m_Zero())) |
| 4658 | return new ZExtInst(CondVal, SelType); |
| 4659 | |
| 4660 | // select C, -1, 0 -> sext C to int |
| 4661 | if (match(V: TrueVal, P: m_AllOnes()) && match(V: FalseVal, P: m_Zero())) |
| 4662 | return new SExtInst(CondVal, SelType); |
| 4663 | |
| 4664 | // select C, 0, 1 -> zext !C to int |
| 4665 | if (match(V: TrueVal, P: m_Zero()) && match(V: FalseVal, P: m_One())) { |
| 4666 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
| 4667 | return new ZExtInst(NotCond, SelType); |
| 4668 | } |
| 4669 | |
| 4670 | // select C, 0, -1 -> sext !C to int |
| 4671 | if (match(V: TrueVal, P: m_Zero()) && match(V: FalseVal, P: m_AllOnes())) { |
| 4672 | Value *NotCond = Builder.CreateNot(V: CondVal, Name: "not." + CondVal->getName()); |
| 4673 | return new SExtInst(NotCond, SelType); |
| 4674 | } |
| 4675 | } |
| 4676 | |
| 4677 | if (Instruction *I = foldSelectNegNot(SI, Builder)) |
| 4678 | return I; |
| 4679 | |
| 4680 | if (Instruction *I = foldSelectAndOrPowerOfTwo(SI, Builder, SQ)) |
| 4681 | return I; |
| 4682 | |
| 4683 | auto *SIFPOp = dyn_cast<FPMathOperator>(Val: &SI); |
| 4684 | |
| 4685 | if (auto *FCmp = dyn_cast<FCmpInst>(Val: CondVal)) { |
| 4686 | FCmpInst::Predicate Pred = FCmp->getPredicate(); |
| 4687 | Value *Cmp0 = FCmp->getOperand(i_nocapture: 0), *Cmp1 = FCmp->getOperand(i_nocapture: 1); |
| 4688 | // Are we selecting a value based on a comparison of the two values? |
| 4689 | if ((Cmp0 == TrueVal && Cmp1 == FalseVal) || |
| 4690 | (Cmp0 == FalseVal && Cmp1 == TrueVal)) { |
| 4691 | // Canonicalize to use ordered comparisons by swapping the select |
| 4692 | // operands. |
| 4693 | // |
| 4694 | // e.g. |
| 4695 | // (X ugt Y) ? X : Y -> (X ole Y) ? Y : X |
| 4696 | if (FCmp->hasOneUse() && FCmpInst::isUnordered(predicate: Pred)) { |
| 4697 | FCmpInst::Predicate InvPred = FCmp->getInversePredicate(); |
| 4698 | Value *NewCond = Builder.CreateFCmpFMF(P: InvPred, LHS: Cmp0, RHS: Cmp1, FMFSource: FCmp, |
| 4699 | Name: FCmp->getName() + ".inv" ); |
| 4700 | // Propagate ninf/nnan from fcmp to select. |
| 4701 | FastMathFlags FMF = SI.getFastMathFlags(); |
| 4702 | if (FCmp->hasNoNaNs()) |
| 4703 | FMF.setNoNaNs(true); |
| 4704 | if (FCmp->hasNoInfs()) |
| 4705 | FMF.setNoInfs(true); |
| 4706 | Value *NewSel = |
| 4707 | Builder.CreateSelectFMF(C: NewCond, True: FalseVal, False: TrueVal, FMFSource: FMF); |
| 4708 | return replaceInstUsesWith(I&: SI, V: NewSel); |
| 4709 | } |
| 4710 | } |
| 4711 | |
| 4712 | if (SIFPOp) { |
| 4713 | // Fold out scale-if-equals-zero pattern. |
| 4714 | // |
| 4715 | // This pattern appears in code with denormal range checks after it's |
| 4716 | // assumed denormals are treated as zero. This drops a canonicalization. |
| 4717 | |
| 4718 | // TODO: Could relax the signed zero logic. We just need to know the sign |
| 4719 | // of the result matches (fmul x, y has the same sign as x). |
| 4720 | // |
| 4721 | // TODO: Handle always-canonicalizing variant that selects some value or 1 |
| 4722 | // scaling factor in the fmul visitor. |
| 4723 | |
| 4724 | // TODO: Handle ldexp too |
| 4725 | |
| 4726 | Value *MatchCmp0 = nullptr; |
| 4727 | Value *MatchCmp1 = nullptr; |
| 4728 | |
| 4729 | // (select (fcmp [ou]eq x, 0.0), (fmul x, K), x => x |
| 4730 | // (select (fcmp [ou]ne x, 0.0), x, (fmul x, K) => x |
| 4731 | if (Pred == CmpInst::FCMP_OEQ || Pred == CmpInst::FCMP_UEQ) { |
| 4732 | MatchCmp0 = FalseVal; |
| 4733 | MatchCmp1 = TrueVal; |
| 4734 | } else if (Pred == CmpInst::FCMP_ONE || Pred == CmpInst::FCMP_UNE) { |
| 4735 | MatchCmp0 = TrueVal; |
| 4736 | MatchCmp1 = FalseVal; |
| 4737 | } |
| 4738 | |
| 4739 | if (Cmp0 == MatchCmp0 && |
| 4740 | matchFMulByZeroIfResultEqZero(IC&: *this, Cmp0, Cmp1, TrueVal: MatchCmp1, FalseVal: MatchCmp0, |
| 4741 | CtxI&: SI, SelectIsNSZ: SIFPOp->hasNoSignedZeros())) |
| 4742 | return replaceInstUsesWith(I&: SI, V: Cmp0); |
| 4743 | |
| 4744 | Type *EltTy = SelType->getScalarType(); |
| 4745 | |
| 4746 | // TODO: Generalize to any ordered / unordered compare. |
| 4747 | if ((Pred == CmpInst::FCMP_ORD || Pred == CmpInst::FCMP_UNO) && |
| 4748 | match(V: Cmp1, P: m_PosZeroFP()) && EltTy->isIEEELikeFPTy()) { |
| 4749 | // Fold out only-canonicalize-non-nans pattern. This implements a |
| 4750 | // wrapper around llvm.canonicalize which is not required to quiet |
| 4751 | // signaling nans or preserve nan payload bits. |
| 4752 | // |
| 4753 | // %hard.canonical = call @llvm.canonicalize(%x) |
| 4754 | // %soft.canonical = fdiv 1.0, %x |
| 4755 | // %ord = fcmp ord %x, 0.0 |
| 4756 | // %x.canon = select i1 %ord, %hard.canonical, %soft.canonical |
| 4757 | // |
| 4758 | // With known IEEE handling: |
| 4759 | // => %x |
| 4760 | // |
| 4761 | // With other denormal behaviors: |
| 4762 | // => llvm.canonicalize(%x) |
| 4763 | // |
| 4764 | // Note the fdiv could be any value preserving, potentially |
| 4765 | // canonicalizing floating-point operation such as fmul by 1.0. However, |
| 4766 | // since in the llvm model canonicalization is not mandatory, the fmul |
| 4767 | // would have been dropped by the time we reached here. The trick here |
| 4768 | // is to use a reciprocal fdiv. It's not a droppable no-op, as it could |
| 4769 | // return an infinity if %x were sufficiently small, but in this pattern |
| 4770 | // we're only using the output for nan values. |
| 4771 | |
| 4772 | if (Pred == CmpInst::FCMP_ORD) { |
| 4773 | MatchCmp0 = TrueVal; |
| 4774 | MatchCmp1 = FalseVal; |
| 4775 | } else { |
| 4776 | MatchCmp0 = FalseVal; |
| 4777 | MatchCmp1 = TrueVal; |
| 4778 | } |
| 4779 | |
| 4780 | bool RcpIfNan = match(V: MatchCmp1, P: m_FDiv(L: m_FPOne(), R: m_Specific(V: Cmp0))); |
| 4781 | bool CanonicalizeIfNotNan = |
| 4782 | match(V: MatchCmp0, P: m_FCanonicalize(Op0: m_Specific(V: Cmp0))); |
| 4783 | |
| 4784 | if (RcpIfNan || CanonicalizeIfNotNan) { |
| 4785 | const fltSemantics &FPSem = EltTy->getFltSemantics(); |
| 4786 | DenormalMode Mode = F.getDenormalMode(FPType: FPSem); |
| 4787 | |
| 4788 | if (RcpIfNan) { |
| 4789 | if (Mode == DenormalMode::getIEEE()) { |
| 4790 | // Special case for the other select operand. Otherwise, we may |
| 4791 | // need to insert freeze on Cmp0 in the compare and select. |
| 4792 | if (CanonicalizeIfNotNan) |
| 4793 | return replaceInstUsesWith(I&: SI, V: Cmp0); |
| 4794 | |
| 4795 | if (isGuaranteedNotToBeUndef(V: Cmp0, AC: &AC, CtxI: &SI, DT: &DT)) { |
| 4796 | // select (fcmp ord x, 0), y, (fdiv 1, x) |
| 4797 | // => select (fcmp ord x, 0), y, x |
| 4798 | // |
| 4799 | // select (fcmp uno x, 0), (fdiv 1, x), y |
| 4800 | // => select (fcmp uno x, 0), x, y |
| 4801 | replaceOperand(I&: SI, OpNum: Pred == CmpInst::FCMP_ORD ? 2 : 1, V: Cmp0); |
| 4802 | return &SI; |
| 4803 | } |
| 4804 | |
| 4805 | auto *FrCmp0 = InsertNewInstBefore( |
| 4806 | New: new FreezeInst(Cmp0, Cmp0->getName() + ".fr" ), |
| 4807 | Old: FCmp->getIterator()); |
| 4808 | |
| 4809 | replaceOperand(I&: *FCmp, OpNum: 0, V: FrCmp0); |
| 4810 | return replaceOperand(I&: SI, OpNum: Pred == CmpInst::FCMP_ORD ? 2 : 1, |
| 4811 | V: FrCmp0); |
| 4812 | } |
| 4813 | } |
| 4814 | |
| 4815 | if (CanonicalizeIfNotNan) { |
| 4816 | // IEEE handling does not have non-canonical values, so the |
| 4817 | // canonicalize can be dropped for direct replacement without |
| 4818 | // looking for the intermediate maybe-canonicalizing operation. |
| 4819 | if (Mode == DenormalMode::getIEEE()) { |
| 4820 | // select (fcmp ord x, 0), canonicalize(x), y |
| 4821 | // => select (fcmp ord x, 0), x, y |
| 4822 | |
| 4823 | replaceOperand(I&: SI, OpNum: Pred == CmpInst::FCMP_ORD ? 1 : 2, V: Cmp0); |
| 4824 | return &SI; |
| 4825 | } |
| 4826 | |
| 4827 | // If denormals may be flushed, we need to retain the canonicalize |
| 4828 | // call. This introduces a canonicalization on the nan path, which |
| 4829 | // we are not free to do as that could change the sign bit or |
| 4830 | // payload bits. We can only do this if there were a no-op like |
| 4831 | // floating-point instruction which may have changed the nan bits |
| 4832 | // anyway. |
| 4833 | |
| 4834 | // Leave the dynamic mode case alone. This would introduce new |
| 4835 | // constraints if the mode may be refined later. |
| 4836 | if (RcpIfNan && (Mode.inputsAreZero() || Mode.outputsAreZero())) |
| 4837 | return replaceInstUsesWith(I&: SI, V: MatchCmp0); |
| 4838 | assert(RcpIfNan || Mode != DenormalMode::getIEEE()); |
| 4839 | } |
| 4840 | } |
| 4841 | } |
| 4842 | } |
| 4843 | } |
| 4844 | |
| 4845 | if (SIFPOp) { |
| 4846 | // TODO: Try to forward-propagate FMF from select arms to the select. |
| 4847 | |
| 4848 | auto *FCmp = dyn_cast<FCmpInst>(Val: CondVal); |
| 4849 | |
| 4850 | // Canonicalize select of FP values where NaN and -0.0 are not valid as |
| 4851 | // minnum/maxnum intrinsics. |
| 4852 | // |
| 4853 | // Note that the `nnan` flag is propagated from the comparison, not from the |
| 4854 | // select. While it's technically possible to transform a `fcmp` + `select |
| 4855 | // nnan` to a `minnum`/`maxnum` call *without* an `nnan`, that would be a |
| 4856 | // pessimization in practice. Many targets can't map `minnum`/`maxnum` to a |
| 4857 | // single instruction, and if they cannot prove the absence of NaN, must |
| 4858 | // lower it to a routine or a libcall. There are additional reasons besides |
| 4859 | // performance to avoid introducing libcalls where none existed before |
| 4860 | // (https://github.com/llvm/llvm-project/issues/54554). |
| 4861 | // |
| 4862 | // As such, we want to ensure that the generated `minnum`/`maxnum` intrinsic |
| 4863 | // has the `nnan nsz` flags, which allow it to be lowered *back* to a |
| 4864 | // fcmp+select if that's the best way to express it on the target. |
| 4865 | if (FCmp && FCmp->hasNoNaNs() && |
| 4866 | (SIFPOp->hasNoSignedZeros() || isSelectZeroSignInsignificant(SI))) { |
| 4867 | Value *X, *Y; |
| 4868 | if (match(V: &SI, P: m_OrdOrUnordFMax(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 4869 | Value *BinIntr = |
| 4870 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::maxnum, LHS: X, RHS: Y, FMFSource: &SI); |
| 4871 | if (auto *BinIntrInst = dyn_cast<Instruction>(Val: BinIntr)) { |
| 4872 | // `ninf` must be propagated from the comparison too, rather than the |
| 4873 | // select: https://github.com/llvm/llvm-project/pull/136433 |
| 4874 | BinIntrInst->setHasNoInfs(FCmp->hasNoInfs()); |
| 4875 | // The `nsz` flag is a precondition, so let's ensure it's always added |
| 4876 | // to the min/max operation, even if it wasn't on the select. This |
| 4877 | // could happen if the select doesn't have `nsz`, but no use of the |
| 4878 | // result can observe the sign of zero. |
| 4879 | BinIntrInst->setHasNoSignedZeros(true); |
| 4880 | // As mentioned above, `nnan` is also a precondition, so we always set |
| 4881 | // the flag. |
| 4882 | BinIntrInst->setHasNoNaNs(true); |
| 4883 | } |
| 4884 | return replaceInstUsesWith(I&: SI, V: BinIntr); |
| 4885 | } |
| 4886 | |
| 4887 | if (match(V: &SI, P: m_OrdOrUnordFMin(L: m_Value(V&: X), R: m_Value(V&: Y)))) { |
| 4888 | Value *BinIntr = |
| 4889 | Builder.CreateBinaryIntrinsic(ID: Intrinsic::minnum, LHS: X, RHS: Y, FMFSource: &SI); |
| 4890 | if (auto *BinIntrInst = dyn_cast<Instruction>(Val: BinIntr)) { |
| 4891 | BinIntrInst->setHasNoInfs(FCmp->hasNoInfs()); |
| 4892 | BinIntrInst->setHasNoSignedZeros(true); |
| 4893 | BinIntrInst->setHasNoNaNs(true); |
| 4894 | } |
| 4895 | return replaceInstUsesWith(I&: SI, V: BinIntr); |
| 4896 | } |
| 4897 | } |
| 4898 | } |
| 4899 | |
| 4900 | // Fold selecting to fabs. |
| 4901 | if (Instruction *Fabs = foldSelectWithFCmpToFabs(SI, IC&: *this)) |
| 4902 | return Fabs; |
| 4903 | |
| 4904 | if (Instruction *I = foldSelectOfOrderedFAbsCmpOfNaNScrubbedValue(SI, IC&: *this)) |
| 4905 | return I; |
| 4906 | |
| 4907 | // See if we are selecting two values based on a comparison of the two values. |
| 4908 | if (CmpInst *CI = dyn_cast<CmpInst>(Val: CondVal)) |
| 4909 | if (Instruction *NewSel = foldSelectValueEquivalence(Sel&: SI, Cmp&: *CI)) |
| 4910 | return NewSel; |
| 4911 | |
| 4912 | if (ICmpInst *ICI = dyn_cast<ICmpInst>(Val: CondVal)) |
| 4913 | if (Instruction *Result = foldSelectInstWithICmp(SI, ICI)) |
| 4914 | return Result; |
| 4915 | |
| 4916 | if (Value *V = foldSelectBitTest(Sel&: SI, CondVal, TrueVal, FalseVal, Builder, SQ)) |
| 4917 | return replaceInstUsesWith(I&: SI, V); |
| 4918 | |
| 4919 | if (Instruction *Add = foldAddSubSelect(SI, Builder)) |
| 4920 | return Add; |
| 4921 | if (Instruction *Add = foldOverflowingAddSubSelect(SI, Builder)) |
| 4922 | return Add; |
| 4923 | if (Instruction *Or = foldSetClearBits(Sel&: SI, Builder)) |
| 4924 | return Or; |
| 4925 | if (Instruction *Mul = foldSelectZeroOrFixedOp(SI, IC&: *this)) |
| 4926 | return Mul; |
| 4927 | |
| 4928 | // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z)) |
| 4929 | auto *TI = dyn_cast<Instruction>(Val: TrueVal); |
| 4930 | auto *FI = dyn_cast<Instruction>(Val: FalseVal); |
| 4931 | if (TI && FI && TI->getOpcode() == FI->getOpcode()) |
| 4932 | if (Instruction *IV = foldSelectOpOp(SI, TI, FI)) |
| 4933 | return IV; |
| 4934 | |
| 4935 | if (Instruction *I = foldSelectIntrinsic(SI)) |
| 4936 | return I; |
| 4937 | |
| 4938 | if (Instruction *I = foldSelectExtConst(Sel&: SI)) |
| 4939 | return I; |
| 4940 | |
| 4941 | if (Instruction *I = foldSelectWithSRem(SI, IC&: *this, Builder)) |
| 4942 | return I; |
| 4943 | |
| 4944 | // Fold (select C, (gep Ptr, Idx), Ptr) -> (gep Ptr, (select C, Idx, 0)) |
| 4945 | // Fold (select C, Ptr, (gep Ptr, Idx)) -> (gep Ptr, (select C, 0, Idx)) |
| 4946 | auto SelectGepWithBase = [&](GetElementPtrInst *Gep, Value *Base, |
| 4947 | bool Swap) -> GetElementPtrInst * { |
| 4948 | Value *Ptr = Gep->getPointerOperand(); |
| 4949 | if (Gep->getNumOperands() != 2 || Gep->getPointerOperand() != Base || |
| 4950 | !Gep->hasOneUse()) |
| 4951 | return nullptr; |
| 4952 | Value *Idx = Gep->getOperand(i_nocapture: 1); |
| 4953 | if (isa<VectorType>(Val: CondVal->getType()) && !isa<VectorType>(Val: Idx->getType())) |
| 4954 | return nullptr; |
| 4955 | Type *ElementType = Gep->getSourceElementType(); |
| 4956 | Value *NewT = Idx; |
| 4957 | Value *NewF = Constant::getNullValue(Ty: Idx->getType()); |
| 4958 | if (Swap) |
| 4959 | std::swap(a&: NewT, b&: NewF); |
| 4960 | Value *NewSI = |
| 4961 | Builder.CreateSelect(C: CondVal, True: NewT, False: NewF, Name: SI.getName() + ".idx" , MDFrom: &SI); |
| 4962 | return GetElementPtrInst::Create(PointeeType: ElementType, Ptr, IdxList: NewSI, |
| 4963 | NW: Gep->getNoWrapFlags()); |
| 4964 | }; |
| 4965 | if (auto *TrueGep = dyn_cast<GetElementPtrInst>(Val: TrueVal)) |
| 4966 | if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false)) |
| 4967 | return NewGep; |
| 4968 | if (auto *FalseGep = dyn_cast<GetElementPtrInst>(Val: FalseVal)) |
| 4969 | if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true)) |
| 4970 | return NewGep; |
| 4971 | |
| 4972 | // See if we can fold the select into one of our operands. |
| 4973 | if (SelType->isIntOrIntVectorTy() || SelType->isFPOrFPVectorTy()) { |
| 4974 | if (Instruction *FoldI = foldSelectIntoOp(SI, TrueVal, FalseVal)) |
| 4975 | return FoldI; |
| 4976 | |
| 4977 | Value *LHS, *RHS; |
| 4978 | Instruction::CastOps CastOp; |
| 4979 | SelectPatternResult SPR = matchSelectPattern(V: &SI, LHS, RHS, CastOp: &CastOp); |
| 4980 | auto SPF = SPR.Flavor; |
| 4981 | if (SPF) { |
| 4982 | Value *LHS2, *RHS2; |
| 4983 | if (SelectPatternFlavor SPF2 = matchSelectPattern(V: LHS, LHS&: LHS2, RHS&: RHS2).Flavor) |
| 4984 | if (Instruction *R = foldSPFofSPF(Inner: cast<Instruction>(Val: LHS), SPF1: SPF2, A: LHS2, |
| 4985 | B: RHS2, Outer&: SI, SPF2: SPF, C: RHS)) |
| 4986 | return R; |
| 4987 | if (SelectPatternFlavor SPF2 = matchSelectPattern(V: RHS, LHS&: LHS2, RHS&: RHS2).Flavor) |
| 4988 | if (Instruction *R = foldSPFofSPF(Inner: cast<Instruction>(Val: RHS), SPF1: SPF2, A: LHS2, |
| 4989 | B: RHS2, Outer&: SI, SPF2: SPF, C: LHS)) |
| 4990 | return R; |
| 4991 | } |
| 4992 | |
| 4993 | if (SelectPatternResult::isMinOrMax(SPF)) { |
| 4994 | // Canonicalize so that |
| 4995 | // - type casts are outside select patterns. |
| 4996 | // - float clamp is transformed to min/max pattern |
| 4997 | |
| 4998 | bool IsCastNeeded = LHS->getType() != SelType; |
| 4999 | Value *CmpLHS = cast<CmpInst>(Val: CondVal)->getOperand(i_nocapture: 0); |
| 5000 | Value *CmpRHS = cast<CmpInst>(Val: CondVal)->getOperand(i_nocapture: 1); |
| 5001 | if (IsCastNeeded || |
| 5002 | (LHS->getType()->isFPOrFPVectorTy() && |
| 5003 | ((CmpLHS != LHS && CmpLHS != RHS) || |
| 5004 | (CmpRHS != LHS && CmpRHS != RHS)))) { |
| 5005 | CmpInst::Predicate MinMaxPred = getMinMaxPred(SPF, Ordered: SPR.Ordered); |
| 5006 | |
| 5007 | Value *Cmp; |
| 5008 | if (CmpInst::isIntPredicate(P: MinMaxPred)) |
| 5009 | Cmp = Builder.CreateICmp(P: MinMaxPred, LHS, RHS); |
| 5010 | else |
| 5011 | Cmp = Builder.CreateFCmpFMF(P: MinMaxPred, LHS, RHS, |
| 5012 | FMFSource: cast<Instruction>(Val: SI.getCondition())); |
| 5013 | |
| 5014 | Value *NewSI = Builder.CreateSelect(C: Cmp, True: LHS, False: RHS, Name: SI.getName(), MDFrom: &SI); |
| 5015 | if (!IsCastNeeded) |
| 5016 | return replaceInstUsesWith(I&: SI, V: NewSI); |
| 5017 | |
| 5018 | Value *NewCast = Builder.CreateCast(Op: CastOp, V: NewSI, DestTy: SelType); |
| 5019 | return replaceInstUsesWith(I&: SI, V: NewCast); |
| 5020 | } |
| 5021 | } |
| 5022 | } |
| 5023 | |
| 5024 | // See if we can fold the select into a phi node if the condition is a select. |
| 5025 | if (auto *PN = dyn_cast<PHINode>(Val: SI.getCondition())) |
| 5026 | if (Instruction *NV = foldOpIntoPhi(I&: SI, PN)) |
| 5027 | return NV; |
| 5028 | |
| 5029 | if (SelectInst *TrueSI = dyn_cast<SelectInst>(Val: TrueVal)) { |
| 5030 | if (TrueSI->getCondition()->getType() == CondVal->getType()) { |
| 5031 | // Fold nested selects if the inner condition can be implied by the outer |
| 5032 | // condition. |
| 5033 | if (Value *V = simplifyNestedSelectsUsingImpliedCond( |
| 5034 | SI&: *TrueSI, CondVal, /*CondIsTrue=*/true, DL)) |
| 5035 | return replaceOperand(I&: SI, OpNum: 1, V); |
| 5036 | |
| 5037 | // We choose this as normal form to enable folding on the And and |
| 5038 | // shortening paths for the values (this helps getUnderlyingObjects() for |
| 5039 | // example). |
| 5040 | if (TrueSI->hasOneUse()) { |
| 5041 | Value *And = nullptr, *OtherVal = nullptr; |
| 5042 | // select(C0, select(C1, a, b), b) -> select(C0&&C1, a, b) |
| 5043 | if (TrueSI->getFalseValue() == FalseVal) { |
| 5044 | And = Builder.CreateLogicalAnd(Cond1: CondVal, Cond2: TrueSI->getCondition(), Name: "" , |
| 5045 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr |
| 5046 | : &SI); |
| 5047 | OtherVal = TrueSI->getTrueValue(); |
| 5048 | } |
| 5049 | // select(C0, select(C1, b, a), b) -> select(C0&&!C1, a, b) |
| 5050 | else if (TrueSI->getTrueValue() == FalseVal) { |
| 5051 | Value *InvertedCond = Builder.CreateNot(V: TrueSI->getCondition()); |
| 5052 | And = Builder.CreateLogicalAnd(Cond1: CondVal, Cond2: InvertedCond, Name: "" , |
| 5053 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr |
| 5054 | : &SI); |
| 5055 | OtherVal = TrueSI->getFalseValue(); |
| 5056 | } |
| 5057 | if (And && OtherVal) { |
| 5058 | replaceOperand(I&: SI, OpNum: 0, V: And); |
| 5059 | replaceOperand(I&: SI, OpNum: 1, V: OtherVal); |
| 5060 | if (!ProfcheckDisableMetadataFixes) |
| 5061 | setExplicitlyUnknownBranchWeightsIfProfiled(I&: SI, DEBUG_TYPE); |
| 5062 | return &SI; |
| 5063 | } |
| 5064 | } |
| 5065 | } |
| 5066 | } |
| 5067 | if (SelectInst *FalseSI = dyn_cast<SelectInst>(Val: FalseVal)) { |
| 5068 | if (FalseSI->getCondition()->getType() == CondVal->getType()) { |
| 5069 | // Fold nested selects if the inner condition can be implied by the outer |
| 5070 | // condition. |
| 5071 | if (Value *V = simplifyNestedSelectsUsingImpliedCond( |
| 5072 | SI&: *FalseSI, CondVal, /*CondIsTrue=*/false, DL)) |
| 5073 | return replaceOperand(I&: SI, OpNum: 2, V); |
| 5074 | |
| 5075 | if (FalseSI->hasOneUse()) { |
| 5076 | Value *Or = nullptr, *OtherVal = nullptr; |
| 5077 | // select(C0, a, select(C1, a, b)) -> select(C0||C1, a, b) |
| 5078 | if (FalseSI->getTrueValue() == TrueVal) { |
| 5079 | Or = Builder.CreateLogicalOr(Cond1: CondVal, Cond2: FalseSI->getCondition(), Name: "" , |
| 5080 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr |
| 5081 | : &SI); |
| 5082 | OtherVal = FalseSI->getFalseValue(); |
| 5083 | } |
| 5084 | // select(C0, a, select(C1, b, a)) -> select(C0||!C1, a, b) |
| 5085 | else if (FalseSI->getFalseValue() == TrueVal) { |
| 5086 | Value *InvertedCond = Builder.CreateNot(V: FalseSI->getCondition()); |
| 5087 | Or = Builder.CreateLogicalOr(Cond1: CondVal, Cond2: InvertedCond, Name: "" , |
| 5088 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr |
| 5089 | : &SI); |
| 5090 | OtherVal = FalseSI->getTrueValue(); |
| 5091 | } |
| 5092 | if (Or && OtherVal) { |
| 5093 | replaceOperand(I&: SI, OpNum: 0, V: Or); |
| 5094 | replaceOperand(I&: SI, OpNum: 2, V: OtherVal); |
| 5095 | if (!ProfcheckDisableMetadataFixes) |
| 5096 | setExplicitlyUnknownBranchWeightsIfProfiled(I&: SI, DEBUG_TYPE); |
| 5097 | return &SI; |
| 5098 | } |
| 5099 | } |
| 5100 | } |
| 5101 | } |
| 5102 | |
| 5103 | // Try to simplify a binop sandwiched between 2 selects with the same |
| 5104 | // condition. This is not valid for div/rem because the select might be |
| 5105 | // preventing a division-by-zero. |
| 5106 | // TODO: A div/rem restriction is conservative; use something like |
| 5107 | // isSafeToSpeculativelyExecute(). |
| 5108 | // select(C, binop(select(C, X, Y), W), Z) -> select(C, binop(X, W), Z) |
| 5109 | BinaryOperator *TrueBO; |
| 5110 | if (match(V: TrueVal, P: m_OneUse(SubPattern: m_BinOp(I&: TrueBO))) && !TrueBO->isIntDivRem()) { |
| 5111 | if (auto *TrueBOSI = dyn_cast<SelectInst>(Val: TrueBO->getOperand(i_nocapture: 0))) { |
| 5112 | if (TrueBOSI->getCondition() == CondVal) { |
| 5113 | replaceOperand(I&: *TrueBO, OpNum: 0, V: TrueBOSI->getTrueValue()); |
| 5114 | Worklist.push(I: TrueBO); |
| 5115 | return &SI; |
| 5116 | } |
| 5117 | } |
| 5118 | if (auto *TrueBOSI = dyn_cast<SelectInst>(Val: TrueBO->getOperand(i_nocapture: 1))) { |
| 5119 | if (TrueBOSI->getCondition() == CondVal) { |
| 5120 | replaceOperand(I&: *TrueBO, OpNum: 1, V: TrueBOSI->getTrueValue()); |
| 5121 | Worklist.push(I: TrueBO); |
| 5122 | return &SI; |
| 5123 | } |
| 5124 | } |
| 5125 | } |
| 5126 | |
| 5127 | // select(C, Z, binop(select(C, X, Y), W)) -> select(C, Z, binop(Y, W)) |
| 5128 | BinaryOperator *FalseBO; |
| 5129 | if (match(V: FalseVal, P: m_OneUse(SubPattern: m_BinOp(I&: FalseBO))) && !FalseBO->isIntDivRem()) { |
| 5130 | if (auto *FalseBOSI = dyn_cast<SelectInst>(Val: FalseBO->getOperand(i_nocapture: 0))) { |
| 5131 | if (FalseBOSI->getCondition() == CondVal) { |
| 5132 | replaceOperand(I&: *FalseBO, OpNum: 0, V: FalseBOSI->getFalseValue()); |
| 5133 | Worklist.push(I: FalseBO); |
| 5134 | return &SI; |
| 5135 | } |
| 5136 | } |
| 5137 | if (auto *FalseBOSI = dyn_cast<SelectInst>(Val: FalseBO->getOperand(i_nocapture: 1))) { |
| 5138 | if (FalseBOSI->getCondition() == CondVal) { |
| 5139 | replaceOperand(I&: *FalseBO, OpNum: 1, V: FalseBOSI->getFalseValue()); |
| 5140 | Worklist.push(I: FalseBO); |
| 5141 | return &SI; |
| 5142 | } |
| 5143 | } |
| 5144 | } |
| 5145 | |
| 5146 | Value *NotCond; |
| 5147 | if (match(V: CondVal, P: m_Not(V: m_Value(V&: NotCond))) && |
| 5148 | !InstCombiner::shouldAvoidAbsorbingNotIntoSelect(SI)) { |
| 5149 | replaceOperand(I&: SI, OpNum: 0, V: NotCond); |
| 5150 | SI.swapValues(); |
| 5151 | SI.swapProfMetadata(); |
| 5152 | return &SI; |
| 5153 | } |
| 5154 | |
| 5155 | if (Instruction *I = foldVectorSelect(Sel&: SI)) |
| 5156 | return I; |
| 5157 | |
| 5158 | // If we can compute the condition, there's no need for a select. |
| 5159 | // Like the above fold, we are attempting to reduce compile-time cost by |
| 5160 | // putting this fold here with limitations rather than in InstSimplify. |
| 5161 | // The motivation for this call into value tracking is to take advantage of |
| 5162 | // the assumption cache, so make sure that is populated. |
| 5163 | if (!CondVal->getType()->isVectorTy() && !AC.assumptions().empty()) { |
| 5164 | KnownBits Known(1); |
| 5165 | computeKnownBits(V: CondVal, Known, CxtI: &SI); |
| 5166 | if (Known.One.isOne()) |
| 5167 | return replaceInstUsesWith(I&: SI, V: TrueVal); |
| 5168 | if (Known.Zero.isOne()) |
| 5169 | return replaceInstUsesWith(I&: SI, V: FalseVal); |
| 5170 | } |
| 5171 | |
| 5172 | if (Instruction *BitCastSel = foldSelectCmpBitcasts(Sel&: SI, Builder)) |
| 5173 | return BitCastSel; |
| 5174 | |
| 5175 | // Simplify selects that test the returned flag of cmpxchg instructions. |
| 5176 | if (Value *V = foldSelectCmpXchg(SI)) |
| 5177 | return replaceInstUsesWith(I&: SI, V); |
| 5178 | |
| 5179 | if (Instruction *Select = foldSelectBinOpIdentity(Sel&: SI, TLI, IC&: *this)) |
| 5180 | return Select; |
| 5181 | |
| 5182 | if (Instruction *Funnel = foldSelectFunnelShift(Sel&: SI, Builder)) |
| 5183 | return Funnel; |
| 5184 | |
| 5185 | if (Instruction *Copysign = foldSelectToCopysign(Sel&: SI, Builder)) |
| 5186 | return Copysign; |
| 5187 | |
| 5188 | if (Instruction *PN = foldSelectToPhi(Sel&: SI, DT, Builder)) |
| 5189 | return replaceInstUsesWith(I&: SI, V: PN); |
| 5190 | |
| 5191 | if (Value *V = foldRoundUpIntegerWithPow2Alignment(SI, Builder)) |
| 5192 | return replaceInstUsesWith(I&: SI, V); |
| 5193 | |
| 5194 | if (Value *V = foldSelectIntoAddConstant(SI, Builder)) |
| 5195 | return replaceInstUsesWith(I&: SI, V); |
| 5196 | |
| 5197 | // select(mask, mload(ptr,mask,0), 0) -> mload(ptr,mask,0) |
| 5198 | // Load inst is intentionally not checked for hasOneUse() |
| 5199 | if (match(V: FalseVal, P: m_Zero()) && |
| 5200 | (match(V: TrueVal, P: m_MaskedLoad(Op0: m_Value(), Op1: m_Specific(V: CondVal), |
| 5201 | Op2: m_CombineOr(Ps: m_Undef(), Ps: m_Zero()))) || |
| 5202 | match(V: TrueVal, P: m_MaskedGather(Op0: m_Value(), Op1: m_Specific(V: CondVal), |
| 5203 | Op2: m_CombineOr(Ps: m_Undef(), Ps: m_Zero()))))) { |
| 5204 | auto *MaskedInst = cast<IntrinsicInst>(Val: TrueVal); |
| 5205 | if (isa<UndefValue>(Val: MaskedInst->getArgOperand(i: 2))) |
| 5206 | MaskedInst->setArgOperand(i: 2, v: FalseVal /* Zero */); |
| 5207 | return replaceInstUsesWith(I&: SI, V: MaskedInst); |
| 5208 | } |
| 5209 | |
| 5210 | Value *Mask; |
| 5211 | if (match(V: TrueVal, P: m_Zero()) && |
| 5212 | (match(V: FalseVal, P: m_MaskedLoad(Op0: m_Value(), Op1: m_Value(V&: Mask), |
| 5213 | Op2: m_CombineOr(Ps: m_Undef(), Ps: m_Zero()))) || |
| 5214 | match(V: FalseVal, P: m_MaskedGather(Op0: m_Value(), Op1: m_Value(V&: Mask), |
| 5215 | Op2: m_CombineOr(Ps: m_Undef(), Ps: m_Zero())))) && |
| 5216 | (CondVal->getType() == Mask->getType())) { |
| 5217 | // We can remove the select by ensuring the load zeros all lanes the |
| 5218 | // select would have. We determine this by proving there is no overlap |
| 5219 | // between the load and select masks. |
| 5220 | // (i.e (load_mask & select_mask) == 0 == no overlap) |
| 5221 | bool CanMergeSelectIntoLoad = false; |
| 5222 | if (Value *V = simplifyAndInst(LHS: CondVal, RHS: Mask, Q: SQ.getWithInstruction(I: &SI))) |
| 5223 | CanMergeSelectIntoLoad = match(V, P: m_Zero()); |
| 5224 | |
| 5225 | if (CanMergeSelectIntoLoad) { |
| 5226 | auto *MaskedInst = cast<IntrinsicInst>(Val: FalseVal); |
| 5227 | if (isa<UndefValue>(Val: MaskedInst->getArgOperand(i: 2))) |
| 5228 | MaskedInst->setArgOperand(i: 2, v: TrueVal /* Zero */); |
| 5229 | return replaceInstUsesWith(I&: SI, V: MaskedInst); |
| 5230 | } |
| 5231 | } |
| 5232 | |
| 5233 | if (Instruction *I = foldSelectOfSymmetricSelect(OuterSelVal&: SI, Builder)) |
| 5234 | return I; |
| 5235 | |
| 5236 | if (Instruction *I = foldNestedSelects(OuterSelVal&: SI, Builder)) |
| 5237 | return I; |
| 5238 | |
| 5239 | // Match logical variants of the pattern, |
| 5240 | // and transform them iff that gets rid of inversions. |
| 5241 | // (~x) | y --> ~(x & (~y)) |
| 5242 | // (~x) & y --> ~(x | (~y)) |
| 5243 | if (sinkNotIntoOtherHandOfLogicalOp(I&: SI)) |
| 5244 | return &SI; |
| 5245 | |
| 5246 | if (Instruction *I = foldBitCeil(SI, Builder, IC&: *this)) |
| 5247 | return I; |
| 5248 | |
| 5249 | if (Instruction *I = foldSelectToCmp(SI)) |
| 5250 | return I; |
| 5251 | |
| 5252 | if (Instruction *I = foldSelectEqualityTest(Sel&: SI)) |
| 5253 | return I; |
| 5254 | |
| 5255 | // Fold: |
| 5256 | // (select A && B, T, F) -> (select A, (select B, T, F), F) |
| 5257 | // (select A || B, T, F) -> (select A, T, (select B, T, F)) |
| 5258 | // if (select B, T, F) is foldable. |
| 5259 | // TODO: preserve FMF flags |
| 5260 | auto FoldSelectWithAndOrCond = [&](bool IsAnd, Value *A, |
| 5261 | Value *B) -> Instruction * { |
| 5262 | if (Value *V = simplifySelectInst(Cond: B, TrueVal, FalseVal, FMF, |
| 5263 | Q: SQ.getWithInstruction(I: &SI))) { |
| 5264 | Value *NewTrueVal = IsAnd ? V : TrueVal; |
| 5265 | Value *NewFalseVal = IsAnd ? FalseVal : V; |
| 5266 | |
| 5267 | // If the True and False values don't change, then preserve the branch |
| 5268 | // metadata of the original select as the net effect of this change is to |
| 5269 | // simplify the conditional. |
| 5270 | Instruction *MDFrom = nullptr; |
| 5271 | if (NewTrueVal == TrueVal && NewFalseVal == FalseVal && |
| 5272 | !ProfcheckDisableMetadataFixes) { |
| 5273 | MDFrom = &SI; |
| 5274 | } |
| 5275 | return SelectInst::Create(C: A, S1: NewTrueVal, S2: NewFalseVal, NameStr: "" , InsertBefore: nullptr, |
| 5276 | MDFrom); |
| 5277 | } |
| 5278 | |
| 5279 | // Is (select B, T, F) a SPF? |
| 5280 | if (CondVal->hasOneUse() && SelType->isIntOrIntVectorTy()) { |
| 5281 | if (ICmpInst *Cmp = dyn_cast<ICmpInst>(Val: B)) |
| 5282 | if (Value *V = canonicalizeSPF(Cmp&: *Cmp, TrueVal, FalseVal, IC&: *this)) { |
| 5283 | return SelectInst::Create( |
| 5284 | C: A, S1: IsAnd ? V : TrueVal, S2: IsAnd ? FalseVal : V, NameStr: "" , InsertBefore: nullptr, |
| 5285 | MDFrom: ProfcheckDisableMetadataFixes ? nullptr : &SI); |
| 5286 | } |
| 5287 | } |
| 5288 | |
| 5289 | return nullptr; |
| 5290 | }; |
| 5291 | |
| 5292 | Value *LHS, *RHS; |
| 5293 | if (match(V: CondVal, P: m_And(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 5294 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, LHS, RHS)) |
| 5295 | return I; |
| 5296 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, RHS, LHS)) |
| 5297 | return I; |
| 5298 | } else if (match(V: CondVal, P: m_Or(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 5299 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, LHS, RHS)) |
| 5300 | return I; |
| 5301 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, RHS, LHS)) |
| 5302 | return I; |
| 5303 | } else { |
| 5304 | // We cannot swap the operands of logical and/or. |
| 5305 | // TODO: Can we swap the operands by inserting a freeze? |
| 5306 | if (match(V: CondVal, P: m_LogicalAnd(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 5307 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ true, LHS, RHS)) |
| 5308 | return I; |
| 5309 | } else if (match(V: CondVal, P: m_LogicalOr(L: m_Value(V&: LHS), R: m_Value(V&: RHS)))) { |
| 5310 | if (Instruction *I = FoldSelectWithAndOrCond(/*IsAnd*/ false, LHS, RHS)) |
| 5311 | return I; |
| 5312 | } |
| 5313 | } |
| 5314 | |
| 5315 | // select Cond, !X, X -> xor Cond, X |
| 5316 | if (CondVal->getType() == SI.getType() && isKnownInversion(X: FalseVal, Y: TrueVal)) |
| 5317 | return BinaryOperator::CreateXor(V1: CondVal, V2: FalseVal); |
| 5318 | |
| 5319 | // For vectors, this transform is only safe if the simplification does not |
| 5320 | // look through any lane-crossing operations. For now, limit to scalars only. |
| 5321 | if (SelType->isIntegerTy() && |
| 5322 | (!isa<Constant>(Val: TrueVal) || !isa<Constant>(Val: FalseVal))) { |
| 5323 | // Try to simplify select arms based on KnownBits implied by the condition. |
| 5324 | CondContext CC(CondVal); |
| 5325 | findValuesAffectedByCondition(Cond: CondVal, /*IsAssume=*/false, InsertAffected: [&](Value *V) { |
| 5326 | CC.AffectedValues.insert(Ptr: V); |
| 5327 | }); |
| 5328 | SimplifyQuery Q = SQ.getWithInstruction(I: &SI).getWithCondContext(CC); |
| 5329 | if (!CC.AffectedValues.empty()) { |
| 5330 | if (!isa<Constant>(Val: TrueVal) && |
| 5331 | hasAffectedValue(V: TrueVal, Affected&: CC.AffectedValues, /*Depth=*/0)) { |
| 5332 | KnownBits Known = llvm::computeKnownBits(V: TrueVal, Q); |
| 5333 | if (Known.isConstant()) |
| 5334 | return replaceOperand(I&: SI, OpNum: 1, |
| 5335 | V: ConstantInt::get(Ty: SelType, V: Known.getConstant())); |
| 5336 | } |
| 5337 | |
| 5338 | CC.Invert = true; |
| 5339 | if (!isa<Constant>(Val: FalseVal) && |
| 5340 | hasAffectedValue(V: FalseVal, Affected&: CC.AffectedValues, /*Depth=*/0)) { |
| 5341 | KnownBits Known = llvm::computeKnownBits(V: FalseVal, Q); |
| 5342 | if (Known.isConstant()) |
| 5343 | return replaceOperand(I&: SI, OpNum: 2, |
| 5344 | V: ConstantInt::get(Ty: SelType, V: Known.getConstant())); |
| 5345 | } |
| 5346 | } |
| 5347 | } |
| 5348 | |
| 5349 | // select (trunc nuw X to i1), X, Y --> select (trunc nuw X to i1), 1, Y |
| 5350 | // select (trunc nuw X to i1), Y, X --> select (trunc nuw X to i1), Y, 0 |
| 5351 | // select (trunc nsw X to i1), X, Y --> select (trunc nsw X to i1), -1, Y |
| 5352 | // select (trunc nsw X to i1), Y, X --> select (trunc nsw X to i1), Y, 0 |
| 5353 | Value *Trunc; |
| 5354 | if (match(V: CondVal, P: m_NUWTrunc(Op: m_Value(V&: Trunc))) && !isa<Constant>(Val: Trunc)) { |
| 5355 | if (TrueVal == Trunc) |
| 5356 | return replaceOperand(I&: SI, OpNum: 1, V: ConstantInt::get(Ty: TrueVal->getType(), V: 1)); |
| 5357 | if (FalseVal == Trunc) |
| 5358 | return replaceOperand(I&: SI, OpNum: 2, V: ConstantInt::get(Ty: FalseVal->getType(), V: 0)); |
| 5359 | } |
| 5360 | if (match(V: CondVal, P: m_NSWTrunc(Op: m_Value(V&: Trunc))) && !isa<Constant>(Val: Trunc)) { |
| 5361 | if (TrueVal == Trunc) |
| 5362 | return replaceOperand(I&: SI, OpNum: 1, |
| 5363 | V: Constant::getAllOnesValue(Ty: TrueVal->getType())); |
| 5364 | if (FalseVal == Trunc) |
| 5365 | return replaceOperand(I&: SI, OpNum: 2, V: ConstantInt::get(Ty: FalseVal->getType(), V: 0)); |
| 5366 | } |
| 5367 | |
| 5368 | if (match(V: CondVal, P: m_Trunc(Op: m_Value(V&: Trunc))) && Trunc->getType() == SelType) { |
| 5369 | if (match(V: FalseVal, P: m_Zero()) && impliesPoison(ValAssumedPoison: TrueVal, V: CondVal) && |
| 5370 | llvm::computeKnownBits(V: TrueVal, Q: SQ.getWithInstruction(I: &SI)) |
| 5371 | .countMaxActiveBits() == 1) |
| 5372 | return BinaryOperator::CreateAnd(V1: Trunc, V2: TrueVal); |
| 5373 | |
| 5374 | if (cast<TruncInst>(Val: CondVal)->hasNoUnsignedWrap() && |
| 5375 | match(V: TrueVal, P: m_One()) && impliesPoison(ValAssumedPoison: FalseVal, V: CondVal) && |
| 5376 | llvm::computeKnownBits(V: FalseVal, Q: SQ.getWithInstruction(I: &SI)) |
| 5377 | .countMaxActiveBits() == 1) { |
| 5378 | return BinaryOperator::CreateOr(V1: Trunc, V2: FalseVal); |
| 5379 | } |
| 5380 | } |
| 5381 | |
| 5382 | Value *MaskedLoadPtr; |
| 5383 | if (match(V: TrueVal, P: m_OneUse(SubPattern: m_MaskedLoad(Op0: m_Value(V&: MaskedLoadPtr), |
| 5384 | Op1: m_Specific(V: CondVal), Op2: m_Value())))) { |
| 5385 | auto *LoadInst = cast<IntrinsicInst>(Val: TrueVal); |
| 5386 | // Keep the load at its original position to avoid crossing writes. The new |
| 5387 | // passthrough must therefore be available there. |
| 5388 | if (DT.dominates(Def: FalseVal, User: LoadInst)) { |
| 5389 | Builder.SetInsertPoint(LoadInst); |
| 5390 | Instruction *In = Builder.CreateMaskedLoad( |
| 5391 | Ty: TrueVal->getType(), Ptr: MaskedLoadPtr, |
| 5392 | Alignment: LoadInst->getParamAlign(ArgNo: 0).valueOrOne(), Mask: CondVal, PassThru: FalseVal); |
| 5393 | In->setAAMetadata(LoadInst->getAAMetadata()); |
| 5394 | return replaceInstUsesWith(I&: SI, V: In); |
| 5395 | } |
| 5396 | } |
| 5397 | |
| 5398 | // Canonicalize sign function ashr pattern: select (icmp slt X, 1), ashr X, |
| 5399 | // bitwidth-1, 1 -> scmp(X, 0) |
| 5400 | // Also handles: select (icmp sgt X, 0), 1, ashr X, bitwidth-1 -> scmp(X, 0) |
| 5401 | unsigned BitWidth = SI.getType()->getScalarSizeInBits(); |
| 5402 | CmpPredicate Pred; |
| 5403 | Value *CmpLHS, *CmpRHS; |
| 5404 | |
| 5405 | // Canonicalize sign function ashr patterns: |
| 5406 | // select (icmp slt X, 1), ashr X, bitwidth-1, 1 -> scmp(X, 0) |
| 5407 | // select (icmp sgt X, 0), 1, ashr X, bitwidth-1 -> scmp(X, 0) |
| 5408 | if (match(V: &SI, P: m_Select(C: m_ICmp(Pred, L: m_Value(V&: CmpLHS), R: m_Value(V&: CmpRHS)), |
| 5409 | L: m_Value(V&: TrueVal), R: m_Value(V&: FalseVal))) && |
| 5410 | ((Pred == ICmpInst::ICMP_SLT && match(V: CmpRHS, P: m_One()) && |
| 5411 | match(V: TrueVal, |
| 5412 | P: m_AShr(L: m_Specific(V: CmpLHS), R: m_SpecificInt(V: BitWidth - 1))) && |
| 5413 | match(V: FalseVal, P: m_One())) || |
| 5414 | (Pred == ICmpInst::ICMP_SGT && match(V: CmpRHS, P: m_Zero()) && |
| 5415 | match(V: TrueVal, P: m_One()) && |
| 5416 | match(V: FalseVal, |
| 5417 | P: m_AShr(L: m_Specific(V: CmpLHS), R: m_SpecificInt(V: BitWidth - 1)))))) { |
| 5418 | |
| 5419 | Function *Scmp = Intrinsic::getOrInsertDeclaration( |
| 5420 | M: SI.getModule(), id: Intrinsic::scmp, OverloadTys: {SI.getType(), SI.getType()}); |
| 5421 | return CallInst::Create(Func: Scmp, Args: {CmpLHS, ConstantInt::get(Ty: SI.getType(), V: 0)}); |
| 5422 | } |
| 5423 | |
| 5424 | return nullptr; |
| 5425 | } |
| 5426 | |